A control method of a tower climbing robot for power inspection, the tower climbing robot, and a computer readable medium

Abstract

The application provides a control method of a tower climbing robot for power inspection, the tower climbing robot and a computer storage medium, the control method is deployed locally and runs, and comprises the following steps: a perception node extracts semantic information and three-dimensional structure features of multi-modal sensing data, and publishes structured data to an asynchronous data channel; a model registration node fuses the structured data and pre-stored standard three-dimensional models of power towers in cross-modal features, calculates a registration transformation matrix to construct and publish a global registration model; a decision node generates and publishes a work strategy instruction in combination with the structured data, model information and state parameters of the decision node; and a control node generates a trajectory and solves a pose according to the work strategy instruction and outputs a motion control instruction. The application realizes localized decoupling operation of nodes, eliminates weak network transmission delay risks, improves system fault isolation capability and climbing stability, and is suitable for a weak network power inspection scene in the wild.

Description

Technical Field

[0001] This invention relates to the field of power maintenance, and more particularly to a control method for a tower-climbing robot used for power inspection, the tower-climbing robot, and a computer-readable medium. Background Technology

[0002] High-voltage transmission towers, as key physical supports for the backbone of the power grid, are exposed to complex outdoor environments for extended periods. Years of exposure to sun, rain, sudden temperature changes, and extreme weather conditions make the towers and their components highly susceptible to corrosion, deformation, and loosening. This not only threatens the safe and stable operation of the lines but also demands extremely frequent routine maintenance and inspections. Traditional methods relying on manual tower climbing for high-altitude work often involve significant personal safety risks and are inefficient and lack standardization. To overcome the limitations of manual inspections, the adoption of power inspection climbing robots to replace manual high-altitude detection and refined maintenance operations has become an inevitable trend and core component of the current intelligent upgrading of power systems.

[0003] However, existing tower climbing robot control systems still face multiple bottlenecks in practical applications. Current technologies focus on two main approaches: a localized approach, employing a highly coupled centralized architecture where core modules such as perception, decision-making, and control are interdependent. If front-end perception experiences processing delays due to environmental interference, it can easily lead to system-wide response lag or even crashes, resulting in weak overall system fault tolerance. The other approach is a cloud-based approach, relying on remote cloud servers for data interaction and command processing. In remote, poorly networked outdoor environments, this cross-platform communication often leads to severe command delays, failing to meet the stringent real-time requirements of dynamic high-altitude operations. Furthermore, existing robot perception methods are mostly limited to fragmented, localized detection, making it difficult to establish an effective spatial mapping between real-time survey images and global standard structures. This limits the device's positioning accuracy on complex towers, resulting in poor collaborative performance in autonomous obstacle crossing and precise alignment.

[0004] In summary, existing climbing robots have certain inherent limitations in terms of software control architecture and underlying computing dependencies, and there is an urgent need for a new control method that can break free from communication constraints and enable flexible and stable operation of system modules. Summary of the Invention

[0005] In order to overcome the above-mentioned technical defects, the purpose of this invention is to provide a control method for a tower climbing robot used for power line inspection, the tower climbing robot, and a computer storage medium.

[0006] This invention discloses a control method for a tower-climbing robot used in power line inspection. The control method is deployed locally on the tower-climbing robot and includes: The sensing node acquires multimodal sensing data collected in real time by the tower climbing robot; extracts semantic features and three-dimensional structural features from the multimodal sensing data; and publishes the structured data composed of semantic features and three-dimensional structural features to the asynchronous data channel. The model registration node listens to the asynchronous data channel and obtains structured data. It then performs cross-modal feature fusion with the pre-stored standard 3D structural model of the power tower, calculates the registration transformation matrix, constructs a global registration model under a unified global coordinate system, and publishes the global registration model information to the asynchronous data channel. The decision node listens to the asynchronous data channel and obtains structured data, global registration model information, and the robot's own state parameters. Based on the above data analysis, it generates operation strategy instructions and then issues the operation strategy instructions to the asynchronous data channel. The control node listens to the asynchronous data channel and obtains global registration model information and operation strategy instructions. Based on a unified global coordinate system, it generates trajectory and calculates pose to generate motion control instructions that drive the actuator of the tower climbing robot.

[0007] Preferably, the structured data is fused with a pre-stored standard 3D structural model of the power tower through cross-modal feature fusion, and the registration transformation matrix is ​​calculated, including: Based on semantic features, components of the same category are retrieved between structured data and standard three-dimensional structural models of power towers to establish an initial set of correspondences in order to complete a coarse semantic matching. Geometric features were extracted from the structured data and the standard three-dimensional structural model of the power tower, and combined with the initial correspondence set, the initial registration transformation matrix was estimated using the random sampling consensus algorithm.

