Wind turbine generator blade automatic inspection route planning method based on laser point cloud

Abstract

The invention relates to the technical field of unmanned aerial vehicle aerial photography data application, in particular to a wind turbine generator blade automatic inspection route planning method based on laser point cloud. According to the technical scheme, the wind turbine generator blade automatic inspection route planning method based on the laser point cloud comprises the working process of wind turbine generator blade automatic inspection route planning; according to the method, breakthrough is achieved through the multi-view annular collection track design and the multi-source point cloud fusion technology, the annular collection track comprises eight evenly-distributed point positions, the staying time of each point position is sufficient, and it is ensured that the whole surface of the blade is completely covered; meanwhile, a multi-line laser radar and single-line laser radar combination scheme is combined, and multi-source point clouds are unified to a WGS84 coordinate system through seven-parameter coordinate transformation, so that registration failure caused by coordinate system difference is avoided, the integrity of point cloud acquisition is improved, and precision loss caused by long-distance scanning is eliminated through multi-view data fusion.

Description

Technical Field

[0001] This invention relates to the field of drone aerial photography data application technology, and in particular to a method for automatic inspection route planning of wind turbine blades based on laser point clouds. Background Technology

[0002] With increasing global attention to environmental issues and market and policy favoring new energy sources, countries are accelerating the construction of reliable alternatives to new energy sources to orderly reduce reliance on traditional energy. Wind power is one of the main new energy sources. With the rapid development of the new energy industry, the technological level of the wind power industry is also constantly improving. The solution of drone inspection is being adopted more and more, gradually replacing the traditional telescope observation and rope descent manual inspection methods. This solves problems such as low inspection efficiency, high cost, long time, high labor intensity for workers, and large loss of power generation during downtime inspection.

[0003] Existing automatic inspection and route planning methods for wind turbine blades cannot directly acquire complete point cloud data of the wind turbine due to the large length of the blades and the insufficient accuracy of lidar when collecting data on narrow objects at long distances. Traditional lidar has a limited scanning range, and the narrow structure of the blades is prone to producing sparse or missing point clouds when scanning at long distances, which directly affects the reliability of subsequent data processing and route planning.

[0004] To address the aforementioned issues, this solution achieves a breakthrough through multi-view circular acquisition trajectory design and multi-source point cloud fusion technology. The circular acquisition trajectory includes eight evenly distributed points, each with sufficient dwell time to ensure complete coverage of the entire blade surface. Simultaneously, by combining multi-line and single-line LiDAR, a seven-parameter coordinate transformation unifies the multi-source point clouds to the WGS84 coordinate system, avoiding registration failures caused by coordinate system differences. This not only improves the integrity of point cloud acquisition but also eliminates the accuracy loss caused by long-distance scanning through multi-view data fusion, ensuring that the generated wind turbine digital twin model has millimeter-level accuracy. This provides a high-precision data foundation for subsequent route planning, significantly enhancing the accuracy and safety of the inspection path. Summary of the Invention

[0005] To overcome the problems of existing automatic inspection and route planning methods for wind turbine blades, which are limited by the large length of wind turbine blades and the insufficient accuracy of lidar when collecting data on narrow objects at long distances, resulting in the inability to directly obtain complete point cloud data of wind turbines, traditional lidar has a limited scanning range, and the narrow structure of the blades is prone to producing sparse or missing point clouds when scanning at long distances, which directly affects the reliability of subsequent data processing and route planning.

[0006] The technical solution of this invention is: an automatic inspection route planning method for wind turbine blades based on laser point clouds, comprising the following steps: S11: By establishing a multi-model parameter database, calibrating high-precision equipment, and configuring an environmental monitoring module, accurate pre-configuration of parameters and equipment calibration for multiple wind turbine models are achieved; S12: By locking the shutdown state, set the ring-shaped acquisition points, simultaneously acquire point clouds from multiple locations and integrate spatiotemporal data, and perform accurate acquisition and synchronization of multi-source point cloud data; S13: Apply statistical filtering, Gaussian smoothing, and vertical reference calibration techniques to optimize noise suppression and preprocessing of point cloud data; S14: Employing ICP coarse registration and NDT fine registration technologies, a multi-view point cloud is integrated to generate a wind turbine digital twin model with millimeter-level precision; S15: Utilize deep learning algorithms to automatically identify tower sections or blades, and extract geometric parameters and deformation benchmark models; S16: Calculate the safety distance in real time and design an adaptive yaw compensation algorithm to generate a dynamic inspection path covering the entire surface of the blade; S17: Through The algorithm optimizes the path to reduce attitude adjustments, combines B-spline curves to smooth the trajectory, and integrates a collision detection module to achieve a safe and efficient inspection path; S18: Automatically generates a complete flight path including photo capture points and return points, adjusts exposure parameters in real time, and simultaneously collects multimodal data; S19: Utilizes AI algorithms to automatically identify defects and generate deformation heat maps, combined with health records to predict the remaining service life of the wind turbine.

[0007] Preferably, the following steps are included when performing multi-model parameter pre-configuration and equipment calibration: S21: A database containing geometric parameters such as tower height, blade length, hub diameter, or number of blades is built based on SQL Server. The database enables fast parameter querying and retrieval via API, and provides multi-model parameter version management and automatic update functionality. S22: Point cloud accuracy is tested using a coordinate measuring machine under a standard target. By adjusting the laser emission power, scanning angle, and sampling frequency parameters, the measurement error is controlled within ±2mm. S23: The PTPv2 protocol is used for microsecond-level time synchronization between the UAV positioning system and the laser scanner, and millimeter-level spatial synchronization is achieved through coordinate transformation algorithms to avoid point cloud distortion caused by timing discrepancies; S24: Based on the blade rotation radius and safety distance standards of different aircraft models, the safe distance for drone inspection is determined through three-dimensional spatial calculation. A three-dimensional electronic fence is set up using geofencing technology to automatically restrict drones from entering dangerous areas where blades rotate. S25: A high-precision total station was used to measure the three-dimensional coordinates of the center point of the tower base. By comparing and analyzing the actual measured values ​​with the design parameters, the least squares method was used to correct the model error, and the tower positioning accuracy was found to be ≤±3mm. S26: Based on the surface material characteristics of blades of different models, the optimal scanning angle and power parameters are determined through experimental testing. The scanning parameters of the laser scanner are adjusted to optimize the point cloud density and quality, and the point cloud resolution is checked to see if it reaches 0.5mm. S27: An automatic parameter configuration software module developed based on C#, which automatically loads the corresponding geometric parameters and scanning parameters according to the selected wind turbine model; S28: By monitoring temperature and humidity changes in real time, an environmental parameter compensation model is established to automatically correct the thermal expansion coefficient of the laser scanner, avoiding measurement errors caused by environmental changes; S29: By combining ground reference station data, the positioning accuracy of UAVs in complex terrain is improved to the centimeter level, avoiding blind spots or repeated inspections caused by positioning errors.

