A mechanical arm control method for distribution network non-power-off operation

Abstract

This invention provides a robotic arm control method for live-line work in power distribution networks, belonging to the field of robotic arm control technology for power distribution network operations. To address the technical problem that existing robots struggle to balance compliance, reliability, and efficiency in power distribution network operations, failing to meet on-site operational requirements, this invention installs a three-dimensional line laser contour sensor at the end of the first robotic arm of a dual-robotic arm system, and a six-dimensional force / torque sensor and a working tool at the end of the second robotic arm. The sensors construct an operational coordinate system and generate a pre-operation guidance trajectory. During operation, the system scans the conductor in real time to obtain the dynamic center point coordinates and offset, and outputs composite commands through a hybrid control model to achieve macroscopic position tracking and high-frequency micro-motion compensation. After contact, the system switches to impedance control, using force feedback to ensure the tool maintains constant force against the conductor. This invention achieves high-precision dynamic tracking, compliant contact, and human-robot collaborative operation, significantly improving the reliability, safety, and efficiency of live-line work.

Description

Technical Field

[0001] This invention provides a robotic arm control method for uninterrupted power distribution network operations, belonging to the field of robotic arm control technology for power distribution network operations. Background Technology

[0002] As the "last mile" of the power system, the power distribution network's reliability directly affects the end-user's electricity experience. Currently, to improve power supply reliability and minimize power outage time for users, live-line work on distribution networks has become the mainstream operating mode. However, traditional manual live-line work methods have long faced severe challenges of "high altitude, high voltage, and high intensity." Workers need to wear heavy insulating protective gear and be lifted to tens of meters in the air by insulated bucket trucks, making close contact with high-voltage lines in a strong electric field environment to complete complex operations such as line connection and equipment replacement. This type of work not only requires extremely high psychological qualities and operational skills from the workers, but also poses significant safety risks such as falls from heights and electric shock injuries. The labor intensity is extremely high, and the work efficiency is also limited by the workers' physiological limits.

[0003] To overcome the limitations of manual live-line work, the technology of live-line working robots for power distribution networks has developed rapidly in recent years. Currently, the mainstream robot operation modes mainly include two types: one is the remote operation mode, in which operators on the ground remotely control every movement of the aerial robotic arm through handles and consoles. Although this mode physically isolates personnel from the high-voltage electric field, the operation efficiency is low, the operator's skill level is extremely high, and there is a lack of force perception and presence, which can easily lead to equipment damage or short circuit accidents due to blind spots or misoperation. The other type is the autonomous operation mode, in which the robot relies entirely on preset programs and environmental perception systems to complete the task autonomously. However, the power distribution network operation environment is complex and changeable, with many uncertain factors such as lighting conditions, conductor posture, and surrounding obstacles. The visual recognition and control algorithms of the pure autonomous mode are not stable enough in the face of extreme situations. Once a perception error or planning failure occurs, there is a lack of effective remedial means, and the success rate of the operation is difficult to meet the requirements for practical application, posing a risk of "uncontrollability".

[0004] While attempts have been made to address these issues through master-slave control or autonomous control, the advantages of both have not been effectively integrated, and the control strategy cannot be dynamically adjusted according to real-time operating conditions during operation. As a result, when dealing with complex and unstructured power distribution network operation environments, the robot's operational compliance, reliability, and efficiency remain difficult to balance, failing to meet the stringent on-site operational requirements. Summary of the Invention

[0005] To address the technical problems existing in the background art, the present invention provides a robotic arm control method for uninterrupted power distribution network operations, comprising the following control steps:

[0006] Step 1: Install a dual robotic arm system, which includes a first robotic arm and a second robotic arm. A three-dimensional line laser profile sensor is installed on the end effector of the first robotic arm, and a six-dimensional force / torque sensor and a working tool are installed on the end effector of the second robotic arm. By controlling the first robotic arm to drive the three-dimensional line laser profile sensor to scan the target working line, three-dimensional point cloud data is obtained, and a six-degree-of-freedom working coordinate system is constructed based on the three-dimensional point cloud data.

[0007] Step 2: Collect the operator's teaching actions through the ground control station, calculate them into the desired pose sequence of the first and second robotic arms, and use the desired pose sequence as the pre-operation guidance trajectory;

[0008] Step 3: During the operation, the first robotic arm is controlled to drive the three-dimensional line laser contour sensor to scan the conductor in real time, generate a micro-contour point cloud, and calculate the dynamic center point coordinates and real-time offset vector of the conductor from the micro-contour point cloud in real time.

[0009] Step 4: Input the real-time offset vector into a hybrid control model, and the hybrid control model outputs a composite control command. The composite control command includes a basic motion command and a micro-motion correction command, wherein:

[0010] The basic motion commands are used to control the second robotic arm to follow a continuous trajectory composed of the coordinates of a dynamic center point.

[0011] The micro-motion correction command is used to drive the end of the second robotic arm to generate a high-frequency, small-amplitude compliant movement to compensate for instantaneous position deviation.

[0012] Step 5: Monitor the feedback data from the six-dimensional force / torque sensor in real time:

[0013] When the working tool contacts the wire and the feedback data exceeds a preset first threshold, the hybrid control model switches the control mode, switches the calculation basis of the micro-motion correction command to the force / torque data fed back by the six-dimensional force / torque sensor, and dynamically adjusts the position correction amount of the end of the second robotic arm according to an impedance control model, so that the working tool adheres to the surface of the wire with the set contact force and slides to the preset position.

[0014] Step 6: During the operation, the conductor deformation data fed back by the three-dimensional line laser profile sensor and the operation reaction force data fed back by the six-dimensional force / torque sensor are integrated in real time and input into an operation efficiency evaluation model.

[0015] If the work status parameters output by the work performance evaluation model show abnormalities, the ground control station will issue an alarm signal and allow the operator to fine-tune the position of the end effector of the second robotic arm through a wearable motion capture device.

[0016] The specific method for constructing a six-degree-of-freedom working coordinate system based on the three-dimensional point cloud data in step 1 is as follows:

[0017] Step 1.1: Control the first robotic arm to drive the three-dimensional line laser contour sensor to perform multi-angle scanning of the target working area, including the conductor, insulator and crossarm, at a preset scanning path and scanning speed, and obtain the original three-dimensional point cloud dataset;

[0018] Step 1.2: The original 3D point cloud dataset is filtered by a data preprocessing unit deployed at the ground control station to remove isolated noise points and outliers caused by abnormal light reflection, thus obtaining the first filtered point cloud data.

[0019] Step 1.3: Perform precise registration between the first filtered point cloud data and the standard tower model pre-stored in the database based on the iterative nearest point algorithm. The iterative nearest point algorithm is used to iteratively solve for an optimal spatial transformation matrix, so as to minimize the root mean square error between the first filtered point cloud data and the standard tower model.

[0020] Step 1.4: Based on the spatial transformation matrix, transform the coordinates of all points in the first filtered point cloud data to a global coordinate system based on the standard tower model to obtain the registered point cloud data;

[0021] Step 1.5: In the registered point cloud data, a segmentation algorithm based on Euclidean clustering is used to extract point cloud clusters belonging to conductors, insulators and crossarms respectively;

[0022] Step 1.6: Perform cylindrical surface fitting on the extracted conductor point cloud clusters to solve the spatial axis equation of the conductor, thereby obtaining the accurate pose and diameter parameters of the conductor. At the same time, construct a three-dimensional bounding box for the point cloud clusters of the insulator and crossarm, and mark the spatial occupancy range of the three-dimensional bounding box in the global coordinate system as the key obstacle area.

[0023] Step 1.7: Using the spatial axis of the conductor as a reference, and combining the conductor diameter parameters and key obstacle areas, establish a six-degree-of-freedom working coordinate system. The origin of the six-degree-of-freedom working coordinate system is located at a preset position on the conductor axis near the starting point of the operation. The Z-axis coincides with the conductor axis, the X-axis points in the direction of the crossarm, and the Y-axis is determined by the right-hand rule.

[0024] The specific method for step 2 is as follows:

[0025] Step 2.1: The operator wears a wearable motion capture device containing multiple inertial measurement units and performs a set of teaching actions simulating the work process; the ground control station collects the three-dimensional angle data of the operator's upper limb shoulder joint, elbow joint and wrist joint in real time at a sampling frequency of more than 100 Hz to obtain the original joint angle sequence.

[0026] Step 2.2: Perform Kalman filtering on the original joint angle sequence to remove Gaussian noise during signal acquisition and obtain a smooth joint angle sequence;

[0027] Step 2.3: Call a pre-generated human-machine kinematics mapping matrix. The human-machine kinematics mapping matrix is ​​obtained by having the operator perform a series of calibration actions during the system calibration phase, while recording the joint angles of the operator and the corresponding end poses of the first and second robotic arms in the six-degree-of-freedom working coordinate system, and fitting the data using the least squares method.

[0028] Step 2.4: Multiply the smooth joint angle sequence with the human-machine kinematics mapping matrix to calculate in real time the expected pose data corresponding to the end effectors of the first and second robotic arms in the six-degree-of-freedom working coordinate system;

[0029] Step 2.5: Arrange the calculated expected pose data at all times in chronological order to form the expected pose sequence. At the same time, apply a preset velocity planning algorithm to the expected pose sequence to smooth the motion velocity between adjacent pose points and generate a continuous, abrupt motion trajectory as a pre-operation guidance trajectory.

[0030] The specific method for step 3 is as follows:

[0031] Step 3.1: After the second robotic arm reaches the preparatory position under the guidance of the pre-operation guide trajectory, activate the visual servo tracking function of the first robotic arm; control the movement of the end effector of the first robotic arm so that the scanning light plane of the three-dimensional line laser contour sensor is always in the same plane perpendicular to the guide axis as the working tool at the end effector of the second robotic arm.

[0032] Step 3.2: The three-dimensional line laser profile sensor continuously scans the section of the conductor to be worked on at a line scanning frequency of more than 1 kHz. Each scan acquires a two-dimensional profile line containing the surface profile of the conductor. Multiple scans are accumulated to form the micro-profile point cloud.

