A method and system for detecting the integrity of a lightning receptor based on three-dimensional positioning of a lightning strike point

Abstract

The application discloses a lightning receptor integrity detection method and system based on a lightning receptor three-dimensional positioning, and relates to the technical field of lightning receptor detection.The method comprises the following steps: a target detection algorithm is used to positionally identify and three-dimensionally map a surface image of a fan blade; a two-dimensional pixel coordinate is used to preliminarily detect the integrity of a lightning receptor image area, and a preliminary corrosion judgment result and a first polishing parameter of the lightning receptor are generated; a corrosion thickness inversion model is used to inversely calculate the corrosion degree of a first resistance detection result, the spatial distribution of the corrosion thickness of the lightning receptor surface is obtained, and a second polishing parameter of the lightning receptor is generated; a second resistance detection result is obtained, and the corrosion degree inversion and polishing are repeatedly executed until the repair qualified condition is met.The application realizes quantitative mapping from an electrical response to a physical corrosion form, and avoids problems such as repair blindness, excessive polishing or insufficient repair during polishing.

Description

Technical Field

[0001] This invention relates to the field of lightning arrester testing technology, and more specifically, to a method and system for detecting the integrity of lightning arresters based on three-dimensional positioning of the lightning arrest point. Background Technology

[0002] With the increasing installed capacity of wind turbines year by year, the operation and maintenance problems exposed during long-term operation of wind turbines are becoming increasingly prominent. Especially in areas with frequent lightning strikes, wind turbine blades that are over 100 meters high are extremely vulnerable to lightning strikes. Once the lightning protection system fails, it can cause minor damage such as ablation and cracking of composite blades, or even electrical short circuits or fires of the entire turbine, resulting in economic losses and safety risks. Therefore, ensuring that the lightning protection system of wind turbines, especially the lightning arresters located at critical blade tips, is in a state of high reliability over the long term has become one of the core tasks of safe operation and maintenance of wind power. Traditional manual inspections or detection methods based on simple resistance tests are difficult to accurately locate the spatial position of the lightning arresters, leading to delayed maintenance or insufficient repairs. Against this background, a lightning arrester integrity detection method based on three-dimensional positioning of the lightning strike point is crucial.

[0003] However, while existing methods for detecting the integrity of lightning arresters can achieve precise spatial location, they still have significant shortcomings in assessing corrosion status. They primarily rely on initial image analysis for judgment, lacking the ability to finely characterize the corrosion status of localized micro-areas of the lightning arrester. Resistance testing typically only provides an overall conduction resistance value, failing to reflect the spatial distribution characteristics of corrosion on the lightning arrester surface, resulting in a lack of basis for setting subsequent grinding parameters. This black-box electrical feedback makes the repair process highly dependent on experience, easily leading to inaccurate grinding areas, unreasonable depth control, and problems such as over-grinding damaging the substrate or under-grinding leaving residual corrosion. This not only affects repair efficiency and quality but may also weaken the structural integrity and lightning protection performance of the lightning arrester, failing to meet the high reliability and long lifespan maintenance requirements of wind power equipment.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a method and system for detecting the integrity of lightning arresters based on three-dimensional positioning of the lightning arrest point, in order to overcome the aforementioned technical problems existing in the prior art.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point is provided, the method comprising the following steps: S1. Acquire the surface image of the wind turbine blade, use the target detection algorithm to locate and identify the surface image of the wind turbine blade, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial position information of the lightning arrester. S2. Perform preliminary integrity detection on the lightning arrester image area in the wind turbine blade surface image based on the two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Generate the initial polishing parameters of the lightning arrester based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning arrester. S3. Obtain the first resistance test result of the lightning arrester after the initial polishing. Use the corrosion thickness inversion model to invert the corrosion degree of the first resistance test result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Generate the secondary polishing parameters of the lightning arrester based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester. S4. Obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

[0007] Furthermore, target detection algorithms are used to locate and identify the surface image of the wind turbine blades, obtaining the two-dimensional pixel coordinates of the lightning arrester. Then, three-dimensional point cloud mapping is performed on these two-dimensional pixel coordinates to obtain the three-dimensional spatial location information of the lightning arrester, including: The surface image of the wind turbine blade is input into a pre-trained target detection algorithm, and the tensor inference acceleration engine is used to accelerate the inference of the surface image of the wind turbine blade to obtain the two-dimensional pixel coordinates of the lightning arrester. The two-dimensional pixel coordinates are matched with the same-view depth map at the pixel level. The depth value corresponding to each pixel is extracted based on the matching result. The two-dimensional pixel points and depth values ​​are back-projected to the camera coordinate system in combination with the camera intrinsic parameter matrix to generate the three-dimensional point cloud of the flash receiver in the camera coordinate system. Outlier filtering and voxel downsampling are performed on the 3D point cloud to obtain the 3D spatial position information of the lightning arrester. Outlier filtering and centering are then performed on the 3D point cloud to obtain a normalized point cloud. The geometric center of the normalized point cloud is calculated to obtain the 3D spatial position information of the lightning arrester in the camera coordinate system.

[0008] Furthermore, the image of the wind turbine blade surface is input into a pre-trained target detection algorithm, and a tensor inference acceleration engine is used to accelerate inference on the wind turbine blade surface image to obtain the two-dimensional pixel coordinates of the lightning arrester, including: A labeled dataset is constructed based on pre-acquired real-world images of wind turbine blade lightning arresters. The labeled dataset is then input into the framework of the object detection algorithm for transfer learning training. The weights of the object detection algorithm are iteratively optimized through the loss function to obtain the pre-trained object detection algorithm. A pre-trained target detection algorithm is used to perform lightning arrester target detection and feature extraction on the surface image of the wind turbine blade to obtain a lightning arrester feature map. The lightning arrester feature map is then input into a tensor inference acceleration engine for model inference and position regression calculation to obtain confidence and coordinate information. Based on the confidence and coordinate information, non-maximum suppression and coordinate decoding are performed to obtain the lightning arrester target detection box. The center position and boundary coordinates of the lightning arrester are extracted based on the target detection box of the lightning arrester, and the two-dimensional pixel coordinates of the lightning arrester are obtained after coordinate normalization and correction calculation.

[0009] Furthermore, based on the two-dimensional pixel coordinates, a preliminary integrity check is performed on the lightning arrester image area in the wind turbine blade surface image, and preliminary corrosion judgment results are generated, including: The lightning arrester image region is cropped and extracted from the wind turbine blade surface image based on the two-dimensional pixel coordinates, and the color space of the lightning arrester image region is converted to obtain the converted lightning arrester image block. A color and brightness comparison analysis was performed on the converted lightning arrester image block to obtain the color deviation and dark area ratio of the lightning arrester surface. The color deviation is compared with the preset color deviation threshold to obtain the color anomaly judgment result, and the dark area ratio is compared with the preset dark area ratio threshold to obtain the rust defect judgment result. Based on the color anomaly judgment result and the rust defect judgment result, the corrosion degree level is determined to obtain the preliminary corrosion judgment result.

