Unmanned aerial vehicle autonomous inspection processing method based on distribution network overhead line equipment

Abstract

The invention provides an unmanned aerial vehicle autonomous inspection processing method based on distribution network overhead line equipment, and relates to the technical field of intelligent control, and the method comprises the steps: generating a self-adaptive compensation factor based on the deformation characteristics of an envelope surface, scanning the surface of a line through the combination of a laser radar and a dual-polarization radar in combination with a hovering state parameter, and carrying out the processing, corrected three-dimensional topology reconstruction data and icing feature vectors are obtained; inputting the corrected three-dimensional topology reconstruction data and the icing feature vector into a pre-trained ResNet deep learning model, identifying icing thickness distribution, insulator damage or foreign matter suspension abnormality, and outputting a quantitative abnormality type, a space coordinate and a confidence score; and according to the quantitative anomaly type, the space coordinates and the confidence score, generating a self-adaptive operation strategy containing tool ID codes and operation parameters. The stability and reliability of the inspection process are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, in particular to a UAV autonomous inspection processing method based on distribution network overhead line equipment. BACKGROUND

[0002] In the inspection scene of the distribution network overhead line in a mountainous area, there are a large number of 10-kilovolt distribution network overhead lines winding and zigzagging through the forest. The terrain around the line is complex, and the terrain in some areas is undulating, making manual inspection extremely difficult.

[0003] The purpose of this inspection is to check whether there are defects such as icing, insulator damage, and suspended foreign matter on the line, in order to ensure the stable operation of the line in bad weather. During the inspection, according to the established method, the UAV first generates a three-dimensional flight trajectory and determines the inspection safety boundary based on the geographic coordinates and topological data of the distribution network overhead line equipment. Then, the UAV successfully arrives at the target ground wire coordinates, realizes centimeter-level dynamic hovering through multi-sensor fusion positioning, and outputs the hovering state parameters.

[0004] In the data acquisition link, the laser radar and the dual-polarization radar are combined to scan the surface of the line. However, due to the complex terrain in the mountainous area, the envelope surface of some areas is irregularly deformed due to terrain interference when generating an adaptive compensation factor based on the envelope surface deformation characteristics. This makes the generated adaptive compensation factor unable to accurately match the actual situation, and some point cloud data in the valley or near the tall trees area is corrected with deviation when correcting the original point cloud data set in space coordinates. For example, near a valley, the conductor point cloud data that should be on the same plane has a height difference of about 5 centimeters after correction, and this deviation is further magnified when constructing the three-dimensional topological reconstruction data. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a UAV autonomous inspection processing method based on distribution network overhead line equipment, which ensures the stability and reliability of the inspection process.

[0006] To solve the above technical problems, the technical solution of the present application is as follows: A UAV autonomous inspection processing method based on distribution network overhead line equipment, the method comprising: generating a three-dimensional flight trajectory and an inspection safety boundary of the UAV based on the geographic coordinates and topological data of the distribution network overhead line equipment; controlling the UAV to arrive at the target ground wire coordinates based on the three-dimensional flight trajectory and the inspection safety boundary, realizing centimeter-level dynamic hovering through multi-sensor fusion positioning, dynamically calibrating 3 spatial reference coordinates in the hovering state, and outputting hovering state parameters including pose accuracy, electromagnetic interference strength, and reference coordinate set; According to the benchmark coordinate set, a dynamic space feature envelope surface is constructed and region rasterization division is carried out, forming a multi-level scanning unit set; based on the envelope surface deformation characteristics, an adaptive compensation factor is generated, combined with the hovering state parameters, the surface is scanned and processed through the combination of the laser radar and the dual-polarization radar, and the modified three-dimensional topological reconstruction data and the icing feature vector are obtained; The modified three-dimensional topological reconstruction data and the icing feature vector are input into a pre-trained ResNet deep learning model, the icing thickness distribution, the insulator damage or the abnormal suspension foreign matter are identified, and the quantitative abnormal type, the spatial coordinates and the confidence score are output; According to the quantitative abnormal type, the spatial coordinates and the confidence score, an adaptive operation strategy containing a tool ID code and operation parameters is generated.

[0007] The above scheme of the present application at least includes the following beneficial effects: By generating a three-dimensional flight trajectory and a safety boundary based on geographic coordinates and topological data, and combining multi-sensor fusion positioning technology, the unmanned aerial vehicle is realized to dynamically hover under complex electromagnetic environment and air flow disturbance at a centimeter level, three spatial benchmark coordinates of dynamic calibration and hovering state parameters are output, which provides a precise spatial positioning benchmark for subsequent scanning operation, effectively solves the problems of missed detection and false detection caused by positioning deviation in traditional unmanned aerial vehicle inspection, and ensures the stability and reliability of the inspection process. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a flowchart of a kind of unmanned aerial vehicle autonomous inspection processing method based on distribution network overhead line equipment provided by the embodiment of the present application.

[0009] Figure 2 is the flowchart of step 1 in the embodiment of the present application Figure 1 . DETAILED DESCRIPTION

[0010] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings; although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0011] As Figure 1 shown, the embodiment of the present application proposes a kind of unmanned aerial vehicle autonomous inspection processing method based on distribution network overhead line equipment, the method includes the following steps: Step 1, based on the geographic coordinates and topological data of distribution network overhead line equipment, the three-dimensional flight trajectory and the inspection safety boundary of unmanned aerial vehicle are generated; Step 2, based on the three-dimensional flight trajectory and the inspection safety boundary output by step 1, control the unmanned aerial vehicle to reach the target ground wire coordinates, achieve centimeter-level dynamic hovering through multi-sensor fusion positioning, dynamically calibrate three spatial reference coordinates in the hovering state, and output hovering state parameters including pose accuracy, electromagnetic interference strength and reference coordinate set; Step 3, according to the reference coordinate set output by step 2, construct a dynamic spatial feature envelope surface and perform regional rasterization division to form a multi-level scanning unit set; based on the envelope surface deformation characteristics, generate an adaptive compensation factor, combine the hovering state parameters, and process the surface through combined scanning of the laser radar and the dual-polarization radar to obtain corrected three-dimensional topological reconstruction data and ice feature vectors; Step 4, input the corrected three-dimensional topological reconstruction data and ice feature vectors into the pre-trained ResNet deep learning model to identify ice thickness distribution, insulator damage or suspended foreign matter abnormalities, and output quantitative abnormal types, spatial coordinates and confidence scores; Step 5, according to the quantitative abnormal types, spatial coordinates and confidence scores output by step 4, generate an adaptive operation strategy including tool ID coding and operation parameters.

[0012] In the embodiment of the present application, the dynamic spatial feature envelope surface construction and rasterization division realize regional differentiated scanning, combined with the adaptive compensation factor to correct geometric distortion, pose drift and electromagnetic interference, so that the three-dimensional topological reconstruction data and ice feature vectors are more consistent with the actual state of the equipment, providing high-precision data support for abnormal identification; the pre-trained ResNet deep learning model is used to fuse the topological structure and ice features to accurately identify abnormalities such as ice thickness distribution, insulator damage and suspended foreign matter, and output quantitative results and confidence, reducing manual interpretation errors and improving the automation and accuracy of defect identification; based on the abnormal type, coordinate and confidence, dynamically match the operation tool, plan the path and optimize the parameters (such as laser power and mechanical arm torque) to form a set of operation instructions with space-time constraints, so that the inspection forms a closed loop from data acquisition, abnormal identification to processing strategy generation, improving the intelligent level of the whole process of distribution network line inspection.

[0013] As shown in Figure 2 in a preferred embodiment of the present application, step 1, based on the geographic coordinates and topological data of the distribution network overhead line equipment, generate the three-dimensional flight trajectory and the inspection safety boundary of the unmanned aerial vehicle, including: Step 11, obtain the GIS geographic coordinates and equipment topological data of the distribution network overhead line, extract the tower spatial position information, ground wire spatial form parameters and adjacent equipment topological relationship; Step 12, input the ground wire spatial form parameters output by step 11 as input, generate the three-dimensional spatial trajectory of the ground wire through the spatial curve fitting algorithm, and calculate the dynamic vertical isolation threshold of the unmanned aerial vehicle based on the electrical safety standard; Step 13, taking the device topology relationship extracted in step 11 and the ground wire three-dimensional space trajectory generated in step 12 as inputs, identifying the vegetation coverage area, building shelter area and electromagnetic interference source distribution, constructing a rotationally symmetric dynamic envelope surface with the ground wire trajectory as the axis, and generating a patrol safety boundary; Step 14, fusing the ground wire three-dimensional space trajectory in step 12 and the rotationally symmetric dynamic envelope surface in step 13, generating a sequence of waypoints and a set of attitude control instructions for the UAV through a piecewise trajectory optimization algorithm, and outputting a three-dimensional flight trajectory and a patrol safety boundary.

[0014] In the embodiment of the present application, the specific implementation process of step 11 is as follows: The operator logs in to the GIS platform module of the target overhead line through the permission login interface of the network management system, inputs the line number or name (such as "XX line #1-#50 tower section"), and initiates a data retrieval request; the output is in a structured file (such as SHP format, CSV format), which contains the basic geographic information of the line passing area, and the specific fields include the unique identification ID of each tower, the latitude and longitude coordinates (accurate to six decimal places), the tower vertex elevation, the coordinates and height data of the surrounding obstacles (such as trees, buildings). The device topology database is output in the form of a relational data table, which includes three sub-tables: the tower information table records the tower model (such as "10kV straight tower" and "35kV strain tower"), material, construction year, and GIS coordinate ID associated with installation location; the ground wire parameter table records the ground wire model (such as "JL / G1A-120 / 20 steel core aluminum stranded wire"), design sag curve parameters (including sag reference values at different temperatures), strain section division (such as #1-#5 towers as a strain section), and the specific height of the wire suspension point on the tower (such as 12 meters from the ground); the connection relationship table records the connection mode of the tower and the ground wire (such as "#1 tower suspends A-phase wire"), the distance between adjacent towers (such as #1-#2 tower distance 150 meters), the connection node of the branch line and the main line (such as "#10 tower T connects branch line #1"), and the connection position of the grounding device and the tower, etc. Topological logic.

[0015] Select all tower records from the GIS geographic coordinate dataset, sort by tower ID, extract the longitude, latitude, and elevation data from each record, and remove invalid values (such as missing coordinates or data outside the line range); associate the extracted three-dimensional coordinates (longitude, latitude, and height) with the model and ID in the tower information table, generate a structured "tower spatial position list", and ensure that the spatial position of each tower can be accurately located; group by the strain section, record the starting / ending tower ID of each strain section, the corresponding sag curve parameters (such as span length and maximum design sag value) of the conductor and ground wire model, and the conductor suspension point height (distinguish different phase sequence conductors, such as the suspension height difference of A phase, B phase, and C phase); calculate the line orientation angle (such as the orientation of #1 to #2 tower is north by east 30°) in each strain section through the coordinates of adjacent towers, and correct the actual shape parameters of the sag curve in combination with the terrain undulation data (such as the tower height difference).

[0016] Based on the connection relationship table, the topology logic of "tower-conductor and ground wire-surrounding equipment" is combed with the tower as the core node: Determine the connection nodes of the tower and the conductor and ground wire: such as the A-phase conductor suspended on the left side of #3 tower and the B-phase conductor suspended on the right side, the conductor is connected to the specific position of the tower cross arm through the insulator string; convert the longitude and latitude coordinates of two towers into horizontal distance (such as the distance between #5 and #6 towers is 200 meters) as the basis for subsequent sag calculation; label the spatial relationship of associated equipment: such as there is a 110kV live line 50 meters away from #8 tower and a building 10 meters below #12 tower, record the relative distance and orientation of these devices to the target line to provide basis for safety boundary construction; cross-check the extracted tower coordinates, conductor and ground wire parameters, and topology relationship: for example, calculate the conductor-to-ground distance through the tower height and suspension point height to ensure compliance with safety standards; verify the reasonableness of the conductor span through the distance between adjacent towers; integrate the checked information into a structured dataset, and output the list of tower spatial positions, the shape parameter table of conductor and ground wire (including sag, suspension point, and orientation angle), and the equipment topology relationship diagram (including adjacent distance and associated equipment position) as input data for step 12.

