An unmanned aerial vehicle-based wind power blade lightning receptor intelligent detection method

Abstract

The present application belongs to the technical field of intelligent detection of wind power blade lightning receptor, and particularly relates to a wind power blade lightning receptor intelligent detection method based on unmanned aerial vehicle, which adopts binocular vision and PMD sensor fusion to collect lightning receptor surface data, constructs a generative partial differential equation (PDE) model to obtain continuous three-dimensional geometry and normal vector field constraints of the lightning receptor; establishes an unmanned aerial vehicle, telescopic rod and swing mechanism coupled ordinary differential equation (ODE) model to obtain system state time evolution constraints; constructs a physical information neural network (PINN) that fuses PDE geometric constraints and ODE dynamic constraints, takes space-time coordinates as input, system state and control instructions as output, and realizes real-time prediction of compliant contact control sequence; controls the unmanned aerial vehicle to carry out vertical compliant contact through the telescopic rod, and completes lightning receptor on-resistance detection and data collection.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection technology for wind turbine blade lightning arresters, and particularly relates to an intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Traditional wind turbine blade lightning arrester inspections often rely on manual high-altitude climbing operations, which not only presents high safety risks, low efficiency, and strong data subjectivity, but also fails to meet the safety and maintenance requirements of wind turbine blade lightning protection systems. Furthermore, conventional non-contact inspections can only observe external defects and cannot perform core electrical performance tests such as continuity resistance measurements. Existing drone-based inspection solutions generally suffer from insufficient sensor data fusion and limited accuracy in surface modeling. Binocular vision and PMD sensors alone cannot simultaneously acquire absolute coordinates and normal vector fields. Additionally, the dynamic coupling modeling of the drone, telescopic mast, and swing mechanism is complex, and outdoor wind field disturbances can easily lead to instability in the inspection attitude. The lack of intelligent control methods with multiple physical constraints makes it difficult to achieve stable, vertical, compliant contact and accurate inspection of the lightning arrester. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for intelligent detection of a three-degree-of-freedom angled wind turbine blade lightning arrester, the method comprising:

[0004] Step 1: Use binocular vision and PMD sensor fusion to collect surface data of the lightning arrester, construct a generative partial differential equation (PDE) model, and obtain the continuous three-dimensional geometry and normal vector field constraints of the lightning arrester.

[0005] Step 2: Establish the ODE model of the coupled dynamics of the UAV, telescopic boom, and swing mechanism to obtain the system state-time evolution constraints;

[0006] Step 3: Construct a physical information neural network PINN that integrates PDE geometric constraints and ODE dynamic constraints, using spatiotemporal coordinates as input and system state and control commands as output, to predict compliant contact control sequences in real time;

[0007] Step 4: Control the drone equipped with the telescopic pole to perform vertical compliant contact, and complete the lightning arrester continuity resistance detection and data acquisition.

[0008] Preferably, in step one, the binocular vision generates a dense point cloud of the lightning arrester to provide absolute three-dimensional coordinates, and the PMD sensor projects an infrared structured grating to obtain a normal vector field with sub-millimeter precision.

[0009] By alternately optimizing the parametric generative equation (PDE) and jointly optimizing the data loss and physical loss, the optimal surface function and PDE parameters that fit the local shape of the lightning arrester are obtained.

[0010] Preferably, the generative PDE is an anisotropic diffusion equation, and a lightweight neural network is used to generate the PDE source terms. The edge preservation coefficient is used to suppress excessive smoothing of the interface between the lightning arrester edge and the substrate, thus preserving the protrusion and edge features.

[0011] Preferably, in step two, the system state vector includes the UAV's center of mass position, attitude quaternion, telescopic rod length, and pitch / yaw angle of the swing mechanism; the UAV rigid body dynamics, the kinematics of the telescopic rod and swing mechanism, the virtual spring, and the damping contact dynamics are coupled to obtain a unified high-order nonlinear ODE.

[0012] Preferably, the contact force is described by a virtual spring-damping model controlled by admittance, and the contact reaction force acts as a coupling term in the UAV dynamic equation; the telescopic rod and swing mechanism servo system are modeled as a second-order system.

[0013] Preferably, in step three, PINN adopts an encoder-decoder multi-head structure, sharing the MLP encoder to extract spatiotemporal features, and the decoder synchronously outputs the UAV system status, control commands, and contact force;

[0014] The total loss function includes data loss, PDE shape constraint loss, ODE dynamic constraint loss, and regularization loss.

[0015] Preferably, the PDE shape constraint loss forces the predicted contact point to be located on the lightning arrester surface characterized by the PDE; the ODE dynamic constraint loss ensures that the state evolution conforms to the coupled dynamic equations through automatic differentiation.

[0016] Regularization loss minimizes the rate and magnitude of change in control commands, improving motion smoothness and safety.

[0017] Preferably, the binocular camera, PMD sensor, force sensor and on-resistance meter are calibrated before the test; the drone hovers 1-2 meters outside the lightning arrester, the binocular and PMD data are fused to reconstruct the local surface, the PINN predictive control sequence is used, the telescopic rod is extended in advance and the angle between the probe and the lightning arrester normal is less than 3°, and the contact is maintained with a constant force of 3-5N to complete the test.

[0018] Preferably, for a hybrid strip and point lightning arrester, an adaptive gated diffusion tensor is used to distinguish the lightning arrester type through semantic segmentation. The PDE model adaptively matches the diffusion parameters of the strip / point lightning arrester, and PINN adaptively switches between flat / ball probes and contact modes.

[0019] Preferably, during the detection process, the contact force, contact position, and conduction resistance data are transmitted back in real time, and secondary compensation is provided for unsuccessful contact points to complete the detection of all points.