[0010] Furthermore, a color and brightness comparison analysis was performed on the converted lightning arrester image blocks to obtain the color deviation and dark area ratio of the lightning arrester surface, including: The pixel value distribution of the saturation and luminance channels is extracted from the converted lightning arrester image block to obtain the color feature vector of the lightning arrester region; The color feature vector is compared with the preset standard color feature template of the lightning arrester by Euclidean distance to obtain the color deviation of the lightning arrester surface. The proportion of dark areas on the surface of the lightning arrester is obtained by statistically analyzing the percentage of pixels with brightness below a preset brightness threshold within the image block of the lightning arrester.

[0011] Furthermore, the corrosion degree is inverted using the corrosion thickness inversion model on the first resistance detection results, yielding the spatial distribution of corrosion thickness on the lightning arrester surface, including: The first resistance detection result is normalized to obtain standardized resistivity data. Spatial interpolation is then used to fill in the blank areas between measurement points in the standardized resistivity data to generate a two-dimensional resistivity pseudo-image. The two-dimensional resistivity pseudo-image is input into the corrosion thickness inversion model to perform resistivity to corrosion degree mapping inversion, and the optimized corrosion thickness distribution map is obtained. Median filtering was used to denoise the optimized corrosion thickness distribution map, resulting in a denoised corrosion thickness distribution map. The scale of the denoised corrosion thickness distribution map was then restored according to the scale parameter during normalization to obtain the spatial distribution of corrosion thickness on the lightning arrester surface.

[0012] Furthermore, the two-dimensional resistivity pseudo-image is input into the corrosion thickness inversion model to perform resistivity-to-corrosion degree mapping inversion, resulting in an optimized corrosion thickness distribution map, including: The two-dimensional resistivity pseudo-image is input into a total variation constrained generative adversarial network for multi-scale spatial feature extraction and fusion to obtain a preliminary corrosion thickness distribution map. The preliminary corrosion thickness distribution map and the preset real corrosion thickness label are input into the discriminator for adversarial discrimination. The network weights of the generator are updated by backpropagation based on the discrimination results, and a corrosion thickness inversion model is generated by combining the total variation regularization term constraint. The weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model are adaptively optimized to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the two-dimensional resistivity pseudo-image, resulting in the optimized corrosion thickness distribution map.

[0013] Furthermore, the two-dimensional resistivity pseudo-image is input into a total variation constrained generative adversarial network for multi-scale spatial feature extraction and fusion, resulting in a preliminary corrosion thickness distribution map, including: The two-dimensional resistivity pseudo-image is input into the generator of the total variation constrained generative adversarial network, and the two-dimensional resistivity pseudo-image is subjected to multi-scale spatial feature extraction from local to global through multi-level convolutional layers and downsampling operations in the encoder to obtain the multi-scale feature map corresponding to the resistivity distribution. By using skip connections, multi-scale feature maps are passed layer by layer to the corresponding levels of the decoder. In the decoder, deconvolution and upsampling operations are used to gradually restore the spatial resolution. The multi-scale features of the same level of the encoder are then fused to generate an intermediate image of the fused erosion thickness. The fused intermediate corrosion thickness map is mapped to a single-channel output using a convolutional layer to obtain a preliminary corrosion thickness distribution map.

[0014] Furthermore, the weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model are adaptively optimized to obtain the optimal hyperparameter combination. This optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the preliminary corrosion thickness distribution map, resulting in the optimized corrosion thickness distribution map, including: The set of loss weights in the generator loss function and the weights of the total variation regularization term are used to construct the hyperparameter combination to be optimized. The tent chaos mapping is used to initialize the population of the improved Osprey optimization algorithm in the preset hyperparameter space of the hyperparameter combination to be optimized, and an initial hyperparameter population set containing multiple sets of hyperparameter combinations is generated. Each set of hyperparameters in the initial hyperparameter population is substituted into the corrosion thickness inversion model, and the two-dimensional resistivity pseudo-image is used as the input of the corrosion thickness inversion model to perform resistivity to corrosion thickness mapping inversion, so as to obtain the initial corrosion thickness distribution map. A fitness function is generated based on the structural similarity between the initial corrosion thickness distribution map and the preset real corrosion thickness label. The optimal hyperparameter combination is obtained by iteratively updating the hyperparameter combination in the initial hyperparameter population set through a combination of the Lévy strategy and the spiral curve strategy. The optimal combination of hyperparameters is substituted into the corrosion thickness inversion model, and the forward propagation calculation of the generator is performed using a two-dimensional resistivity pseudo-image as the initial input of the corrosion thickness inversion model to obtain the optimized corrosion thickness distribution map.

[0015] According to another aspect of the present invention, a lightning arrester integrity detection system based on three-dimensional positioning of the lightning arrest point is provided, the system comprising: The spatial location determination module is used to acquire images of the wind turbine blade surface, use a target detection algorithm to locate and identify the wind turbine blade surface image, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial location information of the lightning arrester. The corrosion preliminary judgment module is used to perform preliminary integrity detection on the lightning rod image area in the wind turbine blade surface image based on two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning rod, the initial polishing parameters of the lightning rod are generated. The secondary corrosion degree judgment module is used to obtain the first resistance detection result of the lightning arrester conduction test after the initial polishing, and to use the corrosion thickness inversion model to invert the corrosion degree of the first resistance detection result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester, the secondary polishing parameters of the lightning arrester are generated. The repair qualification judgment module is used to obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

[0016] The beneficial effects of this invention are as follows: 1. This invention achieves a precise inversion of the corrosion state of local micro-areas of lightning arresters and a quantitative mapping from electrical response to physical corrosion morphology through corrosion degree inversion. This avoids the problems of traditional methods that can only provide a general resistance value and cannot reveal the spatial distribution of corrosion. It also avoids problems such as blind repair, over-polishing, or insufficient repair during polishing. It completely changes the rough mode of traditional one-size-fits-all or trial-and-error polishing, greatly improves the repair quality and service life of lightning arresters, and ensures the long-term reliability of lightning protection systems.

[0017] 2. This invention achieves rapid initial screening of corrosion status and intelligent generation of initial polishing parameters through a preliminary integrity intelligent discrimination mechanism based on color and brightness characteristics. By making full use of the statistical characteristics of saturation and brightness channels, it can effectively distinguish different states such as intact, lightly corroded, and moderately corroded, generate preliminary corrosion judgment results, and generate initial polishing parameters adapted to the current corrosion level. This avoids material waste and structural damage caused by indiscriminate polishing, and improves the targeting and economy of repair.