[0017] The specific implementation process of the above step 12 is as follows: From the ground wire spatial form parameters output in step 11, divide the data set according to the tension section (such as #1-#5 tower as a tension section), each group contains: the three-dimensional coordinates (longitude, latitude, ground wire suspension point height) of the starting tower and the ending tower, span (horizontal distance between two towers), sag reference value at different temperatures (such as-10℃, 25℃, 40℃ corresponding sag), ground wire direction angle, combined with the real-time temperature of the inspection day (such as obtained through the weather data interface), select the sag parameter at the current temperature from the sag reference value as the key input of curve fitting.

[0018] For each tension section, adopt catenary equation fitting method to construct the three-dimensional form of the ground wire: Take the ground wire suspension point of the starting tower and the ending tower as the curve end point, input the span, sag at the current temperature and direction angle, calculate the natural sag of the ground wire in three-dimensional space (considering the height difference of the tower caused by terrain fluctuation); for the line segment with complex terrain or long span (such as more than 300 meters), add intermediate control points (such as the predicted sag position of the midpoint of the span) between the two end points, use cubic spline curve interpolation to supplement the fitting, and ensure smooth transition of the curve.

[0019] Based on the fitted continuous curve, uniformly sample according to the inspection accuracy requirement (such as one sampling point every 5 meters), record the three-dimensional coordinates (longitude, latitude, height from ground) of each sampling point, eliminate abnormal points beyond the range of the tower suspension point connecting line, and correct the curve deviation caused by terrain mutation (such as the actual height adjustment of the ground wire near the hillside); splice all sampling points in the order of tension section to form a complete three-dimensional reference trajectory of the ground wire from the starting tower to the ending tower, and the trajectory data format is: [sampling point serial number, longitude, latitude, height, belonging tension section ID].

[0020] Match the reference value according to the actual voltage level of the line: for example, the reference value for 10kV line is 1.5 meters, and the reference value for 35kV line is 2 meters; adjust the reference value according to the operation state of the line (such as whether it is live inspection): if it is live inspection, add 0.5 meter safety redundancy to the standard reference value; unmanned aerial vehicle flight attitude fluctuation compensation: according to the performance parameters of the selected unmanned aerial vehicle (such as hovering accuracy ±0.3 meters, maximum up and down jitter amplitude ±0.2 meters), the attitude compensation amount is 0.5 meters, combined with the difference between the temperature of the day and the historical extreme temperature (such as the highest temperature in summer 40℃ and the lowest temperature in winter-10℃), according to the thermal expansion and cold contraction characteristics of the ground wire, calculate the maximum sag variation (such as ±0.4 meters); considering factors such as strong wind and air flow disturbance, additionally increase 0.1 meter compensation amount, the final total dynamic compensation amount is 0.5+0.4+0.1=1.0 meters; superimpose the reference value and the dynamic compensation amount: for example, the reference value of 10kV live line is 1.5 meters+dynamic compensation amount 1.0 meters=2.5 meters.

[0021] Set threshold upper and lower limits: the lower limit is 2.5 meters (the unmanned aerial vehicle cannot be closer than this distance to the ground wire), and the upper limit is set according to the inspection requirements (such as 5 meters, to ensure that the lens can clearly shoot the details of the conductor), and finally output the dynamic vertical isolation threshold range; combined with the peripheral equipment topological relationship (such as the height of adjacent buildings, tree distribution) extracted in step 11, verify the rationality of the threshold: for example, if there is a 3-meter-high tree under the line, it is necessary to ensure that the threshold lower limit (2.5 meters) plus the tree height does not exceed the conductor-to-ground safety distance; adjust the threshold for special sections (such as conductor sections crossing highways and rivers) separately: for example, the threshold lower limit for the conductor section crossing the highway is increased to 3 meters to avoid safety risks caused by flying too low; the three-dimensional reference trajectory of the ground wire (including the coordinate sequence of the sampling points) and the dynamic vertical isolation threshold of the unmanned aerial vehicle (including the upper and lower limits) are used as input data for step 13.

[0022] The specific implementation process of the above step 13 is as follows: Retrieve the line peripheral vegetation layer (including tree, shrub distribution coordinates, tree species and growth height records) from the GIS geographic data in step 11, combined with satellite remote sensing images in the past 3 months (to update recent vegetation growth); determine the vegetation safety height threshold according to the line voltage level (for example, within 5 meters on both sides of the 10kV line, the tree height should not exceed 3 meters; the corresponding range of 35kV line tree height should not exceed 2 meters); through spatial distance calculation (horizontal distance between tree coordinates and three-dimensional trajectory of ground wire), filter out tree clusters with horizontal distance ≤5 meters and height exceeding the safety threshold, marked as "vegetation coverage danger zone", record the region boundary coordinates and the highest tree height; obtain the line peripheral building information (coordinates, height, structure type, such as residential buildings, factory buildings) from the GIS geographic data, and focus on extracting building records with horizontal distance ≤15 meters from the line; according to the "Design Regulations for Overhead Distribution Lines", the minimum horizontal safety distance between 10kV line and building is 1.5 meters, and the minimum horizontal safety distance between 35kV line and building is 3 meters; if the building height exceeds 10 meters (may block the line of sight of the unmanned aerial vehicle), the safety distance needs to be increased by 20%.

[0023] Calculate the horizontal distance between the building and the ground wire trajectory. If it is less than the safety distance corresponding to the voltage level, or the building height causes the UAV flight path to be blocked (e.g., the vertical distance between the building top and the ground wire trajectory ≤ 3 meters), mark it as "building shielding danger zone" and record the building boundary and relative orientation to the line. Extract the associated equipment parameters around the line (e.g., transformer location, capacity, switch station coordinates) from the equipment topology database, and retrieve historical electromagnetic interference records from the power dispatching system (e.g., a certain area once caused the UAV communication interruption due to high-frequency signal tower). According to the type of equipment, set the interference radius (e.g., 10kV transformer interference radius 5 meters, high-frequency signal tower interference radius 30 meters), and combine the areas where the interference intensity exceeds the UAV communication threshold (e.g., signal attenuation ≥ 30%) in the historical records to correct the interference range. Draw a circular area centered on the interference source with the calculated interference radius, mark it as "electromagnetic interference danger zone", and label the interference intensity level (e.g., "strong interference zone" "weak interference zone").

[0024] Take the ground wire three-dimensional space trajectory generated in step 12 as the center, and take a sample point every 10 meters along the trajectory. At each sample point, establish a plane perpendicular to the direction of the ground wire (e.g., if the direction of the ground wire is north by east 30°, then the normal direction of the plane is north by east 30°). In each plane, take the ground wire trajectory sampling point as the center, and the inside boundary (close to the ground wire) is strictly equal to the lower limit of the dynamic vertical isolation threshold calculated in step 12 (e.g., 2.5 meters), to ensure the minimum safety distance between the UAV and the ground wire. Open area (no danger zone): the outside boundary (far from the ground wire) of the plane expands 5 meters from the center, forming a symmetrical cylindrical surface segment (inside 2.5 meters + outside 5 meters, total width 7.5 meters). If the plane overlaps with the vegetation danger zone, the outside boundary expansion distance increases to 8 meters (3 meters more than the open area), ensuring that it avoids the possible swinging range of tree tops (calculated based on the historical maximum wind speed of 1.5 meters of tree swing amplitude, with a margin of error), if the plane is close to the building danger zone, the outside boundary expansion distance is adjusted to 10 meters, and it needs to be higher than the building top by 1.5 meters (to avoid the UAV being blocked by the building when flying), if the plane is in the electromagnetic interference zone, the outside boundary needs to be 2 meters beyond the interference source range (e.g., interference radius 30 meters, outside boundary expanded to 32 meters), to ensure that the UAV flies in a low interference environment.

[0025] According to the order of the sampling points of the ground wire trajectory, connect the inside and outside boundaries of each plane to form a continuously rotating symmetrical envelope surface (similar to a "pipe" structure) along the trajectory. When the outside boundary expansion distance of adjacent sampling points differs by more than 2 meters (e.g., from 5 meters in the open area to 10 meters in the building area), add 5 transition sampling points in the middle to gradually change the expansion distance (e.g., 5 meters → 6 meters → 7 meters → 8 meters → 9 meters → 10 meters), to avoid sharp corners in the envelope surface.

[0026] Check whether all the inner side boundaries of the envelope surface satisfy the dynamic vertical isolation threshold (none is less than the lower limit); verify whether the outer side boundaries completely avoid all dangerous areas (the distance from the vegetation, buildings, and interference sources is all ≥0.5 meters of redundancy), and separately check special sections (such as the line crossing the river, road, etc.): the outer side boundary of the envelope surface crossing the road section needs to be 5 meters higher than the road surface (to avoid the influence of low-flying drones on traffic); convert the three-dimensional space range of the envelope surface into structured data: each sampling point corresponds to a set of boundary coordinates (three-dimensional coordinates of the inner side boundary point and the outer side boundary point), and the type of dangerous area to which it belongs is labeled; generate a visual safety boundary model (such as a three-dimensional grid model), which is superimposed on the GIS map to intuitively show the flyable space range of the drone (the inside of the envelope surface is the safety area, and the outside is the no-fly area); finally, output the inspection safety boundary dataset containing the three-dimensional coordinate sequence and the visual model, which is used as the constraint condition for generating the drone flight trajectory in step 14.

[0027] The specific implementation process of the above step 14 is as follows: Based on the ground wire three-dimensional space trajectory (sampling point sequence) generated in step 12, check whether each sampling point is located inside the rotationally symmetric dynamic envelope surface generated in step 13: through spatial coordinate comparison, calculate the distance between the sampling point and the inner side boundary of the envelope surface (≥0.3 meters of safety redundancy) and the distance between the sampling point and the outer side boundary (≥0.5 meters of safety redundancy); if there is a trajectory point that exceeds the envelope surface range (such as a sampling point that is only 0.1 meters away from the inner side boundary), adjust the point: offset 0.2 meters outward along the direction perpendicular to the ground wire trend, to ensure that the adjusted trajectory is completely located inside the safety boundary, and still conforms to the actual shape of the ground wire after offset (the deviation is not more than 0.5 meters).

[0028] The sampling point coordinates (longitude, latitude, and altitude) of the ground wire trajectory are associated with the envelope parameters of the safety boundary (the inner / outer boundary coordinates corresponding to each sampling point) to generate a "trajectory-boundary association table" in the following format: [sampling point serial number, ground wire trajectory coordinates, inner boundary coordinates, outer boundary coordinates, and the type of hazardous area to which it belongs]; two adjacent towers are used as a basic inspection segment (such as tower #1-#2 and tower #2-#3). If the distance between the two towers exceeds 300 meters (such as a large-span segment in a mountainous area), a virtual segmentation point is added in the middle (such as 150 meters from tower #1) to split the long segment into two sub-segments to avoid the difficulty of optimizing a single segment trajectory; key equipment distribution is marked for each segment: key inspection targets within the segment are extracted from the equipment topology relationship in step 11, such as the three groups of insulators between towers #1 and #2 (located on the right side of tower #1, the mid-span wire clamp, and the left side of tower #2), and two lightning arresters, and their precise spatial coordinates are recorded. The initial path is generated along the centerline of the ground conductor trajectory (the midpoint between the ground conductor trajectory and the inner boundary of the envelope surface) with the towers at both ends of the segment as the starting and ending points, ensuring a stable distance from the ground conductor (such as the lower limit of the dynamic vertical isolation threshold + 0.5 meters of redundancy). The A* algorithm is used to search for the shortest path from the starting point to the end point, with path nodes preferentially selected at locations close to the ground conductor but not touching the inner boundary to avoid detours. For the location of key equipment (such as insulators), the path nodes are adjusted to the optimal observation direction of the equipment (such as passing 3 meters diagonally in front of the insulator), ensuring that the angle between the drone camera axis and the normal of the equipment surface is ≤30° (to avoid reflections or obstructions). If the segment contains vegetation, the path needs to be offset 1 meter away from the vegetation (within the envelope). If it passes through an area of ​​weak electromagnetic interference, the path needs to be close to the outer boundary of the envelope surface (away from the center of the interference source).