[0008] Preferably, the following steps are included when collecting and synchronizing multi-source point cloud data: S31: Using a RIEGL VZ-400i laser scanner, set the scanning angle range to 0°-360°, vertical resolution to 0.01°, and horizontal resolution to 0.05°, check if the point cloud density on the blade surface is ≥200 points / cm²; S32: Equipped with a dual-frequency GNSS receiver and an inertial navigation system, it performs microsecond-level time synchronization via the PTPv2 protocol to detect whether the time error of multi-sensor data acquisition is ≤1μs, thus avoiding point cloud distortion caused by timing discrepancies; S33: A circular data acquisition track is set up directly in front of the wind turbine, containing 8 evenly distributed data acquisition points, with each point remaining for at least 30 seconds; S34: The message_filters module of the ROS system is used to align the timestamps of multi-sensor data, and the point cloud data is fused using the PCL library to seamlessly stitch together the main view and the blade tip point cloud to form a complete wind turbine point cloud; S35: Employs a combination of multi-line and single-line lidar, using a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system, thus avoiding registration failures caused by coordinate system differences; S36: Based on a PID controller, a dynamic exposure control algorithm is implemented to monitor the ambient light intensity in real time and automatically adjust the laser emission power. It detects whether the point cloud brightness is uniform and whether there is overexposure / underexposure, while maintaining a constant scanning speed of 5m / s. S37: Establish a data quality assessment module based on the PCL library to monitor point cloud density, noise level, and coordinate accuracy in real time, and automatically remove data frames that do not meet the standards; S38: Uses HDF5 format to store multi-source point cloud data, establishes an indexing mechanism to achieve fast data retrieval, and records metadata such as collection time, location, and environmental parameters.

[0009] Preferably, the point cloud preprocessing and noise suppression process includes the following steps: S41: Using the pass-through filter in the PCL library, determine the ROI region based on the relative position of the lidar and the wind turbine, set the XYZ coordinate threshold to extract the point cloud of the tower and blades, remove interference from the ground and background environment point cloud, and check whether the number of point clouds is reduced by ≥30% after extraction and the key areas are completely preserved; S42: Apply a statistical outlier removal algorithm to calculate the average distance between each point and its k nearest neighbors. Set a standard deviation threshold of 3σ, remove outliers whose average distance exceeds the threshold, and check whether the effective point cloud retention rate is ≥95% and the noise point removal rate is ≥98%. S43: Using a voxel mesh filtering method, the voxel size is set to 0.05m × 0.05m × 0.05m. The centroid of each point within a voxel is used to replace all points, compressing the point cloud data to 10% of the original data, while maintaining the tower cylindricity error ≤ 0.02m and the blade surface smoothness error ≤ 0.01m; S44: Use a Gaussian filter to smooth the point cloud, set the Gaussian kernel radius to 0.1m, suppress high-frequency noise by weighted averaging of neighborhood point coordinates, maintain surface smoothness, and check whether the surface roughness of the point cloud is reduced by more than 30% and the edge features are preserved intact after filtering; S45: A bilateral filtering algorithm is applied to smooth the ordered point cloud. Combining spatial distance weight and gray-level difference weight, the geometric features of the blade edge are preserved while suppressing noise. The positional deviation of the edge point cloud is detected to be ≤0.03m. S46: Set the search radius to 0.3m using a radius filter, count the number of points in the neighborhood of each point, remove isolated points with fewer than 5 neighbors, and check whether the point cloud density uniformity is improved to ≥90% and whether local sparse areas are effectively filled; S47: The RANSAC algorithm was used to fit the cylindrical tower model and the curved blade model. The model fitting error threshold was set to 0.05m. Outliers that were more than 0.05m away from the model were removed. The accuracy of the point cloud matching the model was checked to ensure that the geometric features were not distorted. S48: Calculate the normal direction of each point, and by comparing the consistency of the normal directions of neighboring points, remove outliers with significantly inconsistent normal directions to enhance the geometric consistency of the point cloud, and check whether the normal direction error is ≤5°; S49: Combining point cloud data from LiDAR and depth camera, registration is performed using the ICP algorithm. The high-resolution information from the depth camera supplements the details missing from the LiDAR, improving the integrity of the point cloud. The registration accuracy is checked to ensure it is ≤0.02m and the detail feature retention rate is ≥90%. S410: Embeds a quality monitoring module during preprocessing to calculate point cloud density, noise level, and coordinate accuracy in real time, and automatically adjusts filtering parameters.

[0010] Preferably, the following steps are included when performing multi-scale point cloud registration and fusion: S51: Use the SIFT algorithm in the PCL library to extract multi-scale point cloud feature points, set the number of feature points to ≥500 per region, detect whether the feature points are uniformly distributed on the blade surface and have high discriminative power, and support the coarse registration process; S52: Initial registration is performed using the RANSAC algorithm. The initial transformation matrix is ​​calculated based on feature point matching. The maximum number of iterations is set to 100, and the error threshold is 0.05m. S53: The ICP algorithm is used for fine registration, with a maximum number of iterations of 200. The convergence condition is that the mean square error is <0.01m or the number of iterations reaches the upper limit. The optimal transformation matrix is ​​calculated iteratively using the least squares method. S54: The multi-view point cloud is fused into a complete model using a voxel fusion algorithm. The voxel size is set to 0.02m × 0.02m × 0.02m, and the average coordinates of the points within the voxel are used to replace the original point cloud. S55: Use a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system. Calculate the transformation parameters using control point pairs and check if the coordinate system transformation error is ≤0.01m. S56: A bilateral filtering algorithm is applied during the fusion process, with spatial weight σs=0.3m and grayscale weight σr=0.2, to suppress noise while preserving blade edge features; S57: Establish a registration quality assessment module to calculate point cloud overlap, registration error, and mean square error in real time, and automatically verify whether the fusion results meet the inspection requirements; S58: The Poisson reconstruction algorithm is used to generate a mesh model of the blade surface. The mesh resolution is set to 0.01m. The model surface is checked to see if it is continuous and free of voids, and the error between it and the original point cloud is ≤0.02m. This supports subsequent defect detection.

[0011] Preferably, the intelligent identification and parameter extraction of key components includes the following steps: S61: Using the pass-through filter in the PCL library, set the height threshold Z>1m to remove ground point clouds, retain the main point cloud of the wind turbine, reduce the number of point clouds by ≥40% after processing, and retain the key areas completely; S62: The RANSAC algorithm is applied to fit the point cloud of the tower into a cylindrical model. The maximum number of iterations is set to 1000, and the inward point threshold is set to 0.05m. The three-dimensional coordinates, radius, and height parameters of the tower center point are calculated. S63: Project the point cloud around the wheel hub onto the XY plane, and use a distance transformation algorithm combined with PCA principal component analysis to calculate the center position and main axis direction of the wheel hub. The positioning accuracy is ≤0.02m and the deviation of the main axis direction is ≤0.5°. S64: The DBSCAN clustering algorithm is used to segment the blade point cloud. The neighborhood radius is set to 0.5m and the minimum number of points is 50. The blade point cloud is separated from the tower / hub. The clustering accuracy is ≥95% and the smoothness of the segmentation edge meets the inspection requirements. S65: Determine the axial direction through PCA analysis of blade point cloud, calculate blade span and twist angle, with measurement error ≤0.1° and span calculation accuracy ≤0.01m; S66: The YOLOv5 algorithm was used to detect defects in the point cloud on the blade surface. A confidence threshold of 0.7 was set to identify crack / corrosion defects. The recall rate was ≥90% and the false positive rate was ≤5%. S67: Store the identified tower, hub, and blade parameters in an SQL database, including geometric parameters, location information, and defect data.

[0012] Preferably, the following steps are included when conducting dynamic environmental adaptive route planning: S71: Uses meteorological sensors to collect wind speed, temperature, and humidity data in real time, and publishes environmental parameter topics through the ROS system to provide dynamic input for flight route planning; S72: Based on the wind turbine's operating status and environmental parameters, a dynamic safety distance model is used to calculate the minimum safe distance between the UAV and the blades; S73: Application Improvement The algorithm performs path planning, sets the node expansion cost as a weighted sum of distance and energy consumption, and limits the path search range to a 200m radius around the wind turbine, generating the shortest inspection path that meets safety constraints; S74: Integrated LiDAR obstacle avoidance module, with an obstacle detection distance threshold of 10m. When an obstacle is detected, it automatically triggers path replanning to adjust the flight path and detour around the danger zone; S75: Automatically adjusts camera exposure parameters according to ambient light intensity, ensuring uniform image brightness without overexposure or underexposure; S76: Employs a multi-sensor fusion algorithm to combine GPS positioning data, IMU attitude data, and LiDAR point cloud data to calculate the UAV's real-time position; S77: Real-time verification of flight path safety. It monitors the drone's position through three-dimensional electronic fence technology. When the drone enters a dangerous area, it automatically triggers emergency braking and adjusts the flight path to a safe area.

