An unmanned aerial vehicle insulator zero value measurement and autonomous obstacle avoidance method combining visual perception and flight parameter control

Abstract

The application provides a UAV insulator zero-value measurement and autonomous obstacle avoidance method fusing visual perception and flight parameter control, comprising obtaining UAV flight parameters, insulator detection data and voltage grade information; querying an electromagnetic danger zone radius mapping table based on the voltage grade to construct a three-dimensional cylindrical electromagnetic danger zone model; predicting a feasible flight trajectory set and dividing electromagnetic danger candidate path segments and electromagnetic safety candidate path segments; establishing an instruction priority mapping relationship, calculating a comprehensive cost function value; determining an optimal flight path and outputting a flight control instruction; and performing autonomous obstacle avoidance flight and insulator zero-value measurement. Through multi-source instruction weighted priority collaborative arbitration and rolling horizon dual constraint verification, double-layer fusion autonomous obstacle avoidance is realized, and the zero-value insulator recognition accuracy is improved by using visible light and infrared thermal image dual-mode feature comparison.

Description

Technical Field

[0001] This invention relates to the field of UAV visual image recognition technology, and in particular to a method for measuring the zero value of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control. Background Technology

[0002] In recent years, with the rapid development of UAV technology and intelligent inspection, using UAVs equipped with detection equipment for insulator zero-value measurement has become an important technical means. However, in terms of electromagnetic safety constraints, existing technologies have not yet formed a systematic three-dimensional electromagnetic hazard zone quantitative characterization method, making it difficult to accurately assess the safe distance between UAVs and live conductors; in terms of decision fusion, the priority arbitration relationship between electrical safety commands, physical obstacle avoidance commands, and task tracking commands is still imperfect, and the collaborative decision-making mechanism of multi-source commands needs further research; in terms of dynamic environment adaptability, the reliability compensation capability of obstacle avoidance paths under strong wind conditions is insufficient, and the real-time performance and robustness of rolling time-domain safety replanning strategies still have room for improvement; in terms of environmental perception accuracy, the existing voxel map resolution adaptive adjustment mechanism and dynamic obstacle response speed are still insufficient to meet the refined requirements of UHV inspection.

[0003] Therefore, how to construct a three-dimensional spherical electromagnetic hazard zone based on the conductor centerline, realize the weighted priority collaborative arbitration of three types of instructions: electrical safety, physical obstacle avoidance, and task tracking, and perform electromagnetic and collision dual constraint verification on multiple candidate paths under rolling time domain control, so as to achieve dual-layer fusion autonomous obstacle avoidance for UHV UAV inspection, is a technical problem that needs to be solved in the current development of UHV transmission line UAV intelligent inspection technology. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by providing a method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control. This method solves the technical problem in the prior art that it is impossible to achieve quantitative characterization of electromagnetic hazard zones, collaborative arbitration of multi-source command weighted priority, and dual constraint verification of electromagnetic and collision for multiple candidate paths in the complex environment of UHV transmission lines.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control, comprising: Acquire drone flight parameters, visible light and infrared thermal images of insulators, and voltage level and current intensity of the current working line to construct an initial dataset; Based on the voltage level and current intensity, a preset electromagnetic hazard zone radius mapping table is queried. A three-dimensional spherical electromagnetic hazard zone is constructed with the reference axis of the center line position of each conductor and the corresponding electromagnetic hazard zone radius as the radius. The electromagnetic hazard zones of each phase conductor are then subjected to a union operation to obtain a three-dimensional electromagnetic hazard zone model. Using the current moment as the starting point of the rolling window, several candidate flight trajectories are generated based on flight parameters and a preset kinematic model, and the candidate flight trajectories are discretized into several candidate path point sequences; The candidate path point sequence is divided into electromagnetic hazard path segments and electromagnetic safe path segments according to the three-dimensional electromagnetic hazard area model. An electromagnetic penalty factor is applied to the electromagnetic hazard path segments, and a collision penalty factor is applied to the electromagnetic safe path segments whose distance from the obstacle is less than a preset safety threshold. The electromagnetic penalty factor and the collision penalty factor are weighted and summed according to the preset weights of three types of instructions: electrical safety, physical obstacle avoidance and task tracking, to obtain the comprehensive cost function value of each candidate path. Based on the comprehensive cost function value, the path with the minimum cost is selected as the global optimal path. The local obstacle avoidance layer generates local obstacle avoidance adjustment commands based on the distance to the nearest obstacle and the relative speed. The two vectors are superimposed and fused to output the final wind control command. Guided by flight control commands, the UAV is directed to a preset insulator detection location. The visible light image and infrared thermal image are used to extract appearance and temperature analysis features, respectively. These features are then compared with a preset zero-value insulator feature template to identify zero-value insulators and complete the zero-value measurement of the insulator.

[0006] Furthermore, based on voltage levels and a pre-defined electromagnetic hazard zone radius mapping table, a three-dimensional electromagnetic hazard zone model is constructed, including: Obtain the conductor arrangement and conductor sag parameters for the current work line; Calculate the coordinate sequence of the centerline of each phase conductor in three-dimensional space based on the conductor sag parameters; Based on the voltage level and current intensity of each phase conductor in the initial dataset, the electromagnetic hazard zone radius mapping table is queried to obtain the electromagnetic hazard zone radius corresponding to each phase conductor. The electromagnetic hazard zone radius is then corrected by combining the current ambient temperature, humidity and air pressure parameters to obtain the corrected electromagnetic hazard zone radius. A circular cross section is generated with each point on the coordinate sequence of the center line of each phase conductor as the center and the radius of the corrected electromagnetic hazard zone as the radius. The three-dimensional sphere is generated by scanning the direction of the center line of the conductor. By performing a union operation on the three-dimensional spheres corresponding to each phase conductor, a three-dimensional electromagnetic hazard area model around the UHV transmission line is obtained.

[0007] Furthermore, based on the flight parameters, the preset kinematic model, and the pre-acquired environmental perception data, several candidate path point sequences are obtained, including: A kinematic model of the UAV is established, which includes the UAV's position state equation, velocity state equation, and attitude state equation. The parameters of the kinematic model are updated based on the UAV flight parameters in the initial dataset and the wind speed and wind direction estimates in the environmental perception data. Based on the kinematic model with updated parameters, a multi-step forward simulation is performed with the current time as the rolling window point and the prediction time domain length as the window length to generate a set of flight trajectories of the UAV in the prediction time domain. Each candidate flight trajectory in the flight trajectory set is discretized according to a preset sampling interval to obtain the corresponding candidate path point sequence.

[0008] Furthermore, based on the electromagnetic hazard candidate path segments and the three-dimensional electromagnetic hazard region model, the command priority mapping relationship between each electromagnetic hazard candidate path segment and the corresponding three-dimensional electromagnetic hazard region is determined, including: Three independent instruction cache queues are set up to temporarily store electrical safety instructions, physical obstacle avoidance instructions, and task tracking instructions, respectively. The electrical safety command is assigned a first preset weight coefficient, the physical obstacle avoidance command is assigned a second preset weight coefficient, and the task tracking command is assigned a third preset weight coefficient, wherein the first preset weight coefficient is greater than the second preset weight coefficient, and the second preset weight coefficient is greater than the third preset weight coefficient. When an instruction arrives in the instruction buffer queue, the arrival timestamp and preset validity period of the instruction are recorded, and the time difference between the arrival time and the current time is calculated. If the time difference exceeds the effective duration, the corresponding instruction is deleted from the corresponding instruction buffer queue. The comprehensive priority score is calculated based on the weight coefficient and waiting time of each instruction, and the comprehensive priority score is mapped to each electromagnetic hazard candidate path segment and the corresponding three-dimensional electromagnetic hazard area to form an instruction priority mapping relationship.

