Multi-platform collaborative unmanned inspection task planning method and system for wind power scene

Abstract

The invention provides a multi-platform cooperative unmanned inspection task planning method and system for a wind power scene, and relates to the technical field of wind power plant inspection, and the method comprises the steps: based on a wind power plant three-dimensional scene model, adaptively dividing an inspection region, constructing a reachability path network, calculating an inspection difficulty coefficient according to target detection point information, and calculating a target inspection point; and generating an unmanned aerial vehicle and unmanned vehicle inspection area. A sensor is deployed to collect environment and equipment state data, feature vectors are extracted to establish an inspection constraint model, and main and branch inspection paths are planned on the reachability path network. And performing multi-platform relative positioning by using ultra-wideband radio waves, and dynamically optimizing an inspection path according to the equipment health score and the inspection constraint model. Through multi-platform cooperation, adaptive path planning and intelligent state identification, the wind power plant inspection efficiency and accuracy are improved, and the inspection cost is reduced.

Description

Technical Field

[0001] The present invention relates to wind farm inspection technology, and in particular to a multi-platform collaborative unmanned inspection task planning method and system for wind power scenarios. Background Art

[0002] Wind farms are often located in areas with complex terrain and harsh climates. Traditional manual inspection methods are inefficient and pose safety risks. To improve inspection efficiency and safety, drones and unmanned vehicles are widely used in wind farm inspections. However, due to the widespread distribution of equipment, rugged terrain, and complex environmental variations in wind farms, existing inspection technologies lack the ability to allocate tasks, plan routes, and coordinate across multiple platforms, making it difficult to achieve comprehensive and efficient inspections.

[0003] Furthermore, condition monitoring of wind farm equipment requires precise identification of surface and internal features, but most current inspection systems face technical bottlenecks in environmental perception, dynamic path adjustment, and multi-source data fusion. This makes it difficult for the inspection process to meet the refined requirements of equipment health management, impacting the safety and economic benefits of wind turbine equipment.

[0004] Therefore, there is an urgent need to develop a multi-platform collaborative unmanned inspection task planning method for wind power scenarios, which can improve the efficiency and reliability of wind farm inspections through accurate scenario modeling, multi-source data fusion, intelligent path planning and collaborative control strategies. Summary of the Invention

[0005] The embodiments of the present invention provide a multi-platform collaborative unmanned inspection task planning method and system for wind power scenarios, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, A multi-platform collaborative unmanned inspection task planning method for wind power scenarios is provided, including: Acquire wind farm point cloud data using a dual-band laser scanner, acquire wind farm image data using multi-view oblique photogrammetry, align the wind farm image data based on the wind farm point cloud data, generate a three-dimensional scene model containing wind turbine location information and target detection point information, use an adaptive grid division method based on the three-dimensional scene model to divide the inspection area into multiple inspection blocks, construct an accessibility path network that includes terrain relief coefficient and equipment occlusion coefficient, calculate the inspection difficulty coefficient of each inspection block based on the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; Deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, and obtain environmental parameter data and device status data; use tensor decomposition and feature mapping methods to extract feature vectors from the environmental parameter data and device status data; establish an inspection constraint model based on the three-dimensional scene model; and generate a trunk inspection path and a branch inspection path on the accessibility path network based on the inspection constraint model using a hierarchical progressive path planning algorithm; A multi-platform relative positioning network is constructed based on ultra-wideband radio waves, and the relative position coordinates of drones and unmanned vehicles are obtained in real time on the main inspection path and the branch inspection path. A distributed formation control model is established based on the relative position coordinates. Using the distributed formation control model, the optimal detection angle is calculated based on the target detection point information, and a collaborative detection trajectory of the drone and the unmanned vehicle is generated. Binocular vision cameras and millimeter-wave radars are configured along the collaborative detection trajectory to collect device surface image data and internal feature data; the device surface image data and internal feature data are input into the edge computing node, and the device status characteristics are identified through a deep learning algorithm to generate a device health score; the drone inspection path and the unmanned vehicle inspection path are dynamically optimized and adjusted according to the device health score and the inspection constraint model.

[0007] In an optional embodiment, Acquiring wind farm point cloud data based on a dual-band laser scanner, acquiring wind farm image data using multi-view oblique photogrammetry, registering the wind farm image data based on the wind farm point cloud data, and generating a three-dimensional scene model containing wind turbine location information and target detection point information includes: Acquire wind farm point cloud data based on a dual-band laser scanner, acquire terrain point cloud data using a near-infrared band scanning unit of the dual-band laser scanner, and acquire wind turbine surface point cloud data using a mid-infrared band scanning unit of the dual-band laser scanner. Dynamically adjust the scanning angle according to the terrain complexity of the target area to generate an adaptive density scanning strategy. Perform noise removal and classification processing on the terrain point cloud data and the wind turbine surface point cloud data according to the adaptive density scanning strategy to obtain preprocessed point cloud data. A five-lens tilted camera array is used to acquire multi-view image data of a wind farm. The five-lens tilted camera array includes one vertical view sensor for acquiring vertical view images and four tilt view sensors for acquiring tilt view images. Multiple control points are arranged around the wind turbine and the three-dimensional coordinates of the multiple control points are acquired. The multi-view image data is subjected to distortion correction and color balancing to obtain corrected multi-view image data. Extract image feature points from the corrected multi-view image data and establish a correspondence relationship between the same-name points, and perform bundle adjustment calculation based on the three-dimensional coordinates of multiple control points to obtain camera attitude parameters; extract terrain edge feature points and device contour feature points from the pre-processed point cloud data, and establish a spatial registration relationship between the terrain edge feature points and the device contour feature points and the extracted image feature points; According to the camera posture parameters, the terrain point cloud in the preprocessed point cloud data is roughly aligned with the vertical perspective image to obtain an initial spatial transformation matrix, and the wind turbine surface point cloud is further aligned with the oblique perspective image to obtain an optimized spatial transformation matrix; according to the optimized spatial transformation matrix, the preprocessed point cloud data and the corrected multi-perspective image data are spatially interpolated and texture mapped to generate a three-dimensional scene model, wherein the wind turbine tower coordinates and the nacelle orientation are used as the wind turbine position information, and the blade leading edge position and the hub connection position are marked as target detection point information.

