Power distribution facility inspection method and system based on collaborative operation of unmanned vehicle, robot dog, and robotic arm

Unmanned vehicles, robotic dogs, and robotic arms work together to achieve efficient, accurate, and intelligent inspection of power distribution facilities in high-density cable areas through precise docking, positioning, and dynamic adjustment, solving the problems of low efficiency and high risk in existing technologies.

WO2026232741A1PCT designated stage Publication Date: 2026-11-12ZHEJIANG DAYOU INDUSTRIAL CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/123804
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2025-09-25
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

In existing technologies, the inspection efficiency of power distribution facilities in high-density cable areas is low and the risks are high, making it difficult to achieve full coverage and resulting in insufficient safety.

Method used

An inspection method that employs unmanned vehicles, robot dogs, and robotic arms in a collaborative manner is adopted. Multi-view LiDAR is used to determine the precise parking position of the unmanned vehicle, the robot dog uses a convolutional neural network to accurately locate the fault point, and the robotic arm makes dynamic adjustments and uses a Kalman filter algorithm to perform real-time coordinate offset correction, thereby achieving precise grasping and emergency repair of faulty parts.

Benefits of technology

It improves the efficiency and safety of power distribution facility inspection in high-density cable areas, reduces labor costs and risks, and ensures the stable operation of power distribution facilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025123804_12112026_PF_FP_ABST
    Figure CN2025123804_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A power distribution facility inspection method and system based on collaborative operation of an unmanned vehicle, a robot dog, and a robotic arm. The unmanned vehicle is used to acquire three-dimensional distribution data of cables in a fault area, and accurate parking of the unmanned vehicle in a confined space is implemented by path planning and dynamic adjustment of the height of a chassis; the robot dog is used to detect an occluded area so as to extract features of a fault component and determine accurate coordinates of the fault component; a reachable range of the robotic arm carried on the unmanned vehicle is quantified, a change in the joint angle of the robotic arm is further calculated, and an operation instruction is determined after the position of the unmanned vehicle is finely adjusted to eliminate deviation; and on the basis of a predicted real-time position change trend of the robot dog relative to a fault point, fault position data synchronized with the operation instruction is determined, to generate a grasping scheme by means of the robotic arm so as to accurately perform an emergency repair operation, thereby implementing automatic and accurate emergency repair in a complex cable environment, and improving the operation efficiency and safety.
Need to check novelty before this filing date? Find Prior Art

Description

Methods and systems for inspecting power distribution facilities using unmanned vehicles, robot dogs, and robotic arms in a collaborative manner Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for inspecting power distribution facilities in collaboration with unmanned vehicles, robot dogs, and robotic arms. Background Technology

[0002] Power distribution facility inspection is a core part of power system operation and maintenance. In the current technology, the inspection of power distribution facilities, especially those in complex environments such as high-density cable areas, is usually carried out by manual laborers. However, traditional manual inspection has problems such as low efficiency, high risk, and incomplete coverage.

[0003] Therefore, improving the efficiency and safety of power distribution facilities inspection in high-density cable areas has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention provides a method and system for inspecting power distribution facilities in collaboration with unmanned vehicles, robotic dogs, and robotic arms, addressing the problem of improving the inspection efficiency and safety of power distribution facilities in high-density cable areas.

[0005] To address the aforementioned technical problems, the first aspect of this invention provides a method for inspecting power distribution facilities using a collaborative approach involving an unmanned vehicle, a robotic dog, and a robotic arm, comprising:

[0006] Based on the three-dimensional distribution data of cables in the fault area of ​​the power distribution facility, the initial parking area of ​​the unmanned vehicle is determined, and the movement trajectory of the unmanned vehicle in the fault area is quantified according to the initial parking area to determine the precise parking position of the unmanned vehicle relative to the fault point.

[0007] The occlusion information data of the fault area collected by the robot dog is processed by a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, and the range of motion of the robot arm is quantified based on the relative relationship between the precise docking position and the precise three-dimensional coordinates.

[0008] The real-time coordinate offset data of the robot dog when adjusting its posture is obtained, and the dynamic adjustment parameters of the robotic arm when it is aligned with the fault point are quantified in combination with the range of motion. Based on the dynamic adjustment parameters, operation instructions for the robotic arm are generated.

[0009] The real-time coordinate offset data is processed using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point, and obtain fault position data synchronized with the operation command.

[0010] Based on the operation instructions and the fault location data, capture data is generated to separate the faulty components from the fault area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.

[0011] A second aspect of the present invention provides a power distribution facility inspection system that integrates an unmanned vehicle, a robot dog, and a robotic arm, comprising:

[0012] The location determination module is used to determine the initial parking area of ​​the unmanned vehicle based on the three-dimensional distribution data of the cables in the fault area corresponding to the power distribution facility, and to quantify the movement trajectory of the unmanned vehicle in the fault area according to the initial parking area, so as to determine the precise parking position of the unmanned vehicle relative to the fault point.

[0013] The range quantization module is used to process the occlusion information data of the fault area collected by the robot dog through a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, and to quantify the range of motion of the robotic arm based on the relative relationship between the precise docking position and the precise three-dimensional coordinates.

[0014] The instruction generation module is used to acquire real-time coordinate offset data when the robot dog adjusts its posture, combine the motion range to quantify the dynamic adjustment parameters when the robotic arm aligns with the fault point, and generate operation instructions for the robotic arm based on the dynamic adjustment parameters.

[0015] The position synchronization module is used to process the real-time coordinate offset data using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point, and obtain fault position data synchronized with the operation command.

[0016] The fault repair module is used to generate capture data based on the operation instructions and the fault location data, so as to separate the faulty components from the fault area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.

[0017] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following:

[0018] (1) By processing the three-dimensional distribution data of the cable, the precise stopping area of ​​the unmanned vehicle in the complex fault area is determined; in the occluded environment, the robot dog extracts the contour features of the faulty component through the convolutional neural network, accurately locates the three-dimensional coordinates of the fault point, and dynamically adjusts its posture to avoid obstacles and reduce the inspection risk.

[0019] (2) Seamless collaboration is achieved through the coordinated operation of unmanned vehicles, robot dogs and robotic arms to perform coordinate offset correction, fault location synchronization, etc., shortening the overall operation time and adapting to different terrains and complex environments. This enables efficient, accurate, intelligent and automated inspection and repair of fault areas of power distribution facilities, which not only improves repair efficiency and reduces labor costs and risks, but also ensures the safe and stable operation of power distribution facilities. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the 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.

[0021] Figure 1 is a flowchart of a method for inspecting power distribution facilities in collaboration with an unmanned vehicle, a robot dog, and a robotic arm, according to a certain embodiment of the present invention.

[0022] Figure 2 is a structural diagram of a power distribution facility inspection system that integrates unmanned vehicles, robot dogs, and robotic arms, according to a certain embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0025] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the system or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is merely for describing specific embodiments and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0027] In one embodiment, as shown in FIG1, the first aspect of the present invention provides a method for inspecting power distribution facilities in collaboration with an unmanned vehicle, a robot dog, and a robotic arm, comprising:

[0028] S1. Based on the three-dimensional distribution data of cables in the fault area corresponding to the power distribution facility, determine the initial parking area of ​​the unmanned vehicle, and quantify the movement trajectory of the unmanned vehicle in the fault area according to the initial parking area, so as to determine the precise parking position of the unmanned vehicle relative to the fault point.

[0029] In one embodiment, determining the initial parking area of ​​the unmanned vehicle based on the three-dimensional cable distribution data corresponding to the fault area of ​​the power distribution facility includes:

[0030] The unmanned vehicle uses a multi-view lidar to collect three-dimensional cable distribution data in the fault area corresponding to the power distribution facility; the fault area is a ground space-constrained environment area including the high-density cable crossing area of ​​the power distribution facility.

[0031] The three-dimensional distribution data of the cable is processed using a point cloud distance calculation algorithm to obtain the height data and position coordinates of each cable in the fault area, so as to analyze the density distribution pattern of each cable in the fault area and obtain cable density distribution characteristic data.

[0032] The fault area is divided based on the cable density distribution feature data, and the area where the cable height is lower than a preset height threshold is marked as the parking candidate area of ​​the unmanned vehicle, thus obtaining a set of candidate areas;

[0033] Extract target candidate regions that match the height requirements of the unmanned vehicle from the candidate region set, and use a geometric constraint algorithm to determine the spatial connectivity of each target candidate region to obtain the range of connected regions;

[0034] Based on the range of the connected region and the scanning boundary conditions of the multi-view lidar, the coordinates of the parking area are determined, and a machine learning clustering algorithm is used to optimize and group the target candidate regions within the range of the connected region to generate the initial parking area of ​​the unmanned vehicle.