[0013] Preferably, when performing multi-objective optimization and path smoothing, the following steps are included: S81: Using path length, energy consumption, and safety distance as quantitative indicators, a multi-objective optimization model is constructed. A linear weighted method is used to transform the multi-objective problem into a single-objective problem, with weight coefficients set to 0.4, 0.3, and 0.3 respectively. S82: Application The algorithm performs an initial path search, setting the node expansion cost as a weighted sum of distance and energy consumption, and limiting the path search range to a 200m radius around the wind turbine; S83: Smooth the initial path using a Bézier curve, setting the control point spacing to 0.5m and the curve order to 3; S84: The path is reconstructed using B-spline curves, and the node positions are adjusted to make the curve continuous and satisfy the dynamic constraints of the UAV; S85: Use the particle swarm optimization algorithm to adjust the path parameters, set the number of particles to 50 and the maximum number of iterations to 100, and perform iterative solutions with the dual objectives of minimizing the total path length and minimizing energy consumption; S86: Real-time path safety verification; monitors drone position using 3D electronic fence technology; automatically triggers path replanning when drone enters a dangerous area. S87: Employs multi-sensor fusion positioning, integrating GPS, IMU, and LiDAR data to detect whether the real-time position accuracy of the drone is ≤0.05m.

[0014] As a preferred method, the intelligent inspection and data collection process includes the following steps: S91: The drone automatically takes off and moves to the preset inspection starting point, achieving precise positioning via GPS and IMU, with an initial position error ≤0.3m and stable attitude; S92: Utilizes a PID controller to achieve UAV path tracking, adjusting roll, pitch, and yaw angles to maintain stable flight, with a track deviation ≤0.2m and speed fluctuation ≤0.5m / s; S93: Simultaneously acquires laser point cloud, visible light image, and infrared thermal image data, and performs multi-sensor time synchronization via the PTP protocol; S94: Real-time processing of point cloud data in UAV embedded systems, applying voxel filtering and statistical filtering to remove noise, with a processing latency of ≤80ms; S95: Use the YOLOv5 algorithm to detect blade surface defects in real time, set a confidence threshold of 0.7, mark the defect locations and store them in an SQLite database; S96: Transmits data to ground stations in real time via 5G network, using HDFS distributed storage system; S97: When an abnormal state of the drone is detected, the emergency return-to-home procedure is automatically triggered to plan the shortest path back to the take-off and landing point.

[0015] Preferably, the following steps are included when conducting data analysis and health assessment: S101: Cleans and denoises laser point cloud, visible light image, and infrared thermal image data, and applies the ICP algorithm to achieve spatial alignment of multi-source data; S102: Use the YOLOv5 algorithm to detect and classify surface defects on the blade, set a confidence threshold of 0.7, identify the defect types such as cracks, corrosion, or deformation, and mark their location coordinates; S103: Construct a health assessment model based on the random forest algorithm, inputting defect density, defect area, or runtime parameters, and outputting a health score; S104: The ARIMA time series model is used to predict the remaining service life of the blades. Combined with historical defect growth trend data, the prediction period is set to 6 months, with an error range of ≤5%. S105: Use the Matplotlib library to generate a heatmap of defect distribution and a health score trend curve, and use color mapping to visually display the health status of different areas of the leaf; S106: Store the analysis results in a MySQL database, and design a defect table, health score table, and inspection record table, supporting query and comparative analysis by time / region; S107: The early warning system is automatically triggered when the health score is <60 or a serious defect is detected.

[0016] The beneficial effects of this invention are: 1. Existing methods for automatic inspection route planning of wind turbine blades suffer from limitations. Due to the large length of wind turbine blades and insufficient accuracy of lidar when acquiring narrow objects at long distances, complete point cloud data of the wind turbine cannot be directly obtained. Traditional lidar scanning range is limited, and the narrow structure of the blades is prone to sparse or missing point clouds during long-distance scanning, directly affecting the reliability of subsequent data processing and route planning. This solution achieves a breakthrough through multi-view circular acquisition trajectory design and multi-source point cloud fusion technology. The circular acquisition trajectory includes eight evenly distributed points, each with sufficient dwell time to ensure complete coverage of the entire blade surface. Simultaneously, a combination of multi-line and single-line lidar is used, and a seven-parameter coordinate transformation unifies the multi-source point clouds to the WGS84 coordinate system, avoiding registration failures caused by coordinate system differences. This not only improves the completeness of point cloud acquisition but also eliminates the accuracy loss caused by long-distance scanning through multi-view data fusion, ensuring that the generated wind turbine digital twin model has millimeter-level accuracy. This provides a high-precision data foundation for subsequent route planning, significantly enhancing the accuracy and safety of the inspection path. 2. Existing automatic inspection route planning methods for wind turbine blades suffer from irregular point cloud distribution and high noise levels from lidar-collected data. This presents technical challenges and reliability issues in accurately identifying key components from the chaotic point cloud. Uneven point cloud density, severe noise interference, and difficulties in feature extraction can easily lead to omissions or misidentifications of critical components, affecting the rationality of route planning. This solution improves upon these methods through dual optimization of point cloud preprocessing and intelligent recognition algorithms. In the preprocessing stage, multiple techniques such as statistical filtering, Gaussian smoothing, and voxel grid filtering are used to suppress noise and optimize point cloud density, ensuring uniform point cloud distribution and complete preservation of key area features. In the intelligent recognition stage, deep learning algorithms are used to automatically identify key components such as the tower, hub, and blades. Accurate positioning is achieved through geometric parameter extraction and deformation benchmark model construction. This not only improves the accuracy and reliability of key component identification but also provides a precise basis for dynamic inspection path design through geometric parameter extraction, ensuring that route planning covers all key components and that the path is reasonable. This effectively avoids blind spots or repeated inspections caused by identification errors, improving inspection efficiency and safety. 3. Existing automatic inspection route planning methods for wind turbine blades rely on a single visible light load for inspection. Due to limitations in lighting conditions and viewing angle, these methods struggle to accurately perceive the actual position and morphology of the wind turbine blades, leading to inaccurate data acquisition and affecting the accuracy of route planning. Furthermore, the limitations of single-sensor data can cause blade damage or deformation to be missed, impacting the timeliness of maintenance decisions. This solution achieves a breakthrough through the fusion of multimodal data acquisition and intelligent algorithms. By simultaneously acquiring laser point cloud, visible light image, and infrared thermal image data, and using the PTP protocol to achieve multi-sensor time synchronization, it ensures spatiotemporal alignment of multi-source data. Simultaneously, the YOLOv5 algorithm is used for real-time detection of blade surface defects, combined with health records to predict the remaining service life of the wind turbine. This not only improves the accuracy of data acquisition but also achieves comprehensive perception of the actual position, morphology, and health status of the blades through multimodal data fusion and intelligent algorithms. This ensures that route planning accurately covers damaged areas and adjusts the inspection path in real time, effectively improving inspection results and the accuracy of maintenance decisions, providing a reliable guarantee for the safe operation of wind turbines. Attached Figure Description