[0029] On the optimized segmented trajectory, basic waypoints are set at fixed intervals of 10 meters, and the three-dimensional coordinates of each point are recorded (longitude is accurate to 0.00001°, latitude is accurate to the same, and altitude is accurate to 0.1 meter); near key equipment (such as insulators and wire clamps), the interval is shortened to 3 meters (for example, 1 point is set 3 meters in front of the insulator, 1 point is set in front of the insulator, and 1 point is set in the back of the insulator) to ensure multi-angle shooting; if the equipment is near the tower (for example, within 5 meters of the tower), an additional waypoint is added at the tower (1 meter above the top of the tower). ), used to observe equipment on top of towers; numbered in the format of "segment ID-sequence number" (e.g., the fifth point of the #1-#2 tower segment is marked as "1-2-5"); based on the direction of adjacent waypoints (e.g., the azimuth from point A to point B is 20° north-east), the drone's nose direction is calculated to be consistent with the path direction (yaw angle = azimuth angle) to ensure stable flight direction; if passing through key equipment, the yaw angle is adjusted to the direction of the equipment (e.g., if the equipment is 30° to the right of the path, the yaw angle is increased by 30°).

[0030] Pitch angle: According to the vertical distance between the current waypoint and the guide line (e.g., UAV height 15 meters, guide line height 12 meters), calculate the pitch angle as 15° downward (to ensure the lens is aimed at the guide line); if observing the tower top equipment (height 18 meters), UAV height 17 meters, then adjust the pitch angle to 5° upward.

[0031] Roll angle: Maintain 0° (horizontal state) during normal flight; if the path needs to bypass small obstacles (such as branches), adjust the roll angle according to the turning amplitude (e.g., right turn 10° corresponds to roll angle -5°, left turn corresponds to +5°) to ensure smooth turning; according to the UAV cruising speed (e.g., 5 meters / second), calculate the flight time of adjacent waypoints (e.g., 10 meters interval corresponds to 2 seconds), assign time stamps to each waypoint (e.g., the starting point is 0 seconds, the second point is 2 seconds, and so on), forming a three-dimensional trajectory with time information: [time stamp, longitude, latitude, altitude]; convert the attitude parameters (yaw angle, pitch angle, roll angle) of each waypoint into executable instruction format for the UAV (e.g., yaw angle 30° corresponds to instruction "YAW+30") and mark the instruction effective time (synchronized with the waypoint time stamp); add "take a photo" instruction when passing through key equipment (synchronized with the optimal observation point time stamp), and add "slow down" instruction before turning (speed reduced to 3 meters / second) to ensure flight and observation coordination.

[0032] Final output and verification, output two types of core data: Three-dimensional flight trajectory: a complete set of waypoint coordinates with time sequence, accompanied by a visual trajectory map (superimposed on a GIS map, with safety boundary range marked); attitude adjustment instructions and auxiliary operation instructions sorted by time, format adapted to UAV control systems (e.g., DJI SDK compatible format); simulate UAV flight along the trajectory, check whether all waypoints are within the safety boundary and whether the attitude instructions cause the lens to deviate from the target (deviation ≤5° is qualified), if there are problems return to the optimization stage for re-adjustment.

[0033] The final output is the verified three-dimensional flight trajectory and attitude control instruction set, which serves as the basis for actual UAV inspection, and is accompanied by safety boundary data for real-time verification during flight.

[0034] In the embodiment of the present application, by accurately acquiring and extracting the tower coordinates, ground wire shape parameters and equipment topology relationship, the data source is highly matched with the actual equipment working condition, and the planning error caused by the deviation of the basic data is avoided. The three-dimensional space trajectory conforming to the actual shape of the ground wire is generated, the vertical isolation threshold value is calculated based on the electrical standard and dynamic factors, the safety distance bottom line of the unmanned aerial vehicle and the ground wire is determined from the technical level, the risk of contacting the live equipment is effectively avoided, and the conformability of the trajectory and the ground wire is ensured to improve the inspection pertinence. By identifying dangerous areas such as vegetation, buildings and electromagnetic interference, a rotating symmetrical dynamic envelope surface is constructed to form a safe boundary for inspection, and the precise isolation of the complex environmental risks is realized. The dynamic adjustment characteristics of the boundary ensure that the unmanned aerial vehicle can maintain a safe flight space in different dangerous areas, thereby eliminating the collision, signal interference and other hidden dangers from the spatial range. Through the fusion and segmentation optimization of the trajectory and the boundary, the generated waypoint sequence and attitude instruction set not only strictly follow the safety boundary constraint, but also maximize the observation effect of the key equipment, and reduce the invalid flight mileage.

[0035] In a preferred embodiment of the present application, step 2, based on the three-dimensional flight trajectory and the inspection safety boundary output by step 1, the unmanned aerial vehicle reaches the target ground wire coordinates, and the centimeter-level dynamic hovering is realized through multi-sensor fusion positioning. In the hovering state, three space reference coordinates are dynamically calibrated and the hovering state parameters including the pose accuracy, electromagnetic interference intensity and reference coordinate set are output, including: Step 21, based on the three-dimensional flight trajectory and the inspection safety boundary output by step 1, a multi-degree-of-freedom motion instruction of the unmanned aerial vehicle is generated; Step 22, in response to the unmanned aerial vehicle reaching the target ground wire coordinates, the real-time dynamic carrier phase difference positioning data, the inertial measurement unit data and the visual odometry data are fused to generate a real-time pose feedback stream with centimeter-level accuracy; Step 23, according to the real-time pose feedback stream, three space reference coordinates are dynamically calibrated in the hovering state based on the spatial distribution characteristics of the ground wire, and the environmental electromagnetic field intensity fluctuation is monitored synchronously; Step 24, the real-time pose feedback stream, the space reference coordinate set and the electromagnetic field intensity fluctuation data are aggregated to output the hovering state parameters including the pose accuracy, the electromagnetic interference intensity quantitative value and the reference coordinate set.

[0036] In the embodiment of the present application, the specific implementation process of step 21 is as follows: The target ground wire coordinate corresponding flight trajectory sequence (including longitude, latitude, height and timestamp) is extracted from step 1, and the spatial constraint range (such as the minimum distance of the inner boundary and the maximum distance of the outer boundary) of the region is obtained from the inspection safety boundary data; the waypoint sequence is decomposed into continuous position targets (one position point every 0.1 seconds) and attitude targets (corresponding to the pitch angle, yaw angle and roll angle generated in step 14) according to the timestamp, and the spatial position and body attitude of the unmanned aerial vehicle at each time are determined; according to the difference between the current position and the target position (such as the distance deviation in X, Y and Z axis direction), the horizontal direction (longitude and latitude) and vertical direction (height) speed control instruction (such as "X axis + 0.5 m / s" and "Z axis - 0.2 m / s") is generated, to ensure that the moving speed does not exceed the safety threshold (such as the maximum horizontal speed of 5 m / s and the vertical speed of 2 m / s); according to the angle deviation of the current attitude and the target attitude (such as the yaw angle deviation of 10°), the rotation angular velocity instruction (such as "yaw + 5° / s") is generated, to control the unmanned aerial vehicle head direction and lens angle, before generating the instruction, it is verified whether each target position is located in the inspection safety boundary (such as the distance from the inner boundary is greater than or equal to 0.3 meters), if it exceeds, the instruction is automatically adjusted (such as offsetting 0.5 meters to the safe area), to ensure that the motion process does not touch the no-fly zone.

[0037] The specific implementation process of the above step 22 is as follows: The data is obtained through the GNSS receiver carried by the drone, with an output frequency of 10Hz (once every 0.1 seconds), including latitude and longitude, altitude and positioning precision factor (such as PDOP value), and low-precision data with PDOP>3 is eliminated; the acceleration, angular velocity and attitude angle (pitch, yaw, roll) of the drone are collected, with an output frequency of 100Hz, and high-frequency noise (such as interference caused by fuselage vibration) is removed by low-pass filtering; the ground wire and surrounding environment images are captured by the binocular camera carried by the drone, the feature point displacement of adjacent frames is calculated, and the relative position change is output (such as X-axis movement of 0.1 meters), the output frequency is 30Hz, and abnormal frames with feature point matching success rate <80% are eliminated. The IMU and visual odometer data are interpolated or resampled based on the timestamp of the GNSS data to ensure that the three types of data are aligned at the same time node ( Time deviation ≤ 0.01 seconds); when the GNSS signal is stable (PDOP ≤ 2), GNSS data is used as the main reference, IMU data is used to fill the gaps in the GNSS sampling interval (the position at the intermediate moment is calculated by integration), and visual odometry data is used to correct the accumulated error of the IMU (such as calibration when the drift exceeds 0.05 meters); when the GNSS signal is blocked or interfered with (PDOP > 3), visual odometry data is used as the main reference (based on the relative displacement of the ground wire feature points), and IMU data assists in maintaining short-term stability, while recording the period of signal instability; after fusion, real-time pose data with centimeter-level accuracy is output (position error ≤ 5 cm, attitude angle error ≤ 0.5°), which is sorted by timestamp to form a continuous stream, including the three-dimensional coordinates, attitude angle and data confidence (such as "high confidence" and "medium confidence") at each moment.

[0038] The specific implementation process of the above step 23 is as follows: Based on the real-time pose feedback flow of step 22, when the position fluctuation of the UAV is less than or equal to 3 centimeters and the attitude angle fluctuation is less than or equal to 0.3° within 3 seconds, it is determined that the UAV is in a stable hovering state, and the reference coordinate calibration process is triggered. According to the spatial distribution characteristics of the ground wire (such as the midpoint of the straight section, the connection point of the tension section, and the suspension point of the insulator), three physical points with distinct features (such as the connection node of the ground wire and the insulator, and the position of the wire clamp) are identified by the vision camera. For each feature point, the three-dimensional coordinates of the feature point in the global coordinate system (consistent with the GIS coordinates of step 1) are calculated by combining the current pose of the UAV (position and attitude angle) and the camera intrinsic parameters (focal length and pixel size). The spatial position is calculated by reversing the image pixel coordinates, and the UAV coordinates in the pose feedback flow are converted. Ten sets of coordinate data are collected for each reference point, the two sets with the largest deviation are removed, and the average of the remaining eight sets is taken as the final reference coordinate to ensure the calibration accuracy (error ≤2 centimeters). The electromagnetic sensor carried by the UAV is used to collect the surrounding electromagnetic field strength in real time, with a sampling frequency of 1 Hz. The electric field strength (unit: V / m) and the magnetic field strength (unit: μT) are recorded. The maximum value, the minimum value, and the standard deviation of the electromagnetic field strength within 30 seconds are calculated to determine the fluctuation range (such as "electric field strength fluctuation ±50 V / m"), and compared with the safety threshold (such as the critical value of the anti-interference of the inspection equipment) to mark the interference level (such as "slight interference" and "moderate interference").

[0039] The specific implementation process of the above step 24 is as follows: Integrate the real-time pose feedback stream of step 22 (take the stable data during hovering period), the 3 spatial reference coordinate sets of step 23 (including the confidence score of each coordinate) and the electromagnetic field intensity fluctuation data (including the average value and fluctuation range); check the spatial consistency of the reference coordinate set and the three-dimensional trajectory of the ground wire (such as the distance from the reference point to the trajectory should be ≤10 cm), if the deviation is too large, re-calibrate; verify the sampling integrity of the electromagnetic field data (missing data ≤5%), otherwise supplement; based on the pose feedback stream during the hovering period, calculate the standard deviation of the position coordinates (such as X-axis standard deviation 2.1 cm, Y-axis 1.8 cm, Z-axis 2.3 cm), take the maximum value as the pose accuracy index (such as "pose accuracy: 2.3 cm"); calculate the fluctuation range of the attitude angle (such as pitch angle ±0.2°), as the attitude stability parameter; compare the average value of the electromagnetic field intensity with the preset threshold value (such as 500V / m, the safety threshold value of 10kV line), calculate the relative intensity (such as "measured 320V / m, 64% of the threshold value"); combined with the fluctuation range, output the quantitative value (such as "electromagnetic interference intensity: moderate, fluctuation ±40V / m"); integrate the pose accuracy, electromagnetic interference intensity quantitative value and 3 reference coordinate sets (including coordinate value, confidence) into a structured parameter set, the format example is: pose accuracy: position error ≤2.3 cm, attitude angle fluctuation ≤±0.2°; electromagnetic interference intensity: electric field average 320V / m (threshold 64%), fluctuation ±40V / m; reference coordinate set: P1 (longitude XXX, latitude XXX, height XXX, confidence 98%), P2 (...), P3 (...), finally output the parameter set as the input data of step 3.

