A method for repairing hidden dangers of power transmission lines

Abstract

The present invention relates to a method for repairing hidden dangers of power transmission lines. By constructing a basic generalized robot motion model based on deep learning, combined with the three-dimensional point cloud data of the power transmission line collected by a drone and the fine-grained operation data of remote teaching, accurate repair of hidden dangers of the power transmission line is achieved. The model inputs the hidden danger point cloud data, outputs the angle posture matrix of the operating robot arm, and predicts the angle posture matrix according to the hidden danger point cloud data at the current moment when there is wind. By obtaining the motion decision at the previous moment, adaptive compensation is performed in combination with the posture offset, the motion trajectory of the robot arm is optimized, and high-precision execution of the robot in a complex dynamic environment is ensured. The present invention realizes the robot's fully autonomous and high-precision hidden danger repair operation under wind disaster conditions through unstructured dynamic scene transfer learning and adaptive motion compensation technology, which significantly improves the robot's autonomy and execution efficiency in complex environments.

Description

Technical Field

[0001] The invention belongs to the technical field of line repair operations, and in particular relates to a method for repairing hidden dangers of a power transmission line. Background Art

[0002] With the rapid development of the electric power industry, the operation and maintenance of transmission lines have become increasingly complex and challenging, especially in the case of natural disasters such as wind disasters and lightning strikes. Hidden dangers such as damage to transmission line hardware and wire breakage frequently occur, posing a huge threat to power supply. Traditional transmission line hidden danger repair work usually relies on manual operation, the working environment is relatively harsh, and the repair accuracy is limited by the technical level of manual operation and environmental factors (such as wind speed, climate, etc.), which easily leads to risks and safety hazards in the repair process.

[0003] In this context, a method for repairing hidden dangers in power transmission lines based on robot technology has emerged. Robots can not only replace manual labor in dangerous high-altitude operations and improve the safety of operations, but also achieve efficient repair tasks in complex environments such as wind disasters through precise control systems. However, due to interference from factors such as wind speed and climate, the operating accuracy and stability of robots still face severe challenges.

[0004] At present, some studies have attempted to apply robots to power transmission line repair work. For example, using a robotic arm to perform repair operations through remote teaching, or collecting field data through sensors, and planning and executing repair tasks based on the data. However, these methods usually rely on static environmental assumptions and ignore complex factors of dynamic changes, such as wind speed disturbances and line vibrations. This makes the robot less adaptable in dynamic environments such as wind disasters, and it is difficult to achieve high-precision, low-error repair operations. Summary of the invention

[0005] The purpose of the present invention is to provide a method for repairing hidden dangers of a power transmission line in order to overcome the defects of the prior art.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] The present invention provides a method for repairing hidden dangers of a power transmission line, comprising the following steps:

[0008] When there is no wind, manually operate the teaching robot arm to control the working robot arm to repair hidden dangers, record the teaching robot arm action sequence collected by the corresponding action sensor data, and record the hidden danger point cloud data collected by the depth camera on the working robot arm, wherein the hidden dangers include broken strands and pin falling off, and the action sequence includes the state sequence of the robot arm at each time node, and the state sequence includes the angle posture matrix of the robot arm;

[0009] A basic generalized robot motion model is constructed based on the hidden danger point cloud data obtained when there is no wind and the teaching robot arm motion sequence. The basic generalized robot motion model inputs the hidden danger point cloud data and outputs the angle posture matrix of the robot arm at the next moment;

[0010] When there is wind, the depth camera on the working robot arm collects the hidden danger point cloud data at the current moment, inputs the hidden danger point cloud data at the current moment into the basic generalized robot motion model, and outputs the first angle posture matrix;

[0011] Obtain the motion decision of the operating robot arm at the previous moment, compensate the angle posture matrix of the first robot arm based on the motion decision of the operating robot arm at the previous moment and the hidden danger point cloud data at the current moment, obtain the motion decision of the operating robot arm at the current moment, and control the operating robot arm for automatic repair according to the motion decision of the operating robot arm at the current moment.