[0020] Based on the above, a generative PDE surface model is constructed by fusing binocular vision and PMD sensors. This model can accurately acquire the continuous three-dimensional geometry and high-precision normal vector field of the lightning arrester, effectively suppressing outdoor environmental noise and matching errors. It clearly distinguishes the curvature characteristics of the lightning arrester and the blade substrate, providing strict geometric constraints for contact detection. This significantly improves the accuracy of surface reconstruction and normal alignment, solving the problem of inaccurate lightning arrester positioning and large contact angle deviation under complex blade surfaces from the source. This invention uses UAV-telescopic rod-swing mechanism coupled ODE dynamic modeling and multi-physical constraint PINN control to achieve accurate evolution prediction of system state and compliant contact control under wind disturbance conditions. This ensures that the detection probe is in stable vertical contact with the lightning arrester surface, avoiding blade scratches and probe slippage. At the same time, it greatly improves the detection success rate and the accuracy of conduction resistance measurement, replacing manual high-altitude operations, reducing safety risks, shortening detection time, and adapting to efficient and intelligent detection of different types of lightning arresters such as strips and dots. Attached Figure Description

[0021] Figure 1 This is a system architecture diagram of the intelligent detection system for three-degree-of-freedom angled wind turbine blade lightning arresters provided in an embodiment of the present invention;

[0022] Figure 2 This is a block diagram of the UAV-telescopic pole coupled ODE dynamics for intelligent detection of three-degree-of-freedom angled wind turbine blade lightning arresters provided in this embodiment of the invention;

[0023] Figure 3 This is an outdoor test schematic diagram of the intelligent detection of the three-degree-of-freedom angled wind turbine blade lightning arrester provided in an embodiment of the present invention;

[0024] Figure 4 This is a flowchart of binocular-PMD fusion PDE modeling for intelligent detection of three-degree-of-freedom angled wind turbine blade lightning arresters provided in an embodiment of the present invention;

[0025] Figure 5 This is a schematic diagram of the actuator (telescopic rod + swing) for intelligent detection of a three-degree-of-freedom angled wind turbine blade lightning arrester provided in an embodiment of the present invention;

[0026] Figure 6 This is a PINN structure diagram of a three-degree-of-freedom angled wind turbine blade lightning arrester provided in an embodiment of the present invention. Detailed Implementation

[0027] Phase Measuring Deflectometry (PMD) is a technique for measuring and deflecting phases. Partial Differential Equation (PDE) Ordinary Differential Equation (ODE) Physics-Informed Neural Network (PINN)

[0028] This invention proposes a comprehensive detection system based on a binocular vision-PMD fusion-based generative PDE modeling method for lightning arrester surfaces, an ODE modeling method for UAV-telescopic pole coupled dynamics, and a PINN real-time motion prediction method with multiple physical constraints. This method enables continuous modeling of the lightning arrester surface in complex wind fields, feasible contact trajectory planning for UAVs, and outputs real-time control commands that satisfy physical constraints, ultimately achieving safe and stable detection of lightning arresters by UAVs.

[0029] Obtaining shape information of partial differential equations (PDEs): Generative dynamic PDE modeling based on binocular vision and PMD;

[0030] Several unique challenges arise when using UAVs for intelligent contact-based inspection of lightning arresters on wind turbine blades: a) Surface complexity: Wind turbine blades are composite material surfaces with complex aerodynamic shapes. Lightning arresters are typically attached to them in strip or dot patterns, with significant curvature variations in local areas (such as the leading and trailing edges of the blade). b) Physical contact requirements: For electrical inspections such as continuity resistance measurements, the end of the telescopic rod must maintain stable, perpendicular, and controllable physical contact with the lightning arrester surface. This requires the system to not only know where the surface is, but also understand its local three-dimensional orientation (normal vector) and curvature variation trends. c) The dynamics and instability of the UAV platform: UAVs in outdoor winds are constantly adjusting their attitude and undergoing minor positional movements, resulting in dynamic noise in the observation data. Therefore, while traditional discrete point cloud models (PMDs) can provide positional information, they cannot directly and robustly provide continuous gradient (normal vector) information and are sensitive to noise. PMDs can directly provide a high-precision normal field, but their individual operation cannot obtain absolute three-dimensional coordinates.

[0031] Based on the initial positioning, the UAV flight control system hovers at a distance of approximately 1-2 meters from the target lightning receiver. This distance balances the effective field of view of the binocular camera with the accuracy of the PMD projected pattern. The binocular camera captures high-resolution images, and a stereo matching algorithm is used to generate a dense point cloud of the lightning receiver and its surrounding blade area. This point cloud provides the location and basic outline of the lightning arrester on the macroscopic curved surface of the blade.

[0032] PMD sensors project specific infrared structured gratings onto the same area. Due to the different reflectivity of the lightning arrester surface (typically metal or a conductive coating) and the surrounding fiberglass blade material, the reflected grating deformation modes also differ. By calculating the phase, a surface normal vector field with sub-millimeter precision can be obtained. This is because PMD is extremely sensitive to microscopic undulations on the surface (lightning arrester edges, bolt protrusions), which is essential to ensure that the contact rod is vertically aligned.

[0033] The core of this section is to obtain the continuous, high-precision three-dimensional geometry and normal vector field of the lightning arrester surface, and to describe its shape features using a PDE, providing spatial geometric constraints for subsequent PINN. Traditional methods such as point cloud stitching and surface fitting produce discrete data, lacking analytical physical expression, and are difficult to seamlessly integrate into the differential equation constraint framework of PINN. The specific methods in this section are as follows:

[0034] The surface of the wind turbine blades can be considered as a smooth two-dimensional manifold. Let the height of the parameterized surface in the local region where the lightning arrester is located be... ,in The coordinates are defined on an approximate tangent plane perpendicular to the initial observation direction of the UAV.

[0035] ;

[0036] This naturally leads to a first-order PDE system:

[0037] (PDE-1);

[0038] in, Measured by PMD The calculations yielded a value for the unknown surface directly derived from the PMD data. The first-order PDE system. Under ideal, noise-free conditions, the surface can be obtained by integrating only the PMD data. However, in actual wind turbine blade scenarios, the PMD data may exhibit jumps or noise at the interface between the lightning arrester and the blade substrate, and the micro-movements of the drone can cause... The field exhibits high-frequency disturbances. Therefore, binocular point cloud imaging is required. The provided absolute position information is used for calibration and fusion.