[0008] In an optional embodiment, Based on the three-dimensional scene model, the inspection area is divided into blocks using an adaptive grid division method to obtain multiple inspection blocks. An accessibility path network including the terrain relief coefficient and the equipment occlusion coefficient is constructed. The inspection difficulty coefficient of each inspection block is calculated according to the target detection point information. The generation of drone inspection areas and unmanned vehicle inspection areas includes: Extract terrain elevation data and wind turbine distribution data from the three-dimensional scene model, divide the inspection area into an initial uniform grid, and calculate the elevation gradient value and equipment density value within the initial uniform grid; recursively divide the initial uniform grid based on the elevation gradient value and equipment density value, and subdivide the grid according to a quadtree structure in areas where the elevation gradient value is greater than a preset gradient area and the equipment density value is greater than a preset density threshold, to obtain multiple inspection blocks; Extracting terrain characteristic parameters based on the inspection block, inputting the terrain characteristic parameters into a pre-trained terrain assessment model to obtain a terrain relief coefficient; extracting spatial geometric information of the wind turbine from the three-dimensional scene model, calculating the degree of occlusion of the inspection block by the wind turbine using a ray casting method to obtain a device occlusion coefficient, and constructing an accessibility path network based on the terrain relief coefficient and the device occlusion coefficient; Extract the 3D coordinates and detection azimuth of the target detection point from the 3D scene model, determine the detection distance range and viewing angle range of each target detection point according to the equipment inspection specification, calculate the detection coverage rate of each inspection block using viewpoint accessibility analysis, and calculate the inspection difficulty coefficient based on the detection coverage rate, accessibility path network, and environmental complexity; Based on the deep learning model, features of each inspection block are extracted to obtain the inspection block features. The inspection difficulty coefficient, terrain undulation coefficient and inspection block features are input into a pre-trained regional classification model to obtain the optimal allocation strategy for the inspection blocks. According to the optimal allocation strategy, the inspection blocks are divided into drone inspection areas and unmanned vehicle inspection areas.

[0009] In an optional embodiment, Extracting terrain characteristic parameters based on the inspection block, inputting the terrain characteristic parameters into a pre-trained terrain assessment model to obtain a terrain relief coefficient; extracting spatial geometric information of the wind turbine from the three-dimensional scene model, calculating the degree of occlusion of the inspection block by the wind turbine using a ray casting method to obtain a device occlusion coefficient; constructing an accessibility path network based on the terrain relief coefficient and the device occlusion coefficient includes: An adaptive grid partitioning method is used to divide the inspection area into grid cells. The elevation data of each grid cell is calculated based on the digital elevation model. The elevation gradient between adjacent grid cells is extracted based on the elevation data to obtain the slope change rate. The elevation standard deviation within a specified radius is calculated to obtain the terrain roughness. The terrain position index is obtained based on the relative height difference between the central grid and the surrounding grids. The slope change rate, terrain roughness, and terrain position index are combined into a feature vector, the manually annotated terrain difficulty level is used as a training label to construct a terrain assessment model, a cross-validation method is used to determine the kernel function parameters and penalty factor of the terrain assessment model, the feature vector of each grid cell is input into the terrain assessment model, and a terrain relief coefficient is output; Extract the spatial geometric information of the wind turbine from the three-dimensional scene model and construct a spherical coordinate system at the center of each grid cell; generate a ray direction vector based on the preset azimuth sampling interval and pitch angle sampling interval; use the bounding box acceleration algorithm to calculate the number of intersections between the ray direction vector and the wind turbine, and use the ratio of the number of intersections to the total number of rays as the device occlusion coefficient; Calculating a first weight coefficient of the terrain relief coefficient and a second weight coefficient of the device occlusion coefficient based on the motion characteristics of the UAV, and adding the product of the terrain relief coefficient and the first weight coefficient and the product of the device occlusion coefficient and the second weight coefficient to obtain a comprehensive cost coefficient of the grid unit; An eight-connected undirected graph structure is constructed with the center points of grid cells as nodes. The weighted average of the comprehensive cost coefficients of adjacent grid cells is calculated as the edge cost. The edge cost is used as the input of the minimum spanning tree algorithm. The set of edges that constitute the minimum spanning tree is retained, and the remaining edges are deleted to obtain the final reachability path network.

[0010] In an optional embodiment, Extracting feature vectors from the environmental parameter data and the equipment status data using tensor decomposition and feature mapping methods, establishing an inspection constraint model in combination with the three-dimensional scene model, and generating a trunk inspection path and a branch inspection path on the reachability path network using a hierarchical progressive path planning algorithm based on the inspection constraint model include: Environmental parameter data and device status data are constructed into a third-order tensor according to the sensor quantity dimension, feature dimension, and time step dimension, and the third-order tensor is reduced in dimension using a tensor decomposition method to obtain a core tensor and a modal factor matrix; a radial basis function feature mapping network is trained using the core tensor, and the network parameters of the radial basis function feature mapping network are iteratively optimized to obtain a feature vector by performing nonlinear feature mapping on the core tensor; Constructing a multidimensional constraint model based on the three-dimensional scene model, wherein the multidimensional constraint model includes spatial constraints, time constraints, energy consumption constraints, and safety constraints; calculating a dynamic weight coefficient using the feature vector, multiplying the dynamic weight coefficient with the corresponding constraint and superimposing them to obtain an inspection constraint model; A hierarchical path planning is performed on the accessibility path network. The characteristic vector is used to construct a sampling probability density function on the first layer of grid. Path sampling points are selected based on the sampling probability density function. The path extension step is determined according to the characteristic vector. A trunk inspection path is generated using a path optimization algorithm. The inspection constraint model is used as a path cost function on a sub-grid of the first layer of grid. The inspection point access order is determined based on the inspection point constraint parameters. A branch inspection path connecting each inspection point is generated. The connection between the trunk inspection path and the branch inspection path is smoothed using a spline interpolation method. Through iterative optimization, the smoothed connection is made to meet the kinematic constraint requirements in the spatial constraint.

[0011] In an optional embodiment, A multi-platform relative positioning network is constructed based on ultra-wideband radio waves. The relative position coordinates of drones and unmanned vehicles are obtained in real time on the main inspection path and branch inspection path. A distributed formation control model is established based on the relative position coordinates. The distributed formation control model is used to calculate the optimal detection angle based on the target detection point information, and generate the collaborative detection trajectory of drones and unmanned vehicles, including: Establishing an ultra-wideband radio wave relative positioning network between multiple platforms, using a two-way ranging method to measure the round-trip time delay of radio wave signals, and calculating the relative distance between the UAV and the unmanned vehicle based on the round-trip time delay; Establishing a Kalman filter model to filter the relative distance to obtain a filtered relative distance measurement value, and using trilateration to calculate the relative position coordinates of the UAV and the unmanned vehicle in real time based on the filtered relative distance measurement value; Based on the relative position coordinates, a distributed formation control model is established, in which a desired formation matrix is set in the distributed formation control model, wherein column vectors correspond to the desired relative positions of the UAVs and the UGVs in the formation, and a distributed control law is constructed based on the consistency principle, wherein the distributed control law includes a position error term of the relative position coordinates and a preset trajectory tracking error term; Obtaining target detection point information, constructing an observation utility function based on the target detection point information, and optimizing and solving the observation utility function under pitch angle constraints and azimuth angle constraints to obtain the optimal detection viewing angles of the UAV and the unmanned vehicle; Based on the optimal detection perspective, a trajectory optimization objective function is established. The kinematic equations in the distributed formation control model are used as the dynamic constraints for trajectory optimization, and the expected formation matrix is used as the formation constraint. Based on the dynamic constraints and formation constraints, the trajectory optimization objective function is solved to obtain the expected trajectory that meets the optimal detection perspective. The expected trajectory is used as the preset trajectory, and the preset trajectory is dynamically corrected based on the distributed control law to generate a collaborative detection trajectory for the UAV and the unmanned vehicle.