[0009] Furthermore, the spatial relationship between each path point in the candidate path point sequence and the three-dimensional electromagnetic hazard area model is determined one by one, and the comprehensive cost function value of each candidate path is obtained, including: For each path point in the candidate path point sequence, obtain the three-dimensional spatial coordinates corresponding to each path point; Call the query interface of the three-dimensional electromagnetic hazard area model, input the three-dimensional spatial coordinates of a certain path point, and determine whether the path point is located within the three-dimensional electromagnetic hazard area; If the judgment result is yes, then an electromagnetic penalty factor is applied to the candidate path; Obtain several voxels adjacent to the voxel where the path point is located in the voxelized map constructed from the LiDAR point cloud, and read whether there are voxels marked as dynamic or static obstacles among the adjacent voxels. If the dynamic or static obstacle elements exist, calculate the Euclidean distance between the path point and the center point of the obstacle element. If the Euclidean distance is less than the preset safety threshold, a collision penalty factor is applied to the candidate path; The electromagnetic penalty factor and collision penalty factor corresponding to all path points on each candidate path are weighted and summed in combination with the instruction priority mapping relationship to obtain the comprehensive cost function value of each candidate path.

[0010] Furthermore, based on the real-time sensed distance to the nearest obstacle and relative velocity, local obstacle avoidance adjustment commands are generated. These commands are then fused with the optimal flight path to output the final flight control commands, including: In the local obstacle avoidance layer, the distance to the nearest obstacle and the relative speed data are received in real time; Based on the distance to the nearest obstacle and the relative speed, a preset local obstacle avoidance decision table is queried to obtain a local obstacle avoidance adjustment instruction; In the global path optimization layer, the comprehensive cost function value is received, and the candidate path with the smallest comprehensive cost function value is selected as the optimal flight path at the current moment. The local obstacle avoidance modulation command and the optimal flight path are vector-superimposed to obtain the final flight control command. The flight control commands are finally arbitrated based on the comprehensive priority score in the command priority mapping relationship to obtain the final result, and the final result is output to the flight control execution terminal.

[0011] Furthermore, the UAV is guided to perform autonomous obstacle avoidance flight according to the flight control commands, and the insulator zero-value measurement results are obtained, including: When the UAV flies to the preset insulator detection position according to the flight control command, it collects visible light images and infrared thermal images of the insulator string; The visible light image is denoised and enhanced, and the appearance morphology features of the insulator string are extracted. The appearance morphology features include the integrity of the sheds, the distribution of surface dirt, and the traces of corona discharge. The infrared thermal image is subjected to temperature calibration and background separation processing to extract the temperature distribution characteristics of each insulator sheet in the infrared thermal image. The appearance and temperature distribution characteristics of each insulator piece are compared with the preset zero-value insulator feature template. When the comparison results simultaneously meet the preset temperature anomaly threshold and appearance anomaly threshold, the insulator piece is identified as a zero-value insulator. Based on the distribution parameters of the insulator string and the real-time three-dimensional position coordinates of the current UAV, the measurement results of the geographical coordinates, relative tower offset, and detection confidence of the zero-value insulator are output.

[0012] Furthermore, the resolution of the voxelized map is dynamically adjusted according to the area where the UAV is located, including: The inspection area of ​​the UAV is divided into a central area and an edge area, wherein the central area is the spatial area within a preset radius centered on the power transmission line conductor, and the edge area is the spatial area outside the central area; Based on the distance between the drone and the power transmission line, determine the area where the drone is located at the current moment; If the drone is located in the central region, the resolution of the central region of the voxelized map is set to a first preset resolution value. If the drone is located in the edge region, the resolution of the edge region of the voxelized map is set to a second preset resolution value; Wherein, the first preset resolution value is greater than the second preset resolution value.

[0013] Furthermore, the marking of dynamic obstacles is implemented using a dynamic obstacle response mechanism, including: Identify connected regions of obstacle point clouds in consecutive frames of point cloud data; The positional changes of connected regions of the same obstacle point cloud in consecutive frames are tracked to estimate the motion velocity vector of the obstacle; Based on the obstacle's velocity vector and a preset prediction duration, calculate the predicted position of the obstacle at a future time. In the voxelized map, the voxels corresponding to the obstacles are marked as dynamic obstacles, and the spatial distribution of the dynamic obstacle voxels is updated according to the predicted location; The updated spatial distribution of the dynamic obstacle objects is fed back to the calculation of the collision penalty factor, serving as the basis for the dual constraint verification in the next rolling cycle.

[0014] Furthermore, a computer-readable storage medium is provided for storing computer-readable instructions, which, when read by a computer, enable the execution of the UAV insulator zero-value measurement and autonomous obstacle avoidance method integrating visual perception and flight parameter control.

[0015] The beneficial effects of this invention are as follows: by constructing a three-dimensional cylindrical electromagnetic hazard area based on the centerline of the conductor, and by combining the voltage level with the electromagnetic hazard area radius mapping table for quantitative characterization, a systematic processing of electromagnetic safety constraints is achieved, providing precise electromagnetic safety boundary constraint conditions for UAVs flying around ultra-high voltage transmission lines.

[0016] By establishing a weighted priority collaborative arbitration mechanism among electrical safety commands, physical obstacle avoidance commands, and mission tracking commands, and dynamically adjusting the priority coefficients of various commands according to the electromagnetic risk level, the organic integration of multi-source commands at the decision-making level is achieved, effectively solving the technical problem of the imperfect multi-source command collaborative decision-making mechanism in the existing technology.

[0017] By employing a rolling time-domain control strategy to perform dual electromagnetic and collision constraint verification on multiple candidate paths, and using the comprehensive cost function value to quantitatively evaluate the candidate paths and select the optimal flight path, combined with the fusion calculation of local obstacle avoidance adjustment commands and global paths, a dual-layer fusion autonomous obstacle avoidance function is realized, improving the reliability of obstacle avoidance paths for UAVs in dynamic environments such as strong winds.

[0018] By integrating visual perception and flight parameter control, the system can perform zero-value measurement of insulators while performing autonomous obstacle avoidance flight. By using dual-modal feature extraction and comparison analysis of visible light images and infrared thermal images, the accuracy of zero-value insulator identification is improved, meeting the refined requirements of UHV inspection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance, which integrates visual perception and flight parameter control, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall process of measuring zero value of UAV insulator and autonomous obstacle avoidance by integrating visual perception and flight parameter control, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the multi-source instruction priority mapping relationship structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the principle of comprehensive cost function calculation and global optimal path selection provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the insulator zero-value measurement and autonomous obstacle avoidance architecture provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Please see Figure 1 A method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control includes: Acquire drone flight parameters, insulator inspection data, and the voltage level of the current working line to construct an initial dataset; Based on the mapping table between line voltage level and preset electromagnetic hazard zone radius, a three-dimensional electromagnetic hazard zone model corresponding to each phase conductor is established. Based on the UAV flight parameters, the preset kinematic model and the pre-acquired environmental perception data, the set of future flight trajectories is predicted, and the set of flight trajectories is discretized and sampled to obtain several candidate path point sequences. The selected path point sequence is divided into several electromagnetic hazardous path segments and several electromagnetic safe path segments according to the three-dimensional electromagnetic hazard area model. An electromagnetic penalty factor is applied to the electromagnetic hazardous path segments, and a collision penalty factor is applied to the path segments in the electromagnetic safe path segments where the distance to the obstacle is less than a preset safety threshold. The electromagnetic penalty factor and the collision penalty factor are weighted and summed according to the instruction priority to obtain the comprehensive cost function value of each candidate path. The candidate path with the lowest cost is selected as the global optimal path based on the comprehensive cost function value. The local obstacle avoidance adjustment command generated by real-time perception is fused with the global optimal path to output the final flight control command. Guided by flight control commands, the UAV flies to the insulator detection position, extracts and compares the temperature distribution and appearance features in the visible light and infrared images of the insulator string, identifies zero-value insulators, and completes the zero-value measurement of the insulator. This invention mainly includes the following steps: data acquisition and spatiotemporal alignment, construction of a three-dimensional electromagnetic hazard zone quantitative characterization model, rolling temporal candidate flight trajectory sampling, electromagnetic hazard and electromagnetic safety path segment division and penalty factor application, comprehensive cost function value calculation, global optimal path selection and multi-source command weighted fusion, zero-value insulator identification, measurement quality index (MQI) construction, and hovering pose iterative optimization. Figure 2 As shown, data interaction between each step is achieved through a data bus and communication protocol. The core architecture is based on real-time closed-loop feedback of three types of instructions: electrical safety, physical obstacle avoidance, and task tracking. The hovering pose is iteratively optimized by the Measurement Quality Index (MQI), forming a complete closed loop of data flow and control flow.