[0039] Because the curvature characteristics of different parts of the wind turbine blade vary, we propose that the shape of this local surface is an equilibrium state resulting from the combined effects of its material mechanical properties (stiffness and layup of the blade composite), aerodynamic load history, and the added mass of the lightning arrester. This state should be described by a more general equilibrium equation. PDE-1 is overconstrained because it incorporates data from binocular point clouds. We consider the boundary / interior data. We transform this into an optimization problem: finding a PDE whose solution optimally satisfies both the PMD gradient field and the stereo geometry data. We propose a parameterized generative PDE model:

[0040] (PDE-2);

[0041] in, It is a lightweight neural network (generator) with the following parameters: This PDE is the general form of the anisotropic diffusion equation, capable of describing a wide range of smooth surfaces. Joint optimization solution of the PDE and parameters. We construct a loss function that includes both data loss and physical loss:

[0042] ;

[0043] By fixing optimization Then fix optimization By employing an alternating optimization strategy, we ultimately obtain the optimal surface function simultaneously. Optimal PDE parameters This results in a PDE specifically designed to describe the local shape characteristics of the current lightning arrester. This is equivalent to finding a surface that best matches all observational data, given the current understanding of the physical laws governing surfaces.

[0044] This effectively suppresses random noise from illumination variations or image matching errors in PMD and binocular data, resulting in a cleaner, more continuous surface. Simultaneously, based on the current optimal surface estimate, our understanding of the physical laws governing this surface is redefined. This allows the PDE to better adapt to this specific region, for example, distinguishing the different curvature characteristics of a flat blade skin region versus a raised lightning rod zone region. As a powerful geometrical-physical constraint, it will be constrained into the master model PINN. It tells the control system that regardless of where the telescopic boom is planned to move, the surface point it eventually contacts should approximately satisfy this learned equilibrium law. This greatly constrains the range of possible contact points and improves the reliability of predicting contact points during dynamic flight.

[0045] In the detection of lightning arresters on wind turbine blades by UAVs, dynamic modeling faces multiple challenges: a) Strongly coupled system: The UAV body, telescopic actuator, and swing mechanism constitute a highly coupled dynamic system. The movement of any one component will affect the overall attitude and position, especially when contact forces are generated, this coupling effect is more significant. b) Environmental disturbances: Outdoor wind fields generate continuous and changing aerodynamic disturbances on the UAV, directly affecting its attitude stability and control accuracy. c) Transient and uncertain contact interaction: The contact between the end of the telescopic rod and the lightning arrester surface is a transient event involving collision, force transmission, and possible slippage, which needs to be reasonably described in the dynamic model to ensure the stability and safety of the control system. d) Real-time requirements: The entire model needs to be sufficiently concise to support online prediction, while maintaining sufficient fidelity to accurately reflect system behavior. Therefore, the established ODE will serve as a key time-domain constraint for the Physical Information Neural Network (PINN), ensuring that the control commands predicted by the network conform to the actual physical evolution of the system.

[0046] The core of this section is to establish dynamic equations describing the evolution of the UAV's pose, telescopic boom extension, and swing angle over time, providing time-dependent constraints for PINN. The specific method is as follows:

[0047] System State Definition. Considering the stringent requirements for position and attitude accuracy in the wind turbine blade lightning arrester detection task, we define the system state in a more refined manner. Define the state vector. ,in Represents the position of the UAV's center of mass (world coordinate system); Represents the attitude quaternion of the drone. This represents the length of the telescopic pole (0.5~1 meter). Represents the pitch and yaw angles of a bearing-based oscillating mechanism.

[0048] Unmanned Aerial Vehicle (UAV) Rigid Body Dynamics (Basic ODE). When operating near wind turbine blades, the UAV is affected not only by gravity, rotor thrust, and its own dynamic torque, but also by time-varying wind disturbance and contact reaction forces. Its dynamic equations, simplified in the body coordinate system, are as follows:

[0049] ;

[0050] (ODE-1);

[0051] in The control force / torque under the machine body axis system. The external force / torque caused by the contact force.

[0052] Kinematic extension of the telescopic rod and swing mechanism. Contact rod end point.

[0053] in It is the fixed offset of the rod base in the machine system. By swing angle Decision. Taking the derivative with respect to time, the terminal velocity can be obtained. With state quantity and its derivative Explicit relationships.

[0054] Contact Dynamics and Unified ODE. To prevent damage to the fragile lightning arrester or blade surface and to ensure measurement stability, we employ a virtual spring-damped model based on admittance control to describe the contact force when the tip contacts the lightning arrester surface. Contact force is generated upon contact. Based on the force sensor feedback and the compliant control model (spring-damped model), we have:

[0055] ;

[0056] in This refers to the target location of the surface contact point. Force Reverse Influence on Unmanned Aerial Vehicle Dynamics (ODE-1) Meanwhile, the servo system of the telescopic rod and the swing mechanism can be modeled as a second-order system:

[0057] ;

[0058] (ODE-2);

[0059] ;

[0060] Couple ODE-1 with ODE-2, and... Introduced as a coupling term, it can be organized into a high-order nonlinear ODE describing the entire UAV-robotic arm coupled system:

[0061] ;

[0062] in It is a control input. These are physical parameters such as system inertia and damping.

[0063] In this way, PINN's predictions are no longer purely data-driven fitting, but rather state evolution driven by physical laws. It can naturally deduce how the drone needs to adjust its posture in advance, when the telescopic rod needs to extend, and how the swing mechanism needs to rotate in order to contact a point on the lightning arrester at a specific time and angle.

[0064] In the task of detecting lightning arresters on wind turbine blades, UAVs need to make precise and smooth physical contact with a complex curved surface (lightning arrester) with a known geometry but relatively uncertain position in a dynamic environment (wind disturbance, self-sway). This requires the control system not only to "see" (part one PDE model), but also to "understand" how it moves (part two ODE model), and to decide when and how to act to achieve the target.

[0065] Therefore, construct a spatiotemporal coordinate system. PINN takes the system state, control commands, and contact forces as inputs and outputs them, and trains it using the aforementioned PDE and ODE as strong constraints to enable it to predict optimal maneuvers in real time during flight. The specific implementation process is as follows:

[0066] PINN Network Architecture. We design a multi-head network with an encoder-decoder structure called PINN, denoted as PINN. Its input is normalized spatiotemporal coordinates. The encoder employs a shared multilayer perceptron (MLP) to extract high-dimensional features from spatiotemporal coordinates. The decoder uses a multi-head output:

[0067] Output 1 is (state prediction): This predicts the complete motion trajectory of the drone body and telescopic mechanism from now into the future.