[0040] In the embodiment of the application, the multi-degree-of-freedom motion instruction is generated based on the preset trajectory and the safety boundary, ensuring that the unmanned aerial vehicle can accurately and safely reach the target ground wire coordinate, strictly following the path planning constraints and ensuring flight stability through fine control of speed and attitude; the multi-sensor fusion positioning technology combines the advantages of different devices (global positioning of GNSS, dynamic response of IMU, local accuracy of visual odometry), effectively overcoming the limitations of single sensor (such as GNSS signal blockage, IMU drift), generating real-time pose feedback at the centimeter level; the 3 spatial reference coordinates dynamically calibrated provide reliable spatial anchor points for subsequent three-dimensional reconstruction, ensuring data consistency and accuracy; synchronous monitoring of electromagnetic field fluctuations can timely grasp the environmental interference; the aggregated output of hovering state parameters (pose accuracy, electromagnetic interference, reference coordinates) fully reflects the hovering quality and environmental conditions of the unmanned aerial vehicle, providing key correction basis for subsequent scanning and data processing, and providing quantitative standards for abnormal situation judgment (such as insufficient positioning accuracy, strong interference), ensuring the effectiveness of the inspection data.

[0041] In a preferred embodiment of the present application, step 3, according to the reference coordinate set output by step 2, constructs a dynamic spatial feature envelope surface and performs regional rasterization division to form a multi-level scanning unit set, including: Step 31, taking the reference coordinate set output by step 2 as a sequence of spatial control points, generates a dynamic spatial feature envelope surface through a non-uniform rational B-spline surface reconstruction algorithm; Step 32, performing adaptive regional rasterization division on the dynamic spatial feature envelope surface, generating a multi-level scanning unit set based on the curvature change gradient, and extracting envelope surface deformation characteristic parameters; Step 33, calculating spatial structure adaptive compensation factors according to deformation characteristic parameters, combining pose accuracy and electromagnetic interference intensity quantization values in the hovering state parameters output by step 2, including: Step 331, according to the deformation characteristic parameters output by step 32, extracting the principal curvature distribution gradient and normal vector offset of the envelope surface; calculating the spatial topological distortion intensity induced by the principal curvature distribution gradient, and generating a local coordinate system rotation correction vector based on the normal vector offset; fusing the spatial topological distortion intensity and the local coordinate system rotation correction vector, and outputting the spatial geometric distortion compensation amount through vector composition operation; Step 332, extracting the pose accuracy in the hovering state parameters output by step 2, generating a pose drift compensation amount through a positioning error transfer function; Step 333, extracting the electromagnetic interference intensity quantization value in the hovering state parameters output by step 2, generating a signal propagation compensation amount based on an electromagnetic wave attenuation characteristic model; Step 334, fusing the spatial geometric distortion compensation amount, the pose drift compensation amount, and the signal propagation compensation amount, generating a spatial structure adaptive compensation factor through nonlinear weighted aggregation; performing collaborative scanning on the scanning unit set through laser radar and dual-polarization radar; Step 34, fusing the original point cloud data obtained by scanning, environmental temperature and humidity, wind speed, and electromagnetic field strength, and performing data enhancement processing using wavelet threshold denoising and Kalman filtering, outputting three-dimensional topological reconstruction data and icing feature vectors corrected by the compensation factor.

[0042] In the embodiment of the present application, the specific implementation process of step 31 is as follows: From the hovering state parameters output in step 2, locate the "reference coordinate set" field and extract the complete three-dimensional coordinate data of the three spatial control points, including longitude (accurate to 0.00001°), latitude (same accuracy), and altitude (accurate to 0.01 meter). Simultaneously extract the confidence score corresponding to each coordinate (such as "98%" or "85%"); set the confidence threshold to 90% and compare the confidence of each control point one by one: if the confidence of a point is less than 90% (such as 85%), mark it as an abnormal point and remove it from the control point sequence, and record the reason for removal ("insufficient confidence"); count the number of valid control points remaining after removal: if there are still 3 and the spatial distribution is 5 intervals along the ground line, then If the number of control points is less than 3 or the interval is more than 10 meters due to elimination, interpolation supplementation is triggered. For the missing area, two adjacent valid control points (such as P1 and P3) are selected, and the coordinates of the middle missing point (P2) are calculated by linear interpolation based on the coordinate difference and distance between the two points. For example, if the distance between P1 (X1, Y1, Z1) and P3 (X3, Y3, Z3) is 8 meters, and the missing point P2 is located 4 meters in the middle, then the coordinates of P2 are ((X1+X3) / 2, (Y1+Y3) / 2, (Z1+Z3) / 2), ensuring that the control point sequence after supplementation is evenly distributed along the ground line (interval of 5-10 meters).

[0043] Analysis of the spatial form of the ground wire: if it is a straight segment (the height difference between adjacent towers is less than 5 meters, and the change in the direction is less than 5°), set the NURBS surface order to 3 orders; if it is an arc segment (the height difference is greater than or equal to 5 meters or the change in the direction is greater than or equal to 5°), also use 3 orders (taking into account the smoothness and the shape fitting ability), to ensure that the surface has no sharp angle transition; take 3 control points as the vertices to construct a triangular basic mesh; adjust the grid density according to the complexity of the ground wire: add 1 grid node every 10 meters for the straight segment (generated by interpolation), and add 1 node every 5 meters for the arc segment (encrypt the grid to capture the arc details), and the grid node coordinates are determined by the control point coordinates interpolation; input the control point sequence (including supplementary points) and the control grid parameters into the NURBS surface reconstruction algorithm, and the algorithm generates an initial continuous surface according to the spatial position relationship of the control points, and the algorithm automatically fits the actual distribution characteristics of the ground wire: for the natural sagging area of the ground wire, the surface bends along the sagging trend (such as the midpoint height being lower than both ends); for the area where the insulator is located, the surface fits the concave-convex shape of the shed skirt (locally increases the curvature), and assigns a weight to each control point: the weight of the point with a confidence of 90%-100% is set to 1.0, and the weight of the point with a confidence of 80%-90% is set to 0.8 (points with a confidence of less than 80% have been removed), and the influence of the high-weight control point on the surface shape is greater during the fitting process, reducing the deviation of the control point from the surface; for each control point, calculate the perpendicular distance from its three-dimensional coordinates to the fitted surface (i.e. the shortest distance from the point to the surface), record the deviation values of all points (such as P1 deviation 2 cm, P2 deviation 4 cm), if the maximum deviation is less than or equal to 3 cm, the envelope surface is qualified; if the maximum deviation is greater than 5 cm (such as a point deviation of 6 cm), an intermediate control point is added (1 interpolation control point is added near the point with the maximum deviation), the control grid is regenerated and the surface is fitted; repeat the calculation of the deviation until the deviation of all control points is less than or equal to 3 cm, to ensure that the envelope surface accurately reflects the spatial characteristics of the ground wire and the surrounding equipment.

[0044] The specific implementation process of the above step 32 is as follows: From the dynamic spatial feature envelope generated in step 31, sample and calculate the curvature value (a quantitative indicator of the degree of surface bending) every 2 meters, and record the absolute value of the curvature of each sampling point (such as the curvature of a point on the straight segment of the ground wire is 0.05 / m, and the curvature of a point at the connection of the insulator is 0.2 / m); when the absolute value of the curvature of the sampling point is less than or equal to 0.1 / m, it is determined to be a "flat area" (such as the straight segment of the ground wire and the smooth surface of the tower), and the corresponding grid size is set to 50x50 cm (to reduce the data volume and improve the scanning efficiency); when the absolute value of the curvature of the sampling point is greater than 0.1 / m, it is determined to be a "complex area" (such as the edge of the insulator shed, and the connection between the ground wire and the clamp), and the corresponding grid size is set to 20x20 cm (to increase the sampling and capture the detailed features).

[0045] Along the length direction (the direction of the ground wire) and the width direction (perpendicular to the direction) of the dynamic space feature envelope surface, cutting is performed according to the set grid size: a grid is divided every 50 cm in the flat area, with a coverage of 50 cm (length) x 50 cm (width); a grid is divided every 20 cm in the complex area, with a coverage of 20 cm (length) x 20 cm (width); a unique ID is assigned to each cut grid unit, and the following information is recorded: the three-dimensional coordinates of the unit center (calculated by averaging the coordinates of the four corners of the grid), the area type label (“flat area” or “complex area”, determined based on the curvature threshold of the area to which the unit belongs), and the curvature value corresponding to the center of the unit (extracted from the envelope surface curvature data). All basic scanning units are traversed and grouped according to the area type: flat area units are grouped into one group, and complex area units are grouped into another group; the adjacent units (spatially continuous and without gaps) in each group are aggregated: in the flat area, every 3 continuous 50x50 cm coarse grids are aggregated into a secondary unit, with a coverage of 150 cm (length) x 50 cm (width); in the complex area, every 2 continuous 20x20 cm fine grids are aggregated into a secondary unit, with a coverage of 40 cm (length) x 20 cm (width).

[0046] The basic scanning units and the aggregated secondary scanning units are integrated to form a multi-level scanning unit set of “basic units + secondary units”, and the ID and center coordinates of the basic units included in each secondary unit are labeled; for each grid unit, the maximum curvature and minimum curvature of the unit center are calculated through the envelope surface equation (reflecting the bending degree of the surface in different directions), such as a maximum curvature of 0.3 / m and a minimum curvature of 0.15 / m in the complex area; the normal vector direction (three-dimensional vector) perpendicular to the envelope surface where the unit is located is calculated, which is represented by vector coordinates (such as X-axis component 0.2, Y-axis component 0.1, and Z-axis component 0.9), reflecting the inclination angle of the unit surface; the height difference of the four edges of the unit is measured, and the ratio of the slope difference value to the distance of adjacent edges is calculated to obtain the slope change rate (such as the slope increasing from 0.1 to 0.3, with a change rate of 0.2 / 0.5 m = 0.4 / m), reflecting the change speed of the steepness of the surface; the principal curvature value, normal vector direction, and slope change rate of each unit are associated with the spatial position coordinates (unit center coordinates), and a deformation feature parameter set is formed, ensuring that each parameter can be traced back to a specific grid unit.

[0047] The specific implementation process of the above step 331 is as follows: From the morphological feature parameter set output from step 32, the "principal curvature value" (maximum curvature, minimum curvature) and "normal vector direction" parameters of each scanning unit are screened; for each unit, identify its adjacent 4 units (front, back, left and right), extract the principal curvature value (take the maximum curvature) of the adjacent units; calculate the principal curvature difference value of the current unit and each adjacent unit (such as the current unit curvature 0.2 / m, the adjacent unit curvature 0.1 / m, the difference value is 0.1 / m); divide the difference value by the distance between adjacent units (such as 50 centimeters in a flat area, then the distance is 0.5 meters), to get the curvature change rate (i.e. gradient), take the maximum value of all adjacent gradients as the principal curvature distribution gradient of the unit (such as 0.1 / m ÷ 0.5m = 0.2 / m 2 ).

[0048] Normal vector offset calculation: The angle between the unit normal vector and the global coordinate system Z axis (vertical direction) is extracted, which is the normal vector offset (such as the normal vector and the Z axis angle 15°, then the offset is 15°); set the gradient threshold to 0.05 / m², classify the unit principal curvature distribution gradient: gradient > 0.05 / m² (curvature sudden change area, such as insulator edge): give distortion weight 1.0; gradient ≤ 0.05 / m² (flat area, such as straight line segment of conductor): give distortion weight 0.3. For each unit, multiply its principal curvature distribution gradient by the corresponding weight (such as gradient 0.2 / m² × weight 1.0 = 0.2); calculate the weighted gradient average value of all units as the overall space topology distortion strength (quantify the irregularity of the surface, the higher the value, the more serious the distortion); based on the normal vector offset (the angle between the unit normal vector and the Z axis), calculate the rotation angle: if the offset is 15°, set the rotation angle to 15° (align the Z axis of the local coordinate system with the direction of the unit normal vector).