[0017] Figure 1 The flowchart shown is a method for automatic inspection and route planning of wind turbine blades based on laser point clouds according to the present invention. Figure 2 The diagram shows the workflow of an automatic inspection route planning method for wind turbine blades based on laser point clouds according to the present invention. Figure 3 The diagram shows the point cloud preprocessing and noise suppression process of an automatic inspection route planning method for wind turbine blades based on laser point clouds, according to the present invention. Figure 4 The diagram illustrates the multi-scale point cloud registration and fusion process of an automatic inspection route planning method for wind turbine blades based on laser point clouds, according to the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Please see Figure 1-4 This invention provides an embodiment: an automatic inspection route planning method for wind turbine blades based on laser point clouds, comprising the following steps: S11: By establishing a multi-model parameter database, calibrating high-precision equipment, and configuring an environmental monitoring module, accurate pre-configuration of parameters and equipment calibration for multiple wind turbine models are achieved; S12: By locking the shutdown state, set the ring-shaped acquisition points, simultaneously acquire point clouds from multiple locations and integrate spatiotemporal data, and perform accurate acquisition and synchronization of multi-source point cloud data; S13: Apply statistical filtering, Gaussian smoothing, and vertical reference calibration techniques to optimize noise suppression and preprocessing of point cloud data; S14: Employing ICP coarse registration and NDT fine registration technologies, a multi-view point cloud is integrated to generate a wind turbine digital twin model with millimeter-level precision; S15: Utilize deep learning algorithms to automatically identify tower sections or blades, and extract geometric parameters and deformation benchmark models; S16: Calculate the safety distance in real time and design an adaptive yaw compensation algorithm to generate a dynamic inspection path covering the entire surface of the blade; S17: By optimizing the path using the A* algorithm to reduce attitude adjustments, combining B-spline curves to smooth the trajectory, and integrating a collision detection module, a safe and efficient inspection path is achieved; S18: Automatically generates a complete flight path including photo capture points and return points, adjusts exposure parameters in real time, and simultaneously collects multimodal data; S19: Utilizes AI algorithms to automatically identify defects and generate deformation heat maps, combined with health records to predict the remaining service life of the wind turbine.

[0020] Preferably, the following steps are included when performing multi-model parameter pre-configuration and equipment calibration: S21: A database containing geometric parameters such as tower height, blade length, hub diameter, or number of blades is built based on SQL Server. The database enables fast parameter querying and retrieval via API, and provides multi-model parameter version management and automatic update functionality. S22: Point cloud accuracy is tested using a coordinate measuring machine under a standard target. By adjusting the laser emission power, scanning angle, and sampling frequency parameters, the measurement error is controlled within ±2mm. S23: The PTPv2 protocol is used for microsecond-level time synchronization between the UAV positioning system and the laser scanner, and millimeter-level spatial synchronization is achieved through coordinate transformation algorithms to avoid point cloud distortion caused by timing discrepancies; S24: Based on the blade rotation radius and safety distance standards of different aircraft models, the safe distance for drone inspection is determined through three-dimensional spatial calculation. A three-dimensional electronic fence is set up using geofencing technology to automatically restrict drones from entering dangerous areas where blades rotate. S25: A high-precision total station was used to measure the three-dimensional coordinates of the center point of the tower base. By comparing and analyzing the actual measured values ​​with the design parameters, the least squares method was used to correct the model error, and the tower positioning accuracy was found to be ≤±3mm. S26: Based on the surface material characteristics of blades of different models, the optimal scanning angle and power parameters are determined through experimental testing. The scanning parameters of the laser scanner are adjusted to optimize the point cloud density and quality, and the point cloud resolution is checked to see if it reaches 0.5mm. S27: An automatic parameter configuration software module developed based on C#, which automatically loads the corresponding geometric parameters and scanning parameters according to the selected wind turbine model; S28: By monitoring temperature and humidity changes in real time, an environmental parameter compensation model is established to automatically correct the thermal expansion coefficient of the laser scanner, avoiding measurement errors caused by environmental changes; S29: By combining ground reference station data, the positioning accuracy of UAVs in complex terrain is improved to the centimeter level, avoiding blind spots or repeated inspections caused by positioning errors.

[0021] Preferably, the following steps are included when collecting and synchronizing multi-source point cloud data: S31: Using a RIEGL VZ-400i laser scanner, set the scanning angle range to 0°-360°, vertical resolution to 0.01°, and horizontal resolution to 0.05°, check if the point cloud density on the blade surface is ≥200 points / cm²; S32: Equipped with a dual-frequency GNSS receiver and an inertial navigation system, it performs microsecond-level time synchronization via the PTPv2 protocol to detect whether the time error of multi-sensor data acquisition is ≤1μs, thus avoiding point cloud distortion caused by timing discrepancies; S33: A circular data acquisition track is set up directly in front of the wind turbine, containing 8 evenly distributed data acquisition points, with each point remaining for at least 30 seconds; S34: The message_filters module of the ROS system is used to align the timestamps of multi-sensor data, and the point cloud data is fused using the PCL library to seamlessly stitch together the main view and the blade tip point cloud to form a complete wind turbine point cloud; S35: Employs a combination of multi-line and single-line lidar, using a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system, thus avoiding registration failures caused by coordinate system differences; S36: Based on a PID controller, a dynamic exposure control algorithm is implemented to monitor the ambient light intensity in real time and automatically adjust the laser emission power. It detects whether the point cloud brightness is uniform and whether there is overexposure / underexposure, while maintaining a constant scanning speed of 5m / s. S37: Establish a data quality assessment module based on the PCL library to monitor point cloud density, noise level, and coordinate accuracy in real time, and automatically remove data frames that do not meet the standards; S38: Uses HDF5 format to store multi-source point cloud data, establishes an indexing mechanism to achieve fast data retrieval, and records metadata such as collection time, location, and environmental parameters.

[0022] Preferably, the point cloud preprocessing and noise suppression process includes the following steps: S41: Using the pass-through filter in the PCL library, determine the ROI region based on the relative position of the lidar and the wind turbine, set the XYZ coordinate threshold to extract the point cloud of the tower and blades, remove interference from the ground and background environment point cloud, and check whether the number of point clouds is reduced by ≥30% after extraction and the key areas are completely preserved; S42: Apply a statistical outlier removal algorithm to calculate the average distance between each point and its k nearest neighbors. Set a standard deviation threshold of 3σ, remove outliers whose average distance exceeds the threshold, and check whether the effective point cloud retention rate is ≥95% and the noise point removal rate is ≥98%. S43: Using a voxel mesh filtering method, the voxel size is set to 0.05m × 0.05m × 0.05m. The centroid of each point within a voxel is used to replace all points, compressing the point cloud data to 10% of the original data, while maintaining the tower cylindricity error ≤ 0.02m and the blade surface smoothness error ≤ 0.01m; S44: Use a Gaussian filter to smooth the point cloud, set the Gaussian kernel radius to 0.1m, suppress high-frequency noise by weighted averaging of neighborhood point coordinates, maintain surface smoothness, and check whether the surface roughness of the point cloud is reduced by more than 30% and the edge features are preserved intact after filtering; S45: A bilateral filtering algorithm is applied to smooth the ordered point cloud. Combining spatial distance weight and gray-level difference weight, the geometric features of the blade edge are preserved while suppressing noise. The positional deviation of the edge point cloud is detected to be ≤0.03m. S46: Set the search radius to 0.3m using a radius filter, count the number of points in the neighborhood of each point, remove isolated points with fewer than 5 neighbors, and check whether the point cloud density uniformity is improved to ≥90% and whether local sparse areas are effectively filled; S47: The RANSAC algorithm was used to fit the cylindrical tower model and the curved blade model. The model fitting error threshold was set to 0.05m. Outliers that were more than 0.05m away from the model were removed. The accuracy of the point cloud matching the model was checked to ensure that the geometric features were not distorted. S48: Calculate the normal direction of each point, and by comparing the consistency of the normal directions of neighboring points, remove outliers with significantly inconsistent normal directions to enhance the geometric consistency of the point cloud, and check whether the normal direction error is ≤5°; S49: Combining point cloud data from LiDAR and depth camera, registration is performed using the ICP algorithm. The high-resolution information from the depth camera supplements the details missing from the LiDAR, improving the integrity of the point cloud. The registration accuracy is checked to ensure it is ≤0.02m and the detail feature retention rate is ≥90%. S410: Embeds a quality monitoring module during preprocessing to calculate point cloud density, noise level, and coordinate accuracy in real time, and automatically adjusts filtering parameters.