[0033] Step 3.3: For the two-dimensional contour line obtained in each scan, a sub-pixel extraction algorithm based on the centroid method is used to calculate the gray-scale centroid position in the vertical direction of the contour line. The gray-scale centroid position is an edge point of the conductor on the scan section.

[0034] Step 3.4: Perform least-squares circle fitting on multiple edge points obtained from scanning at the same time to obtain a fitted circle. The coordinates of the center of the fitted circle are the coordinates of the instantaneous geometric center of the conductor on the scanning section in the sensor coordinate system.

[0035] Step 3.5: Using a pre-calibrated hand-eye calibration matrix, transform the coordinates of the instantaneous geometric center in the sensor coordinate system to the six-degree-of-freedom working coordinate system to obtain the coordinates of the dynamic center point;

[0036] Step 3.6: Compare the current dynamic center point coordinates obtained at the current moment with the current pose of the end effector of the second robotic arm, calculate the difference between the two in the X-axis, Y-axis and Z-axis directions in the six-degree-of-freedom working coordinate system, and form a real-time offset vector.

[0037] The specific method for step 4 is as follows:

[0038] Step 4.1: Input the real-time offset vector and its rate of change as input variables into a fuzzy inference engine. The fuzzy inference engine predefines multiple fuzzy rules for different deviation magnitudes and trends.

[0039] Step 4.2: The fuzzy inference engine performs inference based on the input variables and fuzzy rules, and outputs three correction coefficients for adjusting the PID controller parameters, denoted as follows: , , ;

[0040] Step 4.3: A parameter-self-tuning PID controller receives the real-time offset vector as the error signal e(t) and adjusts it according to the correction coefficient. , , Dynamically adjust its proportional, integral, and derivative gains, and calculate a basic control quantity according to the following control law. :

[0041] ;

[0042] in, , , This is the initial gain of the PID controller;

[0043] Step 4.4: The hybrid control model will use the basic control variables. The basic motion commands are mapped and sent to the position loop controller of the second robotic arm to achieve macroscopic position tracking.

[0044] Step 4.5: Simultaneously, the hybrid control model passes the real-time offset vector e(t) through a high-pass filter to extract the high-frequency components with frequencies higher than 5 Hz. ;

[0045] Step 4.6: Convert the high-frequency components Multiply by a preset stiffness coefficient Generate a force correction quantity The force correction amount The code is encoded as the micro-motion correction instruction and directly superimposed on the current loop setpoint of the second robotic arm joint servo driver through a high-speed real-time bus, driving the end effector of the second robotic arm to generate high-frequency micro-motion in Cartesian space.

[0046] The specific method for step 5 is as follows:

[0047] Step 5.1: The six-dimensional force / torque sensor detects the force components in three directions generated when the working tool comes into contact with the wire in real time at a sampling rate of 1 kHz. and torque components in three directions This generates the original force / torque data;

[0048] Step 5.2: Perform gravity compensation calculations on the original force / torque data, subtracting the component of the working tool's own weight under different postures, to obtain the pure contact force / torque data, denoted as... ;

[0049] Step 5.3: The pure contact force / torque data Contact force with a pre-set desired force By comparison, the force / torque error is obtained. ;

[0050] Step 5.4: Calculate the force / torque error. An impedance-controlled filter is input to a mass-spring-damped system, the dynamic characteristics of which are described by a second-order differential equation:

[0051] ;

[0052] in, , , These are the preset target inertia matrix, target damping matrix, and target stiffness matrix, respectively; , , These are the desired position correction, velocity correction, and acceleration correction, respectively.

[0053] Step 5.5: Solve the second-order differential equation to calculate in real time the position correction required by the end effector of the second robotic arm under the current force error. ;

[0054] Step 5.6: Adjust the position correction amount The modified desired trajectory is superimposed onto the continuous trajectory formed by the coordinates of the dynamic center point and sent to the position loop controller of the second robotic arm, so that the working tool slides against the surface of the guide wire with a constant desired contact force.

[0055] The specific method for integrating the operational reaction force data into an operational performance evaluation model in step 6 is as follows:

[0056] Step 6.1: In a laboratory environment, simulate various live-line working conditions, simultaneously collect the conductor deformation data sequence fed back by the three-dimensional line laser profile sensor and the working reaction force data sequence fed back by the six-dimensional force / torque sensor, and manually label the working state corresponding to each moment. The working states include normal wire stripping, normal connection, slight tool jamming, and severe conductor slippage.

[0057] Step 6.2: After normalizing the conductor deformation data sequence and the operation reaction force data sequence, use them as input features and the operation status at the corresponding time as labels to form a training dataset;

[0058] Step 6.3: Construct a multilayer perceptron neural network containing one input layer, three hidden layers, and one output layer; the number of nodes in the input layer is the same as the dimension of the input features, the number of nodes in the output layer is the same as the number of job state categories, and the Softmax function is used as the activation function;

[0059] Step 6.4: Supervised training of the multilayer perceptron neural network is performed using the training dataset. The network weights are optimized through backpropagation algorithm until the classification accuracy of the network on the validation set reaches more than 98%. The trained neural network model is obtained, which is the job performance evaluation model.

[0060] Step 6.5: During the actual operation, the conductor deformation data and operation reaction force data collected at the current moment are input into the operation efficiency evaluation model in real time. After forward calculation, the model outputs a five-dimensional probability vector. Each element in the probability vector represents the probability that the current operation state belongs to the corresponding category. The category with the highest probability value is the operation state parameter.

[0061] The specific method for the ground control station to issue alarm signals and for operators to make fine-tuning interventions in step 6 is as follows:

[0062] Step 6.6: Set a confidence threshold; when the probability value of the corresponding tool slight jamming or wire severe slippage category output by the work performance evaluation model exceeds the confidence threshold, it is determined that an abnormality has occurred in the work process;

[0063] Step 6.7: The ground control station immediately suspends the autonomous operation command output of the hybrid control model and issues an alarm signal through voice and graphical user interface to prompt the operator to intervene;

[0064] Step 6.8: The ground control station switches the control mode of the second robotic arm from the autonomous mode dominated by the hybrid control model to the position increment control mode. In this mode, the small movements made by the operator through the wearable motion capture device will be calculated in real time as the small displacement increment of the end of the second robotic arm in Cartesian space, and sent directly as position commands to the second robotic arm.

[0065] Step 6.9: Operators make fine adjustments by observing the on-site video footage and sensor data until the abnormal state is eliminated; the ground control station monitors the feedback data of the six-dimensional force / torque sensor in real time, and issues a prompt sound again when the contact force returns to the normal range and the normal state probability value output by the work efficiency evaluation model returns to above 0.9.

[0066] Step 6.10: After the operator confirms that the anomaly has been resolved, they can return control of the second robotic arm to the hybrid control model by pressing a confirmation button. The system will then switch back to the autonomous operation mode dominated by the hybrid control model and continue to execute the unfinished tasks.

[0067] Both the first robotic arm and the second robotic arm are redundant degrees of freedom robotic arms with 7 rotary joints, and are mounted on the same insulated lifting platform;

[0068] The three-dimensional line laser profile sensor is a high-speed profile scanner based on the principle of laser triangulation.

[0069] The six-dimensional force / torque sensor is a strain gauge type or an optical sensor;

[0070] The working tools are modular tools that can be automatically replaced, including an electric peeler, a hydraulic connecting pliers, and a bolt fastener.

[0071] The ground control station and the dual robotic arm system exchange data via an optical fiber communication link, which simultaneously carries high-speed video streams, sensor data streams, and control command streams.

[0072] The wearable motion capture device communicates with the ground control station via a wireless local area network;

[0073] All electrical components of the dual robotic arm system are powered by an isolation transformer, and all external communication interfaces are opto-isolated to meet the insulation and electromagnetic compatibility requirements of power distribution network uninterrupted operation sites.

[0074] The present invention has the following advantages over the prior art:

[0075] I. This invention acquires the microscopic contour of a conductor in real time using a three-dimensional line laser contour sensor, and decouples the visual feedback from the basic motion command and the high-frequency micro-motion correction command using a hybrid control model. The micro-motion correction command directly acts on the joint current loop to achieve position compensation with a bandwidth of kilohertz, which can effectively suppress high-frequency disturbances such as conductor wind deflection and robotic arm vibration. Experimental results show that the dynamic tracking error can be reduced from ±3 mm to within ±0.3 mm, which is an order of magnitude better than the traditional visual servo method, laying a solid foundation for subsequent precision operations.

[0076] Second, when the working tool comes into contact with the conductor, the present invention automatically switches to impedance control mode. Based on the real-time feedback of the six-dimensional force / torque sensor, the end position correction is dynamically adjusted to keep the contact force strictly maintained near the preset value (e.g., 5 Newtons), and the contact force overshoot is less than 6.2 Newtons. This mechanism avoids damage to the conductor insulation layer and the tool caused by rigid collisions, while ensuring that stripping, splicing and other processes are completed stably under constant force, which greatly improves the safety and quality of live-line work.

[0077] Third, this invention innovatively introduces a deep learning-based operation efficiency evaluation model, which integrates visual deformation and force data in real time, and can identify abnormal states such as tool jamming and wire slippage in advance. When the autonomous mode fails, the system actively requests human intervention through motion capture equipment for precise fine-tuning. After the anomaly is eliminated, autonomous operation is automatically restored. Comparative tests show that the method has a 100% operation success rate, a 100% anomaly handling success rate, and an average time of only 85 seconds, combining the high efficiency of pure autonomy with the reliability of human intervention. Attached Figure Description

[0078] The present invention will be further described below with reference to the accompanying drawings:

[0079] Figure 1 This is a flowchart of the steps of the robotic arm control method for uninterrupted power supply operation in the power distribution network according to the present invention;

[0080] Figure 2 This is a flowchart illustrating the steps of the present invention to calculate the desired pose sequence for the first robotic arm and the second robotic arm;

[0081] Figure 3 This is a flowchart illustrating the steps of calculating the dynamic center point coordinates of a conductor in real time from a microscopic contour point cloud according to the present invention. Detailed Implementation

[0082] like Figures 1 to 3As shown, this embodiment of the invention provides a robotic arm control method for uninterrupted power distribution network operations, applied to a dual-robotic arm system. The dual-robotic arm system includes a first robotic arm and a second robotic arm. The end effector of the first robotic arm is equipped with a three-dimensional line laser profile sensor, and the end effector of the second robotic arm is equipped with a six-dimensional force / torque sensor and a working tool. The corresponding robotic arm control method mainly includes:

[0083] Step 1: Control the first robotic arm to drive the three-dimensional line laser contour sensor to scan the target operation line, obtain three-dimensional point cloud data, and construct a six-degree-of-freedom operation coordinate system based on the three-dimensional point cloud data.