[0012] In an optional embodiment, Obtain target detection point information, construct an observation utility function based on the target detection point information, and optimize and solve the observation utility function under the conditions of pitch angle constraints and azimuth angle constraints to obtain the optimal detection angles of the UAV and the unmanned vehicle, including: Obtain the three-dimensional spatial coordinates and surface normal vector of the target detection point, establish a ray tracing model based on the three-dimensional spatial coordinates, and calculate the number of obstructions on the detection line of sight; Constructing a multi-target observation utility function, the multi-target observation utility function including a visibility utility term, a resolution utility term, and an occlusion utility term, wherein the visibility utility term is calculated based on the cosine value of the angle between the observation sight line vector and the surface normal vector to obtain a first utility component, the resolution utility term is calculated based on the ratio of the detection distance to a preset distance threshold to obtain a second utility component, and the occlusion utility term is calculated based on the number of obstructions to obtain a third utility component, and performing a weighted combination of the first utility component, the second utility component, and the third utility component to obtain a comprehensive utility value; A spherical coordinate system detection model is established based on the three-dimensional spatial coordinates, the position of the detection platform is mapped into position parameters in the spherical coordinate system, a value range of the position parameters is determined according to the maneuverability of the detection platform, and a constraint condition set including a pitch angle constraint, an azimuth angle constraint, and a detection distance constraint is established; Adopting an adaptive particle swarm optimization algorithm to solve the multi-objective observation utility function, randomly generating an initial particle swarm within the constraint condition set, each particle carries a set of position parameters, and calculating the comprehensive utility value corresponding to each particle; Determining an individual optimal position and a global optimal position according to the comprehensive utility value, calculating a velocity update of the particle swarm based on an adaptive inertia weight, wherein the adaptive inertia weight decays as the number of iterations increases, and superimposing the velocity update with the current position to obtain an updated particle position; Determine whether the updated particle position meets the constraint condition set, perform boundary constraint processing on the particle position that exceeds the constraint range, remap the processed particle position to the spherical coordinate system detection model, repeat the iteration until the preset termination condition is met, and extract the position parameters corresponding to the global optimal position as the optimal detection perspective.

[0013] A second aspect of an embodiment of the present invention provides a multi-platform collaborative unmanned inspection task planning system for wind power scenarios, comprising: The first unit is used to obtain wind farm point cloud data based on a dual-band laser scanner, obtain wind farm image data using multi-view oblique photogrammetry, align the wind farm image data according to the wind farm point cloud data, generate a three-dimensional scene model including wind turbine location information and target detection point information, divide the inspection area into multiple inspection blocks using an adaptive grid division method based on the three-dimensional scene model, construct an accessibility path network including terrain relief coefficient and equipment occlusion coefficient, calculate the inspection difficulty coefficient of each inspection block according to the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; The second unit is configured to deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, obtain environmental parameter data and device status data, extract feature vectors from the environmental parameter data and device status data using tensor decomposition and feature mapping methods, establish an inspection constraint model in combination with the three-dimensional scene model, and generate a trunk inspection path and a branch inspection path on the accessibility path network based on the inspection constraint model using a hierarchical progressive path planning algorithm; The third unit is used to build a multi-platform relative positioning network based on ultra-wideband radio waves, obtain the relative position coordinates of drones and unmanned vehicles on the main inspection path and branch inspection path in real time, establish a distributed formation control model based on the relative position coordinates, use the distributed formation control model to calculate the optimal detection angle based on the target detection point information, generate a collaborative detection trajectory for drones and unmanned vehicles, configure binocular vision cameras and millimeter wave radars along the collaborative detection trajectory, collect device surface image data and internal feature data; input the device surface image data and internal feature data into the edge computing node, identify the device status characteristics through a deep learning algorithm, and generate a device health score; dynamically optimize and adjust the drone inspection path and the unmanned vehicle inspection path based on the device health score and the inspection constraint model.

[0014] The third aspect of the embodiment of the present invention An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] In this embodiment, adaptive grid division and reachability path network construction are used to achieve reasonable division of inspection areas and path planning. Combined with the collaborative operation of drones and unmanned vehicles, a wider inspection area can be covered, improving inspection efficiency. A variety of sensors are used to collect environmental data and equipment status data, and combined with deep learning algorithms to identify equipment status features and generate equipment health scores, thereby more accurately assessing equipment status and achieving precise inspections. Based on the equipment health score and inspection constraint model, the inspection paths of drones and unmanned vehicles are dynamically optimized and adjusted. The inspection plan can be flexibly adjusted according to actual conditions, improving the effectiveness and pertinence of the inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flowchart of a multi-platform collaborative unmanned inspection task planning method for wind power scenarios according to an embodiment of the present invention; Figure 2 This is a structural diagram of a multi-platform collaborative unmanned inspection task planning system for wind power scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1FIG. 1 is a flow chart of a multi-platform collaborative unmanned inspection task planning method for wind power scenarios according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Acquire wind farm point cloud data using a dual-band laser scanner, acquire wind farm image data using multi-view oblique photogrammetry, register the wind farm image data based on the wind farm point cloud data, generate a three-dimensional scene model containing wind turbine location information and target detection point information, segment the inspection area using an adaptive meshing method based on the three-dimensional scene model to obtain multiple inspection blocks, construct an accessibility path network that includes terrain relief coefficients and equipment occlusion coefficients, calculate the inspection difficulty coefficient for each inspection block based on the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; S102. Deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, and obtain environmental parameter data and device status data. Extract feature vectors from the environmental parameter data and device status data using tensor decomposition and feature mapping methods, establish an inspection constraint model based on the three-dimensional scene model, and generate trunk inspection paths and branch inspection paths on the accessibility path network using a hierarchical progressive path planning algorithm based on the inspection constraint model. S103. Construct a multi-platform relative positioning network based on ultra-wideband radio waves, obtain the relative position coordinates of drones and unmanned vehicles on the main inspection path and branch inspection path in real time, establish a distributed formation control model based on the relative position coordinates, use the distributed formation control model to calculate the optimal detection angle based on the target detection point information, generate a collaborative detection trajectory for drones and unmanned vehicles, configure binocular vision cameras and millimeter wave radars along the collaborative detection trajectory, collect device surface image data and internal feature data; input the device surface image data and internal feature data into the edge computing node, identify the device status characteristics through a deep learning algorithm, and generate a device health score; dynamically optimize and adjust the drone inspection path and the unmanned vehicle inspection path based on the device health score and the inspection constraint model.