[0018] 3. This invention introduces a corrosion thickness inversion model based on resistance detection data, which can not only learn the complex nonlinear mapping relationship between resistivity and corrosion thickness, but also maintain a clear corrosion boundary and suppress noise through regularization terms, outputting a high-fidelity corrosion thickness spatial distribution map. This model can truly reflect the remaining metal thickness or corrosion depth in each area of ​​the lightning arrester surface, and can generate highly personalized secondary polishing parameters, thereby achieving precise repair based on physical inversion. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a lightning arrester integrity detection method based on three-dimensional positioning of the lightning arrest point according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a lightning arrester integrity detection system based on three-dimensional positioning of the lightning arrest point according to an embodiment of the present invention. Figure 3 This is a YOLO model inference result diagram in a lightning arrester integrity detection method based on three-dimensional positioning of the lightning arrest point according to an embodiment of the present invention; Figure 4 This is a depth information map in a lightning arrester integrity detection method based on three-dimensional positioning of the lightning arrest point according to an embodiment of the present invention; Figure 5 This is a schematic diagram of three-dimensional coordinates in a lightning arrester integrity detection method based on three-dimensional positioning of the lightning arrest point according to an embodiment of the present invention; Figure 6 This is a diagram showing the result of the normal vector of the lightning point plane in a lightning arrester integrity detection method based on three-dimensional positioning of the lightning point according to an embodiment of the present invention.

[0021] In the picture: 1. Spatial location determination module; 2. Preliminary corrosion assessment module; 3. Secondary corrosion degree assessment module; 4. Repair qualification assessment module. Detailed Implementation

[0022] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0023] According to an embodiment of the present invention, a method and system for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning arrest point is provided.

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning arrest point is provided. The method includes the following steps: S1. Acquire the surface image of the wind turbine blade, use the target detection algorithm to locate and identify the surface image of the wind turbine blade, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial position information of the lightning arrester.

[0025] Specifically, target detection algorithms are used to locate and identify the surface image of the wind turbine blades, obtaining the two-dimensional pixel coordinates of the lightning arrester. Then, three-dimensional point cloud mapping is performed on these two-dimensional pixel coordinates to obtain the three-dimensional spatial location information of the lightning arrester, including: The surface image of the wind turbine blade is input into a pre-trained target detection algorithm, and the tensor inference acceleration engine is used to accelerate the inference of the surface image of the wind turbine blade to obtain the two-dimensional pixel coordinates of the lightning arrester.

[0026] Specifically, the image of the wind turbine blade surface is input into a pre-trained target detection algorithm, and a tensor inference acceleration engine is used to accelerate inference on the wind turbine blade surface image to obtain the two-dimensional pixel coordinates of the lightning arrester, including: A labeled dataset is constructed based on pre-acquired real-world images of wind turbine blade lightning arresters. The labeled dataset is then input into the framework of the object detection algorithm for transfer learning training. The weights of the object detection algorithm are iteratively optimized through the loss function to obtain the pre-trained object detection algorithm. A pre-trained target detection algorithm is used to perform lightning arrester target detection and feature extraction on the surface image of the wind turbine blade to obtain a lightning arrester feature map. The lightning arrester feature map is then input into a tensor inference acceleration engine for model inference and position regression calculation to obtain confidence and coordinate information. Based on the confidence and coordinate information, non-maximum suppression and coordinate decoding are performed to obtain the lightning arrester target detection box. The center position and boundary coordinates of the lightning arrester are extracted based on the target detection box of the lightning arrester, and the two-dimensional pixel coordinates of the lightning arrester are obtained after coordinate normalization and correction calculation.

[0027] The two-dimensional pixel coordinates are matched with the depth map at the same viewpoint at the pixel level. The depth value corresponding to each pixel is extracted based on the matching result. The two-dimensional pixel points and depth values ​​are back-projected to the camera coordinate system in combination with the camera intrinsic parameter matrix to generate the three-dimensional point cloud of the flash receiver in the camera coordinate system.

[0028] Outlier filtering and voxel downsampling are performed on the 3D point cloud to obtain the 3D spatial position information of the lightning arrester. Outlier filtering and centering are then performed on the 3D point cloud to obtain a normalized point cloud. The geometric center of the normalized point cloud is calculated to obtain the 3D spatial position information of the lightning arrester in the camera coordinate system.

[0029] Specifically, this invention is mainly applied to an automated detection and repair process for flashover points on wind turbine blades based on the collaboration of a drone and a robotic arm. The specific process is as follows: Confirm that the blade to be tested is leveled, and shut down the wind turbine and lock it in a locked state; confirm that the drone / mounted equipment is in normal condition, and start the drone for takeoff; use a gimbal camera to take a downward-facing image of the blade to estimate its height, and control the drone to hover approximately 3 meters above the target flashover point; then release the mounted equipment, and use the drone for fine-tuning to ensure a precise landing near the flashover point; activate the air pump to attach the suction cup to the blade surface, completing the mounting equipment fixation; the robotic arm then searches along the flashover point's location. The system simultaneously estimates the orientation of the flash point in the robotic arm's coordinate system using a gimbal camera; a 3D camera detects the flash point in real time and estimates its precise spatial position; the robotic arm calculates its trajectory based on this and controls its end effector to precisely contact the flash point; a grinding motor is activated to grind the flash point; after grinding, the end effector is switched to a probe, which performs resistance circuit measurement on the flash point, followed by corrosion inversion and grinding until the repair is satisfactory; after the operator confirms that the resistance measurement results are correct, the robotic arm resets, the suction cup is released, and the descent system retracts the mounted equipment; finally, it moves to the next measurement point until all measurements are completed, and the drone returns to base.

[0030] Specifically, the automatic lightning arrester identification based on a depth camera utilizes a structured light 3D camera mounted at the end of a robotic arm as the main sensor in practical applications. The structured light camera projects specific light patterns and combines this with a camera to capture the light pattern deformation on the object's surface, achieving 3D scene perception and providing high-resolution color RGB images. Using the RGB images acquired by the camera as input, a visual detection task is performed on the target lightning arrester point to achieve subsequent depth acquisition and 3D localization.

[0031] In flash point detection, this invention employs a YOLO-based deep learning target detection algorithm. The YOLO algorithm boasts high real-time performance and excellent performance in detecting small targets. Through fine-tuning training on a flash point dataset, it can stably identify targets under complex lighting conditions. To meet the high frame rate real-time requirements of the robotic arm's working environment, this invention uses TensorRT (Tensor Inference Acceleration Engine) to accelerate YOLO inference during the deployment phase, achieving a frame rate of 50 FPS. The inference results based on the YOLO model are as follows: Figure 3 As shown, the black box represents the model output.

[0032] This invention utilizes a 3D camera-based structured light system to acquire depth maps and 3D information. After completing the 2D detection of the flash point, the invention employs a structured light 3D camera mounted at the end of a robotic arm to acquire depth information, enabling 3D localization of the flash point. The structured light camera projects specific light patterns and captures the deformation on the target surface. An internal algorithm decodes the light patterns and performs triangulation, directly outputting a high-resolution depth map. The depth value of each pixel corresponds to the projection distance of that pixel onto the camera's optical axis, forming complete scene depth information. By mapping the pixel coordinates of the detected flash points on the RGB image to the depth map, the depth value of the target point can be extracted, achieving rapid conversion from 2D detection to 3D space. The depth image is shown below. Figure 4 As shown, red indicates proximity to the camera, and blue indicates distance from the camera.