[0084] Step 2: Collect the operator's teaching actions through the ground control station, calculate the expected pose sequence of the first robotic arm and the second robotic arm, and use the expected pose sequence as the pre-operation guidance trajectory.

[0085] Step 3: During the operation, the first robotic arm is controlled to drive the three-dimensional line laser contour sensor to scan the conductor in real time, generate a micro-contour point cloud, and calculate the dynamic center point coordinates and real-time offset vector of the conductor from the micro-contour point cloud in real time.

[0086] Step 4: Input the real-time offset vector into a hybrid control model, and output a composite control command from the hybrid control model; the composite control command includes a basic motion command and a micro-motion correction command; the basic motion command is used to control the second robotic arm to follow the continuous trajectory formed by the coordinates of the dynamic center point, and the micro-motion correction command is used to drive the end of the second robotic arm to generate a high-frequency, small-amplitude compliant motion to compensate for instantaneous position deviation.

[0087] Step 5: Monitor the feedback data of the six-dimensional force / torque sensor in real time; when the working tool contacts the wire and the feedback data exceeds a preset first threshold, the hybrid control model switches the control mode, switches the calculation basis of the micro-motion correction command to the force / torque data fed back by the six-dimensional force / torque sensor, and dynamically adjusts the position correction amount of the end of the second robotic arm according to an impedance control model, so that the working tool adheres to the surface of the wire with a set contact force and slides to the preset position.

[0088] Step 6: During the operation, the conductor deformation data fed back by the three-dimensional line laser contour sensor and the operation reaction force data fed back by the six-dimensional force / torque sensor are integrated in real time and input into an operation efficiency evaluation model; if the operation status parameters output by the operation efficiency evaluation model indicate an abnormality, the ground control station issues an alarm signal and allows the operator to fine-tune the position of the end of the second robotic arm through the wearable motion capture device.

[0089] In an embodiment of the present invention, the dual-arm robotic system includes a first robotic arm ARM1 and a second robotic arm ARM2. Both ARM1 and ARM2 are redundant degrees of freedom robotic arms with seven rotary joints, driven by a high-precision harmonic reducer and a permanent magnet synchronous motor, achieving a repeatability accuracy better than ±0.05 mm. The two robotic arms are mounted together on an insulated lifting platform that is raised to a high-altitude working position by an insulated boom lift.

[0090] A three-dimensional line laser profile sensor is mounted on the end effector of the first robotic arm ARM1. This three-dimensional line laser profile sensor It employs a high-speed contour scanner based on the principle of laser triangulation. Its measurement range along the laser line is 100 mm, and its measurement range in the depth direction is 200 mm. Each contour line contains 1280 points, and the highest line scanning frequency can reach 2 kHz.

[0091] The end effector of the second robotic arm ARM2 is equipped with a six-dimensional force / torque sensor. and work tools. The six-dimensional force / torque sensor The strain gauge sensor has a force measurement range of ±200 Newtons in the X, Y, and Z axes, and a torque measurement range of ±20 Newton-meters. Its measurement accuracy is better than 0.5% of full scale, and the data output frequency is 1 kHz. The tool is a modular, automatically replaceable tool that can be replaced with an electric skinning tool, hydraulic pliers, or a bolt fastener depending on the task. In this embodiment, an electric skinning tool is used as an example.

[0092] The ground control station (GCS) is located on the ground and includes a high-performance industrial computer, a graphical user interface, and a wearable motion capture device (MOCAP). The MOCAP contains multiple inertial measurement units (IMUs) fixed to the operator's shoulder, elbow, and wrist joints to collect real-time three-dimensional angle data of the operator's upper limb joints at a sampling frequency of 120 Hz. The GCS communicates with the dual-arm robotic system via a fiber optic link. For data interaction, the optical fiber communication link Simultaneously carrying four high-definition video streams, all sensor data streams, and control command streams, the system employs the UDP protocol to ensure low latency. The wearable motion capture device MOCAP communicates with the ground control station GCS via a wireless local area network (WLAN). All electrical components of the dual-arm robotic system are powered by an isolation transformer, and all external communication interfaces are opto-isolated to meet the insulation and electromagnetic compatibility requirements of uninterrupted power supply operations.

[0093] Furthermore, in step 1, before the operation begins, the operator sends an initialization command to the dual-arm robotic system via the ground control station GCS. First, the first robotic arm ARM1 is controlled to drive the three-dimensional line laser contour sensor. The target work line and its surrounding environment are scanned using a preset scanning path. The preset scanning path is as follows: starting from the suspension clamp of the conductor, it moves at a constant speed of 0.2 meters per second along the conductor's axis, while ensuring the three-dimensional line laser profile sensor... The laser plane remains perpendicular to the conductor axis and swings up and down at a small angle to cover the insulator and crossarm areas. The entire scanning process lasts approximately 15 seconds, acquiring a raw 3D point cloud dataset containing the conductor, insulator, and crossarm.

[0094] Step 1, which involves constructing a six-degree-of-freedom operational coordinate system based on 3D point cloud data, specifically includes the following sub-steps:

[0095] Step 1.1: Control the first robotic arm ARM1 to drive the three-dimensional line laser contour sensor along a preset scanning path and at a preset scanning speed. The target work area, including conductors, insulators and crossarms, is scanned from multiple angles to obtain the original three-dimensional point cloud dataset.

[0096] Step 1.2: The original 3D point cloud dataset is filtered using a data preprocessing unit deployed at the ground control station GCS. First, a statistical filtering algorithm is used to calculate the average distance between each point and its 50 nearest neighbors. Points whose average distance exceeds one standard deviation of the global mean are removed as isolated noise points. Then, a radius filtering algorithm is used. For each point, if fewer than three points are contained within a sphere with a radius of 1 mm centered at that point, it is considered an outlier and removed. After these two steps, the first filtered point cloud data is obtained.

[0097] Step 1.3: Perform precise registration between the first filtered point cloud data and a standard tower model pre-stored in the database based on the Iterative Closest Point Algorithm. The Iterative Closest Point Algorithm iteratively solves for an optimal spatial transformation matrix, minimizing the root mean square error between the first filtered point cloud data after transformation and the standard tower model. In this embodiment, the Iterative Closest Point Algorithm converges after 20 iterations to obtain the optimal spatial transformation matrix.

[0098] Step 1.4: Based on the spatial transformation matrix, transform the coordinates of all points in the first filtered point cloud data to a global coordinate system based on the standard tower model to obtain the registered point cloud data.

[0099] Step 1.5: In the registered point cloud data, a segmentation algorithm based on Euclidean clustering is used to extract point cloud clusters belonging to conductors, insulators, and crossarms respectively. The clustering search radius is set to 3 cm, and the minimum number of cluster points is 50. The algorithm traverses the points in the point cloud sequentially, grouping points with a spatial distance of less than 3 cm into the same cluster, and finally extracts the conductor point cloud cluster, insulator point cloud cluster, and crossarm point cloud cluster.

[0100] Step 1.6: Perform cylindrical fitting on the extracted conductor point cloud clusters to solve for the spatial axis equation of the conductor, thereby obtaining the precise pose and diameter parameters of the conductor. This implementation uses a cylindrical fitting algorithm based on random sampling consistency, setting a maximum of 1000 iterations and a distance threshold of 1 mm to fit a cylinder. The axis of this cylinder is the spatial axis of the conductor, and the radius of the cylinder is the radius of the conductor. In this implementation, the measured conductor radius is 8.5 mm. Simultaneously, for the insulator point cloud clusters and crossarm point cloud clusters, calculate their minimum axis-aligned bounding boxes, and mark the spatial occupancy of these bounding boxes in the global coordinate system as key obstacle regions.

[0101] Step 1.7: Using the spatial axis of the conductor as a reference, and combining the wire diameter parameters and the key obstacle area, establish the six-degree-of-freedom working coordinate system. The origin of the six-degree-of-freedom working coordinate system is located at a preset position on the conductor axis near the starting point of the operation. In this embodiment, the origin is set at 0.5 meters away from the suspension clamp towards the working point. The Z-axis of the six-degree-of-freedom working coordinate system coincides with the conductor axis and points towards the working point; the X-axis is perpendicular to the Z-axis and points towards the crossarm, determined by the centroid of the crossarm point cloud; the Y-axis is determined by the right-hand rule, i.e. This coordinate system will serve as the benchmark for all subsequent motion planning and control.

[0102] Furthermore, after environmental modeling is completed in step 2, the system enters the master-slave collaboration preparation phase. The operator dons the wearable motion capture device MOCAP, stands in front of the ground control station GCS, and observes the 3D reconstructed image of the site through a monitor. The operator performs a set of teaching actions simulating a work process, such as simulating approaching the guide wire, adjusting tool posture, and conforming to the guide wire, lasting approximately 8 seconds.

[0103] Step 2, which calculates the desired pose sequence for the first and second robotic arms, specifically includes the following sub-steps:

[0104] Step 2.1: The operator wears a wearable motion capture device (MOCAP) containing multiple inertial measurement units and performs a set of teaching actions simulating a work process. The ground control station (GCS) collects three-dimensional angle data of the operator's shoulder, elbow, and wrist joints in real time at a sampling frequency of 120 Hz to obtain the original joint angle sequence.

[0105] Step 2.2: Perform Kalman filtering on the original joint angle sequence to remove Gaussian noise from the signal acquisition process, obtaining a smooth joint angle sequence. The process noise covariance matrix Q of the Kalman filter is set as follows: Let the measurement noise covariance matrix R be set as , where I is the identity matrix.