[0021] Wind farm point cloud data is three-dimensional spatial information acquired by laser scanners. It primarily includes terrain undulations, wind turbine spatial location information (such as tower height, hub center position, blade length, and pitch angle), and surrounding environmental features (such as roads, vegetation, and buildings). This data is used to generate a three-dimensional geometric model of the wind farm, providing a foundation for subsequent inspection and path planning.

[0022] Wind farm image data comes from multi-view oblique photography, including high-definition images of the external surface of wind turbines, panoramic views of wind farms, and detailed images of key areas.

[0023] Target inspection point information is the location in the 3D scene model that requires key inspection, including key parts of the wind turbine (such as the blade root, blade tip, tower connection, nacelle surface, etc.). This information is used to plan inspection routes and assign tasks.

[0024] Environmental data is real-time environmental information collected by sensors, including temperature and humidity, wind speed and direction, and light intensity. It impacts the safety of inspection tasks and the quality of data collection. Equipment status data, acquired through vibration sensors, temperature sensors, and other devices, includes operating vibration amplitude, surface temperature, and acoustic characteristics, reflecting the equipment's operating status.

[0025] Equipment health scores are generated based on collected status data and a deep learning algorithm. The scores are used to assess whether equipment requires maintenance or overhaul. Dynamic route optimization adjusts the inspection trajectories of drones and autonomous vehicles based on the scores and real-time environmental conditions, ensuring flexible and efficient execution of tasks.

[0026] In an optional embodiment, obtaining wind farm point cloud data based on a dual-band laser scanner, obtaining wind farm image data using multi-view oblique photogrammetry, registering the wind farm image data based on the wind farm point cloud data, and generating a three-dimensional scene model including wind turbine location information and target detection point information includes: Acquire wind farm point cloud data based on a dual-band laser scanner, acquire terrain point cloud data using a near-infrared band scanning unit of the dual-band laser scanner, and acquire wind turbine surface point cloud data using a mid-infrared band scanning unit of the dual-band laser scanner. Dynamically adjust the scanning angle according to the terrain complexity of the target area to generate an adaptive density scanning strategy. Perform noise removal and classification processing on the terrain point cloud data and the wind turbine surface point cloud data according to the adaptive density scanning strategy to obtain preprocessed point cloud data. A five-lens tilted camera array is used to acquire multi-view image data of a wind farm. The five-lens tilted camera array includes one vertical view sensor for acquiring vertical view images and four tilt view sensors for acquiring tilt view images. Multiple control points are arranged around the wind turbine and the three-dimensional coordinates of the multiple control points are acquired. The multi-view image data is subjected to distortion correction and color balancing to obtain corrected multi-view image data. Extract image feature points from the corrected multi-view image data and establish a correspondence relationship between the same-name points, and perform bundle adjustment calculation based on the three-dimensional coordinates of multiple control points to obtain camera attitude parameters; extract terrain edge feature points and device contour feature points from the pre-processed point cloud data, and establish a spatial registration relationship between the terrain edge feature points and the device contour feature points and the extracted image feature points; According to the camera posture parameters, the terrain point cloud in the preprocessed point cloud data is roughly aligned with the vertical perspective image to obtain an initial spatial transformation matrix, and the wind turbine surface point cloud is further aligned with the oblique perspective image to obtain an optimized spatial transformation matrix; according to the optimized spatial transformation matrix, the preprocessed point cloud data and the corrected multi-perspective image data are spatially interpolated and texture mapped to generate a three-dimensional scene model, wherein the wind turbine tower coordinates and the nacelle orientation are used as the wind turbine position information, and the blade leading edge position and the hub connection position are marked as target detection point information.

[0027] For example, a dual-band laser scanner is first used to acquire wind farm point cloud data. The near-infrared scanning unit is primarily used to acquire terrain point cloud data, such as ground and vegetation; the mid-infrared scanning unit is primarily used to acquire wind turbine surface point cloud data, such as towers, nacelles, and blades. To improve data acquisition efficiency and accuracy, the scanning angle is dynamically adjusted based on the terrain complexity of the target area. In flat terrain, a larger scanning angle and lower scanning density are used; in complex terrain, such as mountains and hills, a smaller scanning angle and higher scanning density are used to obtain more detailed terrain information. This ultimately generates an adaptive density scanning strategy. After data acquisition based on this strategy, noise removal and classification are performed on the terrain point cloud data and wind turbine surface point cloud data. Statistical filtering can be used to remove noise points, and geometric feature-based classification methods can be used to distinguish between ground points, vegetation points, and wind turbine points. The result is clean, classified point cloud data that includes both terrain and wind turbines.

[0028] Next, a five-lens oblique camera array is used to acquire multi-view image data of the wind farm. This array consists of one vertical downward-facing sensor and four oblique-facing sensors. The vertical sensor acquires vertical-view images, primarily used for terrain texture mapping; the four oblique-facing sensors acquire oblique-view images, primarily used for wind turbine texture mapping. To ensure subsequent precise registration, multiple control points are placed around the wind turbines. A high-precision GPS measurement device is used to obtain their three-dimensional coordinates, such as longitude, latitude, and altitude. Distortion correction and color balancing are then performed on the acquired multi-view image data. For example, camera calibration parameters are used to remove lens distortion, and a color balancing algorithm is used to adjust image colors for a more realistic and natural look. The resulting corrected multi-view image data is then obtained.

[0029] Next, image feature points are extracted from the rectified multi-view image data. The SIFT algorithm can be used to extract feature points, establish point correspondences with the same name, and use the RANSAC algorithm to eliminate mismatched points. A bundle adjustment is performed using the 3D coordinates of multiple control points to obtain camera pose parameters, such as the rotation matrix and translation vector. Terrain edge feature points and device outline feature points are extracted from the preprocessed point cloud data. Edge detection algorithms can be used to extract feature points. Spatial registration relationships are established between these point cloud feature points and the extracted image feature points, for example, using the ICP algorithm for iterative closest point registration.

[0030] Next, the terrain point cloud in the preprocessed point cloud data is coarsely aligned with the vertical view image based on the camera pose parameters to obtain an initial spatial transformation matrix. For example, an affine transformation matrix is used for initial alignment. The wind turbine surface point cloud is further aligned with the oblique view image to obtain an optimized spatial transformation matrix. Finally, based on the optimized spatial transformation matrix, spatial interpolation and texture mapping are performed on the preprocessed point cloud data and the corrected multi-view image data to generate a 3D scene model. The 3D coordinates of the center point at the bottom of the wind turbine tower and the orientation angle of the nacelle are used as the wind turbine position information. In the 3D scene model, the leading edge of the blade and the connection point between the blade and the hub are annotated as target detection point information. For example, a point cloud segmentation algorithm is used to identify the blades and hub and extract feature points. For example, a 3D model containing terrain, vegetation, and textured wind turbines is ultimately generated, in which each wind turbine is annotated with position information and target detection point information.