[0033] In this invention, the algorithm back-projects the extracted flash point depth and its corresponding pixel coordinates together into a local point cloud, generating a 3D point in the camera coordinate system of the robotic arm's end effector. The back-projection process utilizes camera intrinsic parameters, namely focal length and principal point coordinates, for precise calculation to ensure the smooth conversion of depth values ​​into 3D coordinates. To improve reliability, outliers or invalid pixels in the depth map are processed using neighborhood mean filtering or interpolation filling methods, thereby reducing the impact of noise on 3D positioning. Furthermore, through statistical error propagation, the depth accuracy of the docking point can be estimated, providing an uncertainty reference for subsequent plane fitting and attitude calculation. A schematic diagram of the 3D coordinates is shown below. Figure 5 As shown.

[0034] Based on 3D information, the planar attitude of the lightning interception point is obtained. After calculating the 3D coordinates of the lightning interception point, this invention further utilizes the 3D point cloud information of the lightning interception point and its surroundings to perform local plane fitting on the target surface to obtain the normal vector of the lightning interception point, thereby controlling the robotic arm to better polish the lightning interception point. A local area is selected at the lightning interception point to form a local point cloud. In the local point cloud, this invention uses the least squares plane fitting method to calculate the plane normal vector. Specifically, eigenvalue decomposition is performed using the covariance matrix of the local point cloud, and the eigenvector corresponding to the smallest eigenvalue is the normal vector of the fitted plane. Simultaneously, the orientation of the flash point relative to the end effector of the robotic arm can be determined by the angle between the normal vector direction and the camera optical axis direction. To reduce the influence of noise, outlier removal is used to filter the local point cloud before calculation, thereby improving the stability of the normal vector estimation. The fitting plane and normal vector information are transformed from the camera coordinate system to the rigid body coordinate system of the robotic arm end effector, forming a three-dimensional spatial attitude description of the local surface of the flash point. The calculation results of the flash point plane normal vector are as follows. Figure 6 As shown, the red arrows represent the two-dimensional direction of the normal vector.

[0035] Specifically, in a scenario of collaborative operation between a drone and a robotic arm, the drone, equipped with a structured light 3D camera, is first controlled to hover approximately 3 meters above the lightning strike point on the wind turbine blade. The structured light 3D camera at the end of the robotic arm then acquires a 1920×1080 resolution RGB image of the wind turbine blade and its corresponding depth map. The camera projects encoded light patterns and decodes deformation features, directly outputting depth data including the projection distance of each pixel. The acquired RGB image is then input into a YOLO object detection model trained through transfer learning. This model constructs a labeled dataset using massive amounts of real-world images of wind turbine blade lightning strike points, iteratively optimizes network weights through a loss function, and after training, uses the TensorRT tensor inference acceleration engine to achieve high-speed inference of 50 FPS. After feature extraction, position regression, and confidence calculation, non-maximum suppression and coordinate decoding are performed on detection results with confidence scores higher than a threshold, resulting in the detected lightning strike point object. The system extracts the center coordinates of the frame and performs normalization and correction calculations to obtain the precise two-dimensional pixel coordinates of the lightning arrester. These coordinates are then matched pixel-by-pixel with the depth map from the same viewpoint to extract the corresponding depth values. Combined with the focal length and principal point coordinate parameters in the camera intrinsic matrix, back-projection calculations convert the two-dimensional pixels and depth values ​​into three-dimensional spatial points in the camera coordinate system, generating the original three-dimensional point cloud of the lightning arrester. This point cloud is then subjected to statistical filtering to remove outliers, voxel downsampling to simplify redundant data, and neighborhood mean filtering and interpolation to fill depth anomalies. After centering and normalization, the geometric center of the normalized point cloud is calculated. Simultaneously, the lightning arrester's normal vector is obtained based on least-squares plane fitting and covariance matrix eigenvalue decomposition of the local point cloud. This provides complete information on the three-dimensional spatial position and attitude of the lightning arrester in the camera coordinate system, offering accurate data support for calculating the robotic arm's grinding trajectory.

[0036] Specifically, the YOLO object detection model used in this invention is a lightweight YOLOv8-n structure, consisting of an input terminal, a C2f backbone feature extraction network, a PANet multi-scale fusion neck network, and a detection head output terminal. Each layer is connected sequentially through convolution, batch normalization, and SiLU activation function. The C2f module uses a dual-branch residual structure to enhance feature extraction capability, PANet achieves high- and low-level feature fusion, and the detection head uses a decoupled head structure to output class confidence and bounding box coordinates separately. The training data comes from different illuminations, angles, and distances collected by the camera at the end of the drone and robotic arm. The real-world images of the wind turbine blade lightning arrester were divided into training and testing sets at a 9:1 ratio. After size normalization, Mosaic data augmentation, and grayscale transformation preprocessing, the images were input into the model. A joint loss function was constructed using CIoU bounding box loss, cross-entropy classification loss, and DFL distribution focal loss. The AdamW optimizer was used for gradient descent updates. In practical use, the optimal training hyperparameters were: input image resolution 640×640, initial learning rate 0.001, learning rate decay coefficient 0.95, batch size 16, number of iterations 200, confidence threshold 0.25, NMS IOU threshold 0.45, and the learning rate was dynamically adjusted using a cosine annealing strategy.

[0037] S2. Perform preliminary integrity detection on the lightning arrester image area in the wind turbine blade surface image based on the two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Generate the initial polishing parameters of the lightning arrester based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning arrester.

[0038] Specifically, based on two-dimensional pixel coordinates, a preliminary integrity check is performed on the lightning arrester image area in the wind turbine blade surface image, and preliminary corrosion judgment results are generated, including: The lightning arrester image region is cropped and extracted from the wind turbine blade surface image based on the two-dimensional pixel coordinates, and the color space of the lightning arrester image region is converted to obtain the converted lightning arrester image block. Color and brightness comparison analysis was performed on the converted lightning arrester image block to obtain the color deviation and dark area ratio of the lightning arrester surface.

[0039] Specifically, a color and brightness comparison analysis was performed on the converted lightning arrester image blocks to obtain the color deviation and dark area ratio of the lightning arrester surface, including: The pixel value distribution of the saturation and luminance channels is extracted from the converted lightning arrester image block to obtain the color feature vector of the lightning arrester region; The color feature vector is compared with the preset standard color feature template of the lightning arrester by Euclidean distance to obtain the color deviation of the lightning arrester surface. The proportion of dark areas on the surface of the lightning arrester is obtained by statistically analyzing the percentage of pixels with brightness below a preset brightness threshold within the image block of the lightning arrester.

[0040] The color deviation is compared with the preset color deviation threshold to obtain the color anomaly judgment result, and the dark area ratio is compared with the preset dark area ratio threshold to obtain the rust defect judgment result. Based on the color anomaly judgment result and the rust defect judgment result, the corrosion degree level is determined to obtain the preliminary corrosion judgment result.