[0035] Specifically, because unmanned vehicles need to move on the ground to carry tools and spare parts during inspections, but the ground space in high-density cable areas is limited and the cable suspension heights are inconsistent, it is difficult for unmanned vehicles to accurately stop near the fault point.

[0036] Based on this, the present invention uses a multi-view LiDAR mounted on an unmanned vehicle to scan the corresponding fault area of ​​the power distribution facility (i.e., the high-density cable crossing area) to obtain three-dimensional cable distribution data from the limited ground space environment of the high-density cable crossing area. For example, in an old industrial park, cables are crisscrossed at different heights and ground space is limited. The multi-view LiDAR installed on the unmanned vehicle scans at a frequency of 200,000 points per second, collecting data from multiple directions such as the side and top, capturing the reflected signals of the cables, and forming three-dimensional point cloud distribution data of the cables. This determines the preliminary spatial distribution information of the cables in the fault area, so as to clearly show the three-dimensional layout of the cables. For example, there are 5 cables in a certain area, located at heights of 1.5 meters, 2.0 meters, 3.0 meters, etc., laying the foundation for subsequent analysis.

[0037] A point cloud distance calculation algorithm is used to extract cable features. Based on the three-dimensional distribution data of cables collected by unmanned vehicles, the algorithm calculates the vertical distance between each cable and the ground through point cloud data, thereby obtaining the height, diameter, and position coordinates of each cable in the fault area. For example, in the aforementioned industrial park scenario, the algorithm analyzes the point cloud in real time to determine that the diameter of a cable with a height of 1.5 meters is about 10 centimeters, and the diameter of a cable with a height of 2.0 meters is 15 centimeters. At the same time, it determines the precise coordinates of each cable, such as (x1, y1, z1) as (10, 20, 1.5), to distinguish the differences in cable height and to provide data support for space occupancy analysis. The fault area is divided into several grids (such as 1m × 1m) to count the number of cables and the average height in each grid. Then, kernel density estimation is used to generate a two-dimensional density heat map to identify the density distribution pattern of each cable in the fault area and obtain cable density distribution feature data. Furthermore, assuming the scanning area is 50m x 50m, the analysis reveals that cables shorter than 2.0m are concentrated in the northwest corner, while cables taller than 3.0m are mostly distributed in the southeast corner. This shows that low-height cables impose a more significant restriction on ground space. For example, under the low cables in the northwest corner, the space height is only 1.5m, restricting the passage of large equipment, while the space in the southeast corner is relatively open. These analysis results can provide a basis for subsequent area division and help optimize the parking planning of unmanned vehicles.

[0038] The fault area is divided based on the cable density distribution characteristics. If the cable height is lower than a preset height threshold, the area below it is marked as a candidate area where the unmanned vehicle can dock. Areas with an area smaller than the minimum turning radius of the unmanned vehicle and areas where the cable height does not meet the standard are removed. Then, a 0.3m buffer zone is extended outward to ensure that the unmanned vehicle maintains a safe distance from the cable when docking, so as to generate a set of candidate areas. The preset height threshold can be 2.5 meters. The division of the fault area utilizes the height distribution characteristics to screen out the potentially usable docking space and improve the space utilization rate.

[0039] The system extracts target candidate regions (i.e., regions that allow the autonomous vehicle to pass smoothly) that match the height requirements of the candidate region set. A geometric constraint algorithm is then used to construct a topology graph model with candidate regions as nodes, reachable paths between regions as edges, and path widths greater than or equal to the width of the autonomous vehicle. Dijkstra's algorithm is then used to search for the longest continuous path and mark connected subgraphs. If a connected region is blocked by a temporary obstacle (such as maintenance equipment), the topology graph is updated in real time using point clouds and recalculated until the spatial connectivity of all target candidate regions is determined. This generates the range of connected regions that the autonomous vehicle can pass through, ensuring the practicality of the parking area and avoiding spatial fragmentation.

[0040] Based on the scanning boundary conditions of the multi-view LiDAR on the autonomous vehicle, the initial parking area coordinates and parking coordinates of the autonomous vehicle can be determined from the connected region range. Features such as area, shape index (aspect ratio), and Euclidean distance from the fault point of candidate regions are extracted. Using the autonomous vehicle's scanning boundary conditions as constraints, the DBSCAN algorithm merges adjacent regions with similar attributes within the connected region range to obtain the initial parking area of ​​the autonomous vehicle. Here, attributes refer to the multi-dimensional features of the candidate regions. These features are used to quantify the region's geometric characteristics (area, shape index), spatial location (the Euclidean distance from the fault point must meet the safety radius constraint), and environmental constraints (minimum turning radius limit). If the cable height in a sub-region suddenly drops to 2.0 meters, clustering operations can identify and adjust the boundary in advance. Furthermore, the parking area can be adjusted in conjunction with real-time environmental changes. For example, if a 2.2-meter-high cable is added to an industrial park, after the LiDAR rescans, the algorithm can dynamically update the candidate region set, removing affected parts to ensure the adaptability of the solution. The entire process, from scanning to optimization, achieves efficient utilization of spatial resources, provides reliable support for autonomous vehicle parking in complex environments, and reduces the cost of manual intervention.

[0041] This invention uses multi-view lidar and density analysis to accurately identify safe stopping points in high-density cable crossing areas, avoiding collision risks; it combines geometric constraints and machine learning clustering to efficiently divide feasible areas in constrained ground environments, improving path planning efficiency; it filters stopping areas based on preset height thresholds to ensure a safe distance between unmanned vehicles and cables, reducing electromagnetic interference risks; and it reduces human intervention through point cloud processing and clustering algorithms, achieving automated and intelligent stopping area selection.

[0042] In one embodiment, quantifying the movement trajectory of the unmanned vehicle in the fault area based on the initial parking area to determine the precise parking position of the unmanned vehicle relative to the fault point includes:

[0043] Based on the initial docking area, the cable density distribution characteristic data, and the height data of each cable, the movement trajectory of the unmanned vehicle under the spatially confined conditions corresponding to the fault area is quantified to obtain an initial trajectory coordinate sequence.

[0044] When it is determined that there is a deviation between the initial trajectory coordinate sequence and the fault area, the model predictive control method is used to generate adjustment parameters for the chassis height of the unmanned vehicle, and the trajectory adjustment coordinate sequence is obtained.

[0045] Based on the trajectory adjustment coordinate sequence and the height data of each cable, the cable height change characteristics within the fault point and its preset area are extracted, so as to determine the area division range of the fault point through a clustering algorithm;

[0046] Based on the defined area and the spatial constraints, a target trajectory adjustment coordinate sequence that meets the connectivity requirements is selected from the trajectory adjustment coordinate sequence and combined to generate a continuous movement trajectory path. This path is then combined with the cable density distribution feature data to determine the precise stopping position of the unmanned vehicle relative to the fault point.

[0047] Specifically, this invention takes the coordinates of the initial docking area, the cable density distribution characteristics of the area, and the height data of the cables contained in the area as inputs. It uses a fast exploration random tree algorithm to transform the cable height data into a three-dimensional obstacle model, restricts the movement height of the unmanned vehicle (e.g., the chassis height must be ≥ cable height + 0.3m), and outputs the movement trajectory of the unmanned vehicle under space-constrained conditions, obtaining an initial trajectory coordinate sequence (including position, velocity, attitude angle, etc.). Other path planning algorithms can also be used. It should be noted that this invention only uses these algorithms and does not improve them. Therefore, the specific steps of the algorithm are not described in detail. For example, in an old industrial park, an unmanned vehicle needs to move from the starting point (0,0,0) to the fault point (40,40,0). Combining the cable height data obtained from the aforementioned LiDAR scan, such as the cable height of 1.5 meters in the northwest corner and 3.0 meters in the southeast corner, the path planning algorithm can analyze the spatial constraints and avoid areas with a height 1.2 meters lower than the chassis of the unmanned vehicle. The algorithm prioritizes the path under the cable with a height of 3.0 meters in the southeast corner, generating an initial trajectory coordinate sequence, such as (0,0,0), (10,10,0), (20,20,0), (30,30,0), (40,40,0). By making full use of the height distribution pattern of the cable, the initial feasibility of the trajectory is ensured.

[0048] A deviation detection algorithm is employed to analyze the degree of matching between the vehicle's trajectory and the actual environment. Specifically, it uses LiDAR SLAM and IMU fusion positioning to calculate trajectory tracking errors (lateral error and heading angle error) and compares them with a preset deviation threshold. If the deviation error exceeds the threshold, a deviation is determined between the initial trajectory coordinate sequence and the fault area. A model predictive control (MMC) method is then used to establish the autonomous vehicle's kinematic model, aiming to minimize tracking errors and chassis height adjustment amplitude (preventing frequent rises and falls). The model outputs chassis height adjustment and trajectory coordinate offset compensation as the trajectory adjustment coordinate sequence. If the trajectory tracking error is not greater than the preset deviation threshold, the original chassis height of the autonomous vehicle is used. Real-time adjustments improve the autonomous vehicle's environmental adaptability and avoid failures caused by static planning.