[0031] In this embodiment, an adaptive density scanning strategy dynamically adjusts the scanning angle and density based on terrain complexity, avoiding redundant data collection and improving data acquisition efficiency. Using a dual-band laser scanner and a five-lens tilting camera array, high-precision, multi-view data can be acquired. Combined with control point information and a precise registration algorithm, a high-precision 3D scene model can be generated. This method not only generates a 3D model containing both terrain and wind turbines, but also annotates wind turbine location information and target detection point information, providing richer information for subsequent inspection and maintenance work.

[0032] In an optional embodiment, the inspection area is divided into blocks based on the three-dimensional scene model using an adaptive grid division method to obtain multiple inspection blocks, and an accessibility path network including a terrain relief coefficient and an equipment occlusion coefficient is constructed. The inspection difficulty coefficient of each inspection block is calculated according to the target detection point information, and the generation of drone inspection areas and unmanned vehicle inspection areas includes: Extract terrain elevation data and wind turbine distribution data from the three-dimensional scene model, divide the inspection area into an initial uniform grid, and calculate the elevation gradient value and equipment density value within the initial uniform grid; recursively divide the initial uniform grid based on the elevation gradient value and equipment density value, and subdivide the grid according to a quadtree structure in areas where the elevation gradient value is greater than a preset gradient area and the equipment density value is greater than a preset density threshold, to obtain multiple inspection blocks; Extracting terrain characteristic parameters based on the inspection block, inputting the terrain characteristic parameters into a pre-trained terrain assessment model to obtain a terrain relief coefficient; extracting spatial geometric information of the wind turbine from the three-dimensional scene model, calculating the degree of occlusion of the inspection block by the wind turbine using a ray casting method to obtain a device occlusion coefficient, and constructing an accessibility path network based on the terrain relief coefficient and the device occlusion coefficient; Extract the 3D coordinates and detection azimuth of the target detection point from the 3D scene model, determine the detection distance range and viewing angle range of each target detection point according to the equipment inspection specification, calculate the detection coverage rate of each inspection block using viewpoint accessibility analysis, and calculate the inspection difficulty coefficient based on the detection coverage rate, accessibility path network, and environmental complexity; Based on the deep learning model, features of each inspection block are extracted to obtain the inspection block features. The inspection difficulty coefficient, terrain undulation coefficient and inspection block features are input into a pre-trained regional classification model to obtain the optimal allocation strategy for the inspection blocks. According to the optimal allocation strategy, the inspection blocks are divided into drone inspection areas and unmanned vehicle inspection areas.

[0033] For example, a 3D scene model of the inspection area is first obtained. This model includes terrain elevation data, the precise location and geometric dimensions of the wind turbines, and other information. Alternatively, high-precision point cloud data is acquired through drone oblique photogrammetry or lidar scanning, and 3D reconstruction is performed to create a 3D scene model. The model records the 3D coordinates and elevation of each point, as well as information such as the wind turbine blade length, tower height, and orientation.

[0034] Then, the inspection area is adaptively gridded. The entire inspection area is initialized to a uniform grid, such as a square grid with a side length of 100 meters. For each initial grid, its elevation gradient value is calculated. The elevation gradient value can be calculated by dividing the maximum elevation difference by the grid side length. At the same time, the density of wind turbines in each grid is calculated. The density of wind turbines represents the number of wind turbines per unit area. A preset elevation gradient threshold, such as 0.1, and a device density threshold, such as 2 units per square kilometer, are set. If the elevation gradient value of a grid is greater than the preset gradient threshold or the device density value is greater than the preset density threshold, the grid is subdivided according to the quadtree structure, that is, divided into four smaller sub-grids. Repeat this process until the elevation gradient values and device density values of all grids are lower than the preset threshold, or the preset minimum grid size is reached. Finally, multiple inspection blocks of different sizes are obtained, which adapt to the complexity of terrain and equipment distribution.

[0035] Next, an accessibility path network is constructed. For each inspection block, terrain characteristic parameters such as average elevation, slope, and roughness are extracted. These parameters are input into a pre-trained terrain assessment model, which can be a deep learning-based regression model. The model outputs a terrain relief coefficient, such as a value between 0 and 1, indicating the degree of terrain relief. At the same time, the spatial geometric information of the wind turbine is extracted from the three-dimensional scene model, and the ray casting method is used to calculate the degree of obstruction of each inspection block by the wind turbine to obtain the equipment obstruction coefficient, such as a value between 0 and 1, indicating the proportion of obstruction. Based on the terrain relief coefficient and the equipment obstruction coefficient, combined with the movement capacity limitations of the inspection equipment, an accessibility path network is constructed. The center point of each inspection block is used as a node. If there are no insurmountable obstacles between the two blocks, such as steep cliffs or densely packed wind turbines, the two nodes are connected. The weight of the edge is determined by the terrain relief coefficient and the equipment obstruction coefficient, which is equal to the weighted average of the two.

[0036] The inspection difficulty coefficient for each inspection block is then calculated. The 3D coordinates and inspection azimuth angles of the target inspection points, such as the specific location of a wind turbine blade, are extracted from the 3D scene model. The inspection distance range and viewing angle range of each target inspection point are determined according to the equipment inspection specifications. Viewpoint accessibility analysis is used to calculate the number of target inspection points that can be covered within each inspection block, resulting in the inspection coverage rate. The inspection difficulty coefficient is calculated by combining the inspection coverage rate, the accessibility path network, and environmental complexity (such as vegetation density and weather conditions).

[0037] Finally, the inspection area is divided. Based on a deep learning model, features are extracted from each inspection area to obtain characteristics such as terrain type and vegetation coverage. The inspection difficulty coefficient, terrain undulation coefficient, and inspection area characteristics are input into a pre-trained regional classification model, such as a support vector machine or random forest model, to determine the optimal allocation strategy for inspection areas. Based on this optimal allocation strategy, the inspection area is divided into areas for drone inspection and areas for unmanned vehicle inspection. For example, areas with high difficulty, undulating terrain, and severe obstruction are allocated to drones, while areas with low difficulty, flat terrain, and good accessibility are allocated to unmanned vehicles.

[0038] In this embodiment, adaptive meshing and reachability path network analysis optimize inspection route planning, reducing inspection time and costs and improving inspection efficiency. Based on the distribution and reachability analysis of target inspection points, all key areas can be covered, improving inspection quality and avoiding omissions. By considering factors such as terrain undulations and equipment obstruction, safety risks during inspections can be reduced, ensuring the safety of both inspectors and equipment.