[0041] It should be noted that when the color deviation exceeds the first threshold or the proportion of dark areas exceeds the second threshold, corrosion is identified and a corrosion level is generated. When the color deviation is less than the first threshold and the proportion of dark areas is less than the second threshold, the condition is identified as intact.

[0042] Specifically, based on the two-dimensional pixel coordinates of the lightning arrester, a 160×160 pixel local image block of the lightning arrester is extracted from the RGB image of the wind turbine blades, with the coordinate center as the origin. This block is then converted from the RGB color space to the HSV color space. The saturation (S) channel and the brightness (V) channel are separated, and the pixel mean and variance within the range of 0~255 for each channel are extracted to form a 16-dimensional color feature vector. The Euclidean distance is calculated between this vector and the pre-calibrated standard feature vector of a healthy lightning arrester. This distance value is used as the color deviation, preferably with a deviation threshold of 35. Values ​​exceeding this threshold are considered color abnormalities. Simultaneously, the pixels in the V channel are statistically analyzed. Pixels with a brightness value below 45 are identified as rusted dark areas, and their proportion is calculated. Preferably, the threshold for the proportion of dark areas is set to 18%. If the proportion exceeds this, it is identified as rust damage. Combining the two judgment results, the surface of the lightning arrester is divided into four levels: no corrosion, light corrosion, moderate corrosion, and severe corrosion, to obtain the preliminary corrosion judgment result. Then, combined with the obtained three-dimensional spatial position and plane normal vector information, the initial grinding parameters including grinding depth, grinding radius, grinding posture, and feed speed are generated. Among them, light corrosion corresponds to a grinding depth of 0.2mm, moderate corrosion to 0.5mm, and severe corrosion to 0.8mm, providing a quantitative control basis for the robotic arm to perform the initial grinding.

[0043] It should be noted that the corrosion level is determined based on the results of the color anomaly assessment and the rust defect assessment, and the rules for determining the corrosion level are shown in Table 1.

[0044] Table 1 Correspondence between Corrosion Degree Levels It should be noted that the initial grinding parameters are generated based on the preliminary corrosion assessment results and the three-dimensional spatial location information of the lightning arrester. The rules for determining the initial grinding parameters are shown in Table 2.

[0045] Table 2. Initial Polishing Parameters Table S3. Obtain the first resistance test result of the lightning arrester after the initial polishing. Use the corrosion thickness inversion model to invert the corrosion degree of the first resistance test result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Generate the secondary polishing parameters of the lightning arrester based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester.

[0046] Specifically, the corrosion degree is inverted using the corrosion thickness inversion model on the first resistance detection results, and the spatial distribution of corrosion thickness on the lightning arrester surface is obtained, including: The first resistance detection result is normalized to obtain standardized resistivity data. Spatial interpolation is then used to fill in the blank areas between measurement points in the standardized resistivity data to generate a two-dimensional resistivity pseudo-image. The two-dimensional resistivity pseudo-image is input into the corrosion thickness inversion model to perform resistivity-to-corrosion degree mapping inversion, resulting in an optimized corrosion thickness distribution map.

[0047] Specifically, the two-dimensional resistivity pseudo-image is input into the corrosion thickness inversion model to perform resistivity-to-corrosion degree mapping inversion, resulting in an optimized corrosion thickness distribution map, including: The two-dimensional resistivity pseudo-image is input into a total variation constrained generative adversarial network for multi-scale spatial feature extraction and fusion to obtain a preliminary corrosion thickness distribution map.

[0048] Specifically, the two-dimensional resistivity pseudo-image is input into a total variation constrained generative adversarial network for multi-scale spatial feature extraction and fusion, resulting in a preliminary corrosion thickness distribution map, including: The two-dimensional resistivity pseudo-image is input into the generator of the total variation constrained generative adversarial network, and the two-dimensional resistivity pseudo-image is subjected to multi-scale spatial feature extraction from local to global through multi-level convolutional layers and downsampling operations in the encoder to obtain the multi-scale feature map corresponding to the resistivity distribution. By using skip connections, multi-scale feature maps are passed layer by layer to the corresponding levels of the decoder. In the decoder, deconvolution and upsampling operations are used to gradually restore the spatial resolution. The multi-scale features of the same level of the encoder are then fused to generate an intermediate image of the fused erosion thickness. The fused intermediate corrosion thickness map is mapped to a single-channel output using a convolutional layer to obtain a preliminary corrosion thickness distribution map.

[0049] The preliminary corrosion thickness distribution map and the preset real corrosion thickness label are input into the discriminator for adversarial discrimination. The network weights of the generator are updated by backpropagation based on the discrimination results, and a corrosion thickness inversion model is generated by combining the total variation regularization term constraint.

[0050] The weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model are adaptively optimized to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the two-dimensional resistivity pseudo-image, resulting in the optimized corrosion thickness distribution map.

[0051] Specifically, the weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model are adaptively optimized to obtain the optimal hyperparameter combination. This optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the preliminary corrosion thickness distribution map, resulting in the optimized corrosion thickness distribution map, including: The set of loss weights in the generator loss function and the weights of the total variation regularization term are used to construct the hyperparameter combination to be optimized. The tent chaos mapping is used to initialize the population of the improved Osprey optimization algorithm in the preset hyperparameter space of the hyperparameter combination to be optimized, and an initial hyperparameter population set containing multiple sets of hyperparameter combinations is generated. Each set of hyperparameters in the initial hyperparameter population is substituted into the corrosion thickness inversion model, and the two-dimensional resistivity pseudo-image is used as the input of the corrosion thickness inversion model to perform resistivity to corrosion thickness mapping inversion, so as to obtain the initial corrosion thickness distribution map. A fitness function is generated based on the structural similarity between the initial corrosion thickness distribution map and the preset real corrosion thickness label. The optimal hyperparameter combination is obtained by iteratively updating the hyperparameter combination in the initial hyperparameter population set through a combination of the Lévy strategy and the spiral curve strategy. The optimal combination of hyperparameters is substituted into the corrosion thickness inversion model, and the forward propagation calculation of the generator is performed using a two-dimensional resistivity pseudo-image as the initial input of the corrosion thickness inversion model to obtain the optimized corrosion thickness distribution map.

[0052] Median filtering was used to denoise the optimized corrosion thickness distribution map, resulting in a denoised corrosion thickness distribution map. The scale of the denoised corrosion thickness distribution map was then restored according to the scale parameter during normalization to obtain the spatial distribution of corrosion thickness on the lightning arrester surface.