[0008] Preferably, the cross-modal feature fusion of structured data and pre-stored standard three-dimensional structural models of power towers, and the calculation of the registration transformation matrix, further include: Based on the initial registration transformation matrix, the iterative nearest point algorithm is used for further iteration; Among them, the objective function of the iterative nearest point algorithm introduces geometric constraint regularization term and semantic constraint regularization term. The geometric constraint regularization term is used to constrain the consistency of normal vector and distance between corresponding point pairs, and the semantic constraint regularization term is used to assign weight penalty according to the consistency of semantic label of corresponding point pairs. When the preset iteration termination condition is met, the result is output as the final registration transformation matrix.

[0009] Preferably, the process of generating work strategy instructions based on the above data analysis includes: We perform weighted fusion and confidence assessment on semantic features, 3D structural features, global registration model information, robot body posture parameters, and equipment operating status parameters; Based on the evaluation results, the system transitions between multiple preset work states and outputs work strategy instructions corresponding to the current work state. The work states include climbing, obstacle crossing, and alignment states under normal working conditions, as well as protection states under abnormal working conditions.

[0010] Preferably, the abnormal operating conditions include: the base adhesion force of the tower climbing robot is lower than the first preset threshold, the body attitude angle of the tower climbing robot exceeds the second preset threshold and continues for the first preset time, or the computing power load in the equipment operating status parameters exceeds the third preset threshold and continues for the second preset time. The operating strategy instructions include shutdown lockout instructions and / or safety reset instructions.

[0011] Preferably, the control method further includes: Configure differentiated transmission priorities for the corresponding asynchronous data channels based on the timeliness requirements and risk levels of different data types published to the asynchronous data channels; When a decision node determines that an abnormal operating condition has occurred, it issues a job strategy instruction to the asynchronous data channel with the highest transmission priority so that it can be received by the nodes that are currently listening.

[0012] Preferably, trajectory generation and pose calculation are performed based on a unified global coordinate system, and the motion control commands for driving the actuator of the tower climbing robot include: Based on the global registration model information and the path instructions in the operation strategy instructions, a global path and critical path points are generated in the grid map of the standard three-dimensional structural model of the power tower. Inverse kinematics and torque allocation are performed based on critical path points to generate joint target values; A three-loop control strategy of position, speed and current is adopted, and a drive signal is output to the actuator according to the target value of the joint.

[0013] Preferably, the three-dimensional structural features are denoised using a combination of direct-pass filtering and statistical outlier filtering algorithms.

[0014] Preferably, the semantic features are obtained by acquiring images using a visual sensor and recognizing them based on the YOLO visual model.

[0015] The second aspect of this application provides a tower-climbing robot for power line inspection, comprising: Actuators are used for movement and operation on power towers; Sensing sensor components are used to acquire multimodal sensing data from the field in real time; An edge computing unit, which is communicatively connected to both the actuator and the sensing sensor assembly, is used to execute any of the aforementioned control methods for a tower-climbing robot used for power inspection.

[0016] Preferably, the sensing sensor assembly further includes an infrared thermal imaging sensor; The edge computing unit is also used to: detect the surface temperature of the power tower and its components based on the infrared thermal imaging sensor; when the deviation between the surface temperature and the ambient temperature exceeds a preset deviation threshold, mark the power tower component with the corresponding surface temperature as a thermally abnormal component, and generate a short-range detection path planning instruction for the control actuator to approach the thermally abnormal component.

[0017] A third aspect of this application provides a computer-readable medium storing a computer program that, when executed by a processor, implements a control method for a tower-climbing robot used for power line inspection as described above.