[0039] In an optional embodiment, terrain characteristic parameters are extracted based on the inspection block, and the terrain characteristic parameters are input into a pre-trained terrain assessment model to obtain a terrain relief coefficient; spatial geometric information of the wind turbine is extracted from the three-dimensional scene model, and a ray casting method is used to calculate the degree of obstruction of the inspection block by the wind turbine to obtain a device obstruction coefficient; and constructing a reachability path network based on the terrain relief coefficient and the device obstruction coefficient includes: An adaptive grid partitioning method is used to divide the inspection area into grid cells. The elevation data of each grid cell is calculated based on the digital elevation model. The elevation gradient between adjacent grid cells is extracted based on the elevation data to obtain the slope change rate. The elevation standard deviation within a specified radius is calculated to obtain the terrain roughness. The terrain position index is obtained based on the relative height difference between the central grid and the surrounding grids. The slope change rate, terrain roughness, and terrain position index are combined into a feature vector, the manually annotated terrain difficulty level is used as a training label to construct a terrain assessment model, a cross-validation method is used to determine the kernel function parameters and penalty factor of the terrain assessment model, the feature vector of each grid cell is input into the terrain assessment model, and a terrain relief coefficient is output; Extract the spatial geometric information of the wind turbine from the three-dimensional scene model and construct a spherical coordinate system at the center of each grid cell; generate a ray direction vector based on the preset azimuth sampling interval and pitch angle sampling interval; use the bounding box acceleration algorithm to calculate the number of intersections between the ray direction vector and the wind turbine, and use the ratio of the number of intersections to the total number of rays as the device occlusion coefficient; Calculating a first weight coefficient of the terrain relief coefficient and a second weight coefficient of the device occlusion coefficient based on the motion characteristics of the UAV, and adding the product of the terrain relief coefficient and the first weight coefficient and the product of the device occlusion coefficient and the second weight coefficient to obtain a comprehensive cost coefficient of the grid unit; An eight-connected undirected graph structure is constructed with the center points of grid cells as nodes. The weighted average of the comprehensive cost coefficients of adjacent grid cells is calculated as the edge cost. The edge cost is used as the input of the minimum spanning tree algorithm. The set of edges that constitute the minimum spanning tree is retained, and the remaining edges are deleted to obtain the final reachability path network.

[0040] For example, the inspection area is first preprocessed. Using digital elevation model (DEM) data, the inspection area is divided into uniformly sized grid cells. The size of each grid cell is adjusted based on actual inspection needs and terrain complexity; for example, it can be set to 5 meters x 5 meters.

[0041] Then, extract the terrain characteristic parameters of each grid cell. Calculate the elevation value of the center point of each grid cell. Calculate the slope change rate based on the difference in the elevation values of the center points of adjacent grid cells. Specifically, the elevation difference between the current grid cell and the eight surrounding adjacent grid cells can be calculated, and the maximum difference divided by the size of the grid cell is the slope change rate. Calculate the standard deviation of the elevation values within a specified radius (for example, a radius of 3 grid cells) to obtain the terrain roughness. Compare the elevation value of the central grid cell with the average of the elevation values of the eight surrounding adjacent grid cells to obtain the terrain position index.

[0042] Next, a terrain assessment model is constructed. The slope change rate, terrain roughness, and terrain position index are combined into a three-dimensional feature vector. The terrain difficulty level of a portion of the grid cells is manually labeled, and the difficulty level can be divided into levels 1 to 5, with level 1 indicating flat terrain and level 5 indicating very rugged terrain. This labeled data is used as training labels to construct a terrain assessment model, such as a support vector machine (SVM) model. Cross-validation is used to determine the model's kernel function parameters and penalty factor. The feature vectors of all grid cells are input into the trained terrain assessment model to obtain the terrain relief coefficient for each grid cell. The terrain relief coefficient ranges from 0 to 1, with larger values indicating greater terrain relief.

[0043] At the same time, the device occlusion coefficient is extracted. The spatial geometric information of equipment such as wind turbines is extracted from the three-dimensional scene model. A spherical coordinate system is constructed at the center of each grid cell. A series of ray direction vectors are generated according to the preset azimuth sampling interval (10 degrees) and pitch angle sampling interval (5 degrees). The bounding box acceleration algorithm is used to calculate the number of intersections between each ray direction vector and the wind turbine. The ratio of the number of intersections to the total number of rays is used as the device occlusion coefficient of the grid cell. The value range of the device occlusion coefficient is between 0 and 1. The larger the value, the higher the degree of device occlusion. If 20 out of 100 rays of a grid cell intersect with the wind turbine, the device occlusion coefficient of the grid cell is 0.2.

[0044] Subsequently, an accessibility path network is constructed. Based on the motion characteristics of the UAV, the weight coefficients of the terrain undulation coefficient and the device occlusion coefficient are determined. If the UAV is sensitive to terrain undulation, the weight of the terrain undulation coefficient can be set to 0.7, and the weight of the device occlusion coefficient can be set to 0.3. Multiply the terrain undulation coefficient of each grid cell by the corresponding weight coefficient, and then multiply the device occlusion coefficient by the corresponding weight coefficient to obtain the comprehensive cost coefficient of the grid cell. Construct an eight-connected undirected graph structure with the center point of the grid cell as the node. Calculate the weighted average of the comprehensive cost coefficients of adjacent grid cells as the cost of the edge between the two nodes. This weighted average reflects the importance of each factor in the final edge cost through a pre-set weight distribution method.

[0045] The edge cost is used as the input to the minimum spanning tree algorithm. The minimum spanning tree is a tree structure that covers all nodes and minimizes the sum of edge costs. It can effectively eliminate redundant connections and retain only the optimal path network. The specific process is as follows: Initialize an empty tree and treat all nodes as independent connected components.

[0046] Sort the edges by cost, adding them to the tree one by one to ensure that no loops are formed.

[0047] Repeat the above steps until all nodes are connected.

[0048] After the minimum spanning tree is constructed, the edge set of the minimum spanning tree is retained and the remaining edges are deleted to obtain the final reachability path network. This network is characterized by a simple structure, high connection efficiency, and guaranteed reachability of all nodes.

[0049] In this embodiment, by considering terrain undulations and equipment obstructions and constructing an accessibility path network, drone inspection routes can be effectively planned, avoiding wasted time in hard-to-reach areas and thus improving inspection efficiency. This method can identify areas with large terrain undulations and severe equipment obstructions and exclude them from the inspection path, thereby reducing the risk of drone collisions or crashes and enhancing inspection safety. Based on refined grid division and terrain feature extraction, combined with terrain assessment models and equipment obstruction coefficients, it is possible to more accurately assess path accessibility, thereby improving the reliability of path planning.