[0049] The system updates the 3D point cloud distribution data of cables collected by the unmanned vehicle based on the trajectory adjustment coordinate sequence. It also extracts the cable height change characteristics (reflecting the degree of undulation) and height gradient (e.g., the height difference between adjacent points > 0.5m is a sudden change point) within the fault point and its preset area (e.g., a radius of 2m) according to the height of each cable. The extracted cable height change characteristics are then processed by the density-based OPTICS clustering algorithm to group the fault point and its preset area. The cable height sudden change area is marked as a potential fault sub-region (e.g., break point, loose joint) to obtain the area division range of the fault point.

[0050] Based on the spatial constraints of the region division and the fault area, the system determines whether the trajectory adjustment coordinate sequence meets the connectivity requirements. This is constrained by cable density not exceeding a preset density threshold and cable height not exceeding a preset height threshold. Target trajectory adjustment coordinate sequences that meet the connectivity requirements are selected from the existing sequence and combined. The combined result is mapped to graph nodes. The existence of reachable paths between nodes is checked, and cubic spline interpolation is used to fill trajectory breakpoints to ensure path continuity. The interpolation step size is reduced (e.g., 0.1m) in densely cabled areas and increased (e.g., 0.5m) in sparse areas, generating a continuous movement trajectory path and its coordinate set. Connectivity verification avoids spatial fragmentation and ensures path integrity.

[0051] Calculate the Euclidean distance between each point on the continuous moving trajectory path and the fault point. Using the weighted sum of the total path length and the maximum distance difference (weight ratio 6:4) as the objective function, optimize the moving trajectory points whose distance difference exceeds a preset threshold using gradient descent or a genetic algorithm, and output the optimized moving trajectory path. Moving trajectory points whose distance difference does not exceed the preset threshold are considered to meet the distance requirements to the fault point and can still be used. Based on the optimized moving trajectory path and cable density distribution characteristics, and ensuring that the stopping position is at least 0.5m away from the nearest cable (to prevent electromagnetic interference and mechanical collisions), select the candidate area with the lowest density (density < 2 cables / m). 2 Using (41,41,0) as a docking point, the center of the autonomous vehicle's chassis is aligned with the coordinate system of the fault point. Fine-tuning is then performed using visual markers (ArUco codes) to obtain and output the precise docking position of the autonomous vehicle relative to the fault point. For example, at (41,41,0), the cable height is 3.0 meters, and the spatial connectivity range is 5 meters × 4 meters, meeting the autonomous vehicle's 3 meters × 2 meters requirement. Therefore, the final docking coordinates are determined to be between (41,41,0) and (43,43,0). This precise positioning of the autonomous vehicle fully utilizes spatial resources and reduces the impact of deviations.

[0052] This invention combines spatial constraints and dynamic adjustments for high-precision trajectory planning, ensuring the safe movement of unmanned vehicles in areas with dense cables; it corrects trajectories in real time through model predictive control to cope with environmental interference; it avoids path breakage or inaccessibility issues through connectivity verification and trajectory optimization; and it uses clustering and interpolation algorithms to quickly fill trajectory gaps, reducing computation time. At the same time, it enables efficient movement and docking of unmanned vehicles in complex environments, significantly improving the level of automation and safety.

[0053] S2. The occlusion information data of the fault area collected by the robot dog is processed by a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, and the movement range of the robot arm is quantified based on the relative relationship between the precise docking position and the precise three-dimensional coordinates.

[0054] In one embodiment, the step of processing the occlusion information data of the fault area collected by the robot dog using a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point includes:

[0055] Based on the precise docking location, infrared sensors and depth cameras deployed on the robot dog are used to acquire occlusion information data of the area under the cable, resulting in a multimodal data set which is then input into a convolutional neural network for processing, outputting a contour feature dataset.

[0056] Based on the contour feature dataset and the parameter information of the depth camera, the three-dimensional coordinate values ​​of the faulty component are determined;

[0057] Based on the three-dimensional coordinate values ​​and the shielding area under the cable, a clustering algorithm is used to divide the region boundary of the faulty component, and a boundary division coordinate sequence is obtained to determine the thermal feature distribution matrix of the faulty component according to the multimodal data set;

[0058] Based on the contour feature dataset, a matching thermal feature distribution matrix is ​​extracted from the thermal feature distribution matrix. When it is determined that there is a positional deviation between the matching thermal feature distribution matrix and the three-dimensional coordinate values, the matching thermal feature distribution matrix is ​​adjusted using a gradient descent algorithm to obtain a thermal feature distribution adjustment matrix. This matrix is ​​then combined with the three-dimensional coordinate values ​​to determine the approximate three-dimensional coordinates of the fault point.

[0059] Specifically, because the robot dog needs to enter the narrow area under the cable to locate the fault during the inspection process, its vision sensor is easily interfered with by the dense cable obstruction, making it difficult to accurately identify the three-dimensional coordinates of the faulty component.

[0060] Based on this, this invention utilizes infrared sensors and depth cameras deployed on the robot dog to acquire occlusion information data of the area under the cable, such as abnormal thermal features like overheated cable joints, and 3D point clouds of the occluded area, according to the precise parking position of the unmanned vehicle. The timestamps of the infrared and depth data are aligned using hardware trigger signals to form a multimodal data set containing temperature and three-dimensional position, thereby improving information comprehensiveness and avoiding the omission of key features by a single sensor. An improved Mask R-CNN is employed, with a ResNet-50-FPN backbone network and an added infrared channel input branch. The improved Mask R-CNN is trained using historical multimodal data sets and their fault labels as training data. The multimodal data set collected by the robot dog based on the precise parking position of the unmanned vehicle is then input into the trained improved Mask R-CNN, outputting a binary mask (initial contour feature) of the faulty component, thus obtaining a contour feature dataset. For multimodal datasets, the initial contour features of faulty components can also be separated from cable obstruction interference using image processing techniques: dense cable areas may obstruct faulty components, and combining high-temperature areas in infrared data with abrupt changes in depth data can initially outline the contour of the faulty components.

[0061] Infrared images and depth point clouds are registered using an extrinsic parameter matrix to generate temperature-labeled 3D point clouds. 100 points are randomly sampled within the contour mask output by the improved MaskR-CNN, and their mean 3D coordinates are calculated as initial 3D coordinate values. Then, with the objective function of minimizing the contour point projection error (from the depth camera coordinate system to the image plane), the translation vector and rotation matrix of the faulty component are iteratively adjusted using the least squares method until convergence, and the optimized 3D coordinate values ​​of the faulty component are obtained as the 3D coordinate values ​​of the faulty component.

[0062] The point cloud data of the area under the cable and the optimized 3D coordinates of that area are taken as input. Density-based DBSCAN clustering is used to process the data with the parameters set (eps=0.05m, min_samples=5). The area boundaries of the faulty component are divided, and the boundary coordinate sequence of the faulty component is output (such as the set of coordinates of the high temperature zone boundary). Based on the clustering results, the thermal feature distribution data of the faulty component collected by the infrared sensor is extracted from the multimodal dataset to generate a thermal feature distribution matrix.

[0063] The contour mask output by Mask R-CNN is mapped onto a thermal feature distribution matrix for contour-temperature alignment. If the distance between the center of the contour and the center of the high-temperature region is greater than a preset center distance, the matching is considered a failure, and the parameters of the thermal feature distribution matrix are optimized using an optimization algorithm. Otherwise, the matching is considered successful, and the thermal feature distribution matrix that has passed the matching and carries the contour mask (from which coordinate data can be calculated) is used as the matching thermal feature distribution matrix to calculate the positional deviation between it and the optimized 3D coordinate values. When the positional deviation exceeds a preset positional deviation threshold, it is determined that there is a positional deviation between the matching thermal feature distribution matrix and the 3D coordinate values. The contour position is translated along the temperature gradient direction using a gradient descent algorithm with an iteration step size of 0.01 meters and a maximum of 50 iterations to adjust the matching thermal feature distribution matrix, resulting in a thermal feature distribution adjustment matrix. When the positional deviation does not exceed the preset positional deviation threshold, the original thermal feature distribution matrix is ​​output as the thermal feature distribution adjustment matrix.