[0018] Compared with existing technologies, the above technical solution has the following advantages: 1. The control method for the tower-climbing robot provided in this application decomposes the overall system into perception nodes, model registration nodes, decision nodes, and control nodes, and deploys the entire control flow on the local edge side of the tower-climbing robot. The system introduces asynchronous data channels, enabling each node to independently handle specific feature extraction, fusion, or computation tasks, with unidirectional data flow between nodes relying on publish and listen mechanisms. This localized and decoupled operating mechanism completely eliminates the high latency and packet loss risks associated with data transmission to the cloud in weak network environments; simultaneously, the operational blockage of a single node will not lead to the paralysis of the entire machine's communication, giving the system excellent fault isolation capabilities and significantly improving the robot's response speed and control stability during dynamic climbing under complex working conditions. 2. In terms of spatial positioning and motion control, this application establishes a complete operational chain from cross-modal registration to hierarchical driving. For measured data with environmental noise, the scheme first uses semantic feature retrieval to establish an initial correspondence, and then combines geometric features to estimate the initial registration transformation matrix, progressively filtering out a large number of invalid matches. On this basis, the system introduces an improved iterative nearest-point algorithm with dual geometric and semantic regularization constraints to force the measured structured data to iteratively align within a specific semantic region, effectively suppressing registration drift and constructing a high-precision unified global coordinate system. Relying on this global digital benchmark, the system adopts a hierarchical architecture from macro-grid map path generation and inverse kinematics calculation to micro-three-loop servo control, gradually transforming abstract operational strategies into smooth and continuous physical actions, ensuring the responsiveness and motion accuracy of the actuator during climbing and obstacle crossing. 3. At the level of intelligent scheduling and safety management, this application constructs a comprehensive decision-making mechanism that integrates external environmental perception and internal physical state. By performing weighted fusion evaluation of visual semantics, 3D structure, global model information, fuselage attitude, and computing load, the system can autonomously switch between climbing, obstacle crossing, alignment, and protection states. To cope with extreme working conditions, the solution sets up multiple safety defenses. Once insufficient adhesion, abnormal attitude, or computing overload is detected, the system will forcibly switch to protection state. At the same time, to solve the problem of bus bandwidth limitation, the system introduces a priority scheduling strategy based on timeliness and danger level to ensure that shutdown lockout or safety reset commands generated under abnormal working conditions can be issued with the highest transmission priority, ensuring that each node obtains and executes safety fallback actions without delay, and comprehensively avoiding the risks of high-altitude overturning and falling. 4. Finally, this application also provides a highly integrated tower-climbing robot. By establishing a local communication connection between the actuator, sensing components, and edge computing unit, the device is completely freed from the constraints of cloud computing power and possesses the hardware foundation for fully autonomous operation in areas without network coverage. Furthermore, the system further enriches its detection dimensions by adding an infrared thermal imaging sensor. The edge computing unit can intuitively determine internal heating phenomena of components that are imperceptible by conventional vision through temperature deviations, and actively generate close-range detection commands to drive the actuator to approach the thermally abnormal components, realizing a transformation from conventional surface inspection to proactive detection of hidden defects. Attached Figure Description

[0019] Figure 1 A schematic diagram of the architecture of each node in the control method for the tower climbing robot used for power inspection provided in this application; Figure 2 A flowchart illustrating the control method for the tower-climbing robot used in power inspection provided in this application; Figure 3 This is a flowchart illustrating the cross-modal feature fusion step in the control method for the tower-climbing robot used for power inspection provided in this application. Detailed Implementation

[0020] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0022] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0024] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0025] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0026] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.

[0027] Please see Figures 1-2 , Figure 1 A schematic diagram of the architecture of each node in the control method for the tower climbing robot used for power inspection provided in this application; Figure 2 This is a flowchart illustrating the control method for the tower-climbing robot used in power line inspection provided in this application.

[0028] like Figures 1-2As shown, the first aspect of the present invention discloses a control method for a tower-climbing robot for power line inspection. The control method is deployed locally on the tower-climbing robot and includes: The sensing node acquires multimodal sensing data collected in real time by the tower climbing robot; extracts semantic features and three-dimensional structural features from the multimodal sensing data; and publishes the structured data composed of semantic features and three-dimensional structural features to the asynchronous data channel. The model registration node listens to the asynchronous data channel and obtains structured data. It then performs cross-modal feature fusion with the pre-stored standard 3D structural model of the power tower, calculates the registration transformation matrix, constructs a global registration model under a unified global coordinate system, and publishes the global registration model information to the asynchronous data channel. The decision node listens to the asynchronous data channel and obtains structured data, global registration model information, and the robot's own state parameters. Based on the above data analysis, it generates operation strategy instructions and then issues the operation strategy instructions to the asynchronous data channel. The control node listens to the asynchronous data channel and obtains global registration model information and operation strategy instructions. Based on a unified global coordinate system, it generates trajectory and calculates pose to generate motion control instructions that drive the actuator of the tower climbing robot.

[0029] This can be understood as follows: To completely resolve the fatal impact of network latency on robot control in the field, the control method of this application deploys all computation processes entirely within the edge computing unit of the robot itself (exemplarily, a Jetson Xavier NX with a computing power of 21 TOPS or an embedded module with higher computing power). Locally, the nodes no longer transmit instructions sequentially, preventing the failure of one node from causing other nodes to stagnate and become dangerous. Instead, an asynchronous data channel is established, through which each node exchanges data by publishing and listening to data information. Each node operates independently and does not interfere with each other's memory.

[0030] For example, the perception node can acquire images in real time at a frame rate of 30fps using an RGB-D depth camera, identify features such as foot spikes and fall arrestors using a lightweight object detection model (such as YOLOv5s), and extract point clouds as 3D structural features using a filtering algorithm. The merged structured data is then packaged and published to the bus. Even if the perception node experiences a brief processing delay due to drastic changes in lighting, the control and decision nodes can still monitor existing data or execute security policies, ensuring that the entire control program will not crash.

[0031] Therefore, the control method for a tower-climbing robot for power line inspection provided in this application decomposes the overall system into perception nodes, model registration nodes, decision nodes, and control nodes, and deploys the entire control flow on the local edge side of the tower-climbing robot. The system introduces asynchronous data channels, enabling each node to independently handle specific feature extraction, fusion, or calculation tasks, with unidirectional data flow between nodes relying on publish and listen mechanisms. This localized and decoupled operating mechanism completely eliminates the high latency and packet loss risks associated with data transmission to the cloud in weak network environments; simultaneously, the blocking of a single node's operation will not cause the entire system's communication to fail, giving the system excellent fault isolation capabilities and significantly improving the robot's response speed and control stability during dynamic climbing under complex working conditions.