[0012] Furthermore, the angle posture matrix of the robotic arm includes the angle posture vectors of each joint of the robotic arm.

[0013] Furthermore, the basic generalized robot motion model includes a deep learning model based on the Transformer architecture, wherein the deep learning model based on the Transformer architecture is trained through a multimodal task framework.

[0014] Furthermore, the basic generalized robot motion model is constructed based on the hidden danger point cloud data and the teaching robot arm motion sequence, specifically including:

[0015] When there is no wind, collect point cloud data of different hidden dangers and teach the robot arm action sequence to form a training data set;

[0016] The hidden danger point cloud data is used as input, the teaching robot arm motion sequence is used as the goal, the deep learning model based on the Transformer architecture is trained, the model parameters are adjusted using the optimization algorithm, the loss function is minimized, and the basic generalized robot motion model is obtained.

[0017] Furthermore, the hidden danger point cloud data is used as input, the teaching robot arm action sequence is used as the target, and the deep learning model based on the Transformer architecture is trained, which specifically includes:

[0018] The hidden danger point cloud data at time t is used as input, and the teaching robot arm angle posture matrix at time t+1 is used as the target to train the deep learning model based on the Transformer architecture.

[0019] Furthermore, the loss function of the basic generalized robot motion model is:

[0020] ,

[0021] in, is the basic generalized robot motion model loss function, N is the total number of joints of the robot arm, is the angle posture vector of the i-th joint at time t+1 predicted by the basic generalized robot motion model, is the angle posture vector of the i-th joint of the teaching robot arm at time t+1, Represents the L2 norm of a vector.

[0022] Furthermore, the motion decision of the working robot arm is: , , described as a mapping rule from the current state S(t) to the next action A(t+1), the motion decision of the working robot arm is Including the current working robot arm angle posture matrix , the end-actuator posture matrix of the robot arm at the current moment .

[0023] Furthermore, the motion decision of the operating robot arm at the previous moment and the hidden danger point cloud data at the current moment are used to compensate the first robot arm angle posture matrix to obtain the motion decision of the operating robot arm at the current moment, which specifically includes:

[0024] According to the first robot arm angle posture matrix Obtain the ideal actuator end attitude matrix through kinematics solution ;

[0025] Obtain the actual actuator end posture matrix based on the hidden danger point cloud data at the current moment ;

[0026] Through the ideal actuator end posture matrix The actual actuator end posture matrix Calculate the pose offset ;

[0027] Get the motion decision of the robot arm at the last moment , based on the motion decision of the robot arm at the last moment Offset from pose The first angle attitude matrix is ​​adjusted by adaptive guided filter Perform compensation to obtain the motion decision of the robot arm at the current moment .

[0028] Furthermore, the input of the adaptive guidance filter is the motion decision of the working robot arm at the previous moment Offset from pose , the output is the motion decision of the robot arm at the current moment, which is described as:

[0029] ,

[0030] in, The motion decision of the robot arm at the current moment. is the compensation function, is the constraint function.

[0031] Furthermore, the motion decision of the robot arm at the current moment Including the second robot arm angle attitude matrix And the attitude matrix of the actuator end in the current working robot arm .

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] (1) The present invention constructs a high-precision training data set, integrates remote teaching operation data and the three-dimensional point cloud map of the power transmission line collected by drones, and accurately restores the key operating steps of hardware replacement and wire repair tasks in wind disaster scenarios. Through unstructured dynamic scene transfer learning, the robot can perform fine-grained hidden danger repair tasks in a dynamic environment, avoiding the problem of insufficient environmental adaptability of static models in traditional methods. This technology effectively improves the generalization ability and operation accuracy of the robot in complex environments, enabling it to handle different types of hidden danger repair tasks, and significantly enhances the robot's application capabilities in dynamic environments such as wind disasters.