[0064] The thermal feature distribution adjustment matrix and the optimized 3D coordinate values ​​(0.4:0.46) are weighted and combined to determine the approximate 3D coordinates of the fault point. This invention utilizes multimodal data fusion from infrared and depth cameras to penetrate cable obstructions for precise fault location; combines CNN contour extraction and least-squares optimization for high-precision 3D reconstruction, improving coordinate accuracy; and enhances robustness under complex environmental interference through joint optimization of thermal features and geometric contours. Cluster-based boundary partitioning reduces reliance on manual annotation and adapts to varying fault morphologies.

[0065] In one embodiment, the step of processing the occlusion information data of the fault area collected by the robot dog using a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point further includes:

[0066] Based on the rough three-dimensional coordinates, the occlusion depth data collected by the depth camera is extracted from the multimodal dataset, and then denoised using an image denoising algorithm to obtain denoised depth data. A convolutional neural network is then used to process the denoised depth data to extract the edge features of the faulty components under cable interference, thereby obtaining an edge feature dataset.

[0067] Determine whether the integrity of the edge feature dataset is lower than a preset integrity threshold, and if the integrity of the edge feature dataset is lower than the preset integrity threshold, optimize the denoised depth data through an iterative optimization algorithm to obtain optimized depth data;

[0068] Based on the optimized depth data and the coarse three-dimensional coordinates, the initial three-dimensional coordinates of the faulty component are determined, and based on the initial three-dimensional coordinates, an interpolation algorithm is used to process the edge feature dataset and the optimized depth data to obtain a continuous three-dimensional coordinate path;

[0069] Based on the continuous three-dimensional coordinate path, the spatial distribution features of the faulty component are extracted from the area under the cable, and a distribution feature dataset is obtained. Then, a clustering algorithm is used to divide the regional boundary of the faulty component, and the precise three-dimensional coordinates of the fault point and its corresponding boundary sequence data are obtained.

[0070] Specifically, when obtaining occluded area data from a depth camera on a robotic dog using coarse coordinates, complex interference often exists beneath the cable, such as dense cables, lighting variations, or dust. This invention extracts occluded depth data from a multimodal dataset corresponding to coarse 3D coordinates, filters it using a nonlocal mean denoising algorithm (21×21 search window, 5×5 similarity window), and fills in missing depth pixels using a conditional generative adversarial network (U-Net generator, PatchGAN discriminator). This removes noise from the occluded depth data, such as depth value jumps caused by cable reflection and electromagnetic interference, resulting in denoised depth data. This denoised depth data is then used as input, and a multi-scale edge detection network is employed to extract edge features of faulty components under cable interference, outputting an edge probability map to generate an edge feature dataset. In one instance, a certain point is shown to have a depth of 2.5 meters, but the surrounding points are all at 2.2 meters. This abrupt change may be caused by reflection from occluded objects. When processing this data using image denoising algorithms, mean filtering can be used to remove isolated noise points. For example, the 2.5-meter depth can be smoothly adjusted to around 2.2 meters to form a denoised depth dataset to improve data smoothness.

[0071] Integrity is defined as the total length of closed edges divided by the theoretical maximum length of closed edges (estimated based on rough 3D coordinates of the component size). This is used to assess the integrity of the edge feature dataset. If the integrity is not lower than a preset integrity threshold, the denoised depth data is output as optimized depth data. Otherwise, a particle swarm optimization algorithm is used to optimize the denoised depth data, using the area and continuity of the closed edge region as the fitness function to generate optimized depth data. Furthermore, for the edge feature dataset, if the integrity is lower than a preset threshold (e.g., 80%), and the edge points only cover 50% of the faulty component's outline, an iterative optimization algorithm is used to adjust the denoised depth dataset. Assuming the initial depth data misses some areas, iterative optimization can supplement the missing data by repeatedly comparing edge points and depth values. For example, the depth near (42.1, 42, 2.2) can be adjusted from an invalid value to 2.18 meters to obtain the optimized depth data set.

[0072] The optimized depth map and the coarse 3D coordinates (from infrared-depth fusion) are registered using the ICP algorithm to align the coordinate system. Harris corner points are extracted from the edge map and back-projected into 3D space. The centroid is calculated as the initial 3D coordinates of the faulty component. Based on the initial 3D coordinates, B-spline interpolation is used to fit the edge feature dataset and the discrete edge points in the optimized depth data into a smooth curve. The density of control points is increased in areas with high curvature (such as corners) to obtain a continuous 3D coordinate path formed by a continuous path point set.

[0073] Based on the positions of each point set along a continuous 3D coordinate path, spatial distribution features of the faulty component, such as geometric and depth features, are extracted from the area under the cable that is obstructed. These features are then combined to generate a distribution feature set. A spectral clustering algorithm is used to process this distribution feature set, and a similarity matrix is ​​constructed based on the Euclidean distance of the feature vectors to delineate the regional boundaries of the faulty component. This yields the precise 3D coordinates of the faulty point and its corresponding boundary sequence data, such as a set of polygon vertex coordinates, to describe the shape of the faulty component. Furthermore, when extracting the spatial distribution features of the faulty component from the obstructed area based on the continuous 3D coordinate path data, the height distribution of the path points can be analyzed. For example, points on the west side are concentrated at 2.2 meters, and those on the east side at 2.5 meters, generating a distribution feature dataset. Further, if an abnormal concentration of depth is found in a certain area, it may indicate the shape characteristics of the faulty component, such as circular or strip-shaped.

[0074] This invention effectively eliminates the impact of cable obstruction and electromagnetic noise on depth data through multi-stage denoising and optimization, improves edge feature integrity, and enhances anti-interference capabilities; it combines interpolation and clustering algorithms for high-precision 3D reconstruction, achieving accurate coordinate positioning of faulty components and breaking through the accuracy limitations of traditional single sensors; it uses iterative optimization based on integrity criteria to ensure the generation of reliable 3D models even in extreme obstruction scenarios; and it combines spatial distribution features with clustering for adaptive boundary division, adapting to the diverse fault morphologies in complex cable-intertwined environments.

[0075] For the range of motion of the robotic arm, i.e., the boundary data of the reachable range of the robotic arm, this invention uses the Euclidean distance and direction vector between the precise stopping position of the unmanned vehicle and the precise three-dimensional coordinates of the fault point to determine it. The Euclidean distance is used to quantify the straight-line distance between the unmanned vehicle and the fault point to guide the planning of the robotic arm's extension length. The direction vector is used to determine the spatial direction that the robotic arm needs to cover to guide the adjustment of joint angles. The maximum extension range of the robotic arm is calculated based on the length of each link and the joint angle limitations as a length constraint. Then, LiDAR point cloud data is combined to identify obstacles in the fault area to plan the obstacle avoidance path for the robotic arm. The DH parameter method is used to establish the kinematic model of the robotic arm to define the transformation relationship between the coordinate systems of each joint. The forward kinematics calculation method is used to calculate the position of the end effector of the robotic arm based on the joint angles to verify whether the fault point is covered. Finally, using the Euclidean distance as the cost function, the A* algorithm is used to plan the shortest path of the robotic arm to cover the fault point, and the gradient projection method is used to adjust the Cartesian path to ensure that the robotic arm can still complete the task when there is a joint fault or obstacle, thus obtaining the range of motion of the robotic arm.

[0076] S3. Obtain the real-time coordinate offset data of the robot dog when adjusting its posture, and combine the motion range to quantify the dynamic adjustment parameters of the robotic arm when it aligns with the fault point, and generate operation instructions for the robotic arm based on the dynamic adjustment parameters.

[0077] In one embodiment, acquiring real-time coordinate offset data when the robot dog adjusts its posture, and combining this with the range of motion to quantify the dynamic adjustment parameters when the robotic arm aligns with the fault point, includes:

[0078] The real-time coordinate offset data of the robot dog when it frequently adjusts its posture is obtained and processed by a filtering algorithm to obtain denoised offset data. This data is then combined with the boundary data of the operating platform in the motion range to calculate the initial value of the dynamic adjustment parameter of the robotic arm at the current posture adjustment frequency, thus obtaining the initial set of dynamic adjustment parameters.

[0079] When the deviation between the initial dynamic adjustment parameter set and the preset adjustment threshold exceeds the preset deviation limit, the gradient descent method is used to iteratively optimize the initial dynamic adjustment parameter set to generate an optimized dynamic adjustment parameter set when the robotic arm is aligned with the fault point.

[0080] Based on the optimized dynamic adjustment parameter set and the real-time coordinate offset data, a clustering algorithm is used to partition the motion range to obtain motion range boundary sequence data, which is then combined with the precise three-dimensional coordinates to calculate the mapping relationship between the optimized dynamic adjustment parameter set and the unmanned vehicle, generating a mapping relationship dataset.