[0068] Output 2 is (control prediction): It outputs the corresponding sequence of control commands, which is the final execution command of the system.

[0069] Output 3 is (contact prediction): It integrates the network's decisions with the lightning arrester geometry established in the first part. Directly connected.

[0070] Construct a multi-physics constraint loss function, the total loss function expression is:

[0071] ;

[0072] Among them, data loss During offline training, we use historically successfully detected flight data (high-fidelity simulation data) as supervision, derived from real-time sensor streaming data. In the online phase, the current sensor reading serves as an instantaneous anchor point, constraining the PINN output. The starting point of the predicted trajectory must match the actual measured value. This ensures that the starting point of the predicted trajectory is real, specifically expressed as:

[0073] ;

[0074] PDE shape constraint loss : Ensure predicted contact points Always located on the lightning arrester surface learned in Part 1 Above, and satisfy its PDE.

[0075] ;

[0076] This means that when planning the contact trajectory, PINN selects contact points guided by the physics of the underlying surface formation, thus avoiding the selection of physically unreasonable (too sharp or abrupt) points for contact.

[0077] ODE dynamic constraint loss The state evolution predicted by the forced neural network satisfies the unified coupled ODE derived in the second part.

[0078] ;

[0079] Where the derivative It can be obtained through the automatic differentiation of the neural network output.

[0080] The significance of this constraint is that PINN no longer merely learns the input-to-output mapping, but is forced to learn the causal physical laws of the system. Implicit forward dynamics simulation is performed internally within the network. When it outputs a set of control commands... It is possible to know how this set of commands will affect the state through ODE. Furthermore, the contact force is influenced by the contact model. Therefore, its output , It is a command that is dynamically feasible and can achieve the expected contact trajectory.

[0081] To ensure smooth, energy-efficient, and safe operation, regularization loss is introduced. Improving control quality is specifically expressed as:

[0082] ;

[0083] By minimizing the rate of change of control commands To avoid violent movements of the motor and mechanism, and to minimize the amplitude penalty of thrust and torque commands, the drone should not be too close to the non-contact area of ​​the blades.

[0084] Minimize using historical flight detection data or high-fidelity simulation data Training PINN parameters During drone flight, the current sensor observations are used as the basis for PINN's operation. By using partial inputs and anchor points that represent the actual data loss at each moment, and through single-step or a few-step network forward propagation and fine-tuning, PINN can directly output the optimal control sequence for a future time period. .

[0085] Action decision: output This includes: the timing and length of the telescopic pole's operation. Directly specify when to extend and how long to extend; the angle of the swing mechanism. Directly provide instructions on how to oscillate to adapt to the surface normal; UAV attitude adjustment Used for coordinated obstacle avoidance and maintaining stable contact.

[0086] Example 1: Contact detection of 83m blade strip lightning rod in an onshore wind farm;

[0087] Test Object and Environment: A rated 3.6MW onshore turbine unit with a blade length of 83m. The lightning arresters are copper / aluminum alloy strip lightning arresters continuously laid near the leading edge of the blades, with a single test section length of approximately 1.2m. The test was conducted in a North China plain wind field with an ambient temperature of 18-24℃, relative humidity of 45%-62%, average wind speed of 3.2-5.8m / s, and gusts not exceeding 7.2m / s. The blades were shut down, feathered, and mechanically locked.

[0088] Hardware components: The UAV adopts the DJI Motrice 350RTK-class hexacopter industrial UAV platform; the binocular sensor uses the Intel RealSense D455, with a baseline of approximately 95mm and outputting a 1280×720 depth map; the PMD uses the pmdflexx2 time-of-flight / phase-modulated depth sensor; the onboard computing unit uses NVIDIA Jetson Orin NX; the telescopic boom is a carbon fiber two-stage telescopic structure, with a length of 0.42-1.10m, a maximum telescopic speed of 0.18m / s, and a repeatability error of ±1.5mm; the oscillation mechanism is a pitch-yaw two-degree-of-freedom gimbal, with an angle range of ±25°, a resolution of 0.1°, and a maximum angular velocity of 60° / s; the end effector is equipped with a 0-50N thin-film force sensor and an elastic gold-plated conductive probe.

[0089] Software and Model Parameters: In this embodiment, the diffusion tensor in the generative PDE model of the lightning arrester surface adopts anisotropic diffusion, specifically expressed as follows:

[0090] ;

[0091] in, Rotation angle along the direction of the lightning arrester strip Constructed two-dimensional rotation matrix, This represents the diffusion coefficient along the direction of the lightning arrester strip. This represents the lateral diffusion coefficient perpendicular to the strip direction. In this embodiment, it is taken as:

[0092] ;

[0093] The edge preservation coefficient is:

[0094] ;

[0095] This is used to characterize the diffusion suppression level at the edges of the lightning arrester, the raised areas, and the interface between the blade substrate, in order to avoid excessive smoothing that could lead to loss of edge features. PDE source term The source term generation network is generated by a lightweight neural network. The source term generation network uses a 2-layer MLP, with each layer containing 32 nodes, and is used to generate surface correction terms based on local binocular point cloud coordinates, PMD normal vectors, and local curvature features.

[0096] The equivalent dynamic parameters in the coupled ODE model of the UAV-telescopic boom-swing mechanism are taken as follows:

[0097] ;

[0098] ;

[0099] in, The equivalent total mass of the drone and its detection actuator. This is the equivalent rotational inertia matrix of the machine body.

[0100] The contact dynamics employ a virtual spring-damped model, with the contact stiffness and damping parameters as follows:

[0101] ;

[0102] ;

[0103] The parameters of the second-order servo system for the telescopic pole are as follows:

[0104] ;

[0105] The parameters of the second-order servo system for the pitch direction of the swing mechanism are:

[0106]

[0107] The parameters of the second-order servo system for the yaw direction of the swing mechanism are as follows:

[0108] ;

[0109] The Multi-Physical Constraint PINN network employs a structure combining Fourier feature encoding and a shared MLP. The input performs Fourier feature mapping on the spatiotemporal coordinates, local surface coordinates, normal vector estimates, and the current state of the UAV, and then feeds these features into the shared backbone network.