[0106] Step 2.3: Invoke a pre-generated human-machine kinematics mapping matrix. This human-machine kinematics mapping matrix is ​​obtained by having the operator perform a series of calibration actions during the system calibration phase, simultaneously recording the operator's joint angles and the corresponding end-effector poses of the first robotic arm ARM1 and the second robotic arm ARM2 in the six-degree-of-freedom working coordinate system, and then fitting the data using the least squares method. In this embodiment, the human-machine kinematics mapping matrix is ​​a 6×9 matrix (9 joint angle inputs, 6-dimensional pose output), obtained through offline calculation and stored in the system.

[0107] Step 2.4: Multiply the smooth joint angle sequence with the human-machine kinematic mapping matrix to calculate in real time the desired pose data corresponding to the end effectors of the first robotic arm ARM1 and the second robotic arm ARM2 in the six-degree-of-freedom working coordinate system.

[0108] Step 2.5: Arrange the calculated desired pose data at all times in chronological order to form the desired pose sequence. Simultaneously, apply a preset velocity planning algorithm to the desired pose sequence to smooth the motion velocity between adjacent pose points, generating a continuous, abrupt motion trajectory as the pre-operation guidance trajectory. The velocity planning algorithm uses trapezoidal velocity planning, setting the maximum linear velocity to 0.1 m / s, the maximum angular velocity to 0.2 radians / s, and the acceleration / deceleration time to 1 second.

[0109] Furthermore, in step 3, when the dual robotic arm system begins to perform the live wire connection operation, the second robotic arm ARM2, under the coarse guidance of the pre-operation guide trajectory, first moves the work tool TOOL to a preparatory position close to the wire, approximately 10 centimeters away. Subsequently, the real-time tracking function of the first robotic arm ARM1 is activated.

[0110] Step 3, which calculates the dynamic center point coordinates of the conductor in real time from the microscopic contour point cloud, specifically includes the following sub-steps:

[0111] Step 3.1: After the second robotic arm ARM2 reaches the pre-position under the guidance of the pre-operation guide trajectory, the visual servo tracking function of the first robotic arm ARM1 is activated. The end effector movement of the first robotic arm ARM1 is controlled, causing the three-dimensional line laser contour sensor... The scanning light plane is always located in the same plane perpendicular to the guide wire axis as the working tool TOOL at the end of the second robotic arm ARM2. This constraint is achieved through real-time calculation of the sensor. The first robotic arm ARM1 is controlled by using the coordinate difference between the tool and the Z-axis coordinates and providing real-time feedback.

[0112] Step 3.2: The three-dimensional line laser profile sensor The section of the conductor to be worked on is continuously scanned at a line scanning frequency of 1 kHz. Each scan acquires a two-dimensional contour line containing the surface contour of the conductor. Each contour line contains 1280 points. Multiple consecutive scans accumulate to form the micro-contour point cloud.

[0113] Step 3.3: For the two-dimensional contour line acquired in each scan, a sub-pixel extraction algorithm based on the centroid method is used to calculate the gray-level centroid position in the vertical direction of the contour line. This gray-level centroid position is an edge point of the conductor on the scanned cross section. Since the contour line has two edges, left and right, two edge points can be extracted from each cross section. The average of these two edge points is taken to obtain the coordinates of the conductor center in the sensor coordinate system of that cross section.

[0114] Step 3.4: Perform least-squares circle fitting on multiple edge points obtained from scanning at the same time to obtain a fitted circle. The coordinates of the center of this fitted circle are the coordinates of the instantaneous geometric center of the conductor on the scanning section in the sensor coordinate system. In this embodiment, 11 sections are fitted, including 5 sections before and after the current time, to improve noise resistance.

[0115] Step 3.5: Using a pre-calibrated hand-eye calibration matrix, transform the coordinates of the instantaneous geometric center in the sensor coordinate system to the six-degree-of-freedom working coordinate system to obtain the coordinates of the dynamic center point. The hand-eye calibration matrix is ​​obtained using the classic Tsai two-step method.

[0116] Step 3.6: Compare the current dynamic center point coordinates obtained at the current moment with the current pose of the end effector of the second robotic arm ARM2, and calculate the differences between the two in the X, Y, and Z axes of the six-degree-of-freedom working coordinate system to form the real-time offset vector. Since the end effector's pose remains essentially unchanged before contact, only positional offset is considered here; therefore, the real-time offset vector is... .

[0117] Furthermore, in step 4, the real-time controller inputs the real-time offset vector into a hybrid control model that combines fuzzy logic and PID control. This model operates on a real-time controller with an FPGA+DSP architecture and a 5 kHz cycle.

[0118] Step 4, which outputs a composite control command from the hybrid control model, specifically includes the following sub-steps:

[0119] Step 4.1: Set the magnitude of the real-time offset vector. The rate of change of e(t) / dt is used as an input variable and fed into a fuzzy inference engine. The fuzzy inference engine predefines several fuzzy rules for different deviation magnitudes and trends, such as: "If e(t) is large and de(t) / dt is positively large, then output a large proportional coefficient correction, a small integral coefficient correction, and a small differential coefficient correction." This implementation uses a triangular membership function to divide e(t) and de(t) / dt into seven fuzzy sets.

[0120] Step 4.2: The fuzzy inference engine performs inference based on the input variables and the fuzzy rules, and outputs three correction coefficients for adjusting the PID controller parameters, denoted as follows: , , The values ​​range from 0.5 to 2.0. The specific correction coefficients are obtained by defuzzifying using the centroid method.

[0121] Step 4.3: A parameter self-tuning PID controller receives the real-time offset vector as an error signal e(t), and adjusts it according to the correction coefficient. , , Dynamically adjust its proportional, integral, and derivative gains, and calculate a basic control quantity according to the following control law. :

[0122] ;

[0123] in, , , The initial gain of the PID controller is set in this embodiment. , , (Units have been dimensionless).

[0124] Step 4.4: The hybrid control model will incorporate the basic control variables. The basic motion commands are mapped and sent to the position loop controller of the second robotic arm ARM2 via the EtherCAT bus to achieve macroscopic position tracking, so that the end effector of the second robotic arm ARM2 follows the continuous trajectory formed by the coordinates of the dynamic center point.

[0125] Step 4.5: Simultaneously, the hybrid control model will use the real-time offset vector After passing through a high-pass filter, high-frequency components with frequencies above 5 Hz are extracted. The cutoff frequency of the high-pass filter is set to 5 Hz.

[0126] Step 4.6: Multiply the high-frequency component by a preset stiffness coefficient. Generate a force correction quantity In this embodiment, the stiffness coefficient Take 2000 Newtons per meter (converted using force control equivalents). Adjust the force correction amount accordingly. The micro-motion correction instructions are encoded and directly superimposed on the current loop setpoints of the servo drivers for each joint of the second robotic arm ARM2 via the FPGA's high-speed parallel bus. This is equivalent to connecting a high-bandwidth force disturbance in parallel outside the position loop, driving the end effector of the second robotic arm ARM2 to generate high-frequency, small-amplitude compliant movements in Cartesian space to compensate for instantaneous position deviations caused by wire deflection or robot body vibration. Experimental results show that when the wire deflection frequency is approximately 2–3 Hz, this high-frequency micro-motion correction can reduce the end effector dynamic tracking error from ±3 mm to within ±0.3 mm.

[0127] Furthermore, in step 5, during the process of the work tool approaching the guide wire, the six-dimensional force / torque sensor is monitored in real time. Feedback data. When the tool (TOOL) contacts the wire, the contact force detected by the sensor increases rapidly. In this embodiment, the preset first threshold is 2.0 Newtons (resultant force value). When the resultant force exceeds 2.0 Newtons, the hybrid control model automatically switches the control mode, changing the calculation basis of the micro-motion correction command from the real-time offset vector to the six-dimensional force / torque sensor. Feedback force / torque data.

[0128] Step 5, which dynamically adjusts the position correction amount of the second robotic arm's end effector based on an impedance control model, specifically includes the following sub-steps:

[0129] Step 5.1: The six-dimensional force / torque sensor The force components in three directions generated when the tool TOOL comes into contact with the wire are detected in real time at a sampling rate of 1 kHz. and torque components in three directions This generates the original force / torque data.

[0130] Step 5.2: Perform gravity compensation calculations on the original force / torque data, subtracting the component of the tool's own weight under different postures, to obtain the pure contact force / torque data, denoted as... Gravity compensation requires prior measurement of the mass m of the tool TOOL and the coordinates of its center of gravity in the sensor coordinate system. In this embodiment, the mass m of the electric peeler is 1.2 kg. Based on the current posture of the robotic arm end (obtained from encoder feedback), the components of gravity in each direction are calculated and subtracted from the original data.

[0131] Step 5.3: The pure contact force / torque data Contact force with a pre-set desired force By comparison, the force / torque error is obtained. In this embodiment, the desired contact force is... Set to 5.0 Newtons, with the direction along the negative Z-axis of the sensor coordinate system (pointing towards the wire).

[0132] Step 5.4: Calculate the force / torque error. The input is fed into an impedance-controlled filter of a mass-spring-damped system. The dynamic characteristics of the impedance-controlled filter are described by a second-order differential equation:

[0133] ;

[0134] in, , , These are the preset target inertia matrix, target damping matrix, and target stiffness matrix, respectively. In this embodiment, to obtain compliant contact behavior, the following are set... It is a diagonal matrix, and the diagonal elements are each 0.5 kg. The diagonal element is taken as 20 Newton-seconds per meter; The diagonal element is taken as 100 Newtons / meter. , , These are the desired position correction, velocity correction, and acceleration correction, respectively.

[0135] Step 5.5: Solve the second-order differential equation to calculate in real time the position correction required at the end of the second robotic arm ARM2 under the current force error. The solution employs the bilinear transform method for discretization, with a sampling time of [missing information]. Second.

[0136] Step 5.6: Adjust the position correction amount The modified desired trajectory is superimposed onto the continuous trajectory formed by the coordinates of the dynamic center point and sent to the position loop controller of the second robotic arm ARM2, thereby enabling the working tool TOOL to slide to a preset position (e.g., 15 cm from the origin of the working coordinate system determined in step 1 along the axis of the conductor) with a constant desired contact force.