[0032] It is important to emphasize that the specific implementation method of this control method is not limited. In one possible implementation, it can be based on the ROS2 (Robot Operating System 2) Topic publish and subscribe mechanism in the actual software architecture. Each node uploads and reads data from the Topic independently. In another possible implementation, it can also be based on the asynchronous information transmission mechanism of other operating systems. This application makes no restrictions here. The above is an explanation of the basic concept of this application. The following will explain in detail the specific implementation of each step.

[0033] First, there are no restrictions on the specific implementation method for acquiring multimodal sensing data in the perception phase.

[0034] For example, the three-dimensional structural features are denoised using a combination of PassThrough Filter and Statistical Outlier Removal (SOR) algorithms.

[0035] Specifically, for point cloud data, a pass-through filter is used to limit the effective depth range to 0.3-5.0m, eliminating invalid depth values ​​that exceed the range. Then, the SOR algorithm (nearest neighbor number k=50, standard deviation multiple μ=1.5) is used to remove floating noise and outlier interference points in the field. Finally, the filtered point cloud is downsampled to voxels (voxel size 5mm) to output a clean and regular scene-aware point cloud, with a single frame processing time of <30ms.

[0036] Similarly, in one possible implementation, semantic features are obtained by acquiring images from a visual sensor and recognizing them based on the YOLO visual model.

[0037] Specifically, the visual sensor in the perception processing node continuously acquires color images (640×480 resolution) and depth point cloud data of the tower surface at a frequency of 30fps. The visual recognition branch uses a lightweight YOLOv5s variant model (7.2M model parameters, inference time <15ms / frame) to complete high-precision recognition of key components such as foot spikes, fall arrestors, and bolts, and outputs component category labels and bounding box coordinates as semantic features.

[0038] Secondly, there are no restrictions on the specific implementation methods of cross-modal feature fusion.

[0039] Please see Figure 3 , Figure 3 This is a flowchart illustrating the cross-modal feature fusion step in the control method for the tower-climbing robot used for power inspection provided in this application.

[0040] like Figure 3 As shown, in one possible implementation, cross-modal feature fusion is performed between structured data and a pre-stored standard 3D structural model of the power tower, and the registration transformation matrix is ​​calculated including: Based on semantic features, components of the same category are retrieved between structured data and standard three-dimensional structural models of power towers to establish an initial set of correspondences in order to complete a coarse semantic matching. Geometric features were extracted from the structured data and the standard three-dimensional structural model of the power tower, and combined with the initial correspondence set, the initial registration transformation matrix was estimated using the random sampling consensus algorithm.

[0041] This can be understood as follows: when performing cross-modal feature fusion between structured data and pre-stored standard three-dimensional structural models of power towers, the initial correspondence is first established by searching and matching based on semantic features, and then geometric features are extracted and combined with the random sampling consensus algorithm to estimate the initial registration transformation matrix.

[0042] The SLAM point cloud maps collected by the robot itself often contain a large amount of environmental floating noise and errors, while the BIM standard model is perfect ideal data, which leads to cross-modal differences. To make these two overlap, this solution adopts a step-by-step alignment strategy. First, semantic coarse matching is performed, in which the system finds similar components through semantic labels. For example, the coordinates of the set marked as bolts in the measured map are used to establish an initial association pair with the coordinates of all bolts in the BIM model, which filters out the possibility of mismatching bolts to the angle steel surface. Then, geometric coarse registration is performed, extracting features such as plane normal vectors and edge segments, and using the Random Sample Consensus (RANSAC) algorithm to calculate an approximate spatial rotation and translation matrix (i.e., the initial registration transformation matrix). This provides an initial registration transformation matrix and also provides a reliable starting benchmark for subsequent precise fine-tuning.

[0043] This scheme combines semantic and geometric features for preliminary calculations. When processing structured data with environmental noise and ideal 3D structural models of power towers, direct spatial comparison can easily lead to local optima. The system first establishes an initial set of semantic correspondences by searching for components of the same category, anchoring a rough pairing range among heterogeneous data. Then, geometric features are extracted and combined with a random sampling consensus algorithm to estimate the initial registration transformation matrix. This progressive logic of semantics first, then geometry, effectively filters out a large number of invalid matching combinations, providing an accurate and reliable initial spatial position for subsequent fine-grained alignment and reducing the probability of the algorithm deviating from the correct solution.

[0044] It is understandable that the accuracy of the initial registration transformation matrix mentioned above is relatively general, and it is suitable for scenarios where the accuracy requirements are not high or the computing power is extremely limited.