[0081] Based on the mapping dataset and the denoised offset data, a linear regression method is used to predict the changing trend of the dynamic adjustment parameters of the robotic arm in the next time period, and a set of predicted adjustment parameters is generated as the output of the dynamic adjustment parameters when the robotic arm is aligned with the fault point.

[0082] Specifically, if the fault is located at a cable junction, the robot dog needs to frequently adjust its posture to avoid obstacles, resulting in a real-time deviation between the fault location data it transmits and the fault coordinates received by the robotic arm.

[0083] Based on this, the present invention utilizes infrared sensors and depth cameras deployed on the robot dog to acquire real-time coordinate offset data when the robot dog frequently adjusts its posture. An adaptive Kalman filter algorithm is then used to process this offset data, resulting in denoised offset data. This denoised offset data is combined with the boundary data of the operating platform within the dynamic range of the robotic arm. Using fuzzy control algorithms or other dynamic adjustment algorithms, the initial values ​​of the joint angular velocity adjustment of the robotic arm at the current posture adjustment frequency are calculated based on the offset, resulting in an initial set of dynamic adjustment parameters. Specifically, the robot dog transmits offset data (0.1, 0.1, 0) and (0.5, 0.5, 0.2) within one second. However, (0.5, 0.5, 0.2) may be an anomaly caused by vibration. Using filtering algorithms such as Kalman filtering to process this data, the anomaly can be smoothly adjusted to (0.2, 0.2, 0.1), forming a denoised offset data set, thereby improving the reliability of the data and providing accurate input for the calculation of dynamic adjustment parameters. When calculating the initial values ​​of the robotic arm's dynamic adjustment parameters using the denoised offset data set and the operation platform boundary data, the current attitude adjustment frequency needs to be considered. For example, if the offset is (0.2, 0.2, 0.1) and the operation platform boundary limits the robotic arm height to within 2.5 meters, and the frequency is 5 times per second, then the initial values ​​can be set as an angle adjustment speed of 10 degrees per second and an extension distance of 0.3 meters per second, forming the initial dynamic adjustment parameter set.

[0084] If the difference between the data in the initial dynamic adjustment parameter set and the preset adjustment threshold exceeds the preset deviation limit, the gradient descent method is used to iteratively optimize the data in the initial dynamic adjustment parameter set to generate an optimized dynamic adjustment parameter set when the robotic arm is aligned with the fault point; otherwise, the initial dynamic adjustment parameter set is output as the optimized dynamic adjustment parameter set.

[0085] Using the optimized dynamic adjustment parameter set and real-time coordinate offset data collected by the robot dog as input, features such as the end-effector trajectory point density, joint angle change rate, and pose offset variance are extracted. A Gaussian Mixture Model (GMM) is used to partition the robot arm's motion range, outputting motion range boundary sequence data to describe the geometric extent of each partition. The end-effector coordinates of the robot arm in each partition are transformed to the same coordinate system as the precise 3D coordinates of the autonomous vehicle. Based on the transformed data, a nonlinear regression model is constructed using a random forest regression algorithm (this model takes the optimized dynamic adjustment parameter set and the precise 3D coordinates of the autonomous vehicle as input, and the end-effector coordinates of the robot arm in each partition as output) to quantify the mapping relationship between the optimized dynamic adjustment parameter set and the precise 3D coordinates of the autonomous vehicle, obtaining a mapping relationship dataset. Alternatively, K-means clustering can be used to divide the reachable adjustment area of ​​the robot arm within the boundary of the operating platform, dividing the offset and boundary data into near-end and far-end groups, generating adjustment area boundary sequence data such as (51,51,1) and (53,53,2).

[0086] Using denoised offset data and mapping relationship dataset as input, linear regression is used to predict the changing trend of the robotic arm's dynamic adjustment parameters in the next time period. For example, the adjustment speed may increase from 10 degrees per second to 12 degrees per second. The predicted adjustment parameter set is obtained as the output of the dynamic adjustment parameters when the robotic arm aligns with the fault point. By predicting the trend, the robotic arm trajectory is corrected in advance, reducing the deviation between the fault position data returned by the robot dog and the fault coordinates received by the robotic arm. It also improves the robotic arm's response foresight, reduces the delay of real-time adjustment, and improves the overall efficiency.

[0087] This invention ensures that the robotic arm can respond quickly even when the robot dog's posture changes frequently through real-time filtering and parameter iteration optimization, thereby improving alignment accuracy. It also effectively suppresses the impact of environmental noise (such as vibration and electromagnetic interference) on the adjustment parameters by combining clustering and regression prediction. Furthermore, it establishes a mapping relationship between the robot dog, robotic arm, and unmanned vehicle to achieve precise coordination under the motion coupling of multiple devices. By predicting trends in advance, it corrects the robotic arm trajectory, reduces action delay, and improves emergency repair efficiency.

[0088] In one embodiment, generating operation commands for the robotic arm based on the dynamically adjusted parameters includes:

[0089] Based on the dynamic adjustment parameters, the joint angle change sequence of the robotic arm is calculated using an inverse kinematics algorithm, and the joint angle change sequence is adjusted according to the parking deviation between the precise parking position and the actual parking position of the unmanned vehicle, to obtain a joint angle adjustment dataset.

[0090] When the joint angle changes in the joint angle adjustment dataset exceed the physical limitations of the robotic arm, the precise docking position is fine-tuned to update the joint angle adjustment dataset, resulting in a set of deviation-eliminated angle changes.

[0091] The set of deviation elimination angle changes is fused with the range of motion to obtain a joint control sequence after dynamic parameter compensation. This sequence is then combined with the finely adjusted precise parking position of the unmanned vehicle and processed by a clustering algorithm to divide the boundary region of the operation command set, thereby generating a preliminary command set.

[0092] Based on the preliminary instruction set and the joint control sequence, the linear regression method is used to predict the trend of joint angle change of the robotic arm in the next time period, and the predicted angle adjustment set is obtained. This set is then combined with the real-time docking deviation of the unmanned vehicle and processed by a filtering algorithm to obtain a denoised instruction set as the operation instruction output for the robotic arm.

[0093] Specifically, since the robotic arm relies on the unmanned vehicle platform for stable support to carry out inspection and repair, if the unmanned vehicle's position is deviated, the robotic arm's operating range is limited, making it difficult to perform repair actions at the fault point.

[0094] Based on this, this invention uses the Jacobian matrix pseudo-inverse method to select the optimal solution by minimizing the weighted norm of the joint angles, according to the obtained dynamic adjustment parameters, to obtain the joint angle change sequence of the robotic arm. It then calculates the docking deviation based on the precise and actual docking positions of the unmanned vehicle, converts the docking deviation into displacement cases in the robotic arm's base coordinate system, updates the joint angle change sequence, and outputs a joint angle adjustment dataset. The core of calculating the robotic arm joint angle change sequence lies in converting the positional requirements of the end effector into angle adjustments for each joint. Assuming the robotic arm end effector needs to be aligned with the fault point coordinates (52, 52, 2), and the current docking position of the unmanned vehicle is (50, 50, 0), inverse kinematics, using the known robotic arm length and joint degrees of freedom, derives the joint angle sequence from the initial posture to the target posture. For example, the shoulder joint adjusts from 30 degrees to 35 degrees, and the elbow joint changes from 45 degrees to 50 degrees. When obtaining compensation data from the parking deviation of the unmanned vehicle, it is assumed that the actual parking position of the unmanned vehicle is offset to (50.2, 50.1, 0.1) due to uneven ground, with a deviation of (0.2, 0.1, 0.1). The compensation data adjusts the initial value of the joint angle through the direction and magnitude of the deviation, so that the end of the robotic arm is aligned with the fault point, thereby improving the adaptability to external disturbances.

[0095] When the joint angle changes in the joint angle adjustment dataset exceed the physical limitations of the robotic arm, the goal is to minimize the docking position adjustment. The joint angles of the robotic arm are subject to behavioral constraints. A sequential quadratic programming method is used to fine-tune the precise docking position to iteratively update the joint angle adjustment dataset. The iteration ends when all joint angles in the updated dataset satisfy the constraints. The final updated joint angle adjustment dataset is then output as the set of angle changes with reduced deviation. This fine-tuning of the autonomous vehicle's position avoids the risk of downtime due to exceeding limits, improving operational continuity.

[0096] Based on the range of motion of the robotic arm, a speed limit is applied to the set of deviation elimination angle changes to obtain the joint control sequence after dynamic parameter compensation. The joint angle, speed, and acceleration statistics (mean, variance) are extracted from the sequence and combined with the precise stopping position of the unmanned vehicle after fine-tuning as features. The K-means algorithm is used to process these features to divide the boundary regions of the operation command set. The clustering algorithm divides the command subsets according to feature similarity. For example, the joint angle data is divided into the proximal region (angle less than 40 degrees) and the distal region (angle greater than 40 degrees), generating boundary points such as (51,51,1) and (53,53,2), and then the preliminary command set is output.