[0023] Preferably, the following steps are included when performing multi-scale point cloud registration and fusion: S51: Use the SIFT algorithm in the PCL library to extract multi-scale point cloud feature points, set the number of feature points to ≥500 per region, detect whether the feature points are uniformly distributed on the blade surface and have high discriminative power, and support the coarse registration process; S52: Initial registration is performed using the RANSAC algorithm. The initial transformation matrix is ​​calculated based on feature point matching. The maximum number of iterations is set to 100, and the error threshold is 0.05m. S53: The ICP algorithm is used for fine registration, with a maximum number of iterations of 200. The convergence condition is that the mean square error is <0.01m or the number of iterations reaches the upper limit. The optimal transformation matrix is ​​calculated iteratively using the least squares method. S54: The multi-view point cloud is fused into a complete model using a voxel fusion algorithm. The voxel size is set to 0.02m × 0.02m × 0.02m, and the average coordinates of the points within the voxel are used to replace the original point cloud. S55: Use a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system. Calculate the transformation parameters using control point pairs and check if the coordinate system transformation error is ≤0.01m. S56: A bilateral filtering algorithm is applied during the fusion process, with spatial weight σs=0.3m and grayscale weight σr=0.2, to suppress noise while preserving blade edge features; S57: Establish a registration quality assessment module to calculate point cloud overlap, registration error, and mean square error in real time, and automatically verify whether the fusion results meet the inspection requirements; S58: The Poisson reconstruction algorithm is used to generate a mesh model of the blade surface. The mesh resolution is set to 0.01m. The model surface is checked to see if it is continuous and free of voids, and the error between it and the original point cloud is ≤0.02m. This supports subsequent defect detection.

[0024] Preferably, the intelligent identification and parameter extraction of key components includes the following steps: S61: Using the pass-through filter in the PCL library, set the height threshold Z>1m to remove ground point clouds, retain the main point cloud of the wind turbine, reduce the number of point clouds by ≥40% after processing, and retain the key areas completely; S62: The RANSAC algorithm is applied to fit the point cloud of the tower into a cylindrical model. The maximum number of iterations is set to 1000, and the inward point threshold is set to 0.05m. The three-dimensional coordinates, radius, and height parameters of the tower center point are calculated. S63: Project the point cloud around the wheel hub onto the XY plane, and use a distance transformation algorithm combined with PCA principal component analysis to calculate the center position and main axis direction of the wheel hub. The positioning accuracy is ≤0.02m and the deviation of the main axis direction is ≤0.5°. S64: The DBSCAN clustering algorithm is used to segment the blade point cloud. The neighborhood radius is set to 0.5m and the minimum number of points is 50. The blade point cloud is separated from the tower / hub. The clustering accuracy is ≥95% and the smoothness of the segmentation edge meets the inspection requirements. S65: Determine the axial direction through PCA analysis of blade point cloud, calculate blade span and twist angle, with measurement error ≤0.1° and span calculation accuracy ≤0.01m; S66: The YOLOv5 algorithm was used to detect defects in the point cloud on the blade surface. A confidence threshold of 0.7 was set to identify crack / corrosion defects. The recall rate was ≥90% and the false positive rate was ≤5%. S67: Store the identified tower, hub, and blade parameters in an SQL database, including geometric parameters, location information, and defect data.

[0025] Preferably, the following steps are included when conducting dynamic environmental adaptive route planning: S71: Uses meteorological sensors to collect wind speed, temperature, and humidity data in real time, and publishes environmental parameter topics through the ROS system to provide dynamic input for flight route planning; S72: Based on the wind turbine's operating status and environmental parameters, a dynamic safety distance model is used to calculate the minimum safe distance between the UAV and the blades; S73: An improved A* algorithm is applied for path planning, with the node expansion cost set as a weighted sum of distance and energy consumption. The path search range is limited to a 200m radius around the wind turbine, generating the shortest inspection path that meets safety constraints. S74: Integrated LiDAR obstacle avoidance module, with an obstacle detection distance threshold of 10m. When an obstacle is detected, it automatically triggers path replanning to adjust the flight path and detour around the danger zone; S75: Automatically adjusts camera exposure parameters according to ambient light intensity, ensuring uniform image brightness without overexposure or underexposure; S76: Employs a multi-sensor fusion algorithm to combine GPS positioning data, IMU attitude data, and LiDAR point cloud data to calculate the UAV's real-time position; S77: Real-time verification of flight path safety. It monitors the drone's position through three-dimensional electronic fence technology. When the drone enters a dangerous area, it automatically triggers emergency braking and adjusts the flight path to a safe area.

[0026] Preferably, when performing multi-objective optimization and path smoothing, the following steps are included: S81: Using path length, energy consumption, and safety distance as quantitative indicators, a multi-objective optimization model is constructed. A linear weighted method is used to transform the multi-objective problem into a single-objective problem, with weight coefficients set to 0.4, 0.3, and 0.3 respectively. S82: The A* algorithm is applied for initial path search, and the node expansion cost is set as the weighted sum of distance and energy consumption. The path search range is limited to a radius of 200m around the wind turbine. S83: Smooth the initial path using a Bézier curve, setting the control point spacing to 0.5m and the curve order to 3; S84: The path is reconstructed using B-spline curves, and the node positions are adjusted to make the curve continuous and satisfy the dynamic constraints of the UAV; S85: Use the particle swarm optimization algorithm to adjust the path parameters, set the number of particles to 50 and the maximum number of iterations to 100, and perform iterative solutions with the dual objectives of minimizing the total path length and minimizing energy consumption; S86: Real-time path safety verification; monitors drone position using 3D electronic fence technology; automatically triggers path replanning when drone enters a dangerous area. S87: Employs multi-sensor fusion positioning, integrating GPS, IMU, and LiDAR data to detect whether the real-time position accuracy of the drone is ≤0.05m.

[0027] As a preferred method, the intelligent inspection and data collection process includes the following steps: S91: The drone automatically takes off and moves to the preset inspection starting point, achieving precise positioning via GPS and IMU, with an initial position error ≤0.3m and stable attitude; S92: Utilizes a PID controller to achieve UAV path tracking, adjusting roll, pitch, and yaw angles to maintain stable flight, with a track deviation ≤0.2m and speed fluctuation ≤0.5m / s; S93: Simultaneously acquires laser point cloud, visible light image, and infrared thermal image data, and performs multi-sensor time synchronization via the PTP protocol; S94: Real-time processing of point cloud data in UAV embedded systems, applying voxel filtering and statistical filtering to remove noise, with a processing latency of ≤80ms; S95: Use the YOLOv5 algorithm to detect blade surface defects in real time, set a confidence threshold of 0.7, mark the defect locations and store them in an SQLite database; S96: Transmits data to ground stations in real time via 5G network, using HDFS distributed storage system; S97: When an abnormal state of the drone is detected, the emergency return-to-home procedure is automatically triggered to plan the shortest path back to the take-off and landing point.