[0022] The data acquisition and spatiotemporal alignment steps obtain raw data from UAV onboard sensors and the ground control station, including UAV real-time position coordinates, flight speed, flight attitude angle, battery level, and other flight parameters; visible light and infrared thermal images of insulators; and voltage level Un, current intensity I of each phase conductor, and environmental parameters of the currently operating line. Specifically, UAV flight parameters are transmitted via the flight control system's CAN bus or MAVLink protocol at a frequency of 50Hz to 100Hz; visible light and infrared thermal images are synchronously acquired by an onboard dual-optical pod, with the acquisition time difference controlled within 100 milliseconds; line voltage level and current intensity are obtained through a real-time communication interface with the power grid dispatching system; and ambient temperature, humidity, and air pressure are acquired in real-time by onboard meteorological sensors, with measurement accuracies of ±0.5℃, ±3% humidity, and ±50Pa, respectively. All acquired multi-source data are spatiotemporally aligned using a unified timestamp, obviously abnormal data points are removed, and missing data is interpolated and imputed. The data is then integrated into an initial dataset according to a predefined data structure, serving as the input benchmark for the entire method.

[0023] In the specific implementation of the three-dimensional electromagnetic hazard zone quantitative characterization model, the first step is to obtain the conductor arrangement and sag parameters of the current operating line. Common conductor arrangements include horizontal, triangular, and vertical arrangements. For horizontally arranged three-phase conductors, the horizontal spacing between phases of a 1000kV UHV AC transmission line is typically 18 to 22 meters. The conductor sag parameters include the position of the lowest point of the conductor sag, the size of the sag, and the height difference of the suspension point, which can be obtained through line design data or UAV LiDAR scanning and reconstruction. Then, based on the conductor sag parameters, the coordinate sequence of the centerline of each phase conductor in three-dimensional space is calculated. Specifically, each phase conductor is discretized along the line direction, and a sampling point is taken every 0.5 to 1 meter. The x, y, and z coordinates of each sampling point in three-dimensional space are then calculated.

[0024] According to the voltage level formula Calculation and voltage level The corresponding baseline electromagnetic hazard zone radius, for example, a 500kV voltage level. 800kV voltage level 1000KV voltage level Then Combined with changes in ambient temperature Humidity change air pressure change and current transient ratio Dynamic correction is performed to obtain the corrected radius of the electromagnetic hazard zone. ,in, For the real-time operating current of each phase conductor, This refers to the rated operating current at the corresponding voltage level. These are preset correction coefficients, with typical values ​​of [value missing]. Transient current ratio The influence of changes in conductor operating current on the electromagnetic field distribution was considered, resulting in a corrected electromagnetic hazard zone radius. It can dynamically adjust according to the line load.

[0025] Using each point on the coordinate sequence of the center lines of each phase conductor as the center of a circle, and the corrected radius of the electromagnetic hazard zone... A circular cross-section is generated for the radius, and a three-dimensional cylinder is generated by scanning along the centerline of the conductor. A union operation is performed on the three-dimensional cylinders corresponding to the A-phase, B-phase, and C-phase conductors to obtain a three-dimensional electromagnetic hazard zone quantification standard model. The three-dimensional electromagnetic hazard zone quantification standard model is stored in voxel form with a spatial resolution set to 0.5 meters, which supports rapid spatial location lookup and serves as the judgment benchmark for subsequent electromagnetic constraint verification.

[0026] A rolling temporal candidate flight trajectory sampling step is used to establish a kinematic model of the UAV, which includes its position state. Speed ​​state and attitude state ,in, For UAV position vector, For velocity vector, For acceleration vector, This is the attitude angle vector (including pitch, roll, and yaw angles). Angular velocity vector To discretize the time step, Indicates the current time, This represents the next discrete time step. The estimated wind speed and direction values ​​from the environmental perception data are used as external disturbance inputs to update the parameters of the kinematic model. The force exerted by the wind on the UAV is modeled as follows: ,in, air density (taken as 1.225 kg / m³) 3 ), For the windward area of ​​the drone, For drag coefficient, This refers to relative wind speed.

[0027] Based on the updated kinematic model, with the current time as the starting point of the rolling window and the look-ahead time domain length... The window length is in seconds, calculated in preset steps. Multi-step forward simulation is performed every second, using a quasi-random sampling method based on Gaussian mixture distribution to generate N=128 candidate flight trajectories. These 128 candidate flight trajectories are divided into altitude-lifting paths, horizontal lateral offset paths, and composite maneuver paths according to a preset ratio, with 32 being altitude-lifting paths, 32 being horizontal lateral offset paths, and 64 being composite maneuver paths. This structured sampling method covers both the two main avoidance directions (vertical and horizontal) and handles oblique obstacle avoidance scenarios through composite maneuver paths. Each of the 128 candidate flight trajectories is processed according to the preset step size. The time is used to discretize the data, resulting in a sequence of candidate path points corresponding to each candidate flight trajectory.

[0028] refer to Figure 3 As shown, the steps of dividing electromagnetic hazard and electromagnetic safety path segments and applying penalty factors perform point-by-point electromagnetic constraint and collision constraint verification on the candidate path point sequence. For each path point in the candidate path point sequence, the query interface of the three-dimensional electromagnetic hazard zone quantification model is called to determine whether the path point is located within the three-dimensional cylindrical electromagnetic hazard zone: if the determination result is yes, the path point is classified into an electromagnetic hazard path segment, and an electromagnetic penalty factor is applied to the candidate path. The value of the electromagnetic penalty factor is a first preset value (e.g., 1000); otherwise, the path point is classified into an electromagnetic safety path segment.

[0029] Acquire several voxels adjacent to the voxel containing the path point in the voxelized map constructed from LiDAR point clouds. If there are voxels marked as dynamic or static obstacles, calculate the Euclidean distance between the path point and the center point of the obstacle voxel. If the Euclidean distance is less than a preset safety threshold (e.g., 1.5 meters), apply a collision penalty factor compensated for wind speed to the corresponding electromagnetically safe path segment in the candidate path. Specifically, a repulsive function is constructed based on an improved artificial potential field algorithm. Where Frep is the repulsive force vector acting on the drone, d is the distance from the path point to the nearest obstacle object, and d0 is a preset repulsive force threshold (e.g., 3 meters). The preset repulsive force gain constant, Wind speed compensation coefficient (with wind speed) (It exhibits a piecewise functional relationship).

[0030] The wind speed compensation coefficient Calculated by segment: When hour =1.0; when hour

[0031] when hour The wind speed The output is obtained from the Kalman fusion of airborne Doppler wind speed sensor and ground meteorological station observations. When the deviation between the Kalman fusion output and the lateral airflow disturbance calculated by the airborne inertial measurement element is greater than 1.5 m / s, the repulsive force gain constant η is additionally multiplied by a safety factor of 1.15. The values ​​of the thresholds (5 m / s, 8 m / s) and slope coefficients (0.06, 0.04) in the piecewise formula are based on the following: Attitude disturbance measurements were conducted on a typical multi-rotor inspection UAV (wheelbase class of 0.45 meters) in the TJ-2 low-speed boundary layer wind tunnel under wind speeds of 0-15 m / s and inflow angles of 0°-30°. The results showed that the average lateral displacement error of the UAV increased linearly from 0.04 meters (5 m / s) to 0.92 meters (15 m / s) with increasing wind speed. 5 m / s was used as the starting point for the linear effect of wind speed, and 8 m / s was used as the starting point for the nonlinear effect (the UAV attitude controller enters the saturated nonlinear range at this wind speed). The slope of 0.06 is the least squares regression value (coefficient of determination) of the repulsive force compensation gain caused by a unit wind speed increment within the 5-8 m / s range. =0.94), and the slope of 0.04 is the log-linear regression value within the interval >8m / s. =0.91); the above coefficients are iteratively updated annually based on the wind tunnel retest data of the matched aircraft model. The repulsion vector generated by the repulsion function is multiplied by the collision penalty factor to obtain the collision penalty factor after wind speed compensation. The base value of the collision penalty factor is a second preset value (e.g., 500).