[0053] Specifically, a four-probe resistance measurement module at the end of the robotic arm is used to set up 64 measurement points (8×8) at 1mm intervals within the measurement area of ​​the lightning arrester after initial polishing. The conduction resistance value of each measurement point is collected sequentially to obtain the first resistance detection result. The measured resistance range is between 0.1Ω and 3.5Ω. This set of resistance data is mapped to the 0-1 interval using the min-max normalization method to obtain standardized resistivity data. The Kriging space interpolation algorithm is used to interpolate and fill the blank areas between the 64 discrete measurement points, generating a two-dimensional resistivity pseudo-image with a resolution of 128×128 pixels. This pseudo-image is then input into a fully variable constrained generative adversarial network generator using a U-Net architecture. In this process, multi-scale depth features of resistivity distribution are extracted through four levels of continuous convolution and 2x downsampling operations in the encoder. Then, the features at each level are passed to the corresponding level of the decoder via skip connections. The feature map resolution is gradually restored by combining deconvolution and upsampling and fusing high and low level features. The initial corrosion thickness distribution map is obtained by mapping it to a single channel through a 1×1 convolutional layer. This distribution map and the preset 0-1.5mm real corrosion thickness labels are input into the PatchGAN discriminator for adversarial discrimination. The total loss function is constructed by weighting and summing the adversarial loss, L1 loss, and histogram matching loss and introducing a total variation regularization term. The generator parameters are updated through backpropagation to obtain the initial corrosion thickness inversion model.

[0054] The total variation regularization weight, adversarial loss weight, L1 loss weight, and histogram matching loss weight are used as the hyperparameter combination to be optimized. A population of 30 individuals using the improved Osprey optimization algorithm is initialized within a predefined search space using the Tent chaotic map. The structural similarity (SSIM) between the inversion results and the true labels is used as the fitness function. After 50 iterations, combining the Levy flight strategy and the spiral search strategy, the optimal combination of hyperparameters is obtained. This optimal combination is then substituted into the inversion model, and forward inference is performed again using a two-dimensional resistivity pseudo-image as input to obtain the optimized result. The corrosion thickness distribution map was obtained. After noise removal using a 3×3 median filter, the actual physical thickness was restored based on the normalized scale parameters. This yielded the spatial distribution data of the corrosion thickness continuously distributed from 0 to 1.5 mm on the lightning arrester surface. Combining the three-dimensional spatial coordinates of the lightning arrester and the surface normal vector, secondary grinding parameters were divided according to the thickness range: 0.2-0.5 mm for fine grinding layer, 0.5-1.0 mm for conventional grinding layer, and greater than 1.0 mm for deep grinding layer. The feed amount for each layer was set to 0.1 mm, 0.2 mm, and 0.3 mm, respectively, forming a quantitative grinding command that can be directly driven by the robotic arm.

[0055] Specifically, the Total Variation Constrained Generative Adversarial Network (TV-CGAN) used in this invention consists of a U-Net architecture generator and a PatchGAN discriminator. The generator encoder contains four convolutional layers, each consisting of a 3×3 convolution, batch normalization, and a GELU activation function, downsampled using 2×2 max pooling. The decoder uses four deconvolutional layers with 2×2 upsampling, fusing features of the same scale from the encoder layer by layer through skip connections, and mapping them to a single-channel erosion thickness map via a 1×1 convolution. The discriminator is a PatchGAN structure consisting of four convolutional layers, using the LeakyReLU activation function. The system outputs a local authenticity discrimination map. Training data is taken from lightning arrester corrosion samples calibrated in the laboratory and measured in the field. Each sample group includes resistivity distribution data and corresponding 0-1.5mm actual corrosion thickness labels. After normalization to the 0-1 interval, random horizontal flipping, and scaling by 0.9-1.1 times, the data is input into a total variation constraint-based generative adversarial network. The total loss function consists of a weighted sum of adversarial loss, L1 loss, histogram matching loss, and a total variation regularization term. The Adam optimizer is used for gradient updates. The optimal training hyperparameters are: input image size 128×128, batch size 16, and initial learning rate 2×10⁻⁶. -4 The algorithm underwent 100 iterations. The total variation regularization term weight search range was 0.001-0.1, and the loss weight search range was 0.1-1.0. The algorithm improved the Osprey optimization algorithm by initializing with Tent chaos and combined Levy flight with spiral search to complete 50 iterations of optimization.

[0056] Specifically, the improved Osprey optimization algorithm used in this invention is based on the standard Osprey optimization algorithm. Improvements are achieved by introducing Tent chaotic mapping, Levy flight strategy, and spiral search strategy. The core architecture consists of four modules: population initialization, prey search, prey attack, and population update. These modules are interconnected through adaptive parameters. Tent chaotic mapping optimizes the initial population distribution, Levy flight strategy enhances global search capability, spiral search strategy improves local optimization accuracy, and the population update module retains the best individuals and eliminates inferior individuals by sorting by fitness values, achieving a dual improvement in optimization efficiency and accuracy. The optimization data is a combination of hyperparameters to be optimized in a fully variable constrained generative adversarial network, covering the weights of the fully variable regularization term, adversarial loss, L1 loss, and histogram matching loss. The preset search spaces for the hyperparameters are as follows: The preprocessing algorithm uses the following parameters: [0.001,0.1], [0.1,1.0], [0.1,1.0], [0.1,1.0]. Boundary constraints are used to eliminate invalid hyperparameter combinations that exceed the search range during preprocessing. The structural similarity (SSIM) between the total variation constraint generative adversarial network inversion results and the actual erosion thickness label is used as the fitness function. An adaptive step size update strategy is used instead of the fixed step size of the standard algorithm, and gradient descent is used to assist in updating to accelerate convergence. The preferred key hyperparameters are: population size of 30, maximum number of iterations of 50, Tent chaotic mapping parameter of 1.9, Levy flight step size coefficient of 1.5, spiral search coefficient of 1.5, and a switching threshold between global and local search of 0.5. An adaptive decay strategy is used for hyperparameter adjustment, with the first 20 iterations focusing on global search and the last 30 iterations focusing on local optimization.

[0057] It should be noted that the secondary grinding parameters of the lightning arrester are generated based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester. The rules for determining the secondary grinding parameters are shown in Table 3.

[0058] Table 3. Correspondence Table of Secondary Polishing Parameters for Lightning Arresters S4. Obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

[0059] Specifically, the second resistance detection result is compared with the preset on-resistance threshold, and the lightning arrester repair qualification is determined based on the resistance comparison result. If the repair is deemed unqualified, S3 is repeated until the repair qualification condition is met or the preset iteration limit is reached, and a lightning arrester integrity detection and evaluation report is generated.