[0028] Preferably, the following steps are included when conducting data analysis and health assessment: S101: Cleans and denoises laser point cloud, visible light image, and infrared thermal image data, and applies the ICP algorithm to achieve spatial alignment of multi-source data; S102: Use the YOLOv5 algorithm to detect and classify surface defects on the blade, set a confidence threshold of 0.7, identify the defect types such as cracks, corrosion, or deformation, and mark their location coordinates; S103: Construct a health assessment model based on the random forest algorithm, inputting defect density, defect area, or runtime parameters, and outputting a health score; S104: The ARIMA time series model is used to predict the remaining service life of the blades. Combined with historical defect growth trend data, the prediction period is set to 6 months, with an error range of ≤5%. S105: Use the Matplotlib library to generate a heatmap of defect distribution and a health score trend curve, and use color mapping to visually display the health status of different areas of the leaf; S106: Store the analysis results in a MySQL database, and design a defect table, health score table, and inspection record table, supporting query and comparative analysis by time / region; S107: The early warning system is automatically triggered when the health score is <60 or a serious defect is detected. Example

[0029] First, laser point cloud data is collected from multiple points directly in front of the wind turbine and at the tips of the three blades. The complete point cloud of the wind turbine is obtained through coarse-to-fine registration and fusion. Next, combining wind turbine parameter information and the geometric features of key components, the point clouds of the tower, hub, and blades are identified. Finally, a flight path for automated inspection of the wind turbine blades is automatically planned. Using the method described in this invention, the wind turbine blade inspection path calculated based on the laser point cloud-based automatic blade inspection path planning method can be applied to the refined inspection of actual wind turbine blades.

[0030] Implementation steps: S111: Set the wind turbine to a locked stop state. In this stop state, the blade stop position is not limited to a positive Y-shape or an inverted Y-shape; it can be any stop position. S112: Collect laser point cloud data of the wind turbine directly in front of it. A-blade laser point cloud data B-blade laser point cloud data C-blade laser point cloud data Three-dimensional mathematical model of wind turbine ,in, This represents the three-dimensional coordinates of the center point of the tower base. Indicates the tower radius, Indicates the height of the tower. Indicates the blade length. This indicates the yaw angle of the wind turbine (the angle between the yaw angle and true north). S113: Unmanned Aerial Vehicle (UAV) Exit Fan Hub The laser point cloud of the wind turbine was collected from a distance of 51.916 meters. Measurements showed that the farthest point cloud distance from the wind turbine hub to the blade was 51.916 meters. However, the blade length of this wind turbine is 60 meters, resulting in a missing point cloud of nearly 10 meters. To ensure the completeness of the frontal laser point cloud data, the laser point cloud at the tip of the wind turbine blade needs to be collected at close range for completion. Therefore, it is necessary to fuse the laser point clouds collected from different locations in stages to obtain a complete frontal point cloud data of the wind turbine without any misalignment. S114: Based on the gimbal attitude and UAV position when collecting laser point clouds from the wind turbine, perform coordinate transformations on the laser point clouds directly in front of the wind turbine and the laser point clouds at the blade tip to the same coordinate system.

[0031] Viewing the collected laser point cloud of the wind turbine unit The collection location is via gimbal attitude And the relative position is converted to the world coordinate system to obtain The formula is expressed as follows: in, For gimbal yaw angle, For the tilt angle of the gimbal, Roll angle of the gimbal; S115: Due to the certain degree of shaking of the UAV in the air, there is an unavoidable deviation in the laser point cloud acquisition. After coarse registration of the point cloud, there is a misalignment phenomenon, so it needs to be corrected. The main idea of ​​fine registration of the point cloud is to register the laser point cloud at the tip of the blade with the laser point cloud directly in front of the wind turbine, calculate its offset, and then compensate and fuse them.

[0032] First, based on the location of the blade data collection, interfering point clouds outside a certain range are removed. Then, point clouds from the tower section are removed from the remaining point clouds to obtain the desired result. Then construct a point cloud. of Tree, The tree construction process is as follows: For point clouds Calculate the variance for each dimension:

[0033] in, It is the first The mean of the dimensions; choose the dimension with the largest variance. As the current segmentation dimension.

[0034] In selected dimensions Above, select the median. As a segmentation value, the point cloud is divided into two subsets:

[0035] right and Repeat the above process until all points are divided or the termination condition is met, thus constructing a point cloud. of Tree.

[0036] Given target point Search for its nearest neighbor. , making the distance minimum, point With point The distance is:

[0037] in, It is the dimension of the point cloud. .

[0038] During the recursive search, after traversing to a leaf node, the target point needs to be determined during backtracking. Is the distance to the current segmentation plane less than the current minimum distance?

[0039] in, It is the partitioning dimension of the current node. It is the splitting value. If this distance is less than the current minimum distance, then we need to enter another subtree to search.

[0040] Calculate separately , , and The minimum distance is the offset between point clouds. Finally, the point clouds that have been compensated for the offset are fused to obtain the final laser point cloud of the wind turbine facing forward. S116: By removing the ground point cloud, tower point cloud, and hub point cloud from the registered and fused wind turbine laser point cloud, the wind turbine blade point cloud can be obtained; S117: The wind turbine base is the origin of the coordinate system. Point clouds with a height less than a certain threshold can be considered ground point clouds, thus ground point cloud removal can be achieved accordingly. The ground point cloud height threshold is: ; S118: Extract a portion of the point cloud within the height threshold range. Calculate the maximum distance between two points in the point cloud. The midpoint between these two points is the center point of the tower, and the maximum distance is the diameter of the tower.

[0041] The tower height threshold is: .

[0042] The tower height threshold range is: ; S119: Project the point cloud around the hub onto the plane along the direction of the wind turbine, and find the center position of the hub through distance transformation.

[0043] The definition of distance transform is to calculate the distance from a non-zero pixel in an image to the nearest zero pixel, which is the shortest distance to the zero pixel. The images processed by distance transform are usually binary images, which are essentially divided into two parts: the background and the object. The object is usually called the foreground target. We usually set the gray value of the foreground target to 255, which is white, and the gray value of the background to 0, which is black. Therefore, the non-zero pixels in the definition are the foreground targets, and the zero pixels are the background. So, the farther the pixels in the foreground target are from the background, the greater the distance.

[0044] Foreground points in a binary image Its Euclidean distance transformation value is: in, It is a set of background points (points with a pixel value of 0); S1110: First, remove the point cloud data at the tower, then remove the point cloud data at the hub, then perform point cloud clustering to obtain the point cloud data of the three blades, and finally calculate the angle between the point cloud data of the three blades and the vertical direction, and sort the blades; S1111: In order to achieve full-coverage and refined inspection of wind turbine units during the drone inspection process, it is necessary to plan flight routes that can be used for automated inspection of wind turbine units to ensure the safety of the drone flight process and the acquisition of images of all wind turbine unit surfaces. Intelligent flight route planning technology can not only automatically generate wind turbine unit inspection routes, but also perform safety checks on the generated wind turbine unit inspection routes.

[0045] Based on the blade identification results, the locations of the blades to be photographed are calculated. Blade inspection is essentially designing flight paths to enable drones to automatically inspect these locations of the blades to be photographed.

[0046] During blade inspection, each blade needs to be inspected on all four sides, front and back. The inspection route includes not only the photo-taking points but also the transfer points from the front to the rear of the wind turbine.