[0032] refer to Figure 4 As shown, the calculation steps for the comprehensive cost function value are based on the cost function. Calculate the comprehensive cost function value for each candidate path. This represents the comprehensive cost function value of the candidate path; The control input vector corresponding to this candidate path, Let be the task reference control input vector; be the squared instruction deviation term; be the energy cost representing the candidate path relative to the task reference trajectory, reflecting the degree of deviation of the candidate path from the task reference trajectory; and be the safety margin integral term. is the minimum distance from the path point to the obstacle element or the boundary of the electromagnetic hazard zone, and is the cumulative obstacle avoidance urgency along the candidate path. The smaller the value, the larger the value of that item; The path smoothing term is obtained by weighting the L2 norm of the velocity vector difference and acceleration vector difference between adjacent path points. Its physical meaning is to suppress high-frequency jitter of the flight trajectory to reduce energy consumption and controller fatigue. and These are the sum of the electromagnetic penalty factor and the collision penalty factor after wind speed compensation on the candidate path, respectively, and their physical meaning is the cost of violating the hard constraint. The preset cost weights, whose physical meanings are task tracking weight, safety margin weight, and trajectory smoothing weight, respectively, typically take values ​​of [values ​​to be filled in]. ; The values ​​are derived from the following: under the fixed obstacle distribution of a 30km standard UHV corridor, the joint objective function prioritizes safety completion rate, followed by smoothness, and minimizes task deviation. A grid search method (step size 0.1) is used to evaluate the results. , , The calibration was obtained by full-range scanning, and the generalization error of the calibrated values ​​is <5% in 500kV / 1000kV scenarios. , These are the preset weights for the electrical safety commands and physical obstacle avoidance commands to be discussed in the next step. , =2), representing the relative importance of the costs of violating hard constraints of different priorities. The relative magnitude of the cost function weights directly determines the optimal path's trade-offs across four dimensions: task deviation, safety margin, trajectory smoothing, and electromagnetic / collision constraints. Increasing the numerical value will cause path selection to favor higher safety margins. Increasing the numerical value will cause path selection to favor higher tracking accuracy and higher smoothness. The comprehensive cost function value... The 128 candidate paths were evaluated and then sorted. The candidate path with the smallest value is selected as the global optimal path.

[0033] The global optimal path selection and multi-source instruction weighted fusion steps involve the global path optimization layer selecting the candidate path with the lowest cost based on the comprehensive cost function value as the global optimal path. Simultaneously, the local obstacle avoidance layer determines the path based on the real-time perceived distance to the nearest obstacle. With relative velocity A local obstacle avoidance adjustment command is generated, which is obtained by querying a preset local obstacle avoidance decision table. The local obstacle avoidance decision table is a two-dimensional lookup table structure, with the row index being the obstacle distance. The three intervals are [0,2) meters, [2,5) meters, and [5,10) meters, with the column index being relative speed. The three intervals ([0,1)m / s, [1,3)m / s, [3,+∞)m / s) correspond to a set of obstacle avoidance strategy codes and obstacle avoidance action quantization parameters: when ∈[0,2) meters and Execute the emergency hovering command when the speed is ≥3m / s, and the target deceleration =-4m / s², target yaw angle change =±90°, hovering height increase Δh=+1.5 meters; when ∈[0,2) meters and When the speed is ∈ [1,3) m / s, an emergency lateral avoidance command is executed, and the lateral velocity is... =±2m / s, target yaw angle change =±60°; when ∈[0,2) meters and When m / s ∈ [0,1), execute the hover command. = =0, maintain the current posture; when ∈[2,5) meters and When the speed is ≥3m / s, a deceleration and steering command is executed, and the target deceleration is... =-2m / s², angular velocity of steering =15° / s, target turning angle =±45°; when ∈[2,5) meters and When the target speed is ∈ [1,3) m / s, the deceleration and obstacle avoidance command is executed, and the target deceleration is... =-1.5m / s², lateral velocity =±1m / s; when ∈[2,5) meters and A small correction command is executed when ∈[0,1)m / s, lateral velocity =±0.5m / s; when ∈[5,10) meters and A deceleration warning command is issued when the speed is ≥3m / s, and the target deceleration is... =-1m / s², Target turning angle =±15°; when ∈[5,10) meters and When the target speed is ∈ [1,3) m / s, the warning tracking command is executed, and the target deceleration is... =-0.5m / s²; when ∈[5,10) meters and When ∈[0,1)m / s, only a warning indicator is triggered, and the speed is not adjusted; when Local obstacle avoidance adjustment commands are not triggered when the distance is ≥10 meters. The specific construction method of the local obstacle avoidance decision table includes: (i) constructing 50,000 Monte Carlo simulation scenarios on the multi-rotor UAV simulation platform AirSim, including 4 types of typical obstacles (bird flocks, hovering workers, temporary towers, other UAVs), wind speed 0-15m / s, and relative speed 0-10m / s; (ii) using a model predictive controller (MPC) for each simulation scenario to solve for the optimal obstacle avoidance action quantification parameters that minimize the distance to the nearest obstacle and the task deviation, as the parameters for that scenario. × (iii) The initial strategies of each element were modified by conducting 800 actual flight tests in the 500kV / 1000kV UHV corridor of a provincial power company, and strategy combinations that caused significant instability or attitude overshoot (pitch or roll angle instantaneously >15°) were eliminated; (iv) The initial strategies of each element were modified quarterly by using least squares regression to analyze the actual flight accident rate indicators (including the event rate of closest approach distance <1.0 meter P-collision-event, emergency hover trigger rate P-hover-trigger, and attitude recovery time Trecovery). , , The obstacle avoidance action quantification parameters are updated. The regression convergence criterion is that the mean change of parameters after two consecutive updates is <5% and P-collision-event <0.3%. Otherwise, iterative correction continues until both criteria are met simultaneously.

[0034] The globally optimal path and the local obstacle avoidance adjustment commands are weighted and fused according to preset weights for three categories of commands: electrical safety, physical obstacle avoidance, and task tracking. A first preset weight is assigned to the electrical safety commands. =3, assigning a second preset weight to the physical obstacle avoidance command. =2, assigning a third preset weight to the task tracking instruction. =1; Set the electrical safety trigger indication function. With physical obstacle avoidance trigger indicator function The value is 1 when the corresponding trigger condition is met, and 0 otherwise; Press Desired velocity vector for electrical safety Physical obstacle avoidance expected velocity vector With the target velocity vector for task tracking Weighted fusion is performed to obtain the final flight control command. When the electrical safety trigger indication function When =1, the weight of the task tracking instruction It is forcibly attenuated to 0 in order to strictly prioritize electrical safety. The 3:2:1 weight ratio was determined based on the following: performance verification was conducted on 15 typical inspection scenarios (including normal inspection, electromagnetic alarm triggering, and dynamic obstacle approach) using weight ratios of 1:1:1, 2:1:1, 3:2:1, 4:2:1, and 5:2:1. The results showed that the 3:2:1 ratio resulted in the lowest electrical safety constraint violation rate (<0.5%) and the smallest mission tracking deviation (<1.8 meters). Furthermore, Lyapunov stability analysis was introduced to verify that this weight ratio would not cause oscillations or divergence in the flight control system.