[0137] Furthermore, in step 6, once the end of the second robotic arm ARM2 reaches the preset stripping or receiving position, the work tool TOOL is activated to begin operation. For example, the electric stripper begins rotating and stripping. During the operation, the three-dimensional line laser contour sensor is fused in real time. Feedback of conductor deformation data and the six-dimensional force / torque sensor The feedback data on the reaction force of the operation is input into an operation performance evaluation model.

[0138] The job performance evaluation model in step 6 is an artificial neural network model based on a multilayer perceptron. The construction and application process of the job performance evaluation model includes:

[0139] Step 6.1: In a laboratory environment, simulate various live-line working conditions and simultaneously acquire data from the three-dimensional line laser profile sensor. The feedback conductor deformation data sequence and the six-dimensional force / torque sensor The feedback reaction force data sequence was collected, and the corresponding operation status at each moment was manually labeled. The operation status included normal wire stripping, normal connection, slight tool jamming, and severe wire slippage. A total of 100,000 valid samples were collected and recorded at a rate of 100 Hz.

[0140] Step 6.2: Normalize the conductor deformation data sequence and the operational reaction force data sequence, and use them as input features. The input feature dimension is 4 (deformation residual + 3 force components), and the operational state at the corresponding time moment is used as the label, and one-hot encoding is performed. Divide the dataset into training set, validation set, and test set in an 8:1:1 ratio.

[0141] Step 6.3: Construct a multilayer perceptron neural network consisting of one input layer, three hidden layers, and one output layer. The input layer has 4 nodes, the three hidden layers have 64, 128, and 64 nodes respectively, and the output layer has 4 nodes (corresponding to the four working states). The Softmax function is used as the activation function. The ReLU activation function is used for the hidden layers.

[0142] Step 6.4: Supervised training of the multilayer perceptron neural network is performed using the training dataset. The network weights are optimized using the backpropagation algorithm until the network achieves a classification accuracy of over 98% on the validation set. This results in the trained neural network model, which is the job performance evaluation model. Training uses the Adam optimizer with a learning rate of 0.001, a batch size of 128, and 50 training epochs. The final accuracy on the test set reaches 98.2%.

[0143] Step 6.5: During the actual operation, the conductor deformation data and operation reaction force data collected at the current moment are input into the operation efficiency evaluation model in real time. The model performs forward calculation and outputs a four-dimensional probability vector. Each element in the probability vector represents the probability that the current operation state belongs to the corresponding category. The category with the highest probability value is the operation state parameter.

[0144] In step 6, if the operation status parameter indicates an abnormality, the ground control station issues an alarm signal and allows the operator to fine-tune the position of the second robotic arm's end effector using the wearable motion capture device. This specifically includes the following sub-steps:

[0145] Step 6.6: Set a confidence threshold, which is 0.9 in this embodiment. When the probability value of the operation performance evaluation model corresponding to the categories of "slight tool jamming" or "severe wire slippage" exceeds 0.9, it is determined that an abnormality has occurred in the operation process.

[0146] Step 6.7: The ground control station GCS immediately suspends the autonomous operation command output of the hybrid control model and issues an alarm signal through voice and graphical user interface to prompt the operator to intervene.

[0147] Step 6.8: The ground control station GCS switches the control mode of the second robotic arm ARM2 from the autonomous mode dominated by the hybrid control model to the position incremental control mode. In this mode, the minute movements made by the operator through the wearable motion capture device MOCAP will be calculated in real time as minute displacement increments of the end effector of the second robotic arm ARM2 in the six-degree-of-freedom working coordinate system, with a maximum movement speed limited to 0.02 m / s, and directly sent to the second robotic arm ARM2 as position commands.

[0148] Step 6.9: The operator performs fine-tuning by observing the on-site video footage and sensor data until the abnormal state is eliminated. The ground control station GCS monitors the six-dimensional force / torque sensor in real time. Feedback data, when the contact force returns to the normal range (e.g.) When the power drop to 5±1 Newtons and the normal state probability value output by the work efficiency evaluation model recovers to above 0.9, a prompt sound will be issued again.

[0149] Step 6.10: After the operator confirms that the abnormality has been eliminated, they return control of the second robotic arm ARM2 to the hybrid control model through a confirmation button. The system then switches back to the autonomous operation mode dominated by the hybrid control model and continues to execute the unfinished tasks.

[0150] At this point, a complete process of the robotic arm control method for uninterrupted power distribution network operation has been completed. Based on the above steps, the robotic arm can perform autonomous and collaborative operations in complex live environments, taking into account efficiency, accuracy and safety.

[0151] In an embodiment of the present invention, the described application scenario is as follows: During a live-line operation on a 10 kV distribution network, a new lead wire needs to be connected to the already energized main conductor L1. The working environment is characterized by a light breeze, causing the conductor to sway randomly by approximately ±5 cm. Insulator strings and crossarms are present near the work site, constituting obstacles.

[0152] I. System Composition:

[0153] This embodiment employs a dual-arm robotic system, comprising a first robotic arm ARM1 and a second robotic arm ARM2. Both arms are redundant degrees of freedom with seven rotary joints, driven by a high-precision harmonic reducer and a permanent magnet synchronous motor, achieving a repeatability accuracy better than ±0.05 mm. The two arms are mounted together on an insulated lifting platform PLAT, which is elevated to the high-altitude working position via an insulated boom lift.

[0154] A three-dimensional line laser profile sensor is installed on the end effector of the first robotic arm ARM1. The sensor It employs a high-speed contour scanner based on the principle of laser triangulation. Its X-axis (along the laser line direction) measurement range is 100 mm, its Z-axis (depth direction) measurement range is 200 mm, the number of X-axis contour points is 1280, and the highest line scanning frequency can reach 2 kHz.

[0155] A six-dimensional force / torque sensor is installed on the end effector of the second robotic arm ARM2. And a tool for completing tasks. A six-dimensional force / torque sensor. The strain gauge sensor has a force measurement range of ±200 Newtons in the X, Y, and Z directions, a torque measurement range of ±20 Newton-meters, a measurement accuracy better than 0.5% of full scale, and a data output frequency of 1 kHz. The tool is a modular, automatically replaceable tool; in this embodiment, an electric peeler is used. It is used to strip the insulation layer of wires.

[0156] The ground control station (GCS) is located on the ground and includes a high-performance industrial computer, a graphical user interface, and a wearable motion capture device (MOCAP). The MOCAP contains multiple inertial measurement units (IMUs), fixed to the operator's shoulder, upper arm, forearm, and hand, respectively, for real-time acquisition of upper limb joint angles at a sampling frequency of 120 Hz. The GCS communicates with the dual-arm robotic system via a fiber optic link. This link enables data interaction. Simultaneously carrying four high-definition video streams, all sensor data streams, and control command streams, the system employs the UDP protocol to ensure low latency. The wearable motion capture device MOCAP communicates with the ground control station GCS via a wireless local area network (WLAN). All electrical components of the entire dual-arm robotic system are powered by an isolation transformer, and all external communication interfaces are opto-isolated to meet the insulation and electromagnetic compatibility requirements of uninterrupted power supply operations.

[0157] II. Implementation steps:

[0158] 1. Environmental modeling and construction of the operational coordinate system:

[0159] Before the operation begins, the operator sends an initialization command to the dual-arm robotic system via the ground control station GCS. First, the first robotic arm ARM1 is controlled to drive the 3D line laser contour sensor. The system scans the target work line and its surrounding environment along a pre-defined scanning path. The scanning path is planned as follows: starting from the suspension clamp of conductor L1, it moves at a constant speed of 0.2 m / s along the conductor axis, while the sensor... The laser plane remains perpendicular to the conductor axis and swings up and down at a small angle to cover the insulator and crossarm areas. The entire scanning process lasts approximately 15 seconds, acquiring a raw 3D point cloud dataset including conductor L1, insulator string INS, and crossarm CROSS. .

[0160] The data preprocessing unit in the ground control station GCS processes the raw 3D point cloud dataset. Filtering is performed as follows: First, a statistical filtering algorithm is used to calculate the average distance between each point and its 50 nearest neighbors. Points whose average distance exceeds one standard deviation of the global mean are removed as isolated noise points. Then, a radius filtering algorithm is used. For each point, if fewer than three points are contained within a sphere centered at that point and with a radius of 1 mm, it is considered an outlier and removed. After these two steps, the first filtered point cloud data is obtained. .

[0161] Next, the first filtered point cloud data Compared with standard tower models pre-stored in the database Perform precise registration based on the Iterative Closest Point (ICP) algorithm. The ICP algorithm iteratively solves for an optimal spatial transformation matrix T (containing rotation and translation) such that... After T-transformation and The root mean square error between them is minimized. In this embodiment, the ICP algorithm converges after 20 iterations, yielding the transformation matrix. .Will Left multiply the coordinates of all points Transformed to a standard tower model The registered point cloud data is obtained in the global coordinate system based on the reference. .

[0162] Registered point cloud data In this study, a point cloud segmentation algorithm based on Euclidean clustering was used: the cluster search radius was set to 3 cm, and the minimum number of cluster points was set to 50. The algorithm iterates through the data sequentially. Points within a spatial distance of less than 3 cm were grouped into the same cluster. Finally, point cloud clusters belonging to conductor L1, insulator string INS, and crossarm CROSS were extracted. , , .

[0163] Converse guide point cloud cluster Cylindrical fitting is performed. A cylinder fitting algorithm based on Random Sample Consensus (RANSAC) is used, with a maximum of 1000 iterations and a distance threshold of 1 mm. The algorithm fits a cylinder, the axis of which is the spatial axis of the conductor. The radius of the cylinder is the same as the radius of the conductor. (The radius measured in this embodiment is 8.5 mm). For insulator point cloud clusters. And the horizontal beam dotted with cloud clusters Calculate the minimum axis-aligned bounding box for each, and mark the spatial occupancy of the bounding box in the global coordinate system as the critical obstacle region. and .