[0097] The initial instruction set and joint control sequence are used as inputs, and the linear regression method is used to predict the trend of joint angle change of the robotic arm in the next time period. The predicted angle adjustment set is output to improve the robotic arm's forward adjustment capability. The predicted angle adjustment set is combined with the real-time docking deviation of the unmanned vehicle and then denoised by Kalman filtering to output a smoothed denoised instruction set as the operation instruction output for the robotic arm.

[0098] This invention achieves high-precision dynamic compensation and high-precision positioning of the robotic arm's end effector through joint optimization of inverse kinematics and docking deviation; it ensures that joint angles do not exceed limits through iterative adjustments, avoiding mechanical damage or motion failure; it improves operational reliability in complex environments by fusing the unmanned vehicle's position with the robotic arm's range of motion; it generates predictive commands by combining trend prediction and filtering noise reduction, reducing motion delay and jitter and enhancing stability; and it adopts a multi-faceted collaborative approach, from deviation compensation to command mapping, progressively supporting the goal of high-precision and high-stability operation. This not only improves the robotic arm's response speed and alignment accuracy but also reduces the probability of failure due to environmental changes, making it suitable for complex maintenance scenarios.

[0099] S4. The real-time coordinate offset data is processed using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point, and obtain fault position data synchronized with the operation command.

[0100] In one embodiment, step S4 includes:

[0101] The real-time coordinate offset data is processed by time series analysis to obtain the real-time offset change trend. The data is then processed by Kalman filtering algorithm to output the predicted coordinates of the robot dog relative to the fault point. These coordinates are then fused with the operation command to obtain a synchronous fault location dataset.

[0102] Clustering algorithm is used to process the synchronous fault location dataset to obtain a spatial distribution dataset of coordinate offsets. This dataset is then combined with the real-time coordinate offset data and processed by a linear regression algorithm to calculate the trend offset of the fault point coordinates, resulting in a trend-adjusted fault coordinate prediction dataset.

[0103] Real-time noise data is extracted from the real-time coordinate offset data, and the real-time noise data is processed using the Kalman filter algorithm. The processing result is used to adjust the fault coordinate prediction dataset to obtain a denoised coordinate prediction dataset.

[0104] The mapping relationship between the denoised coordinate prediction dataset and the operation command is quantified to adjust the operation command and obtain the adjusted operation command. The coordinate prediction continuous sequence of the robot dog relative to the fault point is generated by the interpolation algorithm and used as fault location data synchronized with the operation command.

[0105] Specifically, this invention uses the ARIMA (Autoregressive Integral Moving Average) model, taking the real-time coordinate offset sequence of the robot dog as input, predicting the real-time offset change trend within a preset time period, and processing it with the Kalman filter algorithm to output the predicted coordinates of the fault point and align them with the timestamp of the operation command for fusion, thus obtaining a synchronized fault location dataset.

[0106] Using the predicted coordinates in the synchronous fault location dataset as input, the K-means clustering algorithm is used to divide the region into near and far regions. Assuming that the center of the near region is (52,52,2) and the center of the far region is (53,53,2.2) after the division, the spatial distribution dataset of coordinate offset is determined, the distribution characteristics of the fault point are clarified, and the spatial distribution dataset of coordinate offset is used as the dependent variable and time is used as the independent variable for linear regression modeling. The weights are solved by the least squares method to calculate the trend offset of the fault point coordinates and output the trend-adjusted fault coordinate prediction dataset.

[0107] Real-time noise data is extracted from the real-time coordinate offset data, and the Kalman filter algorithm is used to process the real-time noise data. The processing result is used to correct the fault coordinate prediction dataset, resulting in a denoised coordinate prediction dataset. Assuming the coordinate prediction dataset is (52.18, 52.18, 2.18) and the noise data is 0.03 meters, it is adjusted to 0.01 meters after filtering, generating a denoised fault location dataset, such as (52.17, 52.17, 2.17).

[0108] Define the association function between the operation command and the denoised coordinate prediction dataset to model using a multilayer perceptron. When the change in the predicted coordinates of the denoised coordinate prediction dataset exceeds a preset change threshold, scale the command parameters proportionally to obtain the adjusted operation command. Then, use an interpolation algorithm to generate a continuous sequence of predicted coordinates of the robot dog relative to the fault point as output as fault location data (mainly the location data of the faulty component in the fault point) synchronized with the operation command.

[0109] This invention achieves millisecond-level synchronization between the robot dog's position and operation commands through the dynamic fusion of time series analysis and Kalman filtering, ensuring precise action. Joint clustering and noise modeling effectively suppress the impact of environmental interference (such as vibration and electromagnetic noise) on coordinate prediction. Trend offset calculation based on linear regression adaptively corrects prediction deviations, improving long-term tracking stability. A smooth prediction sequence is generated through interpolation algorithms, avoiding the risk of mechanical jitter or abrupt changes during command execution. The entire process, from coordinate offset acquisition to command optimization, is progressively streamlined, not only improving the accuracy of fault location but also enhancing the system's adaptability to complex environments and reducing the risk of misoperation due to noise or deviation.

[0110] S5. Generate capture data based on the operation instructions and the fault location data to separate the faulty components from the faulty area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.

[0111] In one embodiment, step S5 includes:

[0112] Based on the fault location data, the closed-loop control system of the robotic arm generates the initial grasping angle and force data of the end effector, and obtains the environmental data of the faulty component according to the fault location data to divide the region using a clustering algorithm to obtain the faulty component region.

[0113] The initial grasping angle and force data are processed using an interpolation algorithm to generate continuous end effector motion trajector ...

[0114] Real-time environmental data of the faulty component area is acquired and denoised using a Kalman filter algorithm. The denoising results are then used to adjust the action sequence dataset to obtain a denoised emergency repair action dataset.

[0115] The dynamic environmental change features are extracted from the real-time environmental data, and the fault location data is processed by the interpolation algorithm to generate a continuous fault component separation path to adjust the noise-reducing emergency repair action dataset, so as to obtain the final execution action dataset and use it as the capture data.

[0116] Based on the grasping data, the robotic arm is controlled to separate the faulty component from the faulty area and perform emergency repairs, thereby realizing the inspection of the power distribution facilities.

[0117] Specifically, the core of generating grasping data lies in converting position information into executable mechanical actions. This invention uses fault location data and the geometric parameters (size, material) of the faulty component as input data. The closed-loop control system of the robotic arm, based on the position data of the faulty component and the end effector pose of the robotic arm, solves the initial gripper angle through inverse kinematics and presets the gripping force according to the material hardness (such as rubber / metal) to obtain the initial grasping angle and force data of the end effector of the robotic arm. Based on the position data of the faulty component, the environmental data of the faulty component in the location area (cable distribution and temperature field around the faulty component, etc.) is obtained, and OPTICS clustering is used to divide the safe operating area (low temperature, low density) and the risk area (high temperature, high density) to obtain the faulty component area and its boundary data.

[0118] The discrete grasping angle sequence in the initial grasping angle and force data is converted into a time-angle curve, and cubic spline interpolation is used to ensure the continuity of angle, angular velocity, and angular acceleration to avoid mechanical impact. The clamping force data is linearly interpolated to ensure that the force change rate is within the motor response range, and a continuous end effector motion trajectory is generated to construct an implementation action sequence dataset.

[0119] Real-time environmental data of the faulty component area is acquired and denoised using a Kalman filter algorithm. The resulting filtered environmental data is used to adjust the action sequence dataset, avoiding risky areas, and outputting a denoised emergency repair action dataset. Specifically, when acquiring real-time interference data from a complex environment, it is assumed that wind or vibration causes a positional shift of 0.02 meters. When processing environmental noise using the Kalman filter algorithm, the shift is smoothed from 0.02 meters to 0.01 meters by fusing predicted and observed values, resulting in the denoised emergency repair action dataset.

[0120] The system extracts dynamic environmental features from real-time environmental data, such as spatial features (cable movement speed, displacement vector of faulty components) and temporal features (temperature change rate, vibration spectrum dominant frequency, etc.). It then uses PCA to reduce dimensionality and generate multi-dimensional feature vectors to characterize these dynamic changes. Based on these multi-dimensional feature vectors, a potential field model is constructed around the faulty component, with attractive forces pointing in the separation direction and repulsive forces originating from high-risk areas. B-spline curves are used to fit path points, ensuring continuous curvature and smooth robotic arm joint movements, generating continuous separation paths for the faulty component. Dynamic time warping is then employed to align the separation paths with the action sequence time axis, compensating for robotic arm delays and adjusting the denoised repair action dataset. This results in the final executed action dataset, which serves as the grasping data, containing time-synchronized trajectories and force commands. Finally, based on the grasping data, the robotic arm is controlled to separate the faulty component from the fault area for repair, thereby achieving the inspection of power distribution facilities.