[0110] The shared MLP adopts a 6-layer architecture. 128 nodes;

[0111] The activation function used is the SiLU function: ;

[0112] in, This is the Sigmoid function.

[0113] The total loss function is a weighted combination of data loss, PDE shape constraint loss, ODE dynamic constraint loss, and contact force constraint loss, with the following weight values:

[0114] ;

[0115] Experimental steps: First, complete the four-wire calibration of camera intrinsic and extrinsic parameters, PMD phase zero point, force sensor zero point, and on-resistance meter. The UAV hovers 1.2m outside the lightning arrester and acquires 15 frames of binocular images and PMD phase / intensity frames, which are then fused to generate a local surface. The PDE surface is used to calculate the strip centerline, local normal vector, and candidate contact points. PINN predicts the extension and oscillation control sequence for the next 1.5 seconds. The extension rod extends 0.45 seconds in advance, and the oscillation mechanism ensures that the angle between the probe axis and the local normal is less than 3°. The probe is held with a contact force of 3-5N for 1.0 seconds. The on-resistance is read, and the contact force, contact time, and contact position deviation are transmitted back to complete the retraction and the next point detection.

[0116] Results: A total of 24 detection points were selected. 23 points achieved successful initial contact, and all 24 points were successfully contacted after secondary compensation. The surface reconstruction RMSE was 2.4 mm, and the average normal angle error was 2.1°.

[0117] The lateral deviation of the contact point was 4.2 mm, the average contact force was 3.8 N, and the force fluctuation was ±0.46 N. The conduction resistance deviated by 1.8% from the manual four-wire retest. The average detection time for a single point was 32 s. No probe slippage, blade surface scratches, or accidental contact in non-lightning arrester areas were observed.

[0118] Example 2: Testing of a 108m blade hybrid lightning rod in a mountainous onshore wind field;

[0119] Test Object and Environment: The test involved a 10MW-rated prototype blade, 108m long, with lightning arresters including continuous strips at the blade tip and discrete point-like arresters in the middle section. Significant curvature variations and localized contamination layers were observed. The test was conducted in a mountainous onshore wind field at an altitude of approximately 1450m, with ambient temperatures ranging from 8-16℃, relative humidity from 40%-55%, average wind speeds from 4.0-6.5m / s, turbulence intensity of approximately 0.14, and strong reflectivity on the solar side.

[0120] Hardware components: The UAV uses a FreeflyAltaX-class high-payload multi-rotor platform or an equivalent industrial UAV; the binocular sensor uses dual Baslerace industrial cameras to form an externally synchronized binocular system with a baseline of 120mm and a lens focal length of 8mm; the PMD uses ifmO3D series three-dimensional PMD sensors; the telescopic rod is a carbon fiber three-stage telescopic structure with a length of 0.60-1.50m, a maximum telescopic speed of 0.12m / s, and an axial stiffness of not less than 180N / mm; the swing mechanism has a pitch of ±35°, a yaw of ±35°, an angular velocity of 50° / s, and a replaceable flat / ball-head dual-purpose conductive probe at the end, with a built-in 6-axis force / torque sensor.

[0121] Software and Model Parameters: In this embodiment, the PDE modeling of the lightning arrester surface is designed for a mixed distribution of strip and point lightning arresters, employing an adaptive gated diffusion tensor based on the lightning arrester type. Its expression is:

[0122] ;

[0123] in, This is the lightning arrester type gating factor, used to indicate whether the current local area is closer to a strip lightning arrester or a point lightning arrester. When When the value approaches 1, the model mainly uses strip lightning arresters; when... When the value approaches 0, the model mainly uses the point-like lightning arrester diffusion tensor. .

[0124] For the strip-type lightning arrester region, its anisotropic diffusion parameters are taken as follows:

[0125] ;

[0126] For the point-type lightning arrester region, the polar coordinate direction diffusion parameter is taken as follows:

[0127] ;

[0128] The equivalent dynamic parameters in the coupled ODE model of the UAV-telescopic boom-swing mechanism are taken as follows:

[0129] ;

[0130] ;

[0131] in, The equivalent total mass of the drone and its detection actuator. This is the equivalent rotational inertia matrix of the machine body.

[0132] The contact dynamics employ a virtual spring-damped model, with the contact stiffness and damping parameters as follows:

[0133] ;

[0134] ;

[0135] The parameters of the second-order servo system for the telescopic pole are as follows:

[0136] ;

[0137] The parameters of the second-order servo system for the pitch direction of the swing mechanism are:

[0138] ;

[0139] The parameters of the second-order servo system for the yaw direction of the swing mechanism are as follows:

[0140] ;

[0141] The Multi-Physical Constraint PINN network employs a structure combining Fourier feature encoding and a shared MLP. The input performs Fourier feature mapping on the spatiotemporal coordinates, local surface coordinates, normal vector estimates, and the current state of the UAV, and then feeds these features into the shared backbone network.

[0142] The shared MLP adopts an 8-layer architecture. 192 nodes;

[0143] The activation function uses the GELU function: ;

[0144] in, This is the cumulative distribution function of the standard normal distribution.

[0145] The total loss function is a weighted combination of data loss, PDE shape constraint loss, ODE dynamic constraint loss, and contact force constraint loss, with the following weight values:

[0146] ;

[0147] Experimental Procedure: First, a priori template library for strip / point lightning arresters was established. Binocular image semantic segmentation was used to distinguish the lightning arrester, blade substrate, and contamination layer. Binocular point clouds were used to provide absolute coordinates, PMD normal vectors were used to correct for strip edges and point protrusions, and the PDE-gated network output a continuous surface and type confidence score. PINN simultaneously predicted strip contact modes, flat-head probe modes, point contact modes, and ball-head probe modes, and selected the execution strategy based on the risk score. In areas with high curvature, the oscillating mechanism first performed a ±5° micro-oscillation search, and after confirming that the contact force direction was consistent with the PDE normal, it entered a stable measurement phase. At least three sets of continuity data were collected for each region, and median filtering and a contact force stabilization window were used to filter the final results.