[0034] (2) The adaptive motion compensation algorithm proposed in this invention comprehensively considers wind field disturbance, line shaking and target posture deviation, and adjusts the motion trajectory of the robot arm in real time. The algorithm uses posture offset calculation and adaptive guidance filter to dynamically correct the movement of the robot arm according to historical motion decisions and real-time environmental changes, achieving an accuracy improvement from centimeters to millimeters. Through this technology, the robot can accurately complete the tasks of hardware replacement and wire repair, and can still ensure the operation accuracy even in dynamic environments such as wind disasters, greatly improving the operation safety and repair effect.

[0035] (3) The present invention uses a deep learning model based on the Transformer architecture and trains through a multimodal task framework to optimize the robot's motion decision-making process. The model can predict the angle posture matrix of the operating robot arm at the next moment based on the input of hidden danger point cloud data, and make compensation adjustments with the motion decision at the previous moment. Through this intelligent motion control strategy, the robot can fully autonomously complete complex hidden danger repair tasks, which not only improves the repair efficiency, but also greatly reduces the need for manual intervention, and improves the safety and operability of the operation.

[0036] (4) The present invention integrates the depth camera and other sensors on the working robot arm to collect the current hidden danger point cloud data and actuator end posture information in real time, and uses it as input to make motion decisions and adjust the repair actions in real time. The combination of multimodal data not only improves the environmental perception ability, but also enables the robot to dynamically adjust its motion trajectory according to real-time environmental changes. This greatly improves the stability and accuracy of the robot in complex environments such as wind disasters, ensuring the efficient execution of repair tasks.

[0037] (5) The present invention uses an intelligent feedback mechanism to make adaptive adjustments based on the motion decision of the robot arm at the previous moment and the real-time environmental data. This mechanism enables the robot to dynamically adjust the operation path under the influence of interference factors such as wind disasters, and can adapt to different repair tasks and scene requirements. The robot can not only cope with conventional repair tasks, but also make rapid adjustments based on sudden changes (such as wind speed changes, line vibrations, etc.), ensuring the continuity and efficiency of task execution, and further improving the robustness and stability of the repair operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of obtaining a training data set in a windless environment according to the present invention;

[0039] Figure 2 This is a schematic diagram of the generalized robot motion model that is the basis for constructing the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0041] Example 1.

[0042] This embodiment provides a method for repairing hidden dangers in power transmission lines based on unstructured dynamic scene transfer learning and adaptive motion compensation technology, which can realize fully autonomous and high-precision hidden danger repair operations by robots in dynamic environments such as wind disasters. The specific implementation process is as follows:

[0043] like Figure 1As shown, in windless conditions, the teaching operation is first performed. The manual operation teaching robot arm performs hidden danger repair operations. During the operation, the motion sensor collects and records the action sequence of the robot arm, including the state sequence of the robot arm at each time node. The state sequence includes the angle posture matrix of the robot arm. Through the depth camera, the operating robot arm collects the hidden danger point cloud data of the transmission line. The hidden danger point cloud data includes but is not limited to transmission line hidden dangers such as broken strands and pins falling off. These data are used to construct a training data set to provide a basis for subsequent robot motion model training.

[0044] like Figure 2 As shown in the figure, a basic generalized robot motion model is constructed based on the hidden danger point cloud data obtained in the windless state and the teaching robot arm action sequence. The input of the model is the hidden danger point cloud data, and the output is the angle posture matrix of the robot arm at the next moment. Through the deep learning model based on the Transformer architecture, combined with the multimodal task framework for training, the model parameters are optimized, and by minimizing the loss function, the model can accurately predict the robot's motion decisions in different environments. Specifically, the loss function of the model is:

[0045] ,

[0046] in, is the basic generalized robot motion model loss function, N is the total number of joints of the robot arm, is the angle posture vector of the i-th joint at time t+1 predicted by the basic generalized robot motion model, is the angle posture vector of the i-th joint of the teaching robot arm at time t+1, represents the L2 norm of the vector. The loss function is used to calculate the error between the robot's expected motion and the actual motion, thereby optimizing the model's predictive ability. This training process improves the adaptability and prediction accuracy of the basic generalized robot motion model in various environments.