[0049] According to the direction of the normal vector in the global coordinate system (such as deviating to the X axis or Y axis), the rotation angle is decomposed into rotation components around the X axis and Y axis: if the normal vector deviates to the positive direction of the X axis, mainly allocate the rotation component around the Y axis; deviate to the Y axis, allocate the component around the X axis; convert to a three-dimensional rotation vector (such as rotating 15° around the Y axis, the vector is represented as "X:0, Y:15°, Z:0"), which is used for subsequent coordinate system correction; multiply the topology distortion strength of the unit (such as 0.2) by the basic compensation coefficient (such as 5 centimeters, set according to historical data), to get the compensation amplitude (0.2 × 5 centimeters = 1 centimeter), the stronger the distortion, the larger the amplitude; adjust the compensation direction according to the direction of the rotation correction vector: if the rotation vector is 15° around the Y axis, the compensation direction is along the axis direction of the local coordinate system after rotation (such as the direction inclined at 15°); fuse the amplitude and direction to generate a three-dimensional geometric distortion compensation amount of each unit (such as "X:+0.3 centimeters, Y:+0.8 centimeters, Z:-0.2 centimeters"), which is used to correct the geometric deviation of the scanning data.

[0050] The specific implementation process of each sub-step of step 332 is as follows: From the hovering state parameters output in step 2, extract the specific data under the "pose accuracy" field: position error: X-axis standard deviation 2.1 cm, Y-axis 1.8 cm, Z-axis 2.3 cm (reflecting the positioning fluctuation range of the UAV in three-dimensional directions); attitude angle fluctuation: pitch angle ±0.2°, yaw angle ±0.3°, roll angle ±0.1° (reflecting the fluctuation range of the UAV body inclination).

[0051] The error transfer function is established as follows: The farther the relative distance between the UAV and the scanning unit, the smaller the influence of the position error on the scanning point (the coefficient decreases with the increase of the distance), such as the coefficient is 0.2 when the distance is 5 meters (empirical value based on equipment accuracy test); multiply the position error by the attenuation coefficient, such as Z-axis position error 2.3 cm x 0.2 = 0.46 cm (rounded to 0.5 cm), which means the Z-axis scanning deviation is 0.5 cm at this distance.

[0052] Attitude angle fluctuation transfer calculation: Based on the trigonometric relationship (such as "deviation = distance x sin(angle)"), the angle fluctuation is converted to linear deviation: such as the relative distance between the UAV and the unit is 5 meters, the pitch angle fluctuation is ±0.2°, then the vertical direction deviation = 5 meters x sin(0.2°) ≈ 5 meters x 0.0035 ≈ 0.0175 meters = 1.75 cm, take 0.1 cm (actually scaled down in proportion, because the angle fluctuation range is small).

[0053] The pose drift compensation amount is calculated as follows: Set the weight according to the relative distance between the scanning unit and the UAV: the weight of the near unit is 1.0 when the distance is ≤3 meters, the weight of the medium-distance unit is 0.7 when the distance is 3-5 meters, and the weight of the far unit is 0.4 when the distance is >5 meters (the near unit is more affected by the pose error); add the position error transfer deviation and the attitude angle fluctuation transfer deviation to get the total original deviation of the unit (such as position deviation 0.5 cm + attitude deviation 0.1 cm = 0.6 cm); multiply the total original deviation by the unit weight to get the pose drift compensation amount of each scanning unit (such as near unit 0.6 cm x 1.0 = 0.6 cm, far unit 0.6 cm x 0.4 = 0.24 cm), which is used to correct the scanning deviation caused by inaccurate positioning of the UAV.

[0054] The specific implementation process of step 333 is as follows: From the hovering state parameters output in step 2, extract the core data under the "electromagnetic interference intensity" field: Electric field intensity average: 320 V / m (reflects the average level of interference); fluctuation range: ± 40 V / m (reflects the stability of interference); interference level: “moderate interference” (judged based on preset standards, such as 100-500 V / m as moderate).

[0055] According to the interference level “moderate interference”, match the preset “distance-intensity” attenuation model (this model is suitable for scenarios with moderate and stable interference intensity); calculate through the reference coordinates and scanning unit coordinates in step 2, such as 3 meters (near-distance unit), 8 meters (far-distance unit); electric field intensity value: input the electric field intensity average 320 V / m as the basic parameter for model calculation of attenuation. Based on the “distance-intensity” model, calculate the attenuation deviation according to the following logic: the attenuation amount is proportional to the electric field intensity: the higher the intensity, the greater the attenuation (320 V / m corresponds to a basic attenuation ratio of 10%); the attenuation amount is inversely proportional to the square of the distance: the farther the distance, the smaller the attenuation (3 meters of distance corresponds to a distance coefficient of 1.0, 8 meters of distance corresponds to a coefficient of 0.2); comprehensive calculation: signal attenuation deviation at 3 meters under moderate interference = basic attenuation ratio 10% x distance coefficient 1.0 = 10%; at 8 meters = 10% x 0.2 = 2% (in actual scenarios, fine-tuning needs to be combined with device characteristics, such as laser radar optical signal attenuation being slightly lower, and dual-polarization radar electromagnetic wave attenuation being slightly higher). Convert the attenuation deviation into a correction ratio of signal amplitude: if the attenuation deviation is 10%, then the compensation amount is “+10%” (i.e. increase the measured signal amplitude by 10% to offset the attenuation effect). Unit compensation amount association: adjust according to the distance of the scanning unit and the fluctuation range of the interference: when the fluctuation range is ± 40 V / m, increase the compensation amount of the near-distance unit (3 meters) by 2% (total 12%), and keep the compensation amount of the far-distance unit (8 meters) at 2%, to finally generate the signal propagation compensation amount of each scanning unit (such as “3-meter unit: +12%, 8-meter unit: +2%”).

[0056] The specific implementation process of each sub-step of the above step 334 is as follows: Basic weight setting: set fixed weights according to the degree of deviation influence: spatial geometric distortion compensation amount: weight 40% (dominant device form deviation, most influential); pose drift compensation amount: weight 30% (dominant positioning deviation); signal propagation compensation amount: weight 30% (dominant signal amplitude deviation). Nonlinear weighting adjustment: dynamically increase the weight of the compensation amount if the deviation value exceeds the threshold: if the geometric distortion compensation amount of a unit > 5 cm (exceeds the normal range), its weight increases from 40% to 50%; if the signal propagation compensation amount > 15% (strong attenuation), its weight increases from 30% to 40%. Compensation factor calculation: sum the three compensation amounts according to the maximum weight to generate a spatial structure adaptive compensation factor for each scanning unit (such as “geometric compensation 3 cm x 50% + pose compensation 2 cm x 30% + signal compensation 1.5 cm x 20% = 2.4 cm”).

[0057] The cooperative scanning control (path planning and parameter setting) is as follows: The path is planned based on a multi-level scanning unit set (basic unit + secondary unit): the complex area fine grid unit (20 x 20 cm) is scanned first, and the scanning is pushed forward from the tower to the range midpoint; the gentle area coarse grid unit (50 x 50 cm) is continuously scanned in a straight line segment to reduce the turning back. The laser radar uses high-frequency scanning (100 points per second, 1 point collected every 0.01 seconds to ensure detail coverage), and the dual-polarization radar opens continuous wave detection (continuous signal emission to improve signal-to-noise ratio); the laser radar uses low-frequency scanning (50 points per second, 1 point collected every 0.02 seconds to reduce redundant data), and the dual-polarization radar uses pulse detection (signal emitted once every 0.05 seconds to save power consumption); the synchronization signal is set through the unmanned aerial vehicle control system: the scanning start time deviation of the laser radar and the dual-polarization radar is less than or equal to 0.01 seconds, ensuring that the optical data and electromagnetic wave data of the same scanning unit are aligned in time, and subsequent direct correlation analysis is possible.

[0058] The specific implementation process of the above step 34 is as follows: The original point cloud data of each scanning unit is collected, including three-dimensional coordinates (X / Y / Z accurate to 0.01 meters) and reflection intensity values (0-255 quantization values reflecting the surface reflection ability of the object); the polarization degree (0-1 reflecting the crystal structure of the ice cover) and reflectivity (0-100% reflecting the electromagnetic wave reflection intensity) of each unit are recorded; the temperature and humidity (25°C, 60% RH), wind speed (3 m / s), and electromagnetic field intensity (320 V / m ± 40 V / m) are collected in real time with a sampling frequency consistent with the radar scanning frequency (10 Hz); the laser radar data, dual-polarization radar data, and environmental data are aligned by timestamp (time deviation less than or equal to 0.01 seconds) with the scanning unit ID as the core keyword; a "scanning unit ID-multi-source data" correlation table is established to ensure that the three-dimensional coordinates, reflection characteristics, and environmental parameters of each unit correspond one by one (for example, "unit ID: 101" is associated with its point cloud coordinates, polarization degree, wind speed, etc.).

[0059] Multi-scale wavelet decomposition is performed on the original laser radar point cloud data (e.g., decomposed into 3 layers), dividing the data into low-frequency approximation components (reflecting the main structure) and high-frequency detail components (containing noise); a high-frequency component threshold is set (based on the noise fluctuation range of historical normal data, such as 2 times the standard deviation of the reflection intensity); signals in the high-frequency component that exceed the threshold (such as isolated points where the reflection intensity suddenly jumps from 50 to 200) are marked as noise points and directly excluded; the low-frequency component and the high-frequency component that do not exceed the threshold are retained, and the denoised point cloud data is reconstructed.

[0060] Based on the reflectivity range (e.g., 20%-60%) and polarization degree range (e.g., 0.3-0.7) of normal equipment (such as clean conductors and insulators), a filtering threshold is set: Signals with reflectivity < 10% or > 80% (abnormally low values or clutter caused by electromagnetic interference); signals with polarization < 0.1 or > 0.9 (abnormally high values caused by non-icing or strong interference). Threshold filtering is performed: traverse the dual-polarization radar data, mark and remove signals that meet the above criteria as clutter, and retain valid signals within the normal threshold range. A coordinate smoothing model is established in combination with environmental parameters, the core content is: The greater the wind speed (e.g. 3 m / s), the more intense the conductor oscillation, and the higher the smoothing coefficient (e.g. set the smoothing coefficient to 0.6, and the static state to 0.3); temperature and humidity affect the conductor sag, and the sag is stable at 25°C, so the normal sag trend is retained during smoothing to avoid excessive correction. Dynamic oscillation error correction: For continuous scanning point cloud coordinates (e.g. 10 points per second), a sliding window average method (window size 5 points) is used to calculate according to the smoothing coefficient: current point coordinate = (previous 2 point coordinates x 0.2 + current point coordinate x 0.4 + next 2 point coordinates x 0.2) x smoothing coefficient; After correction, the coordinate fluctuation under 3 m / s wind speed is reduced from ±5 cm to ±2 cm, reducing the positional deviation caused by oscillation. Point cloud coordinate geometric deviation correction: extract the spatial structure adaptive compensation factor of each scanning unit generated in step 334 (e.g. "X: +0.3 cm, Y: +0.8 cm, Z: -0.2 cm"); superimpose the corrected point cloud coordinates: corrected coordinates = original coordinates + compensation factor corresponding component (e.g. original X = 100.5 cm, corrected = 100.5 + 0.3 = 100.8 cm). Radar reflectivity amplitude deviation adjustment: based on signal propagation compensation (e.g. +12%), adjust the dual-polarization radar reflectivity proportionally: corrected reflectivity = original reflectivity x (1 + compensation proportion) (e.g. original reflectivity 50%, corrected = 50% x 1.12 = 56%). The corrected point cloud coordinates are spliced according to the scanning unit to generate the complete three-dimensional structure data of the device, including: the accurate sag shape of the conductor (one coordinate point per meter), the three-dimensional profile of the insulator shed (edge coordinate error ≤2 cm), and the spatial position relationship of the tower connection part.

[0061] Extract key parameters from the corrected dual-polarization radar data to form an icing feature vector: Icing thickness: calculated based on reflectivity and laser radar point cloud thickness (the higher the reflectivity and the greater the point cloud thickness, the thicker the icing); icing density: based on polarization (polarization 0.5-0.7 corresponds to medium density, <0.5 for low density); polarization characteristics: record the trend of polarization with scanning angle (reflecting the uniformity of icing). Finally, output the above data in a structured data set, and label the confidence of each parameter (based on noise reduction and correction effect evaluation, such as "high confidence" and "moderate confidence").