[0045] Therefore, further, in one possible implementation, the cross-modal feature fusion of structured data and pre-stored standard three-dimensional structural models of power towers, and the calculation of the registration transformation matrix, also include: Based on the initial registration transformation matrix, the iterative nearest point algorithm is used for further iteration; Among them, the objective function of the iterative nearest point algorithm introduces geometric constraint regularization term and semantic constraint regularization term. The geometric constraint regularization term is used to constrain the consistency of normal vector and distance between corresponding point pairs, and the semantic constraint regularization term is used to assign weight penalty according to the consistency of semantic label of corresponding point pairs. When the preset iteration termination condition is met, the result is output as the final registration transformation matrix.

[0046] Therefore, based on the aforementioned initial registration transformation matrix, the Iterative Closest Point (ICP) algorithm, which incorporates geometric and semantic constraint regularization terms, is used for fine registration iteration until the termination condition is met. This further provides an accurate registration transformation matrix.

[0047] Specifically, the traditional Iterative Closest Point (ICP) algorithm relies solely on spatial distance to find the nearest point for alignment. When dealing with high-voltage power towers, which have numerous repetitive symmetrical structures (such as densely packed intersecting angle steel), registration slippage or drift is highly likely to occur. Therefore, this application makes targeted improvements to the objective function of the ICP algorithm. For example, the objective function can be expressed as:

[0048] R_geom(T) is the geometric constraint regularization term, constraining the consistency of normal vectors (normal vector deviation threshold δ_n ≤ 5°) and distance consistency (distance deviation between corresponding points < 10% of the feature size) of corresponding point pairs; R_sem(T) is the semantic constraint regularization term, penalizing the weights of corresponding point pairs with inconsistent semantic labels; λ1 and λ2 are regularization coefficients, taking values ​​of 0.3 and 0.2 respectively. The iteration termination condition is: the change in the objective function ΔE < 1e-6 or the number of iterations reaches the upper limit of 200. The final output is the registration transformation matrix. It achieves alignment between scene maps and BIM models under a unified global coordinate system, with a registration accuracy better than 2cm RMSE.

[0049] This can be understood as follows: In the iterative process of continuously moving the measured map to fit the BIM model, not only is the distance between point pairs required to be minimized, but also a double penalty is added: First, the geometric constraint regularization term R_geom(T) forces the normal vector deviation of corresponding point pairs to be extremely small (e.g., deviation threshold < 5°); second, the semantic constraint regularization term R_sem(T) will give a large penalty weight if the algorithm attempts to fit the footplate points onto the crossbeam of the BIM model, thereby forcibly correcting the matching direction of the algorithm. When the convergence change of the objective function is extremely small (e.g., ΔE < 1e-6) or the maximum number of iterations is reached (e.g., 200 times), the final output accuracy can reach a unified global coordinate system within 2cm.

[0050] Therefore, by introducing geometric and semantic constraint regularization terms into the objective function, the algorithm is forced to not only align the normal vectors and distances between point pairs during iteration, but also to accept a weight penalty for semantic label consistency. This ensures that the measured structured data is tightly constrained within the corresponding semantic region of the standard 3D structural model of the power tower, effectively suppressing drift during cross-modal data fusion, constructing a high-precision unified global coordinate system, and providing a reliable digital benchmark for the robot's obstacle crossing and alignment.

[0051] Based on the above two-level registration transformation matrix, cross-modal data fusion can be achieved with high accuracy, and the control of the tower climbing robot can then be realized based on the fused cross-modal data.

[0052] In one possible implementation, generating job strategy instructions based on the above data analysis includes: We perform weighted fusion and confidence assessment on semantic features, 3D structural features, global registration model information, robot body posture parameters, and equipment operating status parameters; Based on the evaluation results, the system transitions between multiple preset work states and outputs work strategy instructions corresponding to the current work state. The work states include climbing, obstacle crossing, and alignment states under normal working conditions, as well as protection states under abnormal working conditions.

[0053] Those skilled in the art will understand that the decision node here not only considers visual images but also reads data from physical sensors at a certain sensor frequency. It then combines the acquired cross-modal data with its own attitude data and operational status, performing weighted fusion and confidence assessment to ultimately switch states within a finite set of conditions, such as climbing, obstacle crossing, and alignment states under normal operating conditions, and protection states under abnormal operating conditions.

[0054] This step establishes a comprehensive evaluation mechanism that integrates external perception and internal state. The system performs weighted fusion and confidence assessment on the input semantic features, 3D structural features, global registration model information, body posture parameters, and equipment operating state parameters, so that the control logic no longer relies on single-dimensional sensor data. Based on the evaluation results, the system smoothly transitions between preset climbing, obstacle crossing, alignment, and protection states under abnormal operating conditions. This state machine transition logic can comprehensively measure the complexity of the external environment and the load-bearing capacity of the equipment itself, giving the robot the autonomous decision-making ability to dynamically adjust its operational behavior according to the actual situation on site.