[0148] Results: A total of 36 detection points were selected. 34 points achieved successful initial contact, and all 36 points were successfully contacted after secondary compensation. The RMSE of the reconstructed hybrid surface was 3.1 mm, the average normal angle error was 2.5°, the average contact force was 4.1 N, the force fluctuation was ±0.52 N, and the average attitude compensation lead was 0.62 s. The deviation between the conduction resistance and manual retest was 2.0%, and the time for sampling the entire blade was approximately 23 minutes. The false trigger rate in highly reflective areas decreased from 8.6% with the binocular-only scheme to 2.1%.

[0149] I. Comparative Verification of Different Detection Schemes and Key Module Ablation. To further verify the improvement effects of the binocular-PMD fusion generative PDE modeling, UAV-actuator ODE coupling modeling, and PINN multi-physics constraint control method proposed in this invention compared with existing technologies, comparative tests were conducted on manual detection, traditional UAV detection, and the scheme of this invention, and univariate ablation experiments were performed on key innovative modules.

[0150] A comparative experiment of core performance indicators across different schemes was conducted. The experimental scenario used a 100m-class wind turbine blade lightning arrester detection scenario, with a total of 50 detection points, including 30 strip-type lightning arresters and 20 point-type lightning arresters. Tests were performed using a manual basket / rope detection method, an existing vision-based UAV-assisted detection method, and the UAV intelligent contact detection method of this invention. The performance comparison results of different wind turbine blade lightning arrester detection schemes are shown in Table 1. Table 1. Performance comparison results of different lightning arrester testing schemes for wind turbine blades;

[0151] Average detection time per point 180-240s 65-90s 28-35s Overall detection efficiency benchmark value An increase of approximately 58% An increase of approximately 82% First contact success rate 92.5% 84.3% 96.8% Success rate after secondary compensation 98% 91.6% 100% Probe-light receiver normal angle error 5.5°-8.0° 4.2°-6.5° 1.8°-2.5° On-resistance detection error ±2.5% ±5.8% ±1.8% Work safety risks High altitude, requiring personnel to approach the blades at high altitude. In China, drones pose a risk of collision. Low-altitude, personnel-free high-altitude operation with compliant control Maximum stable detection wind speed ≤5m / s ≤7m / s ≤10m / s Surface Reconstruction RMSE Not involved 6.5mm 2.4-3.1mm Contact force stability range Manually controlled, with large fluctuations. ±2.0N ±0.5N

[0152] The results above show that although manual inspection can achieve continuity testing, it has problems such as long inspection cycle, high safety risk and insufficient inspection consistency. Existing UAV inspection mainly relies on visual positioning and lacks contact dynamics prediction capability, which can easily cause angular deviation in complex curved surfaces and wind disturbance environments.

[0153] This invention obtains continuous surface constraints through binocular-PMD fusion PDE surface reconstruction and combines it with ODE-PINN predictive control, enabling the probe to adjust its posture in advance and achieve stable vertical contact. Therefore, it is significantly superior to existing solutions in terms of detection efficiency, contact success rate, conduction accuracy and safety.

[0154] II. Single-variable ablation verification experiment for key innovative modules. To verify the necessity of binocular-PMD fusion, PINN control, and PDE / ODE physical constraints, a single-factor control experiment was conducted by changing only the algorithm module while maintaining complete consistency in the UAV platform, detection environment, lightning arrester type, and control hardware. The ablation experiment results based on different combinations of physical constraints are shown in Table 2.

[0155] Test conditions: wind speed: 5-7 m / s; detection distance: 1.2 m; contact force: 4 N; number of test points: 40.

[0156] Table 2 Ablation experimental results based on different combinations of physical constraints;

[0157] Binocular vision only 7.2mm 6.8° 82.5% ±1.8N 5.6% PMD sensor only 5.8mm 4.9° 87.5% ±1.5N 4.3% Binoculars + PMD, no PINN control 3.4mm 3.6° 90.0% ±2.0N 3.5% Binocular + PMD + PINN, without PDE / ODE constraints 3.1mm 3.0° 92.5% ±1.2N 2.8% Complete solution of the present invention 2.4mm 2.1° 97.5% ±0.46N 1.8%

[0158] Experimental results show that while binocular vision alone can obtain absolute spatial position, the local surface gradient error is large due to the influence of blade reflection and insufficient texture, leading to increased contact direction deviation. While using only the PMD sensor can obtain relatively accurate local normal information, the lack of large-scale spatial coordinate constraints results in overall positioning drift. After removing PINN predictive control, the UAV cannot adjust the extension rod length and swing angle in advance based on future conditions; the contact process mainly relies on feedback adjustment, resulting in contact force fluctuations of ±2N.

[0159] After removing the PDE / ODE physical constraints, the network degenerates into a regular data-driven model. Although it can learn control relationships, it cannot guarantee that the prediction results satisfy the geometric continuity of the surface and the feasibility of the dynamics. By adopting the complete solution of this invention, the surface error is reduced by PDE constraints, and the motion prediction is ensured to conform to the real dynamic laws by ODE constraints, so that the contact force is stably controlled within ±0.5N, achieving high-precision compliance detection.

[0160] III. Adaptability Verification in Complex and Extreme Environments. To verify the robustness of the method of this invention under salt spray, high wind speed, strong reflection, pollution, and low temperature environments, extreme environment tests were conducted. The test results of the wind turbine blade lightning arrester detection performance under different complex environments are shown in Table 3: Table 3. Test results of lightning arrester performance of wind turbine blades under different complex environments;

[0161] Standard terrestrial environment Wind speed 3-5 m / s, 20℃ 97.5% 2.4mm 2.1° ±0.46N 1.8% Marine salt spray environment Salt spray concentration 5%, humidity above 90% 96.7% 2.9mm 2.4° ±0.55N 2.1% Strong gusts of wind Gusts of 7-10 m / s 95.8% 3.2mm 2.8° ±0.68N 2.5% Strong solar reflectivity environment High-brightness metallic reflective area 96.7% 3.0mm 2.6° ±0.57N 2.2% Leaf pollution Dust and oil stains cover 10%-20% of the surface. 95.0% 3.5mm 3.0° ±0.72N 2.7% Low temperature environment -20℃ 95.8% 3.3mm 2.9° ±0.70N 2.6%