[0062] The application realizes the comprehensive integration of the three-dimensional form, electromagnetic reflection characteristics of the device and environmental influencing factors by associating lidar, dual-polarization radar and environmental sensor data; the wavelet threshold denoising effectively eliminates isolated noise of the point cloud, the threshold filtering reduces the radar clutter interference, significantly improves the purity of the original data, and avoids feature misjudgment caused by noise (such as mistaking clutter as icing); the coordinates are smoothed in combination with environmental parameters such as wind speed, the dynamic interference such as conductor shaking is reduced, the position fluctuation is controlled within centimeter level, and the stability of the three-dimensional structure reconstruction is ensured; the adaptive compensation factor is applied to correct the geometric deviation and amplitude deviation, the influences of surface distortion, pose drift and electromagnetic attenuation are eliminated, and the three-dimensional topology data of the device is highly consistent with the actual form (error≤2cm).

[0063] In a preferred embodiment of the application, the adaptive compensation factor is generated based on the deformation characteristics of the envelope surface, combined with the hovering state parameters, the surface of the combined scanning line of the lidar and the dual-polarization radar is processed, and the corrected three-dimensional topology reconstruction data and the icing feature vector are obtained, including: Step 335, in response to the cooperative scanning of the lidar and the dual-polarization radar on the scanning unit set, the original point cloud data set and the dual-polarization echo signal stream are obtained, and the environmental temperature and humidity, wind speed and electromagnetic field intensity time series data are synchronously collected; Step 336, based on the original point cloud data set obtained in step 335, the spatial coordinate transformation correction is performed through the spatial structure adaptive compensation factor generated in step 334, and a geometric consistency point cloud set is generated; Step 337, the wavelet threshold denoising processing is performed on the geometric consistency point cloud set output in step 336 to eliminate high-frequency vibration noise; at the same time, the dual-polarization echo signal stream and the environmental time series data obtained in step 335 are fused based on Kalman filtering to generate an enhanced radar reflection feature matrix; Step 338, the dielectric constant distribution spectrum and the surface scattering characteristic vector are extracted from the enhanced radar reflection feature matrix output in step 337, combined with the electromagnetic interference intensity quantitative value in the hovering state parameters output in step 2, and the icing feature vector is constructed; Step 339, the three-dimensional surface topology reconstruction is performed on the denoised geometric consistency point cloud set output in step 337, and the three-dimensional topology reconstruction data corrected by the compensation factor is output.

[0064] In the embodiment of the application, the specific implementation process of the above-mentioned step 335 is as follows: When the laser radar and the dual-polarization radar begin to collaboratively scan the set of scanning units, the system synchronously starts data collection: collects the raw point cloud data output by the laser radar (containing the three-dimensional coordinates and reflection intensity information of each scanning point); collects the echo signal stream of the dual-polarization radar (records the raw signal data of electromagnetic wave reflection); simultaneously triggers the environmental sensors to continuously collect the time sequence data of temperature and humidity (real-time records of temperature and relative humidity values), wind speed (real-time wind speed), and electromagnetic field intensity (records the change values of electromagnetic field intensity in time sequence), to ensure that all data are aligned and associated in the time dimension.

[0065] The specific implementation process of the above step 336 is as follows: Based on the raw point cloud data set obtained in step 335, the three-dimensional coordinates of each point in the raw point cloud are adjusted by calling the spatial structure adaptive compensation factor generated in step 334: according to the corresponding coordinate correction rules in the compensation factor, the X, Y, and Z axis coordinates of the raw point cloud are respectively transformed and corrected to eliminate the coordinate deviation caused by geometric distortion, pose drift, etc.; after correction, all point cloud data are consistent in spatial geometric relationship, forming a geometric consistency point cloud set.

[0066] The specific implementation process of the above step 337 is as follows: For the geometric consistency point cloud set output in step 336, a wavelet threshold denoising method is used: by setting a reasonable threshold, noise points (such as isolated points with abnormal reflection intensity) caused by high-frequency vibration (such as device jitter and environmental interference) in the point cloud are filtered and removed, and effective point cloud data is retained; at the same time, the dual-polarization echo signal stream obtained in step 335 and the environmental time sequence data (temperature and humidity, wind speed, electromagnetic field intensity, etc.) are subjected to data fusion by applying Kalman filtering algorithm: by dynamically adjusting the signal weight through the filtering model, the influence of electromagnetic interference and other noises on the echo signal is suppressed, and the effective signal features are enhanced, to finally generate an enhanced radar reflection feature matrix containing clear reflection characteristics.

[0067] The specific implementation process of the above step 338 is as follows: The enhanced radar reflection feature matrix output in step 337 is subjected to data analysis, the matrix area is divided according to the scanning unit, and the feature parameters related to icing are extracted from each unit: dielectric constant spatial distribution spectrum: by using the electromagnetic wave reflection amplitude and phase change data in the matrix, the dielectric constant at different scanning positions (reflecting the insulation characteristics of the material, the dielectric constant of icing is significantly different from that of air and conductor) is calculated, and the distribution spectrum is formed according to the spatial coordinates (such as dielectric constant 3.5 in a certain area on the surface of the conductor, and dielectric constant 6.0 in the icing area); surface scattering characteristic vector: the reflection direction and energy distribution law of the radar wave on the scanning surface are analyzed, and parameters such as scattering intensity and scattering angle distribution are extracted to form vector data (such as 20% higher scattering intensity on the icing surface than on the clean conductor, and a wider scattering angle range).

[0068] Retrieve the electromagnetic interference intensity quantitative value in the hovering state parameter in step 2: the average value of the electric field intensity (such as 320 V / m), the fluctuation range (±40 V / m), determine the influence degree of the interference on the radar signal (such as medium interference causes the reflection signal amplitude deviation ±5%); correct the extracted dielectric constant distribution spectrum and scattering characteristic vector: subtract the system error value caused by the interference (such as 0.2) from the dielectric constant calculation result, adjust the scattering intensity according to the fluctuation range (such as +5% correction), eliminate the deviation influence of electromagnetic interference on the characteristic parameters; based on the corrected dielectric constant distribution spectrum, identify the icing area (the area with dielectric constant > 5.0), calculate the area thickness (the higher the dielectric constant, the greater the thickness, such as dielectric constant 6.0 corresponds to thickness 5 mm); combined with the surface scattering characteristic vector, analyze the icing surface roughness (the larger the scattering angle range, the rougher the icing), infer the icing density (rough surface corresponds to low density fluffy icing, smooth surface corresponds to high density hard icing); integrate the icing thickness, density, dielectric constant average value, scattering characteristic parameters, etc., arrange them in order according to the scanning unit, and form a complete icing characteristic vector.

[0069] The specific implementation process of the above step 339 is as follows: Pretreat the noise-reduced geometric consistency point cloud set output by step 337: remove a small amount of residual noise points (such as isolated points with a distance deviation of > 3 cm from the surrounding points), group them according to the scanning unit, and mark the spatial coordinate range of each group of point clouds (such as X axis 10-15 meters, Y axis 20-25 meters); splice the point cloud data of adjacent scanning units by matching the feature points (such as continuous points on the conductor surface) in the overlapping area, eliminate the coordinate deviation between units, and form an overall point cloud model; use a surface fitting algorithm (such as moving least squares method) to smooth the spliced point cloud, and construct a continuous three-dimensional surface model: the conductor ground wire part is fitted as a smooth cylindrical surface, and the insulator part is fitted as a curved surface with umbrella skirt concave-convex structure, restoring the actual shape of the equipment.

[0070] Call the spatial structure adaptive compensation factor generated in step 334 to locally correct the fitted three-dimensional surface model: adjust the surface coordinates according to the compensation factor for the curvature distortion area (such as the edge of the insulator umbrella skirt) (such as offset outward by 0.3 cm); correct the height direction deviation (such as adjust upward by 0.2 cm) for the area affected by pose drift (such as the point cloud close to the tower), ensure that the model is consistent with the spatial topological relationship of the actual equipment; the finally generated three-dimensional topological reconstruction data contains: the accurate three-dimensional coordinates of each part of the equipment (error ≤2 cm), the surface curvature distribution, and the connection relationship between parts (such as the connection position of the conductor and the insulator), output in a structured format, and label the data confidence (such as "high confidence area: conductor straight line segment" "medium confidence area: insulator umbrella skirt").

[0071] The present application directly captures the physical characteristic differences (such as dielectric constant, scattering law) of the icing and non-icing areas by analyzing the enhanced radar reflection feature matrix, extracting the dielectric constant distribution spectrum and surface scattering characteristic vector; the feature parameters are corrected in combination with the electromagnetic interference quantitative values in the hovering state parameters, effectively offsetting the interference of the electromagnetic environment on the radar signal, avoiding the misjudgment of the icing characteristics caused by signal deviation (such as misjudging the interference signal as icing); integrating the corrected dielectric characteristics, scattering law and other parameters, forming a feature vector containing key indicators such as icing thickness and density, realizing multi-dimensional quantitative description of the icing state; through point cloud splicing and surface fitting, discrete point cloud data is converted into a continuous three-dimensional surface model, completely restoring the spatial structure form of the line equipment (such as conductor, insulator); in the reconstruction process, an adaptive compensation factor is applied to correct the geometric deviation, eliminate the influence of factors such as surface distortion and pose drift, ensure that the three-dimensional topological reconstruction data is highly consistent with the actual spatial topological relationship of the equipment (error controlled within centimeter level), and finally output the three-dimensional topological reconstruction data containing accurate coordinates, surface curvature and component connection relationship, providing a reliable spatial reference for structure defect identification (such as conductor deformation, insulator damage) and size measurement in line inspection.

[0072] In a preferred embodiment of the present application, step 4, the corrected three-dimensional topological reconstruction data and icing feature vector are input into a pre-trained ResNet deep learning model to identify icing thickness distribution, insulator damage or suspended foreign matter anomaly, and output quantitative anomaly type, spatial coordinates and confidence score, including: Step 41, performing spatial grid normalization processing on the three-dimensional topological reconstruction data corrected by the compensation factor output in step 339 to generate a topological structure tensor; at the same time, embedding the icing feature vector output in step 338 into a complex domain feature space to generate an enhanced icing feature matrix; Step 42, fusing the topological structure tensor and the enhanced icing feature matrix generated in step 41 to construct a multi-modal input data cube through a spatial feature alignment algorithm; Step 43, inputting the multi-modal input data cube into a pre-trained ResNet-101 deep learning model to extract spatial topological features and dielectric property correlation mapping through three residual convolution blocks, and outputting an anomaly probability distribution heat map; Step 44, based on the anomaly probability distribution heat map, identifying the spatial area of icing thickness distribution, insulator damage and suspended foreign matter through Gaussian mixture model clustering, and combining the spatial coordinates in the three-dimensional topological reconstruction data of step 339 to generate quantitative anomaly type, anomaly area center point coordinates and confidence score.

[0073] In the embodiment of the present application, the specific implementation process of step 41 is as follows: Set uniform spatial grid parameters (e.g., grid size of 10x10x5 cm covering the entire inspection area), map the spatial coordinates of the three-dimensional topology reconstruction data output in step 339 to the grid, and ensure that the coordinate range of different device regions is standardized; extract the topological features of each grid unit (e.g., surface curvature, structure type label: conductor / insulator / pole), arrange the feature values in order of grid position to form a topological structure tensor with dimensions [grid length x grid width x grid height x feature number]; perform complex domain mapping on the ice feature vector (including thickness, density, dielectric constant, etc.) output in step 338: decompose each feature value into a real part (e.g., ice thickness absolute value) and an imaginary part (e.g., thickness change rate) to enhance the feature expression dimension; arrange the complex domain features into a matrix in order of the spatial coordinates of the scanning units, with the matrix rows / columns corresponding to spatial positions and the elements being complex feature values, to generate an enhanced ice feature matrix.

[0074] The specific implementation process of the above step 42 is as follows: Take the spatial coordinates of the three-dimensional topology reconstruction data as the reference, and perform coordinate matching on the topological structure tensor and the enhanced ice feature matrix: through the coordinate mapping algorithm, ensure that the topological features (from the tensor) and the ice features (from the matrix) at the same physical position correspond one by one in the spatial dimension, and eliminate position deviation; fuse the aligned topological structure tensor (spatial topological features) and the enhanced ice feature matrix (ice physical properties) in the channel dimension: the topological tensor as the "structure channel" and the ice matrix as the "ice channel", combined to form a multi-modal input data cube with dimensions [spatial length x spatial width x spatial height x modal number (2 kinds)], which completely retains the association between spatial position and multiple features.