[0050] In an optional embodiment, extracting feature vectors from the environmental parameter data and the device status data using a tensor decomposition and feature mapping method, establishing an inspection constraint model in combination with the three-dimensional scene model, and generating a trunk inspection path and a branch inspection path on the reachability path network using a hierarchical progressive path planning algorithm based on the inspection constraint model include: Environmental parameter data and device status data are constructed into a third-order tensor according to the sensor quantity dimension, feature dimension, and time step dimension, and the third-order tensor is reduced in dimension using a tensor decomposition method to obtain a core tensor and a modal factor matrix; a radial basis function feature mapping network is trained using the core tensor, and the network parameters of the radial basis function feature mapping network are iteratively optimized to obtain a feature vector by performing nonlinear feature mapping on the core tensor; Constructing a multidimensional constraint model based on the three-dimensional scene model, wherein the multidimensional constraint model includes spatial constraints, time constraints, energy consumption constraints, and safety constraints; calculating a dynamic weight coefficient using the feature vector, multiplying the dynamic weight coefficient with the corresponding constraint and superimposing them to obtain an inspection constraint model; A hierarchical path planning is performed on the accessibility path network. The characteristic vector is used to construct a sampling probability density function on the first layer of grid. Path sampling points are selected based on the sampling probability density function. The path extension step is determined according to the characteristic vector. A trunk inspection path is generated using a path optimization algorithm. The inspection constraint model is used as a path cost function on a sub-grid of the first layer of grid. The inspection point access order is determined based on the inspection point constraint parameters. A branch inspection path connecting each inspection point is generated. The connection between the trunk inspection path and the branch inspection path is smoothed using a spline interpolation method. Through iterative optimization, the smoothed connection is made to meet the kinematic constraint requirements in the spatial constraint.

[0051] For example, first, environmental parameter data and device status data are collected from various sensors. This data may include temperature, humidity, vibration, current, as well as device operating status and fault information. Assuming there are 10 sensors, each collecting five types of feature data, and recording data once per minute for one hour, a 10x5x60 third-order tensor can be constructed.

[0052] Then, tensor decomposition is performed on this third-order tensor.

[0053] Methods such as CP decomposition (CANDECOMP / PARAFAC decomposition) or Tucker decomposition can be used. CP decomposition is a method for decomposing tensors (multidimensional arrays) that represents high-dimensional tensors as linear combinations of multiple rank-1 tensors. Tucker decomposition is a more flexible tensor decomposition method that represents a low-rank approximation of a tensor by decomposing it into the product of a core tensor and multiple factor matrices. The original tensor is decomposed into a core tensor and multiple modal factor matrices. Through CP decomposition, a 10x5x60 tensor can be decomposed into a smaller core tensor, such as 2x2x2, and three modal factor matrices corresponding to the sensor dimension, feature dimension, and time step dimension, respectively. These modal factor matrices capture the main variation patterns of the original data in each dimension.

[0054] Next, the obtained core tensor is used to train a radial basis function feature mapping network. The elements in the core tensor are used as the network input. By iteratively optimizing the network parameters, the core tensor is mapped to a higher-dimensional feature space to obtain a feature vector. Specifically, a 2x2x2 core tensor can be expanded into an 8-dimensional vector, which is then mapped to a 16-dimensional feature vector through the radial basis function network. This feature vector contains key information about the environment and device status.

[0055] At the same time, a multidimensional constraint model is constructed based on the 3D scene model. This model can be constructed using CAD models, point cloud data, or other methods. It describes the spatial structure of the inspection environment. Based on this model, spatial constraints, temporal constraints, energy constraints, and safety constraints are constructed. Spatial constraints, for example, limit the robot's range of motion and reachable area; temporal constraints, for example, specify the minimum time to complete an inspection task; energy constraints, for example, limit the robot's maximum power consumption; and safety constraints, for example, prevent the robot from entering dangerous areas or colliding with obstacles.

[0056] Dynamic weight coefficients are calculated using the extracted eigenvectors. If a dimension in the eigenvector represents the probability of device failure, a higher weight can be assigned to the safety constraint based on the value of that dimension. These dynamic weight coefficients are multiplied and superimposed with the corresponding constraints to form the final inspection constraint model.

[0057] Hierarchical path planning is performed based on the reachability path network. First, a sampling probability density function is constructed on the first-layer grid using the eigenvectors. This function reflects the importance or attention of different areas. For example, areas with a high density of detection points have a higher sampling probability. Based on this probability density function, sampling points are selected for the path, ensuring that path planning focuses on key areas.

[0058] When determining the path expansion step size, it is adaptively adjusted based on parameters such as terrain complexity and occlusion coefficients contained in the feature vector. The step size is larger in flat areas and smaller in complex areas to improve the efficiency and accuracy of path planning. Subsequently, a path optimization algorithm (such as A* or Dijkstra's algorithm) is used to generate a backbone inspection path that covers the key areas of the inspection task.

[0059] On the subgrids of the first-level grid, the inspection constraint model is used as the path cost function. This cost function evaluates the quality of the path by comprehensively considering factors such as spatial and energy constraints. Based on the constraint parameters of the inspection points, the order of inspection points is determined, for example, devices with lower health scores are prioritized. Branch inspection paths are generated to connect the inspection points, ensuring that they cover each inspection point while satisfying multi-dimensional constraints.

[0060] To achieve smooth connections between the main and branch paths, spline interpolation is used to smooth the connections. This eliminates sudden changes between paths, improving path smoothness and scalability. During path smoothing, an iterative optimization method is used to ensure that the connections meet the kinematic constraints of the spatial constraints, such as the turning radius of the drone and the acceleration limit of the unmanned vehicle.

[0061] The final generated path includes the main inspection path and the branch inspection path, which overall meets the multi-dimensional constraint requirements in the three-dimensional scene and ensures the fluency and safety of the path during execution through smooth optimization.

[0062] In this embodiment, by using feature vectors to guide path planning, the robot can prioritize inspections of key equipment and areas, reducing unnecessary routes and thus improving inspection efficiency. The introduction of a multidimensional constraint model ensures that the robot avoids dangerous areas, avoids collisions, and meets other safety requirements, thereby enhancing inspection safety. Dynamic weight coefficient adjustment allows for dynamic adjustment of inspection strategies based on changes in the environment and equipment status, optimizing resource allocation, such as power and time.