[0035] To prevent starvation caused by low-priority instructions not being executed for extended periods, a comprehensive priority score is further defined. ,in ∈{E,O,T} identifies the instruction category, For the current moment, For the first The timestamp of the instruction class arriving in the instruction buffer queue. The cumulative number of rolling cycles experienced by this type of instruction, ξ is a preset waiting time weighting coefficient (typically ξ=0.5); the comprehensive priority score The priority score increases linearly with the waiting time, ensuring that low-priority instructions can still be preferentially triggered even if they have not been executed for a long time. Through a pre-defined score-dangerous candidate path segment mapping function Associated with each electromagnetic hazard candidate path segment. For a discrete piecewise function based on score intervals: when When the value is greater than 10, it is mapped to the emergency avoidance path segment and an aggravated electromagnetic penalty factor of 5000 is forcibly applied. When the value is less than 5, it is mapped to the emergency avoidance path segment and an aggravated electromagnetic penalty factor of 5000 is forcibly applied. When the value is ≤10, it is mapped to the warning and avoidance path segment and an electromagnetic penalty factor of 2000 is applied. When the value is less than 5, it is mapped to a regular tracking path segment and an electromagnetic penalty factor of 1000 is applied, thereby forming the instruction priority mapping relationship and feeding it back to the comprehensive cost function value. of Cumulative items.

[0036] Furthermore, a safety margin gradient factor is introduced. ,in This is the actual distance from the drone's current position to the nearest conductor centerline. When the safety margin gradient factor g < 0.2, the weight of the electrical safety command... The weight of the electrical safety command is increased from 3 to 5 to trigger an emergency avoidance; when the safety margin gradient factor g > 1.0, the weight of the electrical safety command is increased. Smoothly descend to level 3. The final flight control command. The output is sent to the flight control execution terminal to guide the UAV to perform autonomous obstacle avoidance flight.

[0037] refer to Figure 5As shown, the zero-value insulator identification step is initiated when the UAV flies to the preset insulator detection position according to the flight control command. The visible light image undergoes denoising and contrast-limited adaptive histogram equalization (CLAHE) processing, and appearance morphology features are extracted using edge detection and morphological algorithms. These appearance morphology features include skirt integrity, surface contamination distribution, and corona discharge traces. The infrared thermal image undergoes temperature calibration and background separation processing. Temperature calibration uses a blackbody reference source, and background separation uses a spatial domain-based segmentation method. The temperature distribution features corresponding to each insulator piece in the infrared thermal image are extracted, including the highest temperature, lowest temperature, average temperature, and temperature distribution uniformity index. The appearance morphology features and temperature distribution features of each insulator piece are then fed into a lightweight zero-value insulator identification network. This lightweight zero-value insulator identification network is a MobileNetV3-Small distillation version that supports INT8 quantization and can run at a real-time frame rate of 30fps on an onboard embedded GPU, outputting the identification confidence score for each insulator piece. Discharge stability index (SSI) and signal-to-noise ratio in the corona region When the identification confidence level If the discharge stability index (SSI) is greater than a preset confidence threshold (e.g., 0.8) and the SSI is greater than a preset stability threshold (e.g., 0.6), the insulator is identified as a zero-value insulator; otherwise, it is determined to be transient interference and is not confirmed.

[0038] The training method for the lightweight zero-value insulator identification network includes: (i) the training data comes from the visible light-infrared dual-light registration image database accumulated over three years by the online monitoring system of a provincial power company under the State Grid Corporation of China. This database contains 18,000 images of normal insulator strings and 3,200 images of zero-value insulator strings (1,850 standard zero-value images and 1,350 subcritical zero-value images). All images are manually annotated at four levels (sheath integrity, surface contamination distribution, corona discharge traces, and temperature anomalies) and divided into training, validation, and test sets in a 7:2:1 ratio; (ii) the knowledge distillation method is used to transfer the feature representations of the ResNet-152 teacher network (60.2M parameters, 99.4% accuracy on the test set) to the MobileNetV3-Small student network (2.5M parameters). The distillation loss function is... ,in, For cross-entropy loss, (iii) The training uses the Adam optimizer with a learning rate of 1×10⁻⁴, a batch size of 128, and a total of 80 epochs. An early stopping strategy is adopted (the training is terminated if the accuracy on the validation set does not improve after 5 consecutive epochs) to prevent overfitting. (iv) The input format of the network is a two-stream concatenation of the visible light channel (3×224×224 normalized RGB tensor) and the infrared channel (1×224×224 normalized temperature tensor) (total of 4×224×224 tensors). The output format is a probability vector of length 4 (Confzero, SSI, SNRcorona, Pnormal). (v) INT8 quantization uses TensorRT. The 8.4 toolchain performs min-max calibration layer by layer on the trained floating-point model (the calibration dataset consists of 500 validation images). The weights and activation values ​​are quantized to INT8 (range [-128, 127]). The sensitive layers (the last two fully connected classification layers) retain FP16 precision to avoid precision loss. After quantization, the model size is compressed from 9.3MB to 2.4MB, the single-frame inference latency is reduced from 58ms to 33ms, and the accuracy decreases by only 0.4%. The zero-value insulator feature template is composed of a set of multi-dimensional feature vectors learned by the lightweight zero-value insulator identification network during the training phase. Specifically, it includes: skirt integrity features (skirt defect area ratio ≥3%, edge curvature variance >0.15), surface contamination distribution features (contamination area ≥25% of the insulator sheet surface, HSV color space V channel grayscale variance <0.08), corona discharge trace features (high-frequency Gabor texture energy >0.3, corona area brightness gradient peak >15), and temperature distribution uniformity features (maximum temperature difference between adjacent insulator sheets ≥8℃, temperature gradient variance >2.5℃²). The feature template is stored in the last layer classification weight tensor of the network in the form of a floating-point parameter set, serving as an implicit comparison benchmark for whether each insulator sheet is a zero-value insulator.

[0039] The Measurement Quality Index (MQI) is constructed as follows:

[0040] Construct a Measurement Quality Index (MQI). Among them, The identification confidence level in the zero-value insulator identification results; This is the current drone's shooting angle. The preset optimal shooting angle (e.g.) =15°), The slope coefficient of the Sigmoid function (typical values) =0.3); The signal-to-noise ratio of the corona region is calculated by normalizing the mean gray value of the corona pixels and the variance of the gray value of the background pixels in the visible light image. The distance from the drone to the nearest obstacle. This corresponds to a preset safe distance threshold for the voltage level (e.g., at 1000kV). =12.5 meters); The preset weighting coefficients and Typical value The Quality Index (MQI) is used to jointly characterize the confidence level of zero-value insulator identification, the rationality of the shooting angle, the signal-to-noise ratio of corona imaging, and the obstacle avoidance safety margin, serving as a unified objective function for iterative optimization of hovering pose.

[0041] When the measured quality index (MQI) is lower than a preset threshold (For example When x = 0.75, the hovering pose iterative optimization step is initiated: taking the center pixel of the zero-value insulator in the visible light image as the visual target, and combining the pose output by the UAV's binocular visual odometry and RTK positioning, the optimal hovering position $(x) that maximizes the Measurement Quality Index (MQI) is solved. y , z )$ and optimal hovering posture (x y , z ), where x y z These represent the x-axis coordinates, y-axis coordinates, and z-axis altitude coordinates of the optimal hovering position in the navigation coordinate system. θ ψ These are the roll angle, pitch angle, and yaw angle corresponding to the optimal hovering attitude, respectively; the optimal hovering position satisfies... ,in This represents the current z-axis altitude coordinate of the drone. For the preset electrical safety redundancy height (e.g.) =1.5 meters), to ensure that the hovering height is always outside the electromagnetic hazard zone; the optimal hovering attitude satisfies , , and For the preset attitude stability range (e.g.) Using the reprojection error and the negative gradient of the Measurement Quality Index (MQI) as the cost function, a Bundle Adjustment iterative solution is performed in a nonlinear optimizer. The UAV is then driven to re-execute image acquisition and zero-value insulator identification according to the optimized hovering pose until the MQI converges to a preset threshold MQIth or the number of iterations reaches a preset upper limit (e.g., 10 times). The geographical coordinates, relative tower offset, and final identification result of the zero-value insulator are then output. The geographical coordinates are represented using the WGS84 coordinate system, the relative tower offset includes horizontal and vertical offsets, and the final identification result includes the zero-value insulator number and identification confidence level. Signal-to-noise ratio in the corona region .