[0060] Specifically, after the lightning arrester undergoes secondary polishing, the four-probe resistance measurement module at the end of the robotic arm uses the same 8×8 layout of 64 measuring points as S3, collecting the conduction resistance value of each measuring point at 1mm intervals to obtain the second resistance detection result. The measured resistance data is retained to 3 decimal places, and the range is controlled between 0.05Ω and 2.0Ω. The preset acceptable threshold for the conduction resistance of the lightning arrester is preferably 0.5Ω. At the same time, the preferred standard for judging the repair qualification is: the resistance value of all 64 measuring points is ≤0.5Ω, and the standard deviation of the resistance value of all measuring points is ≤0.1Ω. The preset maximum number of iterations of polishing is 3 times, including the first and second polishing. The collected second resistance detection results are compared point by point with the 0.5Ω conduction resistance threshold, and the standard deviation of the resistance value of all measuring points is calculated. If the resistance of all measuring points meets the condition of ≤0.5Ω, the lightning arrester is considered qualified. If the resistance is 0.5Ω and the standard deviation is ≤0.1Ω, the repair is considered qualified, and an automatic lightning arrester integrity test and evaluation report is generated. The report includes the grinding parameters, resistance test data, corrosion thickness inversion results, and the criteria for qualification. If the resistance at any measuring point is greater than 0.5Ω or the standard deviation is greater than 0.1Ω, the repair is considered unqualified. If the upper limit of 3 iterations has not been reached, the process returns to step S3 and repeats the corrosion thickness inversion, hyperparameter optimization, and secondary grinding parameter generation. The process from S3 to S4 is repeated until the qualification conditions are met. If the upper limit of 3 iterations has been reached and the repair is still unqualified, the lightning arrester cannot be repaired by grinding. The evaluation report marks the abnormal situation and recommends replacing the lightning arrester. At the same time, the resistance test data, grinding parameters, and inversion results of all iterations are recorded to provide complete data support for subsequent maintenance.

[0061] like Figure 2 As shown, according to another embodiment of the present invention, a lightning arrester integrity detection system based on three-dimensional positioning of the lightning arrest point is provided. The system includes: The spatial location determination module 1 is used to acquire images of the wind turbine blade surface, use a target detection algorithm to locate and identify the wind turbine blade surface image, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial location information of the lightning arrester. The corrosion preliminary judgment module 2 is used to perform preliminary integrity detection on the lightning rod image area in the wind turbine blade surface image based on the two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning rod, the initial polishing parameters of the lightning rod are generated. The corrosion degree secondary judgment module 3 is used to obtain the first resistance detection result of the lightning arrester conduction test after the initial polishing, and to use the corrosion thickness inversion model to invert the corrosion degree of the first resistance detection result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Then, the secondary polishing parameters of the lightning arrester are generated based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester. The repair qualification judgment module 4 is used to obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

[0062] In summary, by utilizing the above-mentioned technical solution of this invention, the present invention achieves a precise inversion of the local micro-area corrosion state of the lightning arrester and a quantitative mapping from electrical response to physical corrosion morphology through corrosion degree inversion. This avoids the problems of traditional methods that can only provide a general resistance value and cannot reveal the spatial distribution of corrosion. Furthermore, it avoids problems such as blind repair, over-polishing, or insufficient repair during polishing, completely changing the traditional one-size-fits-all or trial-and-error polishing approach, greatly improving repair quality and lightning arrester service life, and ensuring the long-term reliability of the lightning protection system. Through a preliminary integrity intelligent discrimination mechanism based on color and brightness characteristics, it achieves rapid initial screening of corrosion state and intelligent generation of initial polishing parameters, fully utilizing saturation... The statistical characteristics of the intensity and brightness channels can effectively distinguish between different states such as intact, slightly corroded, and moderately corroded, generating preliminary corrosion judgment results and initial polishing parameters adapted to the current corrosion level. This avoids material waste and structural damage caused by indiscriminate polishing, improving the targeting and economy of the repair. By introducing a corrosion thickness inversion model based on resistance detection data, it can not only learn the complex nonlinear mapping relationship between resistivity and corrosion thickness, but also maintain clear corrosion boundaries and suppress noise through regularization terms, outputting a high-fidelity corrosion thickness spatial distribution map. This map can truly reflect the remaining metal thickness or corrosion depth in each area of ​​the lightning arrester surface, and can generate highly personalized secondary polishing parameters, thereby achieving precise repair based on physical inversion.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point, characterized in that, The method includes: S1. Acquire the surface image of the wind turbine blade, use the target detection algorithm to locate and identify the surface image of the wind turbine blade, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial position information of the lightning arrester. S2. Perform preliminary integrity detection on the lightning arrester image area in the wind turbine blade surface image based on the two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Generate the initial polishing parameters of the lightning arrester based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning arrester. S3. Obtain the first resistance test result of the lightning arrester after the initial polishing. Use the corrosion thickness inversion model to invert the corrosion degree of the first resistance test result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Generate the secondary polishing parameters of the lightning arrester based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester. S4. Obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

2. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 1, characterized in that, The method of using a target detection algorithm to locate and identify the surface image of the wind turbine blades, obtaining the two-dimensional pixel coordinates of the lightning arrester, and then performing three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial position information of the lightning arrester includes: The surface image of the wind turbine blade is input into a pre-trained target detection algorithm, and the tensor inference acceleration engine is used to accelerate the inference of the surface image of the wind turbine blade to obtain the two-dimensional pixel coordinates of the lightning arrester. The two-dimensional pixel coordinates are matched with the same-view depth map at the pixel level. The depth value corresponding to each pixel is extracted based on the matching result. The two-dimensional pixel points and depth values ​​are back-projected to the camera coordinate system in combination with the camera intrinsic parameter matrix to generate the three-dimensional point cloud of the flash receiver in the camera coordinate system. Outlier filtering and voxel downsampling are performed on the 3D point cloud to obtain the 3D spatial position information of the lightning arrester. Outlier filtering and centering are then performed on the 3D point cloud to obtain a normalized point cloud. The geometric center of the normalized point cloud is calculated to obtain the 3D spatial position information of the lightning arrester in the camera coordinate system.

3. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 2, characterized in that, The step of inputting the wind turbine blade surface image into a pre-trained target detection algorithm and using a tensor inference acceleration engine to accelerate inference on the wind turbine blade surface image to obtain the two-dimensional pixel coordinates of the lightning arrester includes: A labeled dataset is constructed based on pre-acquired real-world images of wind turbine blade lightning arresters. The labeled dataset is then input into the framework of the object detection algorithm for transfer learning training. The weights of the object detection algorithm are iteratively optimized through the loss function to obtain the pre-trained object detection algorithm. A pre-trained target detection algorithm is used to perform lightning arrester target detection and feature extraction on the surface image of the wind turbine blade to obtain a lightning arrester feature map. The lightning arrester feature map is then input into a tensor inference acceleration engine for model inference and position regression calculation to obtain confidence and coordinate information. Based on the confidence and coordinate information, non-maximum suppression and coordinate decoding are performed to obtain the lightning arrester target detection box. The center position and boundary coordinates of the lightning arrester are extracted based on the target detection box of the lightning arrester, and the two-dimensional pixel coordinates of the lightning arrester are obtained after coordinate normalization and correction calculation.

4. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 1, characterized in that, The preliminary integrity detection of the lightning arrester image area in the wind turbine blade surface image based on two-dimensional pixel coordinates, and the generation of preliminary corrosion judgment results, include: The lightning arrester image region is cropped and extracted from the wind turbine blade surface image based on the two-dimensional pixel coordinates, and the color space of the lightning arrester image region is converted to obtain the converted lightning arrester image block. A color and brightness comparison analysis was performed on the converted lightning arrester image block to obtain the color deviation and dark area ratio of the lightning arrester surface. The color deviation is compared with the preset color deviation threshold to obtain the color anomaly judgment result, and the dark area ratio is compared with the preset dark area ratio threshold to obtain the rust defect judgment result. Based on the color anomaly judgment result and the rust defect judgment result, the corrosion degree level is determined to obtain the preliminary corrosion judgment result.