[0047] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for automatic inspection route planning of wind turbine blades based on laser point clouds; characterized in that: It includes the following steps: S11: By establishing a multi-model parameter database, calibrating high-precision equipment, and configuring an environmental monitoring module, accurate pre-configuration of parameters and equipment calibration for multiple wind turbine models are achieved; S12: By locking the shutdown state, set the ring-shaped acquisition points, simultaneously acquire point clouds from multiple locations and integrate spatiotemporal data, and perform accurate acquisition and synchronization of multi-source point cloud data; S13: Apply statistical filtering, Gaussian smoothing, and vertical reference calibration techniques to optimize noise suppression and preprocessing of point cloud data; S14: Employing ICP coarse registration and NDT fine registration technologies, a multi-view point cloud is integrated to generate a wind turbine digital twin model with millimeter-level precision; S15: Utilize deep learning algorithms to automatically identify tower sections or blades, and extract geometric parameters and deformation benchmark models; S16: Calculate the safety distance in real time and design an adaptive yaw compensation algorithm to generate a dynamic inspection path covering the entire surface of the blade; S17: Through The algorithm optimizes the path to reduce attitude adjustments, combines B-spline curves to smooth the trajectory, and integrates a collision detection module to achieve a safe and efficient inspection path; S18: Automatically generates a complete flight path including photo capture points and return points, adjusts exposure parameters in real time, and simultaneously collects multimodal data; S19: Utilizes AI algorithms to automatically identify defects and generate deformation heat maps, combined with health records to predict the remaining service life of the wind turbine.

2. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The following steps are included when performing multi-model parameter pre-configuration and equipment calibration: S21: A database containing geometric parameters such as tower height, blade length, hub diameter, or number of blades is built based on SQL Server. The database enables fast parameter querying and retrieval via API, and provides multi-model parameter version management and automatic update functionality. S22: Point cloud accuracy is tested using a coordinate measuring machine under a standard target. By adjusting the laser emission power, scanning angle, and sampling frequency parameters, the measurement error is controlled within ±2mm. S23: The PTPv2 protocol is used for microsecond-level time synchronization between the UAV positioning system and the laser scanner, and millimeter-level spatial synchronization is achieved through coordinate transformation algorithms to avoid point cloud distortion caused by timing discrepancies; S24: Based on the blade rotation radius and safety distance standards of different aircraft models, the safe distance for drone inspection is determined through three-dimensional spatial calculation. A three-dimensional electronic fence is set up using geofencing technology to automatically restrict drones from entering dangerous areas where blades rotate. S25: A high-precision total station was used to measure the three-dimensional coordinates of the center point of the tower base. By comparing and analyzing the actual measured values ​​with the design parameters, the least squares method was used to correct the model error, and the tower positioning accuracy was found to be ≤±3mm. S26: Based on the surface material characteristics of blades of different models, the optimal scanning angle and power parameters are determined through experimental testing. The scanning parameters of the laser scanner are adjusted to optimize the point cloud density and quality, and the point cloud resolution is checked to see if it reaches 0.5mm. S27: An automatic parameter configuration software module developed based on C#, which automatically loads the corresponding geometric parameters and scanning parameters according to the selected wind turbine model; S28: By monitoring temperature and humidity changes in real time, an environmental parameter compensation model is established to automatically correct the thermal expansion coefficient of the laser scanner, avoiding measurement errors caused by environmental changes; S29: By combining ground reference station data, the positioning accuracy of UAVs in complex terrain is improved to the centimeter level, avoiding blind spots or repeated inspections caused by positioning errors.

3. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The following steps are included when collecting and synchronizing multi-source point cloud data: S31: Using a RIEGL VZ-400i laser scanner, set the scanning angle range to 0°-360°, vertical resolution to 0.01°, and horizontal resolution to 0.05°, check if the point cloud density on the blade surface is ≥200 points / cm²; S32: Equipped with a dual-frequency GNSS receiver and an inertial navigation system, it performs microsecond-level time synchronization via the PTPv2 protocol to detect whether the time error of multi-sensor data acquisition is ≤1μs, thus avoiding point cloud distortion caused by timing discrepancies; S33: A circular data acquisition track is set up directly in front of the wind turbine, containing 8 evenly distributed data acquisition points, with each point remaining for at least 30 seconds; S34: The message_filters module of the ROS system is used to align the timestamps of multi-sensor data, and the point cloud data is fused using the PCL library to seamlessly stitch together the main view and the blade tip point cloud to form a complete wind turbine point cloud; S35: Employs a combination of multi-line and single-line lidar, using a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system, thus avoiding registration failures caused by coordinate system differences; S36: Based on a PID controller, a dynamic exposure control algorithm is implemented to monitor the ambient light intensity in real time and automatically adjust the laser emission power. It detects whether the point cloud brightness is uniform and whether there is overexposure / underexposure, while maintaining a constant scanning speed of 5m / s. S37: Establish a data quality assessment module based on the PCL library to monitor point cloud density, noise level, and coordinate accuracy in real time, and automatically remove data frames that do not meet the standards; S38: Uses HDF5 format to store multi-source point cloud data, establishes an indexing mechanism to achieve fast data retrieval, and records metadata such as collection time, location, and environmental parameters.

4. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The point cloud preprocessing and noise suppression process includes the following steps: S41: Using the pass-through filter in the PCL library, determine the ROI region based on the relative position of the lidar and the wind turbine, set the XYZ coordinate threshold to extract the point cloud of the tower and blades, remove interference from the ground and background environment point cloud, and check whether the number of point clouds is reduced by ≥30% after extraction and the key areas are completely preserved; S42: Apply a statistical outlier removal algorithm to calculate the average distance between each point and its k nearest neighbors. Set a standard deviation threshold of 3σ, remove outliers whose average distance exceeds the threshold, and check whether the effective point cloud retention rate is ≥95% and the noise point removal rate is ≥98%. S43: Using a voxel mesh filtering method, the voxel size is set to 0.05m × 0.05m × 0.05m. The centroid of each point within a voxel is used to replace all points, compressing the point cloud data to 10% of the original data, while maintaining the tower cylindricity error ≤ 0.02m and the blade surface smoothness error ≤ 0.01m; S44: Use a Gaussian filter to smooth the point cloud, set the Gaussian kernel radius to 0.1m, suppress high-frequency noise by weighted averaging of neighborhood point coordinates, maintain surface smoothness, and check whether the surface roughness of the point cloud is reduced by more than 30% and the edge features are preserved intact after filtering; S45: A bilateral filtering algorithm is applied to smooth the ordered point cloud. Combining spatial distance weight and gray-level difference weight, the geometric features of the blade edge are preserved while suppressing noise. The positional deviation of the edge point cloud is detected to be ≤0.03m. S46: Set the search radius to 0.3m using a radius filter, count the number of points in the neighborhood of each point, remove isolated points with fewer than 5 neighbors, and check whether the point cloud density uniformity is improved to ≥90% and whether local sparse areas are effectively filled; S47: The RANSAC algorithm was used to fit the cylindrical tower model and the curved blade model. The model fitting error threshold was set to 0.05m. Outliers that were more than 0.05m away from the model were removed. The accuracy of the point cloud matching the model was checked to ensure that the geometric features were not distorted. S48: Calculate the normal direction of each point, and by comparing the consistency of the normal directions of neighboring points, remove outliers with significantly inconsistent normal directions to enhance the geometric consistency of the point cloud, and check whether the normal direction error is ≤5°; S49: Combining point cloud data from LiDAR and depth camera, registration is performed using the ICP algorithm. The high-resolution information from the depth camera supplements the details missing from the LiDAR, improving the integrity of the point cloud. The registration accuracy is checked to ensure it is ≤0.02m and the detail feature retention rate is ≥90%. S410: Embeds a quality monitoring module during preprocessing to calculate point cloud density, noise level, and coordinate accuracy in real time, and automatically adjusts filtering parameters.

5. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The following steps are included in multi-scale point cloud registration and fusion: S51: Use the SIFT algorithm in the PCL library to extract multi-scale point cloud feature points, set the number of feature points to ≥500 per region, detect whether the feature points are uniformly distributed on the blade surface and have high discriminative power, and support the coarse registration process; S52: Initial registration is performed using the RANSAC algorithm. The initial transformation matrix is ​​calculated based on feature point matching. The maximum number of iterations is set to 100, and the error threshold is 0.05m. S53: The ICP algorithm is used for fine registration, with a maximum number of iterations of 200. The convergence condition is that the mean square error is <0.01m or the number of iterations reaches the upper limit. The optimal transformation matrix is ​​calculated iteratively using the least squares method. S54: The multi-view point cloud is fused into a complete model using a voxel fusion algorithm. The voxel size is set to 0.02m × 0.02m × 0.02m, and the average coordinates of the points within the voxel are used to replace the original point cloud. S55: Use a seven-parameter coordinate transformation to unify multi-source point clouds into the WGS84 coordinate system. Calculate the transformation parameters using control point pairs and check if the coordinate system transformation error is ≤0.01m. S56: A bilateral filtering algorithm is applied during the fusion process, with spatial weight σs=0.3m and grayscale weight σr=0.2, to suppress noise while preserving blade edge features; S57: Establish a registration quality assessment module to calculate point cloud overlap, registration error, and mean square error in real time, and automatically verify whether the fusion results meet the inspection requirements; S58: The Poisson reconstruction algorithm is used to generate a mesh model of the blade surface. The mesh resolution is set to 0.01m. The model surface is checked to see if it is continuous and free of voids, and the error between it and the original point cloud is ≤0.02m. This supports subsequent defect detection.

6. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The intelligent identification and parameter extraction of key components includes the following steps: S61: Using the pass-through filter in the PCL library, set the height threshold Z>1m to remove ground point clouds, retain the main point cloud of the wind turbine, reduce the number of point clouds by ≥40% after processing, and retain the key areas completely; S62: The RANSAC algorithm is applied to fit the point cloud of the tower into a cylindrical model. The maximum number of iterations is set to 1000, and the inward point threshold is set to 0.05m. The three-dimensional coordinates, radius, and height parameters of the tower center point are calculated. S63: Project the point cloud around the wheel hub onto the XY plane, and use a distance transformation algorithm combined with PCA principal component analysis to calculate the center position and main axis direction of the wheel hub. The positioning accuracy is ≤0.02m and the deviation of the main axis direction is ≤0.5°. S64: The DBSCAN clustering algorithm is used to segment the blade point cloud. The neighborhood radius is set to 0.5m and the minimum number of points is 50. The blade point cloud is separated from the tower / hub. The clustering accuracy is ≥95% and the smoothness of the segmentation edge meets the inspection requirements. S65: Determine the axial direction through PCA analysis of blade point cloud, calculate blade span and twist angle, with measurement error ≤0.1° and span calculation accuracy ≤0.01m; S66: The YOLOv5 algorithm was used to detect defects in the point cloud on the blade surface. A confidence threshold of 0.7 was set to identify crack / corrosion defects. The recall rate was ≥90% and the false positive rate was ≤5%. S67: Store the identified tower, hub, and blade parameters in an SQL database, including geometric parameters, location information, and defect data.

7. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The following steps are included in dynamic environment adaptive route planning: S71: Uses meteorological sensors to collect wind speed, temperature, and humidity data in real time, and publishes environmental parameter topics through the ROS system to provide dynamic input for flight route planning; S72: Based on the wind turbine's operating status and environmental parameters, a dynamic safety distance model is used to calculate the minimum safe distance between the UAV and the blades; S73: Application Improvement The algorithm performs path planning, sets the node expansion cost as a weighted sum of distance and energy consumption, and limits the path search range to a 200m radius around the wind turbine, generating the shortest inspection path that meets safety constraints; S74: Integrated LiDAR obstacle avoidance module, with an obstacle detection distance threshold of 10m. When an obstacle is detected, it automatically triggers path replanning to adjust the flight path and detour around the danger zone; S75: Automatically adjusts camera exposure parameters according to ambient light intensity, ensuring uniform image brightness without overexposure or underexposure; S76: Employs a multi-sensor fusion algorithm to combine GPS positioning data, IMU attitude data, and LiDAR point cloud data to calculate the UAV's real-time position; S77: Real-time verification of flight path safety. It monitors the drone's position through three-dimensional electronic fence technology. When the drone enters a dangerous area, it automatically triggers emergency braking and adjusts the flight path to a safe area.

8. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: When performing multi-objective optimization and path smoothing, the following steps are included: S81: Using path length, energy consumption, and safety distance as quantitative indicators, a multi-objective optimization model is constructed. A linear weighted method is used to transform the multi-objective problem into a single-objective problem, with weight coefficients set to 0.4, 0.3, and 0.3 respectively. S82: Application The algorithm performs an initial path search, setting the node expansion cost as a weighted sum of distance and energy consumption, and limiting the path search range to a 200m radius around the wind turbine; S83: Smooth the initial path using a Bézier curve, setting the control point spacing to 0.5m and the curve order to 3; S84: The path is reconstructed using B-spline curves, and the node positions are adjusted to make the curve continuous and satisfy the dynamic constraints of the UAV; S85: Use the particle swarm optimization algorithm to adjust the path parameters, set the number of particles to 50 and the maximum number of iterations to 100, and perform iterative solutions with the dual objectives of minimizing the total path length and minimizing energy consumption; S86: Real-time path safety verification; monitors drone position using 3D electronic fence technology; automatically triggers path replanning when drone enters a dangerous area. S87: Employs multi-sensor fusion positioning, integrating GPS, IMU, and LiDAR data to detect whether the real-time position accuracy of the drone is ≤0.05m.

9. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The intelligent inspection and data collection process includes the following steps: S91: The drone automatically takes off and moves to the preset inspection starting point, achieving precise positioning via GPS and IMU, with an initial position error ≤0.3m and stable attitude; S92: Utilizes a PID controller to achieve UAV path tracking, adjusting roll, pitch, and yaw angles to maintain stable flight, with a track deviation ≤0.2m and speed fluctuation ≤0.5m / s; S93: Simultaneously acquires laser point cloud, visible light image, and infrared thermal image data, and performs multi-sensor time synchronization via the PTP protocol; S94: Real-time processing of point cloud data in UAV embedded systems, applying voxel filtering and statistical filtering to remove noise, with a processing latency of ≤80ms; S95: Use the YOLOv5 algorithm to detect blade surface defects in real time, set a confidence threshold of 0.7, mark the defect locations and store them in an SQLite database; S96: Transmits data to ground stations in real time via 5G network, using HDFS distributed storage system; S97: When an abnormal state of the drone is detected, the emergency return-to-home procedure is automatically triggered to plan the shortest path back to the take-off and landing point.

10. The method for automatic inspection route planning of wind turbine blades based on laser point clouds according to claim 1, characterized in that: The following steps are included when conducting data analysis and health assessments: S101: Cleans and denoises laser point cloud, visible light image, and infrared thermal image data, and applies the ICP algorithm to achieve spatial alignment of multi-source data; S102: Use the YOLOv5 algorithm to detect and classify surface defects on the blade, set a confidence threshold of 0.7, identify the defect types such as cracks, corrosion, or deformation, and mark their location coordinates; S103: Construct a health assessment model based on the random forest algorithm, inputting defect density, defect area, or runtime parameters, and outputting a health score; S104: The ARIMA time series model is used to predict the remaining service life of the blades. Combined with historical defect growth trend data, the prediction period is set to 6 months, with an error range of ≤5%. S105: Use the Matplotlib library to generate a heatmap of defect distribution and a health score trend curve, and use color mapping to visually display the health status of different areas of the leaf; S106: Store the analysis results in a MySQL database, and design a defect table, a health score table, and an inspection record table, supporting query and comparative analysis by time / region; S107: The early warning system is automatically triggered when the health score is less than 60 or a serious defect is detected.

Images