[0055] Furthermore, in one possible implementation, abnormal operating conditions include: the base adhesion force of the tower climbing robot is lower than a first preset threshold, the body attitude angle of the tower climbing robot exceeds a second preset threshold and continues for a first preset time, or the computing power load in the equipment operating status parameters exceeds a third preset threshold and continues for a second preset time. The operating strategy instructions include shutdown lockout instructions and / or safety reset instructions.

[0056] Specifically, abnormal operating conditions are defined as specific situations. For example, when the suction force of either base, fed back by the dual base pressure sensors, falls below a first preset threshold (exemplarily, 1000N), it means that detachment may occur; or when the built-in IMU detects that the pitch / roll angle exceeds a second preset threshold (exemplarily, ±15°) for a first preset time (e.g., 500ms), it means that the fuselage is experiencing overturning and instability; or even when the GPU computing power load of the edge computing device exceeds a third preset threshold (e.g., 95%) for more than 3 seconds, it means that the operation may be about to freeze. In any of the above extreme cases, the system will unconditionally cut off the current operation, the state machine will immediately transition to the protection state, and directly issue a safety reset command to stop the brake or reverse along the original path.

[0057] Multiple safety defenses are established at both physical and computational boundaries. When the base adhesion force is detected to be low, the body attitude angle is continuously too large, or the computational load is continuously exceeded, the decision node can immediately identify potential risks of mechanical instability or system crash. At this time, the state machine will forcibly switch to protection mode and directly issue a shutdown lockout command or a safety reset command. This underlying hardware and computational status monitoring logic can forcibly take over the equipment when conventional planning fails or external environmental changes occur, avoiding serious accidents such as overturning and falling that may occur when the robot is climbing at heights.

[0058] When a dangerous situation occurs, each node should ideally be able to obtain the danger information immediately or as quickly as possible, so that the tower climbing robot can respond quickly.

[0059] Therefore, in one possible implementation, the control method also includes: Configure differentiated transmission priorities for the corresponding asynchronous data channels based on the timeliness requirements and risk levels of different data types published to the asynchronous data channels; When a decision node determines that an abnormal operating condition has occurred, it issues a job strategy instruction to the asynchronous data channel with the highest transmission priority so that it can be received by the nodes that are currently listening.

[0060] This can be understood as follows: the internal communication bus bandwidth of the edge computing unit is limited. Under normal conditions, massive point cloud data and high-definition images occupy a large number of communication channels. This application utilizes the underlying Quality of Service (QoS) control mechanism to allocate different priority levels to different topics. When the decision node triggers the aforementioned emergency operation policy instructions such as shutdown and locking, the instruction will be labeled with the highest danger level. The bus will suspend the transmission of low-priority data (such as environmental log recording or dense point cloud publishing), allowing this life-saving control instruction to reach the control node and execution mechanism instantly with the highest priority, ensuring that the overall emergency response cycle of the system is strictly reduced to a threshold, for example, within 200ms.

[0061] Considering the bus bandwidth limitations of edge computing units under dense concurrency, different data types objectively have different requirements for real-time transmission. By evaluating timeliness requirements and configuring differentiated transmission priorities based on hazard levels, priority passage of critical data flows can be guaranteed. Especially in abnormal operating conditions, the job policy instructions output by the decision node are issued with the highest transmission priority, ensuring that all monitoring nodes can obtain and execute the corresponding protection actions without delay, thus solving the problem of instruction transmission congestion under high load conditions.

[0062] Finally, based on the above information, the tower-climbing robot can be controlled in detail.

[0063] In one possible implementation, trajectory generation and pose calculation are performed based on a unified global coordinate system, generating motion control commands to drive the actuator of the tower-climbing robot, including: Based on the global registration model information and the path instructions in the operation strategy instructions, a global path and critical path points are generated in the grid map of the standard three-dimensional structural model of the power tower. Inverse kinematics and torque allocation are performed based on critical path points to generate joint target values; A three-loop control strategy of position, speed and current is adopted, and a drive signal is output to the actuator according to the target value of the joint.

[0064] This can be understood as follows: the robot's motion execution is scientifically broken down into three levels, from macroscopic to microscopic. The upper layer is the trajectory planning layer, which discretizes the high-precision BIM model into a grid map and uses... The pathfinding algorithm generates global path points with a planned step size (e.g., 40cm) to address how to bypass obstacles. The middle layer is the motion planning layer, which, after obtaining the path points, substitutes them into the robot arm's DH parameter table to perform inverse kinematics calculations, determining how many degrees each motor should rotate (i.e., the joint target value), thus addressing the hand and foot positioning. The bottom layer is the servo control layer, employing a classic three-loop PID control strategy (position loop, velocity loop, and current loop) to output drive current signals to the motors at an extremely high frequency (e.g., 1kHz response rate), addressing the stability of the motion. This layered and progressive control architecture gradually transforms abstract operational strategies into smooth, continuous physical movements, ensuring the servo accuracy and responsiveness of the actuators during operation.