[0075] The specific implementation process of the above step 43 is as follows: Input the multi-modal input data cube into the pre-trained ResNet-101 model, extract the basic spatial features (e.g., device edges, contours) through the first layer of residual convolution blocks (including 3x3 convolution, batch normalization, and ReLU activation); the second layer of residual blocks strengthens the local detail features (e.g., insulator shed texture, ice surface undulation); the third layer of residual blocks fuse global features to establish the association mapping between spatial topological features (e.g., conductor sag shape) and dielectric properties (e.g., ice dielectric constant distribution); after processing by the three layers of residual blocks, calculate the abnormal probability (0-1, the higher the value, the greater the abnormality possibility) of each spatial grid unit through the model output layer (including 1x1 convolution and sigmoid activation function), generate a two-dimensional abnormal probability distribution heat map according to the grid position, and intuitively display the potential abnormal area.

[0076] The specific construction process of the above pre-trained ResNet-101 model is as follows: The multi-source annotation data of the distribution network line inspection is collected, including: three-dimensional topology reconstruction data (including conductors and insulators in normal and abnormal states three-dimensional structure), icing feature vector (annotating icing thickness, density and other parameters), abnormal label data (manually annotating icing exceeding area, insulator damage position, hanging foreign matter type and coordinates), ensuring that the proportion of normal samples and abnormal samples is balanced (such as 1:1); performing enhancement processing on the original data: randomly rotating (±10°), scaling (0.8-1.2 times), and locally cropping (focusing on key areas such as insulator sheds) the three-dimensional topology data; adding slight noise (±5%) to the icing feature vector to simulate measurement error and expand the diversity of the data set; normalizing the spatial coordinates of the three-dimensional topology data to a unified range (such as [-1, 1]), and normalizing the icing feature vector according to the mean-standard deviation standard, to ensure that the input data has consistent scales; dividing the training set (70%), the validation set (20%) and the test set (10%); setting the input layer to receive a multi-modal data cube, matching the structure output by step 42 (such as [64x64x32x2], where 64x64x32 is the spatial grid size and 2 is the number of modes: topology structure + icing feature), and mapping the input channel to a 64-dimensional feature channel through a convolution layer; adding a 7x7 convolution layer (step 2) to preliminarily extract features from the input data and capture basic edges and contour information; connecting a 3x3 max pooling layer (step 2) to compress the spatial dimension and reduce the amount of calculation, and outputting a feature map size of 1 / 4 of the input.

[0077] Residual block structure: two kinds of residual blocks are designed: Normal residual block: containing 2 3x3 convolution layers (both with batch normalization and ReLU activation), the input and output channel numbers are the same; the jump connection directly adds the input features and the convolution output to alleviate the gradient disappearance problem.

[0078] Down-sampling residual block: the first 3x3 convolution layer is set to step 2 (to realize down-sampling), and the input channel number is 1 / 2 of the output; the jump connection is added after adjusting the channel number through a 1x1 convolution layer and then added to the output to ensure dimension matching.

[0079] Network stage division: ResNet-101 contains 4 residual convolution stages, and the number of residual blocks is set to 3, 4, 23 and 3 in sequence: The first stage: 3 normal residual blocks, output feature channel number 256; The second stage: 4 residual blocks (including 1 down-sampling block), output feature channel number 512; The third stage: 23 residual blocks (including 1 down-sampling block), output feature channel number 1024 (core feature extraction stage, capturing complex topology and icing associated features); The fourth stage: 3 residual blocks (including 1 down-sampling block), output feature channel number 2048.

[0080] After 4 residual stages, a global average pooling layer is added to compress the high-dimensional feature map into a 2048-dimensional global feature vector, which integrates the global correlation information of spatial topology and dielectric properties.

[0081] Output layer construction: connect the fully connected layer and the classification head: The first fully connected layer maps the 2048-dimensional features to 1024-dimensional features, followed by ReLU activation and dropout (probability 0.5) to prevent overfitting. The output layer uses a multi-label classification design, which outputs three parallel convolutional layers to output: icing thickness anomaly probability map, insulator damage probability map, and suspended foreign object probability map, each with a dimension corresponding to the input spatial grid (e.g., 64x64).

[0082] Pre-training process implementation: Loss function selection: use a weighted cross-entropy loss function to assign higher weights to abnormal samples (icing, damage, and foreign objects) (e.g., normal sample weight 1.0, abnormal sample weight 2.0) to address class imbalance. Use the Adam optimizer with an initial learning rate of 0.001 and cosine annealing scheduling (learning rate decays to 1 / 2 every 10 epochs). Set the batch size to 32 and train for a total of 100 epochs. Evaluate the performance on the validation set after each training round (calculate the abnormal recognition accuracy and recall rate), and save the model with the lowest validation set loss as the pre-trained base model. If the validation set performance does not improve for 10 consecutive rounds, terminate the training early. Fine-tune the pre-trained model on the specific data set for network inspection: freeze the parameters of the first 2 residual stages (preserve the general feature extraction capability), and only train the last 2 stages and the output layer. Reduce the learning rate to 0.0001 and train for 20 epochs to enhance the model's ability to identify specific features of line equipment abnormalities. Verify the performance of the fine-tuned model on the test set. If the recognition accuracy of a certain type of abnormality (e.g., small-sized foreign objects) is low, increase the training weight of that type of sample accordingly, and repeat the fine-tuning until the test set accuracy is ≥90%. Finally, determine the pre-trained ResNet-101 model for step 4 abnormality recognition.

[0083] The specific implementation process of the above step 44 is as follows: Based on the abnormal probability distribution heat map, a probability threshold (such as a region with a probability greater than 0.6 is considered as a candidate abnormal region) is set, and a Gaussian mixture model is used to cluster the candidate region: divided into 3 categories (icing abnormality, insulator damage, and suspended foreign matter) according to feature differences (icing abnormality corresponds to high dielectric constant + continuous spatial distribution characteristics, insulator damage corresponds to local structure mutation + low reflection characteristics, and suspended foreign matter corresponds to isolated high probability point + non-device structure characteristics); for each clustered region, determine the abnormal type (match the preset feature template: such as "high icing thickness" corresponds to icing type); calculate the region center point coordinates (take the average value of all grid coordinates in the region); calculate the average value of the abnormal probability in the region as the confidence score (such as an average probability of 0.85 corresponds to a confidence of 85%); combine the accurate coordinates in the three-dimensional topological reconstruction data of step 339 to output the structured results: quantitative abnormal type (such as "icing thickness exceeds standard" and "insulator damage"), abnormal region center point three-dimensional coordinates (longitude / latitude / height), and confidence score.

[0084] In a preferred embodiment of the present application, step 5, according to the quantitative abnormal type, spatial coordinates and confidence score output by step 4, generates an adaptive operation strategy containing tool ID code and operation parameters, including: Step 51, analyze the quantitative abnormal type, abnormal region center point coordinates and confidence score output by step 44, and match the tool ID code through the preset abnormal-tool mapping rule library; Step 52, based on the abnormal region center point coordinates, generate the UAV approach path combined with the three-dimensional flight trajectory of step 1, and dynamically adjust the operation safety distance threshold according to the confidence score; Step 53, fuse the abnormal type of step 44 and the electromagnetic interference intensity quantitative value in the hovering state parameter of step 2, and calculate the laser power, mechanical arm operation torque and deicing time through the operation parameter optimization function; Step 54, aggregate the tool ID code, UAV approach path, operation safety distance threshold and operation parameters to generate an adaptive operation strategy instruction set containing space-time constraints; calculate the laser power, mechanical arm operation torque and deicing time through the operation parameter optimization function, including: The quantitative abnormality type output by the analysis step 44 is parsed, and based on a preset abnormality physical characteristic mapping rule, a physical action parameter reference value of the target tool is determined; the electromagnetic interference intensity quantitative value in the hovering state parameter output by the extraction step 2 is extracted, and the physical action parameter reference value is dynamically corrected through an electromagnetic attenuation compensation model to generate an anti-interference operation parameter intermediate value; according to the confidence score output by the step 44, an operation parameter dynamic scaling coefficient is calculated through a safety margin adjustment function; the anti-interference operation parameter intermediate value and the dynamic scaling coefficient are weighted and fused; for the icing abnormality type, the dielectric constant distribution spectrum in the icing feature vector of the step 338 is combined to output a laser power parameter and a deicing time length; for the insulator damage or suspended foreign matter type, the surface curvature feature in the three-dimensional topological reconstruction data of the step 339 is combined to output a mechanical arm operation torque parameter; the laser power, the mechanical arm operation torque and the deicing time length parameters are aggregated to generate an operation parameter optimization result.

[0085] In the embodiment of the present application, the specific implementation process of the above step 51 is as follows: Key information is extracted from the output result of step 44: quantitative abnormality type (such as "icing thickness exceeds standard", "insulator damage", "suspended foreign matter"), abnormal area center point coordinates (three-dimensional coordinate values) and confidence score (such as 85%), classified and arranged by type (such as all "icing abnormalities" are classified into one category); a preset "abnormality-tool mapping rule library" is called, which stores the association between abnormality type and corresponding operation tool (such as "icing abnormality" corresponds to laser deicing tool, code "TOOL-ICE-001"; "suspended foreign matter" corresponds to mechanical arm grabbing tool, code "TOOL-ARM-002"); according to the analyzed abnormality type, the corresponding tool ID code is accurately matched, and the associated tool code of each abnormality type is recorded.

[0086] The specific implementation process of the above step 52 is as follows: The abnormal area center point coordinates output by step 44 are taken as target points, and the three-dimensional flight trajectory generated in step 1 is called as a basic path framework; the shortest approach path from the current hovering point to the target point is planned in the trajectory framework through a path planning algorithm (such as A* algorithm), and the path needs to avoid the safety boundary in step 1 (such as dangerous areas away from towers and conductors), to ensure smooth and no sudden turning of the path; a basic safety distance threshold (such as 5 meters in a normal state) is set; the threshold is adjusted according to the confidence score of step 44: when the confidence score is greater than or equal to 90%, the threshold is lowered by 10% (such as 4.5 meters, to improve operation accuracy); when the score is less than 70%, the threshold is increased by 20% (such as 6 meters, to improve safety redundancy); when the score is between 70% and 90%, the basic threshold is maintained, and the operation safety distance threshold of each abnormal area is finally determined.

[0087] The specific implementation process of the above step 53 is as follows: Determine the type of anomaly (e.g., "icing anomaly" or "insulator damage") in step 44, and call the electromagnetic interference intensity quantification value (e.g., electric field intensity 320 V / m, fluctuation range ±40 V / m) in the hovering state parameter in step 2; through the operation parameter optimization function, combine the icing thickness (from the icing feature vector in step 338) and the electromagnetic interference intensity to calculate the laser power (the stronger the interference, the power is appropriately increased to offset the attenuation) and the deicing time length (the thicker the thickness, the longer the time length); combine the structural intensity of the abnormal area (from the three-dimensional topological reconstruction data in step 339) and the electromagnetic interference intensity to calculate the mechanical arm operation torque (avoid excessive torque under strong interference to cause equipment damage), and ensure smooth operation.

[0088] The specific implementation process of step 54 is as follows: Collect the tool ID code in step 51, the unmanned aerial vehicle approach path and the operation safety distance threshold in step 52, and the operation parameters (laser power, mechanical arm torque, and deicing time length) in step 53, and group them according to the association of "anomaly type-tool-path-parameter"; add time and space constraints to each group of information: in terms of time, sort the operation priority according to the anomaly confidence (high confidence anomaly is given priority to execute); in terms of space, determine the start and end time of the approach path and the effective range of the safety distance; finally, integrate them into a structured adaptive operation strategy instruction set, including tool calling sequence, path execution node, parameter adjustment opportunity, and other executable information.