[0164] Finally, using the conductor axis A six-degree-of-freedom working coordinate system is established based on this. . origin Set on the conductor axis Above, located 0.5 meters from the clamp of the plumb line towards the work point. Z-axis With the axis of the conductor Coincident, pointing towards the work point. X-axis Perpendicular to And it points in the direction of the crossarm CROSS (determined by the centroid of the crossarm point cloud). Y-axis Determined by the right-hand rule, i.e. This coordinate system This will serve as the benchmark for all subsequent motion planning and control.

[0165] 2. Human-machine teaching and pre-job guidance trajectory generation:

[0166] After environmental modeling is completed, the system enters the master-slave collaboration preparation phase. Operators don wear MOCAP wearable motion capture devices and stand in front of the ground control station GCS, observing the 3D reconstructed image of the site on a monitor. The operator performs a series of teaching actions simulating wire connection: first, extending the right arm forward to simulate a robotic arm approaching the wire; then rotating the wrist to simulate adjusting the stripper's posture; finally, slightly lowering the arm to simulate placing the stripper against the wire surface. The entire teaching process lasts approximately 8 seconds.

[0167] The ground control station GCS acquires three-dimensional angle data from the inertial measurement units of the operator's shoulder, elbow, and wrist joints in real time at a sampling frequency of 120 Hz, obtaining the original joint angle sequence. Using a Kalman filter to... For filtering, the process noise covariance matrix Q of the Kalman filter is set as follows: Let the measurement noise covariance matrix R be set as Obtain smooth joint angle sequence .

[0168] The system pre-generates the human-machine kinematic mapping matrix. The matrix The generation process is as follows: During the system calibration phase, the operator performs a series of calibration actions, while simultaneously recording the joint angles of the operator and the first robotic arm ARM1 and the second robotic arm ARM2 in the working coordinate system. For the corresponding end-effector pose, at least 200 sets of data are collected, and a 6×9 mapping matrix is ​​fitted using the least squares method (9 joint angle inputs, 6-dimensional pose output). In this embodiment, The specific values ​​are obtained through offline calculations and then stored in the system.

[0169] Smooth joint angle sequence With mapping matrix Multiplication, real-time calculation in the working coordinate system Below, the desired pose data corresponding to the end effectors of the first robotic arm ARM1 and the second robotic arm ARM2. and The calculated desired pose data at all times are arranged in chronological order to form the desired pose sequence. .

[0170] To obtain smooth motion, a trapezoidal velocity planning algorithm is applied to the desired pose sequence. The algorithm interpolates the linear motion between adjacent pose points, setting the maximum linear velocity to 0.1 m / s, the maximum angular velocity to 0.2 radians / s, and the acceleration / deceleration time to 1 second. After velocity planning, a continuous, abrupt motion trajectory is generated. , as a pre-operation guidance trajectory.

[0171] 3. Real-time visual tracking and dynamic center point calculation:

[0172] When the dual robotic arm system begins to perform live wire connection work, the second robotic arm ARM2 first guides the work along the pre-operation trajectory. Under the rough guidance, the electric peeler Move to a preparatory position close to the guide wire L1, approximately 10 centimeters away from the guide wire. Then, activate the real-time tracking function of the first robotic arm ARM1.

[0173] To achieve precise line-of-sight alignment, the end effector of the first robotic arm ARM1 is controlled to enable the 3D line laser contour sensor to... The scanning light plane is always aligned with the electric peeler at the end of the second robotic arm ARM2. Located on the same perpendicular to the conductor axis Within the plane. This constraint is calculated in real-time by the sensor. and tools exist The first robotic arm ARM1 is controlled by measuring the coordinate difference in the direction and providing real-time feedback.

[0174] 3D line laser profile sensor The section of the conductor to be worked on is continuously scanned at a line scanning frequency of 1 kHz. Each scan acquires a two-dimensional contour line containing the surface profile of the conductor, and each contour line contains 1280 points. Continuous scans accumulate to form a microscopic contour point cloud. .

[0175] For each scanned 2D contour line, a sub-pixel extraction algorithm based on the centroid method is used to extract the conductor edge points. Specifically, on the contour line, regions where the laser reflection intensity changes abruptly correspond to conductor edges. The algorithm calculates the grayscale centroid position along the vertical direction in these intensity abrupt regions, which is taken as a conductor edge point on that scanned section. Since the contour line has two edges, left and right, two edge points can be extracted for each section. The average of these two edge points is taken to obtain the coordinates of the conductor center in the sensor coordinate system for that section.

[0176] Subsequently, the center points of the conductors from multiple cross-sections obtained at the same time (in this embodiment, 5 cross-sections before and after the current time, for a total of 11 cross-sections) are fitted using least-squares circle fitting. The fitting algorithm is based on SVD decomposition to find an optimal fitted circle, whose center coordinates are... That is, the instantaneous geometric center of the conductor on the scanning section in the sensor coordinate system. The coordinates below.

[0177] Using a pre-calibrated hand-eye calibration matrix The instantaneous geometric center coordinates are changed from the sensor coordinate system. Transform to the working coordinate system Below, the coordinates of the dynamic center point are obtained. .

[0178] Finally, the dynamic center point coordinates obtained at the current moment are... Current pose of the ARM2 end effector of the second robotic arm Compare and calculate the two in the working coordinate system. The differences in the X, Y, and Z axes form a real-time offset vector. Since the end effector of the robotic arm remains essentially unchanged before contact, only positional offset is considered here.

[0179] 4. Hybrid control model outputs composite instructions:

[0180] The real-time controller inputs the real-time offset vector E(t) into a hybrid control model that combines fuzzy logic and PID control. This model operates on a real-time controller with an FPGA+DSP architecture and a 5 kHz cycle.

[0181] First, the magnitude e(t) = |E(t)| of the real-time offset vector E(t) and its rate of change de(t) / dt are input variables into a fuzzy inference engine. Fuzzy rules are defined in the fuzzy inference engine, for example:

[0182] If e(t) is "large" and de(t) / dt is "positive large", then the output will have a large proportional coefficient correction, a small integral coefficient correction, and a small differential coefficient correction.

[0183] If e(t) is "small" and de(t) / dt is "zero", then output a moderate proportional coefficient correction, a moderate integral coefficient correction, and a large differential coefficient correction.

[0184] This embodiment uses a triangular membership function to divide e(t) and de(t) / dt into seven fuzzy sets. The fuzzy inference engine uses the centroid method to defuzzify the data based on the rules activated by the current input, and outputs three correction coefficients for adjusting the PID controller parameters. , , The values ​​range from 0.5 to 2.0.

[0185] The parameter-self-tuning PID controller receives the three components of the real-time offset vector E(t) as error signals and dynamically adjusts its proportional, integral, and derivative gains according to the correction coefficients. Taking the X-axis direction as an example, the control law is:

[0186] ;

[0187] Among them, the initial gain , , (Units have been dimensionless). The Y-axis is similar to the Z-axis. The control quantities in the three directions are combined to obtain the basic control quantity. Hybrid control models will The basic motion commands are mapped and sent to the position loop controller of the second robotic arm ARM2 via the EtherCAT bus to achieve macroscopic position tracking, enabling the end effector of the robotic arm to follow the continuous trajectory formed by the dynamic center point.

[0188] Meanwhile, the hybrid control model passes each of the three components of the real-time offset vector E(t) through a high-pass filter with a cutoff frequency of 5 Hz to extract the high-frequency components with frequencies higher than 5 Hz. High-frequency components Multiply by a preset stiffness coefficient Newtons per meter (converted via force control equivalents) generate a force correction quantity. Correct the force amount The code is encoded as a micro-motion correction instruction and directly superimposed on the current loop setpoints of the servo drivers of each joint of the second robotic arm ARM2 via the FPGA's high-speed parallel bus. This is equivalent to connecting a high-bandwidth force disturbance in parallel outside the position loop, driving the robotic arm end effector to generate high-frequency micro-motions with frequencies up to several hundred hertz and amplitudes on the millimeter level in Cartesian space, thereby compensating in real time for instantaneous position deviations caused by wire deflection or robot body vibration. In this embodiment, the measured wire deflection frequency is approximately 2–3 hertz. Through this high-frequency micro-motion correction, the end effector dynamic tracking error can be reduced from ±3 mm to within ±0.3 mm.

[0189] 5. Force-sensory contact and resistance control:

[0190] Electric peeler During the approach to the conductor, a six-dimensional force / torque sensor is used for real-time monitoring. Feedback data from the electric peeler. When contact occurs with wire L1, the contact force detected by the sensor increases rapidly. This embodiment presets a first threshold. Newtons (resultant force value). When the resultant force exceeds 2.0 Newtons, the hybrid control model automatically switches the control mode, changing the calculation basis of the micro-motion correction command from the real-time offset vector to the force / torque data fed back by the force / torque sensor.

[0191] Six-dimensional force / torque sensor Real-time detection of force components in three directions at a sampling rate of 1 kHz and torque components in three directions This generates the original force / torque data. Because the sensor is mounted at the end of the robotic arm, its measurements include the tool's own weight. Therefore, gravity compensation is necessary: ​​the electric peeler must be pre-measured. The mass m = 1.2 kg, and the coordinates of its center of gravity in the sensor coordinate system are known. Based on the current posture of the robotic arm's end effector (obtained from encoder feedback), calculate the components of gravity in each direction. Subtracting from the middle yields the pure contact force / torque data. .

[0192] Pure contact force / torque data With desired contact force A comparison is made. In this embodiment, the desired contact force is set to... Newton's law, directed along the negative Z-axis of the sensor coordinate system (pointing towards the conductor). Force error. .

[0193] Force error An impedance-controlled filter is input to a mass-spring-damped system. The dynamic characteristics of this filter are described by a second-order differential equation:

[0194] ;

[0195] in, , , These are the preset target inertia matrix, target damping matrix, and target stiffness matrix, respectively. In this embodiment, to obtain compliant contact behavior, the following are set... It is a diagonal matrix, and the diagonal elements are each 0.5 kg. The diagonal element is taken as 20 Newton-seconds per meter; The diagonal element is taken as 100 Newtons / meter. , , These are the desired position correction, velocity correction, and acceleration correction, respectively.