[0162] Test results show that while the contact success rate of this invention decreases to some extent with increasing environmental complexity, it remains above 95% overall, indicating that the invention has high environmental adaptability. Specifically, in a marine salt spray environment, the spatial geometric information provided by binocular vision and the continuous surface constraint of the PDE effectively compensate for local measurement errors due to the attenuation of the PMD reflection signal, maintaining a contact success rate above 96%. In a 7–10 m / s gust environment, the contact success rate decreases slightly due to continuous wind disturbance to the UAV, but ODE dynamic prediction and PINN control can correct the UAV attitude and actuator motion in advance, keeping the contact process stable. Even in environments with strong reflection, pollution, and low temperatures, the invention maintains a high detection success rate, fully verifying the robustness and engineering applicability of the method in complex environments.

[0163] Therefore, this invention is not only applicable to ordinary onshore wind farms, but can also meet the intelligent detection needs of wind turbine blade lightning arresters in complex environments such as offshore, high-altitude and high-wind-speed environments.

[0164] Figure 3 This invention provides an intelligent detection system for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs), comprising a wind turbine tower 1, wind turbine blades 2, lightning arresters 3 disposed on the surface of the wind turbine blades, a UAV platform 4, a telescopic detection rod 5 mounted at the front end of the UAV platform, a conductive contact probe 6 disposed at the end of the telescopic detection rod, a binocular vision camera 7 disposed near the conductive contact probe, a PMD (Phase Measuring Deflectometry) sensor 8, a swing mechanism 12 disposed at the connection between the telescopic detection rod and the UAV platform, and a six-dimensional force / torque sensor 13 disposed at the output end of the swing mechanism or the mounting end of the conductive contact probe. The UAV platform 4 establishes a wireless communication connection with a ground control terminal 11, which is used for flight mission planning, detection mission management, data reception, detection result display, and model parameter updates.

[0165] Among them, the binocular vision camera 7 is used to acquire binocular images of the wind turbine blade 2 and the lightning arrester 3 area, and obtains the dense three-dimensional point cloud of the lightning arrester area through a stereo matching algorithm, providing the system with the absolute spatial coordinate information of the target area; the PMD sensor 8 is used to project a structured grating onto the surface of the lightning arrester and acquire the reflection phase information, and obtains the high-precision normal vector field and local curvature information of the local area of ​​the lightning arrester through phase calculation, providing gradient constraints for subsequent generative partial differential equation (PDE) modeling.

[0166] The telescopic detection rod 5 can extend and retract axially in the direction shown by arrow 9 in the figure to achieve precise adjustment of the distance between the detection probe and the lightning arrester. The swing mechanism 12 is set between the telescopic detection rod and the UAV platform. It adopts a pitch-yaw two-degree-of-freedom structure to drive the telescopic detection rod to adjust its attitude around the pitch axis and yaw axis. This allows the conductive contact probe 6 to automatically adjust its contact angle according to the local normal direction of the lightning arrester, achieving a vertical and compliant contact between the probe and the surface of the lightning arrester, improving contact stability and continuity detection accuracy.

[0167] A six-dimensional force / torque sensor 13 is positioned between the output of the swing mechanism 12 and the conductive contact probe 6. It measures the contact force and torque in three directions during real-time contact between the probe and the lightning arrester, and feeds the detected force / torque information back to the control system. Based on the feedback from the six-dimensional force / torque sensor 13, the control system, combined with the admittance control model, adjusts the extension / retraction of the telescopic detection rod 5 and the attitude of the swing mechanism 12 in real time. This achieves closed-loop control of the contact force, ensuring that the contact force remains within a preset range during detection. This prevents damage to the lightning arrester or blade surface due to excessive contact force, or failure of the continuity test due to insufficient contact force.

[0168] Arrow 10 in the figure indicates the wind direction and airflow direction in an outdoor wind field environment. Wind disturbance can cause dynamic changes in the attitude and position of the UAV platform 4, thereby affecting the stability of the contact detection process. This invention establishes the coupled dynamic ordinary differential equation (ODE) of the UAV-telescopic detection rod-swing mechanism and combines it with the generative partial differential equation (PDE) of the local surface of the lightning arrester to construct a physical information neural network (PINN) that integrates the geometric constraints of PDE and the dynamic constraints of ODE. This enables UAV attitude prediction, telescopic control, swing compensation, and compliant contact control under wind disturbance conditions, improving the stability, robustness, and detection accuracy of the contact detection process.

[0169] The ground control terminal 11 is used to receive binocular images, PMD measurement data, six-dimensional force / torque data, conduction resistance data and UAV flight status information uploaded by the UAV platform in real time, and to complete the detection data storage, result visualization display, task scheduling management and PINN model parameter update. At the same time, it sends flight control commands and detection control commands to the UAV platform to realize remote monitoring and closed-loop control of the entire intelligent detection process of wind turbine blade lightning arresters.

[0170] Figure 5This invention provides an actuator for intelligent detection of lightning arresters on wind turbine blades, installed at the front end of an unmanned aerial vehicle (UAV) platform 21. It includes a telescopic rod mounting base 22, a pitch / yaw swing mechanism 23, a primary telescopic rod 24, a secondary telescopic rod 25, a six-dimensional force / torque sensor 26, a conductive contact probe mounting joint 27, and conductive contact probes. A lightning arrester 29 is provided on the surface of the wind turbine blade 28.

[0171] The telescopic rod mounting base 22 is fixedly connected to the UAV platform 21 to support the entire detection execution mechanism and transmit the control commands of the UAV platform to the swing mechanism 23. The swing mechanism 23 is located between the telescopic rod mounting base 22 and the first-stage telescopic rod 24. It preferably adopts a pitch-yaw two-degree-of-freedom gimbal structure, which can adjust the attitude around the pitch axis and yaw axis respectively to compensate for the angle error caused by the attitude change of the UAV and wind disturbance, so that the telescopic detection mechanism always remains facing the local normal direction of the lightning arrester.