[0047] In a windy environment, the robot arm collects the hidden danger point cloud data at the current moment in real time through the depth camera and inputs it into the basic generalized robot motion model. The model generates the first angle posture matrix based on the input point cloud data. At this time, under the influence of wind field disturbance and line shaking, the robot generates preliminary motion decisions based on the current environmental state. This technology enables the robot to still have a certain degree of motion ability in a complex environment and can quickly make adaptive adjustments.

[0048] After obtaining the motion decision of the operating robot arm at the previous moment, the first angle posture matrix is ​​compensated in combination with the hidden danger point cloud data at the current moment to obtain a more accurate motion decision at the current moment. Through this compensation method based on the motion decision of the previous moment and the current hidden danger point cloud data, the robot can adaptively adjust to environmental factors such as wind disasters, thereby maintaining the continuity and accuracy of the repair task. This process significantly improves the robot's autonomous repair capability and reduces the need for human intervention. The key technology in the compensation process is the adaptive guidance filter, which adjusts the robot's motion trajectory according to the posture offset, allowing the robot to stably perform tasks in a dynamic environment.

[0049] In order to further improve the accuracy of motion compensation, the embodiment adopts the posture offset calculation technology. By comparing the ideal actuator end posture matrix with the actual end posture matrix, the posture offset is calculated. , the formula is:

[0050] ,

[0051] in, is the ideal actuator end posture matrix, is the actual actuator end posture matrix, the ideal actuator end posture matrix According to the first robot arm angle posture matrix The actual actuator end attitude matrix is ​​obtained by kinematic solution. Obtained based on the hidden danger point cloud data at the current moment.

[0052] This offset reflects the difference between the robot's actual execution and expectations. The adaptive motion compensation algorithm adjusts the robot's motion trajectory through this posture offset, thereby improving the operation accuracy, enabling the robot to achieve millimeter-level accuracy in hardware replacement and wire repair operations in a storm environment.

[0053] During the entire repair process, the motion decision of the robot arm is based on the motion decision at the previous moment. and real-time environment data (such as pose offset ) is dynamically adjusted. The input of the adaptive guidance filter is the motion decision and posture offset of the previous moment, and the output is the optimized motion decision of the current moment. The decision process is achieved by optimizing the compensation function To minimize the deviation of the end effector's posture, thus achieving precise adjustment of the motion trajectory. Adaptive compensation enables the robot to maintain operating accuracy and stability in unstable environments such as wind speed changes and line shaking, ensuring the smooth completion of the repair task.

[0054] The input of the adaptive guidance filter is the motion decision of the robot arm at the previous moment. Offset from pose , the output is the motion decision of the robot arm at the current moment, which is described as:

[0055] ,

[0056] in, The motion decision of the robot arm at the current moment. is the compensation function, is the constraint function.

[0057] In addition, the present invention achieves efficient environmental perception and intelligent decision-making by building a multimodal task framework and integrating deep learning and motion control technology. Under this framework, the operating robot arm can predict the motion state at future moments based on the input hidden danger point cloud data and historical action sequences, thereby intelligently performing repair tasks. This highly intelligent control method enables the robot to adaptively adjust and automatically perform hidden danger repair operations based on real-time data, thereby improving the robot's autonomy and flexibility in complex dynamic environments.

[0058] Through the above implementation, the power transmission line hidden danger repair method of the present invention not only improves the adaptability of the robot in a dynamic environment, but also greatly improves the robot's repair accuracy and task execution efficiency in complex situations such as wind disasters. Through adaptive motion compensation, deep learning optimization and multimodal data fusion, the robot can realize fully autonomous hidden danger repair operations and achieve high-precision repair effects at the millimeter level, which has significant technical advantages.

[0059] Example 2.

[0060] A depth camera is used to obtain point cloud data of the entire wire section and the damaged area of ​​the wire. The point cloud data of the entire wire section is used as the basis of the training data set, and the point cloud data of the damaged area of ​​the wire is used to locate the damaged area and place the robotic arm where it needs to be repaired.