[0196] Solve the above second-order differential equation (discretization using the bilinear transform method, sampling time...). (seconds) to calculate in real time the position correction that the end effector of the second robotic arm ARM2 needs to make under the current force error. .Will This is superimposed onto the continuous trajectory formed by the coordinates of the dynamic center point, generating a corrected desired trajectory, which is then sent to the position loop controller. In this way, the electric peeler... Apply a constant contact force of approximately 5 Newtons to the surface of the conductor and slide it along the surface of the conductor to the preset stripping position (15 cm from the origin of the working coordinate system determined in step 1 along the conductor axis).

[0197] 6. Work performance evaluation and anomaly handling:

[0198] Once the end of the second robotic arm ARM2 reaches the preset stripping position, the electric stripper is activated. The peeling process begins with rotation. During peeling, a 3D line laser contour sensor is fused in real-time. Feedback of conductor deformation data (i.e., the deviation of the conductor profile from the ideal circle, calculated as the fitting residual) and a six-dimensional force / torque sensor. The feedback data on the reaction force of the operation is input into an operation performance evaluation model.

[0199] The performance evaluation model for this task is an artificial neural network based on a multilayer perceptron. Its construction process is as follows:

[0200] First, in a laboratory environment, various live-line working conditions were simulated, and numerous wire stripping tests were conducted using the same robotic arm and sensor system. Simultaneously, conductor deformation data sequences (root mean square value of the fitted residual for each scan) and operational reaction force data sequences (forces in three directions) were acquired and recorded at a rate of 100 Hz. Meanwhile, experienced operators manually annotated the corresponding operational status at each moment through video playback, categorizing it into four types: normal wire stripping, normal connection, slight tool jamming, and severe conductor slippage. A total of 100,000 valid samples were collected.

[0201] The conductor deformation data and operational reaction force data were normalized (mean 0, variance 1) and used as input features. The input feature dimension was 4 (deformation residual + 3 force components). The operational state at the corresponding time point was used as the label and one-hot encoded. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio.

[0202] A multilayer perceptron neural network was constructed, consisting of one input layer (4 nodes), three hidden layers (64, 128, and 64 nodes respectively), and one output layer (4 nodes). The hidden layer activation function was ReLU, and the output layer used the Softmax function. Supervised training was performed on the training set using the Adam optimizer, a learning rate of 0.001, a batch size of 128, and 50 training epochs. After training, the classification accuracy reached 98.5% on the validation set and 98.2% on the test set. This trained neural network model was then deployed to the real-time inference engine of the ground control station's GCS.

[0203] In actual operation, the conductor deformation data (denoted as d(t)) and the operation reaction force data collected at the current moment are transmitted in real time. The input is fed into the job performance evaluation model. After forward computation, the model outputs a four-dimensional probability vector. These represent the probabilities of normal operation, slight jamming, severe slippage, and other abnormalities, respectively. The category with the highest probability value is the operational status parameter.

[0204] Set confidence threshold When the model outputs or When the value exceeds 0.9, an anomaly is determined to have occurred in the operation process. For example, during a peeling process, sensor feedback indicates a sudden increase in torque, abnormal wire deformation, and the model output... The threshold is exceeded.

[0205] The ground control station GCS immediately transmitted the signal via fiber optic link. The autonomous operation command output of the hybrid control model is suspended, and an alarm signal is issued by voice broadcasting "Tool stuck, please intervene" and by flashing a red alarm box on the graphical interface.

[0206] Simultaneously, the ground control station GCS switched the control mode of the second robotic arm ARM2 from the autonomous mode dominated by the hybrid control model to the position incremental control mode. In this mode, the minute movements made by the operator through the wearable motion capture device MOCAP will be calculated in real time as the position of the second robotic arm ARM2 end effector in the working coordinate system. The operator generates minute displacement increments (maximum movement speed limited to 0.02 m / s) and sends them directly as position commands to the second robotic arm ARM2. The operator then performs fine-tuning operations by observing the high-definition video feed and force / torque data curves, such as gently lifting the robotic arm to disengage the peeler from its stuck position.

[0207] Ground control station GCS real-time monitoring of six-dimensional force / torque sensors Feedback data, when the contact force returns to the normal range (e.g.) The probability of a normal state is reduced to 5±1 Newtons and the output of the work efficiency evaluation model. When the error value recovers to above 0.9, a "Resolved Anomaly" prompt will sound again. After the operator confirms that the anomaly has been resolved, they can return control of the second robotic arm ARM2 to the hybrid control model via a confirmation button on the graphical interface. The system will then switch back to autonomous operation mode and continue performing the unfinished wire stripping task until the insulation layer is completely stripped, preparing for subsequent wire splicing.

[0208] III. Comparison of Results:

[0209] To verify the effectiveness of the method described in this embodiment, a comparative experiment was conducted under the same 10 kV distribution network live-line connection operation scenario. The comparison objects included:

[0210] Comparison Method 1 (pure remote operation): The operator controls the end positions of the two robotic arms using two 3D mice and completes the wire connection operation by relying on multiple video feeds.

[0211] Comparison Method 2 (purely autonomous): The system relies entirely on vision servoing and autonomously completes the wire connection operation according to the preset program, without any human intervention mechanism.

[0212] Each method was repeated 20 times, and the success rate, average operation time, contact force overshoot (reflecting the impact on the conductor), and success rate of handling abnormal situations were recorded. The results are shown in Table 1.

[0213]

[0214] Table 1 Comparison of the effects of different control methods

[0215] As shown in Table 1, the method in this embodiment achieves a 100% success rate while maintaining high operational efficiency (85 seconds), with minimal contact force overshoot (6.2 Newtons), and can handle 100% of abnormal situations that occur during the operation. While the purely remote operation method can handle abnormalities, it is inefficient, highly dependent on operator skill, and has a low success rate; the purely autonomous method is fast, but cannot autonomously recover from abnormalities, resulting in a low overall success rate. The comparative results demonstrate that the method of this invention effectively combines the high efficiency of autonomous operation with the reliability of human-machine collaboration.

[0216] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robotic arm control method for uninterrupted power distribution network operation, characterized in that: The control steps include the following: Step 1: Install a dual robotic arm system, which includes a first robotic arm and a second robotic arm. A three-dimensional line laser profile sensor is installed on the end effector of the first robotic arm, and a six-dimensional force / torque sensor and a working tool are installed on the end effector of the second robotic arm. By controlling the first robotic arm to drive the three-dimensional line laser profile sensor to scan the target working line, three-dimensional point cloud data is obtained, and a six-degree-of-freedom working coordinate system is constructed based on the three-dimensional point cloud data. Step 2: Collect the operator's teaching actions through the ground control station, calculate them into the desired pose sequence of the first and second robotic arms, and use the desired pose sequence as the pre-operation guidance trajectory; Step 3: During the operation, the first robotic arm is controlled to drive the three-dimensional line laser contour sensor to scan the conductor in real time, generate a micro-contour point cloud, and calculate the dynamic center point coordinates and real-time offset vector of the conductor from the micro-contour point cloud in real time. Step 4: Input the real-time offset vector into a hybrid control model, and the hybrid control model outputs a composite control command. The composite control command includes a basic motion command and a micro-motion correction command, wherein: The basic motion commands are used to control the second robotic arm to follow a continuous trajectory composed of the coordinates of a dynamic center point. The micro-motion correction command is used to drive the end of the second robotic arm to generate a high-frequency, small-amplitude compliant movement to compensate for instantaneous position deviation. Step 5: Monitor the feedback data from the six-dimensional force / torque sensor in real time: When the working tool contacts the wire and the feedback data exceeds a preset first threshold, the hybrid control model switches the control mode, switches the calculation basis of the micro-motion correction command to the force / torque data fed back by the six-dimensional force / torque sensor, and dynamically adjusts the position correction amount of the end of the second robotic arm according to an impedance control model, so that the working tool adheres to the surface of the wire with the set contact force and slides to the preset position. Step 6: During the operation, the conductor deformation data fed back by the three-dimensional line laser profile sensor and the operation reaction force data fed back by the six-dimensional force / torque sensor are integrated in real time and input into an operation efficiency evaluation model. If the work status parameters output by the work performance evaluation model show abnormalities, the ground control station will issue an alarm signal and allow the operator to fine-tune the position of the end effector of the second robotic arm through a wearable motion capture device.

2. The robotic arm control method for uninterrupted power distribution network operation according to claim 1, characterized in that: The specific method for constructing a six-degree-of-freedom working coordinate system based on the three-dimensional point cloud data in step 1 is as follows: Step 1.1: Control the first robotic arm to drive the three-dimensional line laser contour sensor to perform multi-angle scanning of the target working area, including the conductor, insulator and crossarm, at a preset scanning path and scanning speed, and obtain the original three-dimensional point cloud dataset; Step 1.2: The original 3D point cloud dataset is filtered by a data preprocessing unit deployed at the ground control station to remove isolated noise points and outliers caused by abnormal light reflection, thus obtaining the first filtered point cloud data. Step 1.3: Perform precise registration between the first filtered point cloud data and the standard tower model pre-stored in the database based on the iterative nearest point algorithm. The iterative nearest point algorithm is used to iteratively solve for an optimal spatial transformation matrix, so as to minimize the root mean square error between the first filtered point cloud data and the standard tower model. Step 1.4: Based on the spatial transformation matrix, transform the coordinates of all points in the first filtered point cloud data to a global coordinate system based on the standard tower model to obtain the registered point cloud data; Step 1.5: In the registered point cloud data, a segmentation algorithm based on Euclidean clustering is used to extract point cloud clusters belonging to conductors, insulators and crossarms respectively; Step 1.6: Perform cylindrical surface fitting on the extracted conductor point cloud clusters to solve the spatial axis equation of the conductor, thereby obtaining the accurate pose and diameter parameters of the conductor. At the same time, construct a three-dimensional bounding box for the point cloud clusters of the insulator and crossarm, and mark the spatial occupancy range of the three-dimensional bounding box in the global coordinate system as the key obstacle area. Step 1.7: Using the spatial axis of the conductor as a reference, and combining the conductor diameter parameters and key obstacle areas, establish a six-degree-of-freedom working coordinate system. The origin of the six-degree-of-freedom working coordinate system is located at a preset position on the conductor axis near the starting point of the operation. The Z-axis coincides with the conductor axis, the X-axis points in the direction of the crossarm, and the Y-axis is determined by the right-hand rule.