[0172] The primary telescopic rod 24 and the secondary telescopic rod 25 constitute a multi-stage telescopic actuator, which achieves axial extension and retraction under the drive of an electric lead screw, ball screw, linear motor or other linear drive device. It is used to adjust the working distance between the conductive contact probe and the lightning arrester, so as to achieve precise positioning and contact control at different detection distances.

[0173] A six-dimensional force / torque sensor 26 is installed at the front end of the secondary telescopic rod 25 to measure the contact force and contact torque in three directions of the conductive contact probe during contact with the lightning arrester in real time, and feeds the measurement results back to the control system. The control system adjusts the telescopic rod extension and retraction amount and the swing mechanism posture in real time according to the contact force feedback to achieve closed-loop control of the contact force, ensuring that the conductive contact probe always acts stably on the surface of the lightning arrester with a preset contact force.

[0174] The conductive contact probe mounting joint 27 is located at the front end of the six-dimensional force / torque sensor 26 and is connected to the conductive contact probe. It is used to provide a certain range of small angle compensation capability, so that the conductive contact probe can further adapt to the local curvature changes of the lightning arrester, improve contact stability, and avoid probe slippage or local stress concentration caused by local attitude errors.

[0175] The wind turbine blade 28 is the object of inspection, and its surface is equipped with strip-type or point-type lightning arresters 29. During the inspection process, after the binocular vision camera and PMD sensor complete the three-dimensional reconstruction and normal estimation of the lightning arrester, the Physical Information Neural Network (PINN) combines the generative partial differential equation (PDE) surface constraints and the UAV-actuator coupled dynamic ordinary differential equation (ODE) to predict the future control sequence, control the swing mechanism 23, the first-stage telescopic rod 24 and the second-stage telescopic rod 25 to move in coordination, so that the conductive contact probe gradually approaches the target contact point along the local normal direction of the lightning arrester, and completes compliant contact under the real-time feedback of the six-dimensional force / torque sensor 26, realizing the detection of the lightning arrester's conduction resistance and the acquisition of related detection data.

[0176] This represents the position vector of the UAV's centroid in the world coordinate system. Indicates the velocity of the drone's center of mass; This indicates the acceleration of the drone's center of mass; Represents the attitude quaternion The corresponding rotation matrix; This indicates the control thrust generated by the unmanned aerial vehicle (UAV) system. This indicates external disturbance forces, including wind disturbance and contact reaction forces; This represents the equivalent total mass of the drone and its actuators.

[0177] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

[0178] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for intelligent detection of lightning arresters on wind turbine blades based on unmanned aerial vehicles (UAVs), characterized in that: Includes the following steps: Step 1: Use binocular vision and PMD sensor fusion to collect surface data of the lightning arrester, construct a generative partial differential equation (PDE) model, and obtain the continuous three-dimensional geometry and normal vector field constraints of the lightning arrester. Step 2: Establish the ODE model of the coupled dynamics of the UAV, telescopic boom, and swing mechanism to obtain the system state-time evolution constraints; Step 3: Construct a physical information neural network PINN that integrates PDE geometric constraints and ODE dynamic constraints, using spatiotemporal coordinates as input and system state and control commands as output, to predict compliant contact control sequences in real time; Step 4: Control the drone equipped with the telescopic pole to perform vertical compliant contact, and complete the lightning arrester continuity resistance detection and data acquisition.

2. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: In step one, binocular vision generates a dense point cloud of the lightning arrester to provide absolute three-dimensional coordinates, and the PMD sensor projects an infrared structured grating to obtain a normal vector field with sub-millimeter precision. By alternately optimizing the parametric generative equation (PDE) and jointly optimizing the data loss and physical loss, the optimal surface function and PDE parameters that fit the local shape of the lightning arrester are obtained.

3. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that: The generative PDE is an anisotropic diffusion equation. A lightweight neural network is used to generate the PDE source terms. The edge preservation coefficient is used to suppress excessive smoothing of the interface between the lightning arrester edge and the substrate, thus preserving the protrusion and edge features.

4. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: In step two, the system state vector includes the UAV's center of mass position, attitude quaternion, telescopic rod length, and pitch / yaw angle of the swing mechanism; the UAV rigid body dynamics, the kinematics of the telescopic rod and swing mechanism, the virtual spring, and the damped contact dynamics are coupled to obtain a unified high-order nonlinear ODE.

5. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 4, characterized in that: The contact force is described by a virtual spring-damping model controlled by admittance, and the contact reaction force acts as a coupling term in the UAV dynamic equations; the telescopic rod and swing mechanism servo system are modeled as a second-order system.

6. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: In step three, PINN adopts an encoder-decoder multi-head structure, sharing the MLP encoder to extract spatiotemporal features, and the decoder to synchronously output the UAV system status, control commands, and contact force; The total loss function includes data loss, PDE shape constraint loss, ODE dynamic constraint loss, and regularization loss.

7. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 6, characterized in that: PDE shape constraint loss forces the prediction that the contact point is located on the lightning arrester surface characterized by PDE; ODE dynamic constraint loss ensures that the state evolution conforms to the coupled dynamic equations through automatic differentiation. Regularization loss minimizes the rate and magnitude of change in control commands, improving motion smoothness and safety.

8. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: Before testing, the binocular camera, PMD sensor, force sensor and on-resistance meter are calibrated; the drone hovers 1-2 meters outside the lightning arrester, the binocular and PMD data are fused to reconstruct the local surface, the PINN predictive control sequence is used, the telescopic rod is extended in advance and the angle between the probe and the lightning arrester normal is less than 3°, and the contact is maintained with a constant force of 3-5N to complete the test.

9. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: For mixed strip and point lightning arresters, an adaptive gated diffusion tensor is adopted, and the lightning arrester type is distinguished by semantic segmentation. The PDE model adaptively matches the diffusion parameters of strip / point lightning arresters, and PINN adaptively switches between flat / ball probes and contact modes.

10. The intelligent detection method for wind turbine blade lightning arresters based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that: During the testing process, the contact force, contact position, and conduction resistance data are transmitted back in real time. Secondary compensation is provided for unsuccessful contact points to complete the full-point testing.

Images