[0065] The above is a complete and detailed description of the control method for the tower climbing robot used in power inspection provided in this application.

[0066] The second aspect of this application provides a tower-climbing robot for power line inspection, comprising: Actuators are used for movement and operation on power towers; Sensing sensor components are used to acquire multimodal sensing data from the field in real time; An edge computing unit, which is communicatively connected to both the actuator and the sensing sensor assembly, is used to execute any of the aforementioned control methods for a tower-climbing robot used for power inspection.

[0067] By integrating actuators, sensing sensor components, and edge computing units with communication connectivity, a distributed hardware physical entity is provided. Through the localized execution of the aforementioned control methods by the edge computing unit, the device is freed from the constraints of cloud servers, possessing the hardware foundation for autonomous detection and operation throughout the entire process in high-voltage transmission line areas without network coverage, thus having direct engineering deployment and practical value.

[0068] Furthermore, in one possible implementation, the sensing sensor assembly also includes an infrared thermal imaging sensor; The edge computing unit is also used to: detect the surface temperature of the power tower and its components based on the infrared thermal imaging sensor; when the deviation between the surface temperature and the ambient temperature exceeds a preset deviation threshold, mark the power tower component with the corresponding surface temperature as a thermally abnormal component, and generate a short-range detection path planning instruction for the control actuator to approach the thermally abnormal component.

[0069] By adding an infrared thermal imaging sensor, the robot's environmental exploration capabilities have been further enhanced. The edge computing unit compares the surface temperature of components with the ambient temperature, enabling it to intuitively identify internal heating phenomena that are imperceptible to conventional vision and structured point clouds. After marking components with thermal anomalies, the system generates near-range detection path planning instructions to control the actuators to approach, extending its functionality from routine surface inspection to proactively approaching and exploring hidden thermal defects.

[0070] It is understood that the specific implementation of the tower climbing robot provided in this application is not limited. To facilitate understanding by those skilled in the art, three embodiments are provided below for illustration.

[0071] Example 1: The hardware platform uses a Jetson Xavier NX embedded edge computing module (6-core ARM CPU + 384-core GPU, computing power 21 TOPS), equipped with an Intel RealSense D455i RGB-D depth camera (effective depth range 0.3-5m, frame rate 30fps) as the environmental acquisition sensor. The entire system is deployed on the dual-base alternating adsorption type power tower climbing robotic arm. The software system is based on the ROS2 Humble Hawksbill distribution to build a distributed node architecture, independently deploying perception processing nodes, model registration nodes, motion control nodes, and collaborative decision-making nodes. Each node achieves asynchronous data interaction and collaborative scheduling through ROS2 topics and services.

[0072] Example 2: The hardware platform uses a Jetson Orin Nano embedded module (6-core ARM CPU + 1024-core GPU, 40 TOPS computing power), suitable for medium-sized climbing robot platforms with relatively sufficient computing resources. Due to the improved GPU computing power, the perception processing node can use a YOLOv5m model with a larger number of parameters (21.2M parameters), improving recognition accuracy (mAP improvement of approximately 3%) while maintaining inference latency <25ms / frame. In the improved ICP fine registration in Phase 3, the model registration node adds a parallel computing branch for normal vector consistency verification, optimizing the registration iteration time from <100ms to <60ms, further improving the system's real-time performance.

[0073] Example 3: Based on Example 1, an infrared thermal imaging sensor (640×512 resolution, 30Hz frame rate) is added for detecting abnormal temperature rises in tower components. The collaborative decision-making node adds thermal anomaly judgment logic: when the surface temperature of a component deviates from the ambient temperature by more than 15K, it is marked as a thermally abnormal component, and priority is given to close-range, refined detection. The motion control node correspondingly adds a close-range approach path planning strategy for thermally abnormal components, controlling the approach distance within the range of 0.5-1.0m to ensure that the thermal imaging resolution meets the defect identification requirements. This example is suitable for thermal defect inspection scenarios in high-temperature summer seasons or high-load transmission lines.

[0074] A third aspect of this application provides a computer-readable medium storing a computer program that, when executed by a processor, implements a control method for a tower-climbing robot used in power line inspection as described above. This provides practical digital media support for the installation, deployment, program distribution, and version iteration of the control method.