[0061] The ground engineer will operate the teaching robot arm to provide The data are annotated as examples of the dataset, and the trained model is used as the basic generalized robot motion model.

[0062] When working, the robot arm adopts the control method of low-rank adaptive fine-tuning technology (LORA), taking the three-dimensional point cloud data collected by the depth camera as input, and the actual working joint state sequence of the operating robot arm is the output of the motion model, and this output is consistent with the teaching value The decay function can be fitted as:

[0063] ,

[0064] in, is the basic generalized robot motion model loss function, N is the total number of joints of the robot arm, is the angle posture vector of the i-th joint at time t+1 predicted by the basic generalized robot motion model, is the angle posture vector of the i-th joint of the teaching robot arm at time t+1, Represents the L2 norm of a vector.

[0065] Execution strategy for robot motion model output The ideal motion sequence , the actual working posture is estimated to be , the pose offset can be calculated .in is a seven-element array consisting of They are the spatial position, posture information and rotation angle of the end of the robot arm. After the output posture offset and ideal motion sequence are introduced into the adaptive compensation algorithm, the actual motion sequence is finally output. .

[0066] In this embodiment, a depth camera will be installed on the drone and at the end of the robotic arm. The depth camera will be used to take RGB images and point cloud data of the robotic arm at two angles. Relying on the pre-trained 2D basic model, DINO (Visual Transformer) and Segment Anything Model (SAM) are used for object segmentation and matching. SAM is used to generate all possible objects and masks in the target scene, and DINO is used to calculate the similarity between objects and select the best matching object. After determining the best matching object, the classic two-viewpoint geometry method is used to estimate the relative posture and obtain the actual operation posture estimate. .

[0067] Basic generalized robot motion model: A model based on the transformer architecture, using a VLA-based multimodal task framework.

[0068] The open source big model OpenVLA is used as the main decision-making body, and a large amount of point cloud data and teaching data are used to fine-tune the big model to make the output of the big model close enough . Ensure the reliability of model output.

[0069] Movement Decision It is a core concept in motion control or task planning, which can describe a mapping rule from state S(t) to the next action A(t+1), that is, , The data in includes the angle attitude matrix , the actuator end posture matrix , path point spatial coordinates P, time constraints, space constraints and other information.

[0070] Pose offset calculation: The input of the pose offset calculation is the ideal end pose and the actual terminal posture , The data is recorded in middle, The data is acquired in real time by the depth camera equipped by the robot, and the difference between the two can be used to obtain the posture offset .

[0071] Adaptive motion compensation algorithm: The main input of the adaptive motion compensation algorithm is The ideal working sequence in , and the pose offset The adaptive motion control algorithm is divided into two parts: the adaptive direction guidance filter part and the other closed-loop compensation part.

[0072] The input of the adaptive guidance filter is the actual motion decision at the previous moment and pose offset , the description of the adaptive guided filter is:

[0073] ,

[0074] Meaning the movement decision at the last moment and the compensation function at the next moment The sum of is the step size of the compensation algorithm. is the constraint of the previous operation, which means to minimize the difference between the actual end posture and the ideal end posture. The output is the motion sequence at the future moment .

[0075] Motion sequence for future moments and the ideal motion sequence entered Accumulate and finally output the actual motion sequence , and at the same time generate the motion decision at that time for the compensation calculation at the next moment.