[0042] The present invention also includes a preferred embodiment for dynamically adjusting the resolution of the voxelized map: the inspection area of ​​the UAV is divided into a central region (e.g., a cylindrical space with a radius of 15 meters centered on the power transmission line conductor) and an edge region outside the central region. When the UAV is located in the central region, the resolution of the central region of the voxelized map is set to a first preset resolution value (e.g., a high resolution of 0.1 m × 0.1 m × 0.1 m); when the UAV is located in the edge region, the resolution of the edge region of the voxelized map is set to a second preset resolution value (e.g., a lower resolution of 0.5 m × 0.5 m × 0.5 m). The first preset resolution value is greater than the second preset resolution value. The dynamic adjustment of the resolution is achieved through resampling of the voxel grid, which ensures the accuracy of obstacle detection while reducing the amount of data processing. The specific method for constructing the voxelized map is as follows: (i) The Velodyne VLP-16 16-line lidar mounted on the UAV collects raw point cloud data at a frequency of 10Hz. Each frame of the point cloud contains approximately 30,000 three-dimensional spatial points (x, y, z) and reflection intensity information, and is aligned with the UAV's body coordinate system through an external parameter calibration matrix; (ii) The point cloud space is divided into voxel units at a basic resolution of 0.5 meters using an octotree (OctoMap) data structure, and each voxel maintains an occupancy probability. , Binary Bayesian Filtering Perform continuous frame fusion update; among which, =0.7 represents the prior probability of a point cloud hitting a voxel. =0.4 is the prior probability of crossing the voxel); (iii) for cumulative Voxels with a value >0.65 are labeled as occupying voxels and participating in obstacle voxel classification. Voxels with a value <0.35 are designated as free-space voxels. The resolution values ​​of 0.1 meters for the central region and 0.5 meters for the edge region are based on the radius of the electromagnetic hazard zone. +1 meter serves as the upper limit for the center area radius. A resolution of 0.1 meters in the center area matches approximately 1 / 4 of the drone's wheelbase (typically 0.45 meters), allowing it to distinguish small obstacles such as twigs, cables, and bird nests at the centimeter level. Corresponding to the zero-tolerance safety requirements in the vicinity; the 0.5-meter resolution in the edge region matches the size of the smallest detectable obstacle (individual bird in a flock) (approximately 0.3 meters), reducing storage overhead by 41.2% and spatial query latency by 36.5% while ensuring recognition accuracy. The dynamic switching of the resolution is performed in real time through the merging / splitting operation of child nodes in an octree, triggered when the real-time distance between the UAV and the nearest guide centerline crosses the center-edge region interface, with an average switching latency of 12ms, meeting the real-time requirements of 10Hz rolling time-domain control.

[0043] Another preferred embodiment of the present invention is a dynamic obstacle response mechanism: In point cloud data of consecutive frames, a clustering algorithm based on Euclidean distance is used to identify connected regions of obstacle point clouds, grouping point clouds with a spatial distance less than a preset threshold (e.g., 0.8 meters) into the same obstacle connected region; the positional changes of the connected regions of the same obstacle point cloud in consecutive frames are tracked, and the motion velocity vector of the obstacle is estimated using a Kalman filter or particle filter algorithm; the predicted position of the obstacle at a future time is calculated based on the motion velocity vector and a preset prediction duration (e.g., 2 seconds); the voxels corresponding to the obstacle in the voxelized map are marked as dynamic obstacles, and the occupancy probability of the corresponding voxels is increased according to the update method of the occupied grid map based on the predicted position. The spatial distribution of the updated dynamic obstacle voxels is fed back to the calculation stage of the collision penalty factor, serving as the basis for the dual constraint verification of the electromagnetic safety path segment in the next rolling cycle.

[0044] As one of the preferred embodiments of the present invention, the electromagnetic hazard zone radius mapping table is divided into multiple table cells according to the line voltage level U, and each table cell contains the reference electromagnetic hazard zone radius. The corresponding rated current And the corrected electromagnetic hazard zone radius under typical operating conditions. The range value, specifically disclosed for the 110kV voltage level, is as follows: =600A、 =1.0 meter ∈ [0.95, 1.20] meters; at a voltage level of 220kV =1200A、 =1.5 meters ∈ [1.45, 1.80] meters; at a voltage level of 330kV =1800A、 =2.0 meters ∈ [1.95, 2.40] meters; at a voltage level of 500kV =2500A、 =2.5 meters ∈ [2.45, 3.10] meters; at 800kV voltage level =3500A、 =6.4 meters ∈ [6.30, 7.85] meters; at a voltage level of 1000kV =4000A、 =10.0 meters ∈[9.85, 12.50] meters. The specific construction method of the electromagnetic hazard zone radius mapping table includes: (i) taking the specified safe distance for live-line work at different voltage levels as the initial benchmark value, and according to the benchmark electromagnetic hazard zone radius formula. = / 1000 to obtain the theoretical reference radius for each voltage level; (ii) using an electric field strength probe in the UHV test hall of the China Electric Power Research Institute, the power frequency electric field strength E at 0.5-meter intervals in the horizontal and vertical directions around the three-phase conductors of each voltage level is measured. The theoretical reference radius is verified by using E=4kV / m as the human tolerance threshold, and a conservative value is taken as the final value. (iii) according to the modified formula Using initial values, 100,000 Monte Carlo simulations were conducted under a full range of operating conditions, including temperatures from -20℃ to +45℃, humidity from 20%RH to 95%RH, air pressure from 97kPa to 104kPa, and current transient ratios from 0.4 to 1.2. The results were statistically analyzed for each voltage level. The 95% confidence interval is used as the mapping table. ∈[lower bound, upper bound] entries; (iv) every three years, the correction factor is adjusted based on newly added live-line work accident data and the measured current statistics of the power grid dispatching system. Least squares regression is performed for updating, with the regression convergence criterion being that the mean change in the correction coefficient between two consecutive updates is less than 3%. When calculated in real-time according to the aforementioned correction formula... When the data exceeds the upper limit of the specified interval, a data anomaly alarm is triggered, and the upper limit of the interval is used as a backup threshold to ensure the robustness of electromagnetic safety constraints. The electromagnetic hazard zone radius mapping table is stored in the UAV's onboard non-volatile memory and is periodically updated synchronously by the ground control station to adapt to the gradual changes in parameters after the UHV transmission line is put into operation.

[0045] In a scenario involving the inspection of densely packed ultra-high voltage (UHV) transmission lines, suppose a drone needs to perform zero-value measurements on the insulator strings of a 1000kV AC transmission line tower. The line conductors are horizontally arranged with a phase-to-phase spacing of 20 meters, and the A-phase current... =2800A, rated current =4000A, current transient ratio =0.7; meteorological conditions were: temperature 35℃, humidity 55%, air pressure 101.0kPa, wind speed 8m / s, easterly wind direction. The real-time position coordinates of the UAV were (x: 1056.3m, y: 234.7m, z: 45.2m), flight speed 12m / s, attitude angles were pitch angle 3°, roll angle -2°, and yaw angle 45°; the above flight parameters were acquired via CAN bus at a transmission frequency of 80Hz, and visible light images and infrared thermal images of the insulator string were acquired simultaneously (acquisition time difference ≤80ms). The voltage level U and the current intensity of each phase were read from the power grid dispatch system. All multi-source data were integrated into an initial dataset after being aligned with a unified timestamp.

[0046] Substituting the formula for the radius of the baseline electromagnetic hazard zone and the modified formula into... The reference radius at a voltage level of 1000kV was calculated. =10 meters and the corrected radius =11.80 meters, and based on this, a three-dimensional electromagnetic hazard zone quantitative characterization model was constructed. Based on this model and the kinematic model, in Within a 3-second look-ahead window, the 128 candidate flight trajectories (including 32 altitude rise paths, 32 horizontal lateral offset paths, and 64 composite maneuver paths) are sampled and generated. Each candidate trajectory is then discretized at Δt=0.1 seconds to obtain the corresponding candidate path point sequence.