[0075] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A control method for a tower-climbing robot used for power line inspection, characterized in that, The control method is deployed locally on the tower-climbing robot and includes: The sensing node acquires multimodal sensing data collected in real time by the tower climbing robot; extracts semantic features and three-dimensional structural features from the multimodal sensing data; and publishes the structured data composed of the semantic features and three-dimensional structural features to the asynchronous data channel. The model registration node listens to the asynchronous data channel and acquires the structured data. It then performs cross-modal feature fusion of the structured data with a pre-stored standard 3D structural model of the power tower to calculate the registration transformation matrix, thereby constructing a global registration model in a unified global coordinate system. The global registration model information is then published to the asynchronous data channel. Specifically, the process of fusing the structured data with the pre-stored standard 3D structural model of the power tower and calculating the registration transformation matrix includes: Based on the semantic features, components of the same category are retrieved between the structured data and the standard three-dimensional structural model of the power tower to establish an initial set of correspondences in order to complete the semantic coarse matching. Geometric features are extracted from the structured data and the standard three-dimensional structural model of the power tower, respectively. Combined with the initial correspondence set, the initial registration transformation matrix is ​​estimated using a random sampling consensus algorithm. Based on the initial registration transformation matrix, the iterative nearest point algorithm is used for further iteration; The objective function of the iterative nearest point algorithm incorporates geometric constraint regularization and semantic constraint regularization. The geometric constraint regularization is used to constrain the consistency of the normal vector and distance between corresponding point pairs, while the semantic constraint regularization is used to assign weight penalties based on the consistency of the semantic labels of the corresponding point pairs. When the preset iteration termination condition is met, the result is output as the final registration transformation matrix. The decision node listens to the asynchronous data channel and obtains the structured data, the global registration model information, and the robot's own state parameters. Based on the above data analysis, it generates operation strategy instructions and then issues the operation strategy instructions to the asynchronous data channel. The control node listens to the asynchronous data channel and obtains the global registration model information and the operation strategy instructions. Based on the unified global coordinate system, it generates trajectory and calculates pose to generate motion control instructions that drive the climbing robot's actuator.

2. The control method for the tower-climbing robot used for power line inspection as described in claim 1, characterized in that, The process of generating work strategy instructions based on the above data analysis includes: The semantic features, the three-dimensional structural features, the global registration model information, the body posture parameters of the tower climbing robot, and the equipment operating status parameters are weighted and fused together with confidence assessment. Based on the evaluation results, the system performs state transitions between multiple preset operation states and outputs the operation strategy instruction corresponding to the current operation state; wherein, the operation states include climbing, obstacle crossing, and alignment states under normal operating conditions, and protection states under abnormal operating conditions.

3. The control method for the tower-climbing robot used for power line inspection as described in claim 2, characterized in that, The abnormal operating conditions include: the base adhesion force of the tower climbing robot is lower than the first preset threshold, the body attitude angle of the tower climbing robot exceeds the second preset threshold and continues for a first preset time, or the computing power load in the equipment operating status parameters exceeds the third preset threshold and continues for a second preset time. The operation strategy instructions include shutdown lockout instructions and / or safety reset instructions.

4. The control method for the tower-climbing robot used for power line inspection as described in claim 3, characterized in that, The control method further includes: Based on the timeliness requirements and risk levels of different data types published to the asynchronous data channels, configure differentiated transmission priorities for the corresponding asynchronous data channels; When the decision node determines that an abnormal operating condition has occurred, it issues a job strategy instruction to the asynchronous data channel with the highest transmission priority so that it can be obtained by the nodes that are listening.

5. The control method for the tower-climbing robot used for power line inspection as described in claim 1, characterized in that, The process of generating trajectory and calculating pose based on the unified global coordinate system, and generating motion control commands to drive the actuator of the tower climbing robot, includes: Based on the global registration model information and the path instructions in the operation strategy instructions, a global path and critical path points are generated in the grid map of the standard three-dimensional structure model of the power tower. Based on the key path points, inverse kinematics solution and torque allocation are performed to generate joint target values; A three-loop control strategy of position, speed and current is adopted, and a drive signal is output to the actuator according to the target value of the joint.

6. The control method for the tower-climbing robot used for power line inspection as described in claim 1, characterized in that, The three-dimensional structural features are denoised using a combination of direct-pass filtering and statistical outlier filtering algorithms.

7. The control method for the tower-climbing robot used for power line inspection as described in claim 1, characterized in that, The semantic features are obtained by acquiring images using a visual sensor and recognizing them based on the YOLO visual model.

8. A tower-climbing robot for power line inspection, characterized in that, include: Actuators are used for movement and operation on power towers; Sensing sensor components are used to acquire multimodal sensing data from the field in real time; An edge computing unit is communicatively connected to the actuator and the sensing sensor assembly, respectively, and is used to execute the control method for the tower climbing robot for power inspection as described in any one of claims 1-7.

9. The tower-climbing robot for power line inspection as described in claim 8, characterized in that, The sensing sensor assembly also includes an infrared thermal imaging sensor. The edge computing unit is also used to: detect the surface temperature of the power tower and its components based on the infrared thermal imaging sensor; when it is determined that the deviation between the surface temperature and the ambient temperature exceeds a preset deviation threshold, mark the power tower component corresponding to the surface temperature as a thermally abnormal component, and generate a near-range detection path planning instruction to control the actuator to approach the thermally abnormal component.

10. A computer-readable medium, characterized in that, The computer-readable medium stores a computer program that, when executed by a processor, implements the control method for a tower-climbing robot used in power inspection as described in any one of claims 1-7.

Images