5. The lightning arrester integrity detection method based on three-dimensional positioning of the lightning strike point according to claim 4, characterized in that, The step of performing color and brightness comparison analysis on the converted lightning arrester image block to obtain the color deviation and dark area ratio of the lightning arrester surface includes: The pixel value distribution of the saturation and luminance channels is extracted from the converted lightning arrester image block to obtain the color feature vector of the lightning arrester region; The color feature vector is compared with the preset standard color feature template of the lightning arrester by Euclidean distance to obtain the color deviation of the lightning arrester surface. The proportion of dark areas on the surface of the lightning arrester is obtained by statistically analyzing the percentage of pixels with brightness below a preset brightness threshold within the image block of the lightning arrester.

6. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 1, characterized in that, The process of using a corrosion thickness inversion model to invert the corrosion degree of the first resistance detection result, and obtaining the spatial distribution of corrosion thickness on the lightning arrester surface, includes: The first resistance detection result is normalized to obtain standardized resistivity data. Spatial interpolation is then used to fill in the blank areas between measurement points in the standardized resistivity data to generate a two-dimensional resistivity pseudo-image. The two-dimensional resistivity pseudo-image is input into the corrosion thickness inversion model to perform resistivity to corrosion degree mapping inversion, and the optimized corrosion thickness distribution map is obtained. Median filtering was used to denoise the optimized corrosion thickness distribution map, resulting in a denoised corrosion thickness distribution map. The scale of the denoised corrosion thickness distribution map was then restored according to the scale parameter during normalization to obtain the spatial distribution of corrosion thickness on the lightning arrester surface.

7. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 6, characterized in that, The step of inputting the two-dimensional resistivity pseudo-image into the corrosion thickness inversion model to perform resistivity-to-corrosion degree mapping inversion, and obtaining the optimized corrosion thickness distribution map includes: The two-dimensional resistivity pseudo-image is input into a total variation constrained generative adversarial network for multi-scale spatial feature extraction and fusion to obtain a preliminary corrosion thickness distribution map. The preliminary corrosion thickness distribution map and the preset real corrosion thickness label are input into the discriminator for adversarial discrimination. The network weights of the generator are updated by backpropagation based on the discrimination results, and a corrosion thickness inversion model is generated by combining the total variation regularization term constraint. The weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model are adaptively optimized to obtain the optimal hyperparameter combination. The optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the two-dimensional resistivity pseudo-image, resulting in the optimized corrosion thickness distribution map.

8. The lightning arrester integrity detection method based on three-dimensional positioning of the lightning strike point according to claim 7, characterized in that, The step of inputting the two-dimensional resistivity pseudo-image into a fully variable constrained generative adversarial network for multi-scale spatial feature extraction and fusion to obtain a preliminary corrosion thickness distribution map includes: The two-dimensional resistivity pseudo-image is input into the generator of the total variation constrained generative adversarial network, and the two-dimensional resistivity pseudo-image is subjected to multi-scale spatial feature extraction from local to global through multi-level convolutional layers and downsampling operations in the encoder to obtain the multi-scale feature map corresponding to the resistivity distribution. By using skip connections, multi-scale feature maps are passed layer by layer to the corresponding levels of the decoder. In the decoder, deconvolution and upsampling operations are used to gradually restore the spatial resolution. The multi-scale features of the same level of the encoder are then fused to generate an intermediate image of the fused erosion thickness. The fused intermediate corrosion thickness map is mapped to a single-channel output using a convolutional layer to obtain a preliminary corrosion thickness distribution map.

9. The method for detecting the integrity of a lightning arrester based on three-dimensional positioning of the lightning strike point according to claim 7, characterized in that, The process involves adaptively optimizing the weights of the total variation regularization term and the hyperparameters of the loss function in the corrosion thickness inversion model to obtain the optimal hyperparameter combination. This optimal hyperparameter combination is then substituted into the corrosion thickness inversion model to perform resistivity-to-corrosion thickness mapping inversion on the preliminary corrosion thickness distribution map, resulting in the optimized corrosion thickness distribution map, including: The set of loss weights in the generator loss function and the weights of the total variation regularization term are used to construct the hyperparameter combination to be optimized. The tent chaos mapping is used to initialize the population of the improved Osprey optimization algorithm in the preset hyperparameter space of the hyperparameter combination to be optimized, and an initial hyperparameter population set containing multiple sets of hyperparameter combinations is generated. Each set of hyperparameters in the initial hyperparameter population is substituted into the corrosion thickness inversion model, and the two-dimensional resistivity pseudo-image is used as the input of the corrosion thickness inversion model to perform resistivity to corrosion thickness mapping inversion, so as to obtain the initial corrosion thickness distribution map. A fitness function is generated based on the structural similarity between the initial corrosion thickness distribution map and the preset real corrosion thickness label. The optimal hyperparameter combination is obtained by iteratively updating the hyperparameter combination in the initial hyperparameter population set through a combination of the Lévy strategy and the spiral curve strategy. The optimal combination of hyperparameters is substituted into the corrosion thickness inversion model, and the forward propagation calculation of the generator is performed using a two-dimensional resistivity pseudo-image as the initial input of the corrosion thickness inversion model to obtain the optimized corrosion thickness distribution map.

10. A lightning arrester integrity detection system based on three-dimensional positioning of the lightning strike point, used to implement the lightning arrester integrity detection method based on three-dimensional positioning of the lightning strike point as described in any one of claims 1-9, characterized in that, The system includes: The spatial location determination module is used to acquire images of the wind turbine blade surface, use a target detection algorithm to locate and identify the wind turbine blade surface image, obtain the two-dimensional pixel coordinates of the lightning arrester, and perform three-dimensional point cloud mapping on the two-dimensional pixel coordinates to obtain the three-dimensional spatial location information of the lightning arrester. The corrosion preliminary judgment module is used to perform preliminary integrity detection on the lightning rod image area in the wind turbine blade surface image based on two-dimensional pixel coordinates, and generate preliminary corrosion judgment results. Based on the preliminary corrosion judgment results and the three-dimensional spatial position information of the lightning rod, the initial polishing parameters of the lightning rod are generated. The secondary corrosion degree judgment module is used to obtain the first resistance detection result of the lightning arrester conduction test after the initial polishing, and to use the corrosion thickness inversion model to invert the corrosion degree of the first resistance detection result to obtain the spatial distribution of corrosion thickness on the surface of the lightning arrester. Based on the spatial distribution of corrosion thickness and the three-dimensional spatial position information of the lightning arrester, the secondary polishing parameters of the lightning arrester are generated. The repair qualification judgment module is used to obtain the second resistance detection result of the lightning arrester conduction test after secondary polishing, compare it with the preset conduction resistance threshold, and repeat the corrosion degree inversion and polishing according to the resistance comparison result until the repair qualification conditions are met.

Images