[0076] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0077] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for repairing hidden dangers of a power transmission line, characterized in that: The following steps are involved: When there is no wind, manually operate the teaching robot arm to control the working robot arm to repair hidden dangers, record the teaching robot arm action sequence collected by the corresponding action sensor data, and record the hidden danger point cloud data collected by the depth camera on the working robot arm, wherein the hidden dangers include broken strands and pin falling off, and the action sequence includes the state sequence of the robot arm at each time node, and the state sequence includes the angle posture matrix of the robot arm; A basic generalized robot motion model is constructed based on the hidden danger point cloud data obtained when there is no wind and the teaching robot arm motion sequence. The basic generalized robot motion model inputs the hidden danger point cloud data and outputs the angle posture matrix of the robot arm at the next moment; When there is wind, the depth camera on the working robot arm collects the hidden danger point cloud data at the current moment, inputs the hidden danger point cloud data at the current moment into the basic generalized robot motion model, and outputs the first angle posture matrix; Obtain the motion decision of the operating robot arm at the previous moment, compensate the angle posture matrix of the first robot arm based on the motion decision of the operating robot arm at the previous moment and the hidden danger point cloud data at the current moment, obtain the motion decision of the operating robot arm at the current moment, and control the operating robot arm to automatically repair according to the motion decision of the operating robot arm at the current moment; The motion decision of the working robot arm is: , , described as a mapping rule from the current state S(t) to the next action A(t+1), the motion decision of the working robot arm is Including the current working robot arm angle posture matrix , the end-actuator posture matrix of the robot arm at the current moment ; The method of compensating the first robot arm angle posture matrix based on the motion decision of the operating robot arm at the previous moment and the hidden danger point cloud data at the current moment to obtain the motion decision of the operating robot arm at the current moment specifically includes: According to the first robot arm angle posture matrix Obtain the ideal actuator end attitude matrix through kinematics solution ; Obtain the actual actuator end posture matrix based on the hidden danger point cloud data at the current moment ; Through the ideal actuator end posture matrix The actual actuator end posture matrix Calculate the pose offset ; Get the motion decision of the robot arm at the last moment , based on the motion decision of the robot arm at the last moment Offset from pose The first angle attitude matrix is ​​adjusted by adaptive guided filter Perform compensation to obtain the motion decision of the robot arm at the current moment .

2. A method for repairing hidden dangers of a power transmission line according to claim 1, characterized in that: The angle posture matrix of the robotic arm includes the angle posture vectors of each joint of the robotic arm.

3. A method for repairing hidden dangers of a power transmission line according to claim 1, characterized in that: The basic generalized robot motion model includes a deep learning model based on a Transformer architecture, wherein the deep learning model based on the Transformer architecture is trained through a multimodal task framework.

4. A method for repairing hidden dangers of a power transmission line according to claim 1, characterized in that: The basic generalized robot motion model is constructed based on the hidden danger point cloud data and the teaching robot arm motion sequence, specifically including: When there is no wind, collect point cloud data of different hidden dangers and teach the robot arm action sequence to form a training data set; The hidden danger point cloud data is used as input, the teaching robot arm motion sequence is used as the goal, the deep learning model based on the Transformer architecture is trained, the model parameters are adjusted using the optimization algorithm, the loss function is minimized, and the basic generalized robot motion model is obtained.

5. A method for repairing hidden dangers of a power transmission line according to claim 4, characterized in that: The method uses the hidden danger point cloud data as input, teaches the robot arm action sequence as the target, and trains the deep learning model based on the Transformer architecture, specifically including: The hidden danger point cloud data at time t is used as input, and the teaching robot arm angle posture matrix at time t+1 is used as the target to train the deep learning model based on the Transformer architecture.

6. A method for repairing hidden dangers of a power transmission line according to claim 1, characterized in that: The basic generalized robot motion model loss function is: ， in, is the basic generalized robot motion model loss function, N is the total number of joints of the robot arm, is the angle posture vector of the i-th joint at time t+1 predicted by the basic generalized robot motion model, is the angle posture vector of the i-th joint of the teaching robot arm at time t+1, Represents the L2 norm of a vector.

7. A method for repairing hidden dangers of a power transmission line according to claim 1, characterized in that: The input of the adaptive guidance filter is the motion decision of the robot arm at the previous moment Offset from pose , the output is the motion decision of the robot arm at the current moment, which is described as: ， in, The motion decision of the robot arm at the current moment. is the compensation function, is the constraint function, For actual operation posture estimation.

8. A method for repairing hidden dangers of a power transmission line according to claim 7, characterized in that: The motion decision of the operating robot arm at the current moment Including the second robot arm angle attitude matrix And the attitude matrix of the actuator end in the current working robot arm .

Images