3. The robotic arm control method for uninterrupted power distribution network operation according to claim 2, characterized in that: The specific method for step 2 is as follows: Step 2.1: The operator wears a wearable motion capture device containing multiple inertial measurement units and performs a set of teaching actions simulating the work process; the ground control station collects the three-dimensional angle data of the operator's upper limb shoulder joint, elbow joint and wrist joint in real time at a sampling frequency of more than 100 Hz to obtain the original joint angle sequence. Step 2.2: Perform Kalman filtering on the original joint angle sequence to remove Gaussian noise during signal acquisition and obtain a smooth joint angle sequence; Step 2.3: Call a pre-generated human-machine kinematics mapping matrix. The human-machine kinematics mapping matrix is ​​obtained by having the operator perform a series of calibration actions during the system calibration phase, while recording the joint angles of the operator and the corresponding end poses of the first and second robotic arms in the six-degree-of-freedom working coordinate system, and fitting the data using the least squares method. Step 2.4: Multiply the smooth joint angle sequence with the human-machine kinematics mapping matrix to calculate in real time the expected pose data corresponding to the end effectors of the first and second robotic arms in the six-degree-of-freedom working coordinate system; Step 2.5: Arrange the calculated expected pose data at all times in chronological order to form the expected pose sequence. At the same time, apply a preset velocity planning algorithm to the expected pose sequence to smooth the motion velocity between adjacent pose points and generate a continuous, abrupt motion trajectory as a pre-operation guidance trajectory.

4. The robotic arm control method for uninterrupted power distribution network operation according to claim 3, characterized in that: The specific method for step 3 is as follows: Step 3.1: After the second robotic arm reaches the preparatory position under the guidance of the pre-operation guide trajectory, activate the visual servo tracking function of the first robotic arm; control the movement of the end effector of the first robotic arm so that the scanning light plane of the three-dimensional line laser contour sensor is always in the same plane perpendicular to the guide axis as the working tool at the end effector of the second robotic arm. Step 3.2: The three-dimensional line laser profile sensor continuously scans the section of the conductor to be worked on at a line scanning frequency of more than 1 kHz. Each scan acquires a two-dimensional profile line containing the surface profile of the conductor. Multiple scans are accumulated to form the micro-profile point cloud. Step 3.3: For the two-dimensional contour line obtained in each scan, a sub-pixel extraction algorithm based on the centroid method is used to calculate the gray-scale centroid position in the vertical direction of the contour line. The gray-scale centroid position is an edge point of the conductor on the scan section. Step 3.4: Perform least-squares circle fitting on multiple edge points obtained from scanning at the same time to obtain a fitted circle. The coordinates of the center of the fitted circle are the coordinates of the instantaneous geometric center of the conductor on the scanning section in the sensor coordinate system. Step 3.5: Using a pre-calibrated hand-eye calibration matrix, transform the coordinates of the instantaneous geometric center in the sensor coordinate system to the six-degree-of-freedom working coordinate system to obtain the coordinates of the dynamic center point; Step 3.6: Compare the current dynamic center point coordinates obtained at the current moment with the current pose of the end effector of the second robotic arm, calculate the difference between the two in the X-axis, Y-axis and Z-axis directions in the six-degree-of-freedom working coordinate system, and form a real-time offset vector.

5. The robotic arm control method for uninterrupted power distribution network operation according to claim 4, characterized in that: The specific method for step 4 is as follows: Step 4.1: Input the real-time offset vector and its rate of change as input variables into a fuzzy inference engine. The fuzzy inference engine predefines multiple fuzzy rules for different deviation magnitudes and trends. Step 4.2: The fuzzy inference engine performs inference based on the input variables and fuzzy rules, and outputs three correction coefficients for adjusting the PID controller parameters, denoted as follows: , , ; Step 4.3: A parameter-self-tuning PID controller receives the real-time offset vector as the error signal e(t) and adjusts it according to the correction coefficient. , , Dynamically adjust its proportional, integral, and derivative gains, and calculate a basic control quantity according to the following control law. : ； in, , , This is the initial gain of the PID controller; Step 4.4: The hybrid control model will use the basic control variables. The basic motion commands are mapped and sent to the position loop controller of the second robotic arm to achieve macroscopic position tracking. Step 4.5: Simultaneously, the hybrid control model passes the real-time offset vector e(t) through a high-pass filter to extract the high-frequency components with frequencies higher than 5 Hz. ; Step 4.6: Convert the high-frequency components Multiply by a preset stiffness coefficient Generate a force correction quantity The force correction amount The code is encoded as the micro-motion correction instruction and directly superimposed on the current loop setpoint of the second robotic arm joint servo driver through a high-speed real-time bus, driving the end effector of the second robotic arm to generate high-frequency micro-motion in Cartesian space.

6. The robotic arm control method for uninterrupted power distribution network operation according to claim 5, characterized in that: The specific method for step 5 is as follows: Step 5.1: The six-dimensional force / torque sensor detects the force components in three directions generated when the working tool comes into contact with the wire in real time at a sampling rate of 1 kHz. and torque components in three directions This generates the original force / torque data; Step 5.2: Perform gravity compensation calculations on the original force / torque data, subtracting the component of the working tool's own weight under different postures, to obtain the pure contact force / torque data, denoted as... ; Step 5.3: The pure contact force / torque data Contact force with a pre-set expectation By comparison, the force / torque error is obtained. ; Step 5.4: Calculate the force / torque error. An impedance-controlled filter is input to a mass-spring-damped system, the dynamic characteristics of which are described by a second-order differential equation: ； in, , , These are the preset target inertia matrix, target damping matrix, and target stiffness matrix, respectively; , , These are the desired position correction, velocity correction, and acceleration correction, respectively. Step 5.5: Solve the second-order differential equation to calculate in real time the position correction required by the end effector of the second robotic arm under the current force error. ; Step 5.6: Adjust the position correction amount The modified desired trajectory is superimposed onto the continuous trajectory formed by the coordinates of the dynamic center point and sent to the position loop controller of the second robotic arm, so that the working tool slides against the surface of the guide wire with a constant desired contact force.

7. The robotic arm control method for uninterrupted power distribution network operation according to claim 6, characterized in that: The specific method for integrating the operational reaction force data into an operational performance evaluation model in step 6 is as follows: Step 6.1: In a laboratory environment, simulate various live-line working conditions, simultaneously collect the conductor deformation data sequence fed back by the three-dimensional line laser profile sensor and the working reaction force data sequence fed back by the six-dimensional force / torque sensor, and manually label the working state corresponding to each moment. The working states include normal wire stripping, normal connection, slight tool jamming, and severe conductor slippage. Step 6.2: After normalizing the conductor deformation data sequence and the operation reaction force data sequence, use them as input features and the operation status at the corresponding time as labels to form a training dataset; Step 6.3: Construct a multilayer perceptron neural network containing one input layer, three hidden layers, and one output layer; the number of nodes in the input layer is the same as the dimension of the input features, the number of nodes in the output layer is the same as the number of job state categories, and the Softmax function is used as the activation function; Step 6.4: Supervised training of the multilayer perceptron neural network is performed using the training dataset. The network weights are optimized through backpropagation algorithm until the classification accuracy of the network on the validation set reaches more than 98%. The trained neural network model is obtained, which is the job performance evaluation model. Step 6.5: During the actual operation, the conductor deformation data and operation reaction force data collected at the current moment are input into the operation efficiency evaluation model in real time. After forward calculation, the model outputs a five-dimensional probability vector. Each element in the probability vector represents the probability that the current operation state belongs to the corresponding category. The category with the highest probability value is the operation state parameter.

8. The robotic arm control method for uninterrupted power distribution network operation according to claim 7, characterized in that: The specific method for the ground control station to issue alarm signals and for operators to make fine-tuning interventions in step 6 is as follows: Step 6.6: Set a confidence threshold; when the probability value of the corresponding tool slight jamming or wire severe slippage category output by the work performance evaluation model exceeds the confidence threshold, it is determined that an abnormality has occurred in the work process; Step 6.7: The ground control station immediately suspends the autonomous operation command output of the hybrid control model and issues an alarm signal through voice and graphical user interface to prompt the operator to intervene; Step 6.8: The ground control station switches the control mode of the second robotic arm from the autonomous mode dominated by the hybrid control model to the position increment control mode. In this mode, the small movements made by the operator through the wearable motion capture device will be calculated in real time as the small displacement increment of the end of the second robotic arm in Cartesian space, and sent directly as position commands to the second robotic arm. Step 6.9: Operators make fine adjustments by observing the on-site video footage and sensor data until the abnormal state is eliminated; the ground control station monitors the feedback data of the six-dimensional force / torque sensor in real time, and issues a prompt sound again when the contact force returns to the normal range and the normal state probability value output by the work efficiency evaluation model returns to above 0.

9. Step 6.10: After the operator confirms that the anomaly has been resolved, they can return control of the second robotic arm to the hybrid control model by pressing a confirmation button. The system will then switch back to the autonomous operation mode dominated by the hybrid control model and continue to execute the unfinished tasks.

9. The robotic arm control method for uninterrupted power distribution network operation according to claim 1, characterized in that: Both the first robotic arm and the second robotic arm are redundant degrees of freedom robotic arms with 7 rotary joints, and are mounted on the same insulated lifting platform; The three-dimensional line laser profile sensor is a high-speed profile scanner based on the principle of laser triangulation. The six-dimensional force / torque sensor is a strain gauge type or an optical sensor; The working tools are modular tools that can be automatically replaced, including an electric peeler, a hydraulic connecting pliers, and a bolt fastener.

10. The robotic arm control method for uninterrupted power distribution network operation according to claim 1, characterized in that: The ground control station and the dual robotic arm system exchange data via an optical fiber communication link, which simultaneously carries high-speed video streams, sensor data streams, and control command streams. The wearable motion capture device communicates with the ground control station via a wireless local area network; All electrical components of the dual robotic arm system are powered by an isolation transformer, and all external communication interfaces are opto-isolated to meet the insulation and electromagnetic compatibility requirements of power distribution network uninterrupted operation sites.

Images