[0089] The specific implementation process of the "operation parameter optimization function" in step 54 is as follows: The quantitative anomaly type of the analysis step 44 is called to preset "abnormal physical characteristic mapping rule"; the laser power reference value (such as 30W) and the deicing time length reference value (such as 60 seconds) are set according to the ice thickness level (such as "5-10mm thick ice"); the mechanical arm operation torque reference value (such as 5N.m) is set according to the anomaly size (such as "diameter 5cm foreign matter"); the electromagnetic interference intensity quantitative value of the step 2 is extracted, and the reference value is corrected through an electromagnetic attenuation compensation model: the laser power reference value is increased by 5% (offset signal attenuation) and the mechanical arm torque reference value is reduced by 3% (avoiding misoperation caused by electromagnetic interference) for every 100V / m increase of the electromagnetic interference intensity, to generate an anti-interference operation parameter intermediate value; the scaling coefficient is calculated according to the confidence score of the step 44: the coefficient is 1.1 (to improve operation efficiency) when the score is above 90%, the coefficient is 1.0 (to maintain the standard) when the score is between 70% and 90%, and the coefficient is 0.9 (to reduce intensity to ensure safety) when the score is below 70%; the anti-interference laser power intermediate value x the scaling coefficient is combined with the dielectric constant distribution spectrum (the higher the dielectric constant, the power fine adjustment is increased by 2%) in the ice characteristic vector of the step 338 to determine the final laser power; the deicing time length is calculated and output according to "thickness x scaling coefficient"; the anti-interference mechanical arm torque intermediate value x the scaling coefficient is combined with the surface curvature (the greater the curvature, the torque fine adjustment is reduced by 1%) of the three-dimensional topological reconstruction data of the step 339 to determine the final operation torque, and the laser power, the mechanical arm torque and the deicing time length are integrated to generate the operation parameter optimization result.

[0090] The present application combines three-dimensional trajectory to generate the optimal approaching path, dynamically adjusts the safety distance according to the confidence, shortens the distance to improve the precision when the confidence is high, increases the distance to reduce the risk when the confidence is low, balances the efficiency and safety, optimizes the parameters by fusing the electromagnetic interference intensity, ice characteristics and topological characteristics, offsets the electromagnetic attenuation by the laser power, and adapts the surface curvature by the mechanical arm torque, to ensure the stable and reliable operation effect in complex environment; the parameter reference value is corrected by the electromagnetic attenuation compensation model, effectively offsets the influence of electromagnetic interference on the laser power and the mechanical arm torque, and avoids the operation failure caused by environmental interference.

[0091] The above is the preferred embodiment of the present application, and it should be noted that those skilled in the art can make some improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for autonomous inspection of overhead line equipment using a UAV, characterized in that: The method comprises: Generate the drone's three-dimensional flight trajectory and inspection safety boundary based on the geographic coordinates and topological data of the distribution network overhead line equipment; Based on the three-dimensional flight trajectory and inspection safety boundary, the drone is controlled to reach the target ground line coordinates, and multi-sensor fusion positioning is used to achieve centimeter-level dynamic hovering. In the hovering state, three spatial reference coordinates are dynamically calibrated and the hovering state parameters are output; Based on the reference coordinate set, a dynamic spatial feature envelope is constructed and rasterized to form a multi-level scanning unit set. An adaptive compensation factor is generated based on the deformation characteristics of the envelope. Combined with the hovering state parameters, the surface of the line is scanned and processed using a combination of lidar and dual-polarization radar to obtain corrected 3D topological reconstruction data and ice cover feature vectors. The corrected 3D topology reconstruction data and ice feature vectors are fed into a pre-trained ResNet deep learning model to identify ice thickness distribution, insulator damage, or suspended foreign object anomalies, and output quantitative anomaly type, spatial coordinates, and confidence score. Generate an adaptive operation strategy including tool ID code and operation parameters based on the quantitative anomaly type, spatial coordinates and confidence score.

2. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 1, characterized in that: Generate the drone's 3D flight trajectory and inspection safety boundary based on the geographic coordinates and topological data of the distribution network overhead line equipment, including: Obtain GIS geographic coordinates and equipment topology data of distribution network overhead lines, extract tower spatial location information, ground wire spatial morphological parameters, and adjacent equipment topology relationships; The ground wire spatial morphological parameters are used as input, a three-dimensional trajectory of the ground wire is generated through a spatial curve fitting algorithm, and a dynamic vertical isolation threshold of the UAV is calculated based on electrical safety standards; The device topology and the three-dimensional trajectory of the ground wire are used as input to identify vegetation coverage, building shielding areas, and the distribution of electromagnetic interference sources. A rotationally symmetric dynamic envelope with the ground wire trajectory as the axis is constructed to generate an inspection safety boundary. The three-dimensional spatial trajectory of the ground wire and the rotationally symmetric dynamic envelope surface are integrated, and the UAV's waypoint sequence and attitude control instruction set are generated through the segmented trajectory optimization algorithm, outputting the three-dimensional flight trajectory and inspection safety boundary.

3. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 2, characterized in that: Based on the three-dimensional flight trajectory and inspection safety boundary, the drone is controlled to reach the target ground line coordinates. Multi-sensor fusion positioning is used to achieve centimeter-level dynamic hovering. In the hovering state, three spatial reference coordinates are dynamically calibrated and the hovering state parameters including posture accuracy, electromagnetic interference intensity and reference coordinate set are output, including: Generate multi-degree-of-freedom motion instructions for the UAV based on the 3D flight trajectory and inspection safety boundaries; In response to the drone arriving at the target ground line coordinates, it fuses real-time dynamic carrier phase differential positioning data, inertial measurement unit data, and visual odometry data to generate a real-time pose feedback stream with centimeter-level accuracy; According to the real-time posture feedback flow, the three spatial reference coordinates are dynamically calibrated based on the spatial distribution characteristics of the ground wire in the hovering state, and the fluctuation of the environmental electromagnetic field intensity is simultaneously monitored; Aggregate real-time posture feedback stream, spatial reference coordinate set and electromagnetic field intensity fluctuation data, and output hovering state parameters including posture accuracy, electromagnetic interference intensity quantization value and reference coordinate set.

4. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 3, characterized in that: According to the reference coordinate set, a dynamic spatial feature envelope is constructed and the region is rasterized to form a multi-level scanning unit set, including: The reference coordinate set is used as a sequence of spatial control points, and the dynamic spatial feature envelope surface is generated by the non-uniform rational B-spline surface reconstruction algorithm. Adaptive regional rasterization is performed on the dynamic spatial feature envelope surface, a multi-level scanning unit set is generated based on the curvature change gradient, and the envelope surface deformation feature parameters are extracted at the same time; The spatial structure adaptive compensation factor is calculated based on the deformation characteristic parameters. Combined with the pose accuracy and electromagnetic interference intensity quantization value in the hovering state parameters, the scanning unit set is collaboratively scanned by the laser radar and dual-polarization radar. The original point cloud data, ambient temperature and humidity, wind speed, and electromagnetic field intensity obtained by the fusion scan are processed, and wavelet threshold noise reduction and Kalman filtering are used for data enhancement. The three-dimensional topological reconstruction data and ice cover feature vector corrected by the compensation factor are output.

5. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 4, characterized in that: The adaptive compensation factor of the spatial structure is calculated based on the deformation characteristic parameters, combined with the posture accuracy and electromagnetic interference intensity quantization value in the hovering state parameters, including: Based on the deformation feature parameters, the principal curvature distribution gradient and normal vector offset of the envelope surface are extracted; the curvature-induced spatial topological distortion intensity is calculated based on the principal curvature distribution gradient, and the local coordinate system rotation correction vector is generated based on the normal vector offset; the spatial topological distortion intensity and the local coordinate system rotation correction vector are fused, and the spatial geometric distortion compensation is output through vector synthesis operation; Extract the pose accuracy from the hovering state parameters and generate the pose drift compensation through the positioning error transfer function; Extract the quantitative value of electromagnetic interference intensity from the hovering state parameters and generate the signal propagation compensation value based on the electromagnetic wave attenuation characteristic model; The spatial geometric distortion compensation, posture drift compensation and signal propagation compensation are integrated to generate a spatial structure adaptive compensation factor through nonlinear weighted aggregation.

6. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 5, characterized in that: Based on the deformation characteristics of the envelope surface, an adaptive compensation factor is generated. Combined with the hovering state parameters, the line surface is scanned and processed by a combination of lidar and dual-polarization radar to obtain corrected 3D topological reconstruction data and ice cover feature vectors, including: In response to the coordinated scanning of the scanning unit set by the laser radar and the dual-polarization radar, the original point cloud data set and the dual-polarization echo signal stream are obtained, and the time series data of the ambient temperature and humidity, wind speed and electromagnetic field intensity are simultaneously collected; Based on the original point cloud dataset, spatial coordinate transformation correction is performed through the spatial structure adaptive compensation factor to generate a geometrically consistent point cloud set; Wavelet threshold denoising is performed on the geometrically consistent point cloud set to eliminate high-frequency vibration noise. At the same time, the dual-polarization echo signal stream and environmental time series data are fused using Kalman filtering to generate an enhanced radar reflection feature matrix. The dielectric constant distribution spectrum and surface scattering characteristic vector are extracted from the enhanced radar reflection characteristic matrix, and the ice cover characteristic vector is constructed by combining the electromagnetic interference intensity quantitative value in the hovering state parameter. Perform 3D surface topology reconstruction on the denoised geometrically consistent point cloud set and output 3D topology reconstruction data corrected by the compensation factor.

7. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 6, characterized in that: The corrected 3D topology reconstruction data and ice feature vectors are fed into a pre-trained ResNet deep learning model to identify ice thickness distribution, insulator damage, or suspended foreign object anomalies. The model then outputs the quantitative anomaly type, spatial coordinates, and confidence score, including: The three-dimensional topological reconstruction data corrected by the compensation factor is subjected to spatial grid normalization to generate a topological structure tensor. At the same time, the ice feature vector is embedded into the complex domain feature space to generate an enhanced ice feature matrix. The topological structure tensor and the enhanced ice cover feature matrix are integrated to construct a multimodal input data cube through a spatial feature alignment algorithm. The multimodal input data cube is fed into a pre-trained ResNet-101 deep learning model. The three-layer residual convolutional block extracts the spatial topological features and dielectric property correlation mapping, and outputs an anomaly probability distribution heat map. Based on the anomaly probability distribution heat map, Gaussian mixture model clustering is used to identify the spatial areas of ice thickness distribution, insulator damage and hanging foreign objects. Combined with the spatial coordinates in the three-dimensional topological reconstruction data, quantitative anomaly types, coordinates of the center point of the anomaly area and confidence scores are generated.

8. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 7, characterized in that: Generate an adaptive operation strategy containing tool ID code and operation parameters based on the quantitative anomaly type, spatial coordinates and confidence score, including: Analyze and quantify the anomaly type, the coordinates of the center point of the anomaly area, and the confidence score, and match the tool ID code through the preset anomaly-tool mapping rule library; Based on the coordinates of the center point of the abnormal area and combined with the three-dimensional flight trajectory, the drone's approach path is generated, and the operating safety distance threshold is dynamically adjusted according to the confidence score; The anomaly type is integrated with the quantitative value of the electromagnetic interference intensity in the hovering state parameters, and the laser power, manipulator operating torque, and de-icing time are calculated through the operation parameter optimization function. Aggregate tool ID code, drone approach path, operation safety distance threshold and operation parameters to generate an adaptive operation strategy instruction set with spatiotemporal constraints.

9. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 8, characterized in that: The laser power, robot arm operating torque, and de-icing time are calculated using the operation parameter optimization function, including: Analyze and quantify the anomaly type, and determine the physical parameter baseline value of the target tool based on the preset anomaly physical characteristic mapping rules; Extracting the quantitative value of electromagnetic interference intensity from the hovering state parameters, dynamically correcting the physical action parameter reference value through the electromagnetic attenuation compensation model, and generating an intermediate value of the anti-interference operation parameter; According to the confidence score, the dynamic scaling factor of the operation parameters is calculated through the safety margin adjustment function; Perform weighted fusion of the intermediate value of the anti-interference operation parameter and the dynamic scaling coefficient: According to the abnormal type of icing, combined with the dielectric constant distribution spectrum in the icing feature vector, the laser power parameters and deicing time are output; For damaged insulators or hanging foreign objects, the robot arm operating torque parameters are output based on the surface curvature characteristics in the 3D topological reconstruction data. Aggregate laser power, robotic arm operating torque and deicing time parameters to generate operation parameter optimization results.

10. The method for autonomous inspection of distribution network overhead line equipment by a UAV according to claim 1, characterized in that: Hovering state parameters include posture accuracy, electromagnetic interference intensity and reference coordinate set.

Images