[0047] During the drone's flight, the lidar detected a flock of birds as a dynamic obstacle 12 meters ahead, and the drone was 18 meters away from the electromagnetic hazard zone. A candidate path point (x: 1072.5m, y: 241.2m, z: 42.8m) was located within the electromagnetic hazard zone and was classified as an electromagnetic hazard path segment with an electromagnetic penalty factor of 1000. Another path point had an Euclidean distance of 1.2 meters from the center of the obstacle object (less than the 1.5-meter safety threshold), and was compensated for by a wind speed factor. =1.0+0.06×(8-5)=1.18 After calculation, a collision penalty factor of 590 after wind speed compensation is applied to the corresponding electromagnetic safety path segment in the candidate path. Substituting into the cost function formula, the global path optimization layer selects the candidate path with the smallest comprehensive cost function value as the global optimal path; the local obstacle avoidance layer generates a deceleration warning command based on the nearest obstacle distance of 12 meters and the relative speed of 6 m / s; the current safety margin gradient factor g=(18-11.80) / 11.80=0.525>0.2, the weight ωE of the electrical safety command remains at 3, and the weighted fusion formula is applied to... The weighted fusion yields the final flight control command. It is then output to the flight control execution terminal.

[0048] After the drone arrives at the preset insulator detection location, it simultaneously acquires visible light and infrared thermal images. After CLAHE preprocessing and temperature calibration, the features are fed into the lightweight zero-value insulator identification network. A specific insulator segment is output with a recognition confidence score. 0.91, Discharge Stability Index (SSI) = 0.78, Corona Region Signal-to-Noise Ratio =0.85, identified as a zero-value insulator; substituting into the MQI formula, the calculated MQI is 0.681, which is lower than the preset threshold. =0.75. The hover pose iterative optimization is initiated, and BundleAdjustment is performed using the MQI negative gradient joint reprojection error as the cost function to obtain the optimized hover pose (z). =max(45.2,11.80+1.5)=45.2 meters, θ =14.8° =-0.5°), driving the UAV to re-execute image acquisition and recognition. After 3 iterations, the Measurement Quality Index (MQI) converged to 0.83. The final output includes the geographical coordinates of the zero-value insulator (34.2589°N, 108.8212°E), the relative tower offset (horizontal offset 1.8 meters, vertical offset 0.5 meters), and the final recognition result. =0.93、 =0.88).

[0049] During the aforementioned detection process, when the drone flies to a position 12 meters away from the guide wire, the resolution of the central region of the voxelized map is adjusted from 0.5 meters to 0.1 meters; when it flies to a position 18 meters away from the guide wire, it is adjusted back to 0.5 meters. Simultaneously, the obstacle's motion velocity vector (vx: 3 m / s, vy: 1 m / s, vz: 0.5 m / s) is detected in the continuous frame point cloud data. The predicted position of the obstacle is calculated based on a prediction duration of 2 seconds, and the spatial distribution of the dynamic obstacle voxels is updated. This information is fed back to the next rolling cycle as the basis for the dual constraint verification of the electromagnetic safety path segment.

[0050] To verify the technical effectiveness of the method described in this invention compared to existing technologies, a systematic field test and hardware-in-the-loop (HIL) simulation verification were conducted in a 500kV / 1000kV UHV transmission corridor of a provincial power company. The test control group included the pure artificial potential field method, the pure fast search random tree method (RRT-Connect), and a single-frame zero-value insulator identification method without MQI iteration. Regarding autonomous obstacle avoidance performance, the statistical results of 2000 randomized inspection tasks in dense corridor scenarios (horizontal distance between adjacent lines ≤ 25 meters, typical obstacles being flocks of birds and drone interference) showed that the method described in this invention achieved a safety completion rate of 99.2%, and the minimum closest approach distance to the conductor was 11.4 meters (≥ 11.4 meters). The safety completion rate of the pure artificial potential field method control group was 85.3%, and the minimum closest approach distance was only 3.6 meters (multiple entries into electromagnetic hazard areas); the safety completion rate of the RRT-Connect control group was 88.7%, and the minimum closest approach distance was 7.2 meters. The average path planning time of the method described in this invention is 22.3 ms (running on a Jetson Xavier NX embedded platform), a 70.5% decrease compared to RRT-Connect's 75.6 ms, meeting the real-time requirement of 10Hz rolling time domain control.

[0051] Regarding compensation performance under strong wind conditions, in a 12 m / s crosswind test environment, the mean trajectory tracking error along the candidate path for the control group without wind speed compensation was 0.78 meters, and the 95% confidence upper bound was 1.36 meters; the method described in this invention introduces the wind speed compensation coefficient. Subsequently, the mean trajectory tracking error decreased to 0.21 meters, and the 95% confidence upper bound was compressed to 0.42 meters, representing a tracking error reduction of approximately 73.1%. Regarding the accuracy of zero-value insulator identification, the comparative experiment on 1500 normal insulators and 320 zero-value insulators yielded the following results: the accuracy of pure single-frame visible light identification was 76.4%, and the recall rate was 81.5%; the accuracy of single-frame identification using visible light and infrared fusion increased to 88.2%, and the recall rate to 91.7%; after introducing MQI-driven hovering pose iteration optimization in the method described in this invention, the identification accuracy reached 97.8%, the recall rate reached 99.1%, and the false alarm rate was only 0.7%; MQI converged to... The average number of iterations for values ​​above 0.75 is 2.7, and the convergence rate within the upper limit of 10 iterations reaches 99.6%. Regarding engineering deployment efficiency, the described voxelized map resolution dynamic adjustment mechanism reduces storage overhead by 41.2% and spatial query latency by 36.5% compared to a fixed 0.1-meter full-field resolution scheme. The MobileNetV3-Small distilled version, after INT8 quantization, has a model size of only 2.4MB and a single-frame inference latency of 33ms, which can stably meet the real-time recognition requirements of 30fps on an onboard embedded GPU. The above comparative experimental data fully demonstrate the technical advantages of the method described in this invention over existing technologies in multiple dimensions, including autonomous obstacle avoidance safety, real-time planning performance, robustness under strong wind conditions, accuracy of zero-value insulator identification, and engineering deployment efficiency.

[0052] Through the detailed description of the specific application scenarios above, it can be seen that the technical solution provided by the present invention can effectively solve the technical problems faced by UAVs in the inspection of UHV transmission lines, such as electromagnetic safety constraints, multi-source command arbitration, dynamic environmental adaptability, obstacle avoidance reliability in strong wind environments, and fine measurement of zero-value insulators. It realizes hovering posture closed-loop optimization driven by the Measurement Quality Index (MQI) and safe autonomous flight of UAVs in complex electromagnetic environments.

[0053] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

Claims

1. A method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control, characterized in that, include: Acquire drone flight parameters, visible light and infrared thermal images of insulators, and voltage level and current intensity of the current working line to construct an initial dataset; Based on the voltage level and current intensity, a preset electromagnetic hazard zone radius mapping table is queried. A three-dimensional spherical electromagnetic hazard zone is constructed with the reference axis of the center line position of each conductor and the corresponding electromagnetic hazard zone radius as the radius. The electromagnetic hazard zones of each phase conductor are then subjected to a union operation to obtain a three-dimensional electromagnetic hazard zone model. Using the current moment as the starting point of the rolling window, several candidate flight trajectories are generated based on flight parameters and a preset kinematic model, and the candidate flight trajectories are discretized into several candidate path point sequences; The candidate path point sequence is divided into electromagnetic hazard path segments and electromagnetic safe path segments according to the three-dimensional electromagnetic hazard area model. An electromagnetic penalty factor is applied to the electromagnetic hazard path segments, and a collision penalty factor is applied to the electromagnetic safe path segments whose distance from the obstacle is less than a preset safety threshold. The electromagnetic penalty factor and the collision penalty factor are weighted and summed according to the preset weights of three types of instructions: electrical safety, physical obstacle avoidance and task tracking, to obtain the comprehensive cost function value of each candidate path. Based on the comprehensive cost function value, the path with the minimum cost is selected as the global optimal path. The local obstacle avoidance layer generates local obstacle avoidance adjustment commands based on the distance to the nearest obstacle and the relative speed. The two vectors are superimposed and fused to output the final wind control command. Guided by flight control commands, the UAV is directed to a preset insulator detection location. The visible light image and infrared thermal image are used to extract appearance and temperature analysis features, respectively. These features are then compared with a preset zero-value insulator feature template to identify zero-value insulators and complete the zero-value measurement of the insulator.

2. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 1, characterized in that: Based on voltage levels and a pre-defined electromagnetic hazard zone radius mapping table, a three-dimensional electromagnetic hazard zone model is constructed, including: Obtain the conductor arrangement and conductor sag parameters for the current work line; Calculate the coordinate sequence of the centerline of each phase conductor in three-dimensional space based on the conductor sag parameters; Based on the voltage level and current intensity of each phase conductor in the initial dataset, the electromagnetic hazard zone radius mapping table is queried to obtain the electromagnetic hazard zone radius corresponding to each phase conductor. The electromagnetic hazard zone radius is then corrected by combining the current ambient temperature, humidity and air pressure parameters to obtain the corrected electromagnetic hazard zone radius. A circular cross section is generated with each point on the coordinate sequence of the center line of each phase conductor as the center and the radius of the corrected electromagnetic hazard zone as the radius. The three-dimensional sphere is generated by scanning the direction of the center line of the conductor. By performing a union operation on the three-dimensional spheres corresponding to each phase conductor, a three-dimensional electromagnetic hazard area model around the UHV transmission line is obtained.

3. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 2, characterized in that: Based on the flight parameters, the preset kinematic model, and the pre-acquired environmental perception data, several candidate path point sequences are obtained, including: A kinematic model of the UAV is established, which includes the UAV's position state equation, velocity state equation, and attitude state equation. The parameters of the kinematic model are updated based on the UAV flight parameters in the initial dataset and the estimated wind speed and wind direction values ​​in the environmental perception data. Based on the kinematic model with updated parameters, a multi-step forward simulation is performed with the current time as the rolling window point and the prediction time domain length as the window length to generate a set of flight trajectories of the UAV in the prediction time domain. Each candidate flight trajectory in the flight trajectory set is discretized according to a preset sampling interval to obtain the corresponding candidate path point sequence.

4. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 5, characterized in that: Based on the electromagnetic hazard candidate path segment and the three-dimensional electromagnetic hazard region model, the command priority mapping relationship between each electromagnetic hazard candidate path segment and the corresponding three-dimensional electromagnetic hazard region is determined, including: Three independent instruction cache queues are set up to temporarily store electrical safety instructions, physical obstacle avoidance instructions, and task tracking instructions, respectively. The electrical safety command is assigned a first preset weight coefficient, the physical obstacle avoidance command is assigned a second preset weight coefficient, and the task tracking command is assigned a third preset weight coefficient, wherein the first preset weight coefficient is greater than the second preset weight coefficient, and the second preset weight coefficient is greater than the third preset weight coefficient. When an instruction arrives in the instruction buffer queue, the arrival timestamp and preset validity period of the instruction are recorded, and the time difference between the arrival time and the current time is calculated. If the time difference exceeds the effective duration, the corresponding instruction is deleted from the corresponding instruction buffer queue. The comprehensive priority score is calculated based on the weight coefficient and waiting time of each instruction, and the comprehensive priority score is mapped to each electromagnetic hazard candidate path segment and the corresponding three-dimensional electromagnetic hazard area to form an instruction priority mapping relationship.

5. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 4, characterized in that: For each path point in the candidate path point sequence, determine its spatial relationship with the 3D electromagnetic hazard area model, and obtain the comprehensive cost function value for each candidate path, including: For each path point in the candidate path point sequence, obtain the three-dimensional spatial coordinates corresponding to each path point; Call the query interface of the three-dimensional electromagnetic hazard area model, input the three-dimensional spatial coordinates of a path point, and determine whether the path point is located within the three-dimensional electromagnetic hazard area; If the judgment result is yes, then an electromagnetic penalty factor is applied to the candidate path; Obtain several voxels adjacent to the voxel where the path point is located in the voxelized map constructed from the LiDAR point cloud, and read whether there are voxels marked as dynamic or static obstacles among the adjacent voxels. If the dynamic or static obstacle elements exist, calculate the Euclidean distance between the path point and the center point of the obstacle element. If the Euclidean distance is less than the preset safety threshold, a collision penalty factor is applied to the candidate path; The electromagnetic penalty factor and collision penalty factor corresponding to all path points on each candidate path are weighted and summed in combination with the instruction priority mapping relationship to obtain the comprehensive cost function value of each candidate path.

6. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 5, characterized in that, Based on the real-time sensed distance to and relative velocity of the nearest obstacle, local obstacle avoidance adjustment commands are generated. These commands are then fused with the optimal flight path to output the final flight control commands, including: In the local obstacle avoidance layer, the distance to the nearest obstacle and the relative speed data are received in real time; Based on the distance to the nearest obstacle and the relative speed, a preset local obstacle avoidance decision table is queried to obtain a local obstacle avoidance adjustment instruction; In the global path optimization layer, the comprehensive cost function value is received, and the candidate path with the smallest comprehensive cost function value is selected as the optimal flight path at the current moment. The local obstacle avoidance modulation command and the optimal flight path are vector-superimposed to obtain the final flight control command. The flight control commands are finally arbitrated based on the comprehensive priority score in the command priority mapping relationship to obtain the final result, and the final result is output to the flight control execution terminal.

7. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 6, characterized in that, The flight control commands guide the UAV to perform autonomous obstacle avoidance flight and acquire insulator zero-value measurement results, including: When the UAV flies to the preset insulator detection position according to the flight control command, it collects visible light images and infrared thermal images of the insulator string; The visible light image is denoised and enhanced, and the appearance morphology features of the insulator string are extracted. The appearance morphology features include the integrity of the sheds, the distribution of surface dirt, and the traces of corona discharge. The infrared thermal image is subjected to temperature calibration and background separation processing to extract the temperature distribution characteristics of each insulator sheet in the infrared thermal image. The appearance and temperature distribution characteristics of each insulator piece are compared with the preset zero-value insulator feature template. When the comparison results simultaneously meet the preset temperature anomaly threshold and appearance anomaly threshold, the insulator piece is identified as a zero-value insulator. Based on the distribution parameters of the insulator string and the real-time three-dimensional position coordinates of the current UAV, the measurement results of the geographical coordinates, relative tower offset, and detection confidence of the zero-value insulator are output.

8. The method for measuring zero values ​​of UAV insulators and autonomous obstacle avoidance by integrating visual perception and flight parameter control according to claim 7, characterized in that: The resolution of the voxelized map is dynamically adjusted according to the area where the UAV is located, including: The inspection area of ​​the UAV is divided into a central area and an edge area, wherein the central area is the spatial area within a preset radius centered on the power transmission line conductor, and the edge area is the spatial area outside the central area; Based on the distance between the drone and the power transmission line, determine the area where the drone is located at the current moment; If the drone is located in the central region, the resolution of the central region of the voxelized map is set to a first preset resolution value. If the drone is located in the edge region, the resolution of the edge region of the voxelized map is set to a second preset resolution value; Wherein, the first preset resolution value is greater than the second preset resolution value.

9. A method for measuring zero-value of UAV insulators and autonomous obstacle avoidance that integrates visual perception and flight parameter control, as described in claim 8, is characterized in that... The marking of dynamic obstacles is implemented using a dynamic obstacle response mechanism, including: Identify connected regions of obstacle point clouds in consecutive frames of point cloud data; The positional changes of connected regions of the same obstacle point cloud in consecutive frames are tracked to estimate the motion velocity vector of the obstacle; Based on the obstacle's velocity vector and a preset prediction duration, calculate the predicted position of the obstacle at a future time. In the voxelized map, the voxels corresponding to the obstacles are marked as dynamic obstacles, and the spatial distribution of the dynamic obstacle voxels is updated according to the predicted location; The updated spatial distribution of the dynamic obstacle objects is fed back to the calculation of the collision penalty factor, serving as the basis for the dual constraint verification in the next rolling cycle.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable instructions, which, when read by a computer, enable the execution of the UAV insulator zero-value measurement and autonomous obstacle avoidance method as described in any one of claims 1 to 9, which integrates visual perception and flight parameter control.

Images