[0121] This invention achieves high-precision grasping and separation of faulty components through closed-loop control and dynamic path planning, avoiding secondary damage to surrounding cables; real-time environmental perception and data denoising ensure stable operation even in complex electromagnetic interference and vibration environments; smooth separation paths are generated based on interpolation and feature extraction, reducing robotic arm jitter and improving action continuity; the entire process from positioning to separation is automated, which can significantly shorten the power distribution facility fault recovery time.

[0122] This application proposes a collaborative inspection method for power distribution facilities using an unmanned vehicle, a robotic dog, and a robotic arm, addressing the issue of improving the inspection efficiency and safety of power distribution facilities in high-density cable areas. The method determines the precise stopping position of the unmanned vehicle based on the three-dimensional cable distribution data collected by the unmanned vehicle in the high-density cable fault area. The obstruction information data collected by the robotic dog in the narrow area under the cable in the fault area is processed by a convolutional neural network to eliminate cable obstruction interference, thereby accurately identifying the three-dimensional coordinates of the faulty component and determining the range of motion of the robotic arm, which relies on the stable support provided by the unmanned vehicle platform. The method also utilizes the coordinate offset data transmitted back by the robotic dog when it frequently adjusts its posture to avoid obstacles (i.e., the coordinate offset data transmitted by the robotic dog). The system combines the fault location data returned by the robot dog with the fault coordinates received by the robot arm. By integrating the robot arm's range of motion, the dynamic adjustment parameters when the robot arm aligns with the fault point are quantified, thereby generating operation commands for the robot arm. Time series analysis and filtering algorithms are used to process the coordinate offset data returned by the robot dog to predict the accurate position of the robot dog relative to the fault point and synchronized with the operation command time when the unmanned vehicle needs to temporarily adjust its position to adapt to the robot arm's needs. This prediction is then combined with the operation command to control the robot arm to generate a grasping scheme to separate and repair faulty components in the fault area, thereby achieving effective inspection of power distribution facilities. This solution can achieve automated and precise emergency repair in complex cable environments, improving work efficiency and safety.

[0123] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0124] In another embodiment, as shown in FIG2, a second aspect of the present invention provides a power distribution facility inspection system that integrates an unmanned vehicle, a robot dog, and a robotic arm, comprising:

[0125] The location determination module 10 is used to determine the initial parking area of ​​the unmanned vehicle based on the three-dimensional distribution data of the cables in the fault area corresponding to the power distribution facility, and to quantify the movement trajectory of the unmanned vehicle in the fault area according to the initial parking area, so as to determine the precise parking position of the unmanned vehicle relative to the fault point.

[0126] The range quantization module 20 is used to process the occlusion information data of the fault area collected by the robot dog through a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, and to quantify the range of motion of the robotic arm based on the relative relationship between the precise docking position and the precise three-dimensional coordinates.

[0127] The instruction generation module 30 is used to acquire real-time coordinate offset data when the robot dog adjusts its posture, so as to combine the motion range to quantify the dynamic adjustment parameters when the robotic arm is aligned with the fault point, and generate operation instructions for the robotic arm based on the dynamic adjustment parameters.

[0128] The position synchronization module 40 is used to process the real-time coordinate offset data using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point and obtain fault position data synchronized with the operation command.

[0129] The fault repair module 50 is used to generate capture data based on the operation instructions and the fault location data, so as to separate the faulty components from the fault area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.

[0130] It should be noted that each module in the aforementioned power distribution facility inspection system involving unmanned vehicles, robotic dogs, and robotic arms can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module. For specific limitations regarding the power distribution facility inspection system involving unmanned vehicles, robotic dogs, and robotic arms, please refer to the limitations of the power distribution facility inspection method involving unmanned vehicles, robotic dogs, and robotic arms described above; both have the same function and role, and will not be repeated here.

[0131] In summary, this invention relates to the field of information technology and discloses a method and system for inspecting power distribution facilities using an unmanned vehicle, a robot dog, and a robotic arm in collaboration. The method utilizes an unmanned vehicle to acquire three-dimensional distribution data of cables in a faulty area, and achieves precise parking of the unmanned vehicle in confined spaces through path planning and dynamic adjustment of the chassis height. The robot dog detects obstructed areas to extract features of faulty components and determine their precise coordinates. The reachable range of the robotic arm mounted on the unmanned vehicle is quantified, and the joint angle changes of the robotic arm are calculated. After the unmanned vehicle fine-tunes its position to eliminate deviations, the operation command is determined. Based on the predicted real-time position change trend of the robot dog relative to the fault point, fault location data synchronized with the operation command is determined, allowing the robotic arm to generate a grasping scheme for precise implementation of emergency repair actions. This achieves automated and precise emergency repair in complex cable environments, improving operational efficiency and safety.

[0132] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0133] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for inspecting power distribution facilities using a collaborative approach involving an unmanned vehicle, a robot dog, and a robotic arm, characterized in that: include: Based on the three-dimensional distribution data of cables in the fault area of ​​the power distribution facility, the initial parking area of ​​the unmanned vehicle is determined, and the movement trajectory of the unmanned vehicle in the fault area is quantified according to the initial parking area to determine the precise parking position of the unmanned vehicle relative to the fault point. The occlusion information data of the fault area collected by the robot dog is processed by a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, so as to quantify the range of motion of the robotic arm with the precise stopping position. The real-time coordinate offset data of the robot dog when adjusting its posture is obtained, and the dynamic adjustment parameters of the robotic arm when it is aligned with the fault point are quantified in combination with the range of motion. Based on the dynamic adjustment parameters, operation instructions for the robotic arm are generated. The real-time coordinate offset data is processed using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point, and obtain fault position data synchronized with the operation command. Based on the operation instructions and the fault location data, capture data is generated to separate the faulty components from the fault area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.

2. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The determination of the initial parking area for the unmanned vehicle based on the three-dimensional cable distribution data corresponding to the fault area of ​​the power distribution facility includes: The unmanned vehicle uses a multi-view lidar to collect three-dimensional cable distribution data in the fault area corresponding to the power distribution facility; the fault area is a ground space-constrained environment area including the high-density cable crossing area of ​​the power distribution facility. The three-dimensional distribution data of the cable is processed using a point cloud distance calculation algorithm to obtain the height data and position coordinates of each cable in the fault area, so as to analyze the density distribution pattern of each cable in the fault area and obtain cable density distribution characteristic data. The fault area is divided based on the cable density distribution feature data, and the area where the cable height is lower than a preset height threshold is marked as the parking candidate area of ​​the unmanned vehicle, thus obtaining a set of candidate areas; Extract target candidate regions that match the height requirements of the unmanned vehicle from the candidate region set, and use a geometric constraint algorithm to determine the spatial connectivity of each target candidate region to obtain the range of connected regions; Based on the range of the connected region and the scanning boundary conditions of the multi-view lidar, the coordinates of the parking area are determined, and a machine learning clustering algorithm is used to optimize and group the target candidate regions within the range of the connected region to generate the initial parking area of ​​the unmanned vehicle.

3. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 2, characterized in that, The step of quantifying the movement trajectory of the unmanned vehicle in the fault area based on the initial parking area to determine the precise parking position of the unmanned vehicle relative to the fault point includes: Based on the initial docking area, the cable density distribution characteristic data, and the height data of each cable, the movement trajectory of the unmanned vehicle under the spatially confined conditions corresponding to the fault area is quantified to obtain an initial trajectory coordinate sequence. When it is determined that there is a deviation between the initial trajectory coordinate sequence and the fault area, the model predictive control method is used to generate adjustment parameters for the chassis height of the unmanned vehicle, and the trajectory adjustment coordinate sequence is obtained. Based on the trajectory adjustment coordinate sequence and the height data of each cable, the cable height change characteristics within the fault point and its preset area are extracted, so as to determine the area division range of the fault point through a clustering algorithm; Based on the defined area and the spatial constraints, a target trajectory adjustment coordinate sequence that meets the connectivity requirements is selected from the trajectory adjustment coordinate sequence and combined to generate a continuous movement trajectory path. This path is then combined with the cable density distribution feature data to determine the precise stopping position of the unmanned vehicle relative to the fault point.

4. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The process of processing the occlusion information data of the fault area collected by the robot dog using a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point includes: Based on the precise docking location, infrared sensors and depth cameras deployed on the robot dog are used to acquire occlusion information data of the area under the cable, resulting in a multimodal data set which is then input into a convolutional neural network for processing, outputting a contour feature dataset. Based on the contour feature dataset and the parameter information of the depth camera, determine the three-dimensional coordinate values ​​of the faulty component; Based on the three-dimensional coordinate values ​​and the shielding area under the cable, a clustering algorithm is used to divide the region boundary of the faulty component, and a boundary division coordinate sequence is obtained to determine the thermal feature distribution matrix of the faulty component according to the multimodal data set; Based on the contour feature dataset, a matching thermal feature distribution matrix is ​​extracted from the thermal feature distribution matrix. When it is determined that there is a positional deviation between the matching thermal feature distribution matrix and the three-dimensional coordinate values, the matching thermal feature distribution matrix is ​​adjusted using a gradient descent algorithm to obtain a thermal feature distribution adjustment matrix. This matrix is ​​then combined with the three-dimensional coordinate values ​​to determine the approximate three-dimensional coordinates of the fault point.

5. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 4, characterized in that, The step of processing the occlusion information data of the fault area collected by the robot dog through a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point also includes: Based on the rough three-dimensional coordinates, the occlusion depth data collected by the depth camera is extracted from the multimodal dataset, and then denoised using an image denoising algorithm to obtain denoised depth data. A convolutional neural network is then used to process the denoised depth data to extract the edge features of the faulty components under cable interference, thereby obtaining an edge feature dataset. Determine whether the integrity of the edge feature dataset is lower than a preset integrity threshold, and if the integrity of the edge feature dataset is lower than the preset integrity threshold, optimize the denoised depth data through an iterative optimization algorithm to obtain optimized depth data; Based on the optimized depth data and the coarse three-dimensional coordinates, the initial three-dimensional coordinates of the faulty component are determined, and based on the initial three-dimensional coordinates, an interpolation algorithm is used to process the edge feature dataset and the optimized depth data to obtain a continuous three-dimensional coordinate path; Based on the continuous three-dimensional coordinate path, the spatial distribution features of the faulty component are extracted from the area under the cable, and a distribution feature dataset is obtained. Then, a clustering algorithm is used to divide the regional boundary of the faulty component, and the precise three-dimensional coordinates of the fault point and its corresponding boundary sequence data are obtained.

6. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The step of acquiring real-time coordinate offset data when the robot dog adjusts its posture, and combining this with the range of motion to quantify the dynamic adjustment parameters when the robotic arm aligns with the fault point, includes: The real-time coordinate offset data of the robot dog when it frequently adjusts its posture is obtained and processed by a filtering algorithm to obtain denoised offset data. This data is then combined with the boundary data of the operating platform in the motion range to calculate the initial value of the dynamic adjustment parameter of the robotic arm at the current posture adjustment frequency, thus obtaining the initial set of dynamic adjustment parameters. When the deviation between the initial dynamic adjustment parameter set and the preset adjustment threshold exceeds the preset deviation limit, the gradient descent method is used to iteratively optimize the initial dynamic adjustment parameter set to generate an optimized dynamic adjustment parameter set when the robotic arm is aligned with the fault point. Based on the optimized dynamic adjustment parameter set and the real-time coordinate offset data, a clustering algorithm is used to partition the motion range to obtain motion range boundary sequence data, which is then combined with the precise three-dimensional coordinates to calculate the mapping relationship between the optimized dynamic adjustment parameter set and the unmanned vehicle, generating a mapping relationship dataset. Based on the mapping dataset and the denoised offset data, a linear regression method is used to predict the changing trend of the dynamic adjustment parameters of the robotic arm in the next time period, and a set of predicted adjustment parameters is generated as the output of the dynamic adjustment parameters when the robotic arm is aligned with the fault point.

7. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The step of generating operation commands for the robotic arm based on the dynamically adjusted parameters includes: Based on the dynamic adjustment parameters, the joint angle change sequence of the robotic arm is calculated using an inverse kinematics algorithm, and the joint angle change sequence is adjusted according to the parking deviation between the precise parking position and the actual parking position of the unmanned vehicle, to obtain a joint angle adjustment dataset. When the joint angle changes in the joint angle adjustment dataset exceed the physical limitations of the robotic arm, the precise docking position is fine-tuned to update the joint angle adjustment dataset, resulting in a set of deviation-eliminated angle changes. The set of deviation elimination angle changes is fused with the range of motion to obtain a joint control sequence after dynamic parameter compensation. This sequence is then combined with the finely adjusted precise parking position of the unmanned vehicle and processed by a clustering algorithm to divide the boundary region of the operation command set, thereby generating a preliminary command set. Based on the preliminary instruction set and the joint control sequence, the linear regression method is used to predict the trend of joint angle change of the robotic arm in the next time period, and the predicted angle adjustment set is obtained. This set is then combined with the real-time docking deviation of the unmanned vehicle and processed by a filtering algorithm to obtain a denoised instruction set as the operation instruction output for the robotic arm.

8. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The process of using time series analysis and Kalman filtering to process the real-time coordinate offset data to predict the real-time position change trend of the robot dog relative to the fault point, and obtaining fault position data synchronized with the operation command, includes: The real-time coordinate offset data is processed by time series analysis to obtain the real-time offset change trend. The data is then processed by Kalman filtering algorithm to output the predicted coordinates of the robot dog relative to the fault point. These coordinates are then fused with the operation command to obtain a synchronous fault location dataset. Clustering algorithm is used to process the synchronous fault location dataset to obtain a spatial distribution dataset of coordinate offsets. This dataset is then combined with the real-time coordinate offset data and processed by a linear regression algorithm to calculate the trend offset of the fault point coordinates, resulting in a trend-adjusted fault coordinate prediction dataset. Real-time noise data is extracted from the real-time coordinate offset data, and the real-time noise data is processed using the Kalman filter algorithm. The processing result is used to adjust the fault coordinate prediction dataset to obtain a denoised coordinate prediction dataset. The mapping relationship between the denoised coordinate prediction dataset and the operation command is quantified to adjust the operation command and obtain the adjusted operation command. The coordinate prediction continuous sequence of the robot dog relative to the fault point is generated by the interpolation algorithm and used as fault location data synchronized with the operation command.

9. The method for inspecting power distribution facilities using a collaborative approach of unmanned vehicle, robot dog, and robotic arm as described in claim 1, characterized in that, The process of generating capture data based on the operation instructions and the fault location data to isolate faulty components from the fault area and perform emergency repairs, thereby achieving the inspection of the power distribution facilities, includes: Based on the fault location data, the closed-loop control system of the robotic arm generates the initial grasping angle and force data of the end effector, and obtains the environmental data of the faulty component according to the fault location data to divide the region using a clustering algorithm to obtain the faulty component region. The initial grasping angle and force data are processed using an interpolation algorithm to generate continuous end effector motion trajector ... Real-time environmental data of the faulty component area is acquired and denoised using a Kalman filter algorithm. The denoising results are then used to adjust the action sequence dataset to obtain a denoised emergency repair action dataset. The dynamic environmental change features are extracted from the real-time environmental data, and the fault location data is processed by the interpolation algorithm to generate a continuous fault component separation path to adjust the noise-reducing emergency repair action dataset, so as to obtain the final execution action dataset and use it as the capture data. Based on the grasping data, the robotic arm is controlled to separate the faulty component from the faulty area and perform emergency repairs, thereby realizing the inspection of the power distribution facilities.

10. A power distribution facility inspection system that integrates an unmanned vehicle, a robot dog, and a robotic arm, characterized in that: include: The location determination module is used to determine the initial parking area of ​​the unmanned vehicle based on the three-dimensional distribution data of the cables in the fault area corresponding to the power distribution facility, and to quantify the movement trajectory of the unmanned vehicle in the fault area according to the initial parking area, so as to determine the precise parking position of the unmanned vehicle relative to the fault point. The range quantization module is used to process the occlusion information data of the fault area collected by the robot dog through a convolutional neural network to obtain the precise three-dimensional coordinates of the fault point, and to quantify the range of motion of the robotic arm based on the relative relationship between the precise docking position and the precise three-dimensional coordinates. The instruction generation module is used to acquire real-time coordinate offset data when the robot dog adjusts its posture, combine the motion range to quantify the dynamic adjustment parameters when the robotic arm aligns with the fault point, and generate operation instructions for the robotic arm based on the dynamic adjustment parameters. The position synchronization module is used to process the real-time coordinate offset data using time series analysis and Kalman filtering algorithm to predict the real-time position change trend of the robot dog relative to the fault point, and obtain fault position data synchronized with the operation command. The fault repair module is used to generate capture data based on the operation instructions and the fault location data, so as to separate the faulty components from the fault area and carry out emergency repairs, thereby realizing the inspection of the power distribution facilities.