[0063] In an optional embodiment, a multi-platform relative positioning network is constructed based on ultra-wideband radio waves, and the relative position coordinates of the UAV and the unmanned vehicle are obtained in real time on the main inspection path and the branch inspection path. A distributed formation control model is established based on the relative position coordinates. The distributed formation control model is used to calculate the optimal detection angle based on the target detection point information, and the generation of the collaborative detection trajectory of the UAV and the unmanned vehicle includes: Establishing an ultra-wideband radio wave relative positioning network between multiple platforms, using a two-way ranging method to measure the round-trip time delay of radio wave signals, and calculating the relative distance between the UAV and the unmanned vehicle based on the round-trip time delay; Establishing a Kalman filter model to filter the relative distance to obtain a filtered relative distance measurement value, and using trilateration to calculate the relative position coordinates of the UAV and the unmanned vehicle in real time based on the filtered relative distance measurement value; Based on the relative position coordinates, a distributed formation control model is established, in which a desired formation matrix is set in the distributed formation control model, wherein column vectors correspond to the desired relative positions of the UAVs and the UGVs in the formation, and a distributed control law is constructed based on the consistency principle, wherein the distributed control law includes a position error term of the relative position coordinates and a preset trajectory tracking error term; Obtaining target detection point information, constructing an observation utility function based on the target detection point information, and optimizing and solving the observation utility function under pitch angle constraints and azimuth angle constraints to obtain the optimal detection viewing angles of the UAV and the unmanned vehicle; Based on the optimal detection perspective, a trajectory optimization objective function is established. The kinematic equations in the distributed formation control model are used as the dynamic constraints for trajectory optimization, and the expected formation matrix is used as the formation constraint. Based on the dynamic constraints and formation constraints, the trajectory optimization objective function is solved to obtain the expected trajectory that meets the optimal detection perspective. The expected trajectory is used as the preset trajectory, and the preset trajectory is dynamically corrected based on the distributed control law to generate a collaborative detection trajectory for the UAV and the unmanned vehicle.

[0064] For example, first, ultra-wideband radio wave modules are installed on a drone and an unmanned vehicle to establish a relative positioning network. The round-trip time delay of the radio wave signal is measured using two-way ranging. Specifically, the drone transmits a radio wave signal to the unmanned vehicle. Upon receiving the signal, the unmanned vehicle immediately returns it to the drone. The drone records the time it takes to send and receive the signal, thereby calculating the round-trip time delay. Assuming the signal propagates at the speed of light, the relative distance between the drone and the unmanned vehicle can be calculated based on the round-trip time delay and the speed of light.

[0065] To improve ranging accuracy, a Kalman filter model is used to filter relative distance measurements. The Kalman filter smoothes noisy measurements through two steps: prediction and update, resulting in a more accurate relative distance estimate. Specifically, the relative distance at the next moment is predicted based on the historical motion of the drone and vehicle. This prediction is then corrected based on the current measurement to obtain the filtered relative distance. For example, if the measured relative distance is 10 meters, the Kalman filter yields a filtered relative distance of 9.8 meters.

[0066] Using filtered relative distances, combined with trilateration, the relative position coordinates of drones and unmanned vehicles can be calculated in real time. Three reference nodes with known locations are selected. By measuring the relative distances between the drone / unmanned vehicle and these three reference nodes, the 2D or 3D coordinates of the drone / unmanned vehicle relative to the reference nodes can be calculated. Assuming the distances between the drone and the three reference nodes are 5 meters, 7 meters, and 9 meters, respectively, the drone's position coordinates (x, y, z) in the reference coordinate system can be calculated based on these distances.

[0067] Next, a distributed formation control model is established based on the calculated relative position coordinates. Within this model, a desired formation matrix is set to describe the desired relative positions of the drones and unmanned vehicles within the formation. A 2x2 matrix can be set to represent the desired lateral distance of 2 meters and longitudinal distance of 1 meter between the drones and unmanned vehicles. Simultaneously, a distributed control law is constructed based on the consistency principle. This control law incorporates a position error term for the relative position coordinates and a preset trajectory tracking error term. By adjusting the parameters in the control law, the drones and unmanned vehicles can be controlled to maintain the desired formation and track the preset trajectory.

[0068] Obtain target detection point information, such as the target's location, size, and shape. Based on this information, construct an observation utility function to evaluate observation performance at different viewing angles. Within the constraints of pitch and azimuth angles, optimize the observation utility function, for example, using gradient descent to find its maximum value. This yields the optimal detection angles for both drones and autonomous vehicles. Assuming the target is 5 meters in front of you at an altitude of 2 meters, the optimal pitch angle for the drone can be calculated to be -30 degrees, and the azimuth angle to be 0 degrees.

[0069] Based on the optimal detection viewpoint, a trajectory optimization objective function is established. The kinematic equations in the distributed formation control model serve as dynamic constraints for trajectory optimization, and the desired formation matrix serves as a formation constraint. For example, the requirement is that the UAVs and vehicles reach the optimal detection viewpoint while maintaining the desired formation. Based on the dynamic and formation constraints, the trajectory optimization objective function is solved to obtain the desired trajectory that satisfies the optimal detection viewpoint.

[0070] Finally, the calculated desired trajectory is used as the preset trajectory and input into the distributed control law. The preset trajectory is dynamically corrected to generate the collaborative detection trajectory for the drone and unmanned vehicle. The control law adjusts the control instructions of the drone and unmanned vehicle in real time based on their current states and the preset trajectory, ensuring that they fly / drive along the desired trajectory and ultimately reach the optimal detection angle.

[0071] In this embodiment, the relative positioning network of ultra-wideband radio waves enables high-precision positioning between multiple platforms, thereby ensuring that drones and unmanned vehicles are always in the correct position when performing tasks. By setting up a distributed formation control model and the desired formation matrix, drones and unmanned vehicles can work together and still perform tasks efficiently in complex environments. The calculation of the optimal detection perspective can ensure that the target detection points are fully covered, improving inspection efficiency. In addition, the trajectory optimization process ensures that each path can be adjusted in real time in a dynamic environment, maximizing the efficiency and accuracy of task execution. Distributed control laws and dynamic corrections enable drones and unmanned vehicles to work together efficiently, ensure the smooth completion of tasks, and have strong adaptability to cope with different task scenarios and changes.

[0072] In an optional embodiment, target detection point information is obtained, an observation utility function is constructed based on the target detection point information, and the observation utility function is optimized and solved under pitch angle constraints and azimuth angle constraints to obtain the optimal detection viewing angles of the UAV and the unmanned vehicle, including: Obtain the three-dimensional spatial coordinates and surface normal vector of the target detection point, establish a ray tracing model based on the three-dimensional spatial coordinates, and calculate the number of obstructions on the detection line of sight; Constructing a multi-target observation utility function, the multi-target observation utility function including a visibility utility term, a resolution utility term, and an occlusion utility term, wherein the visibility utility term is calculated based on the cosine value of the angle between the observation sight line vector and the surface normal vector to obtain a first utility component, the resolution utility term is calculated based on the ratio of the detection distance to a preset distance threshold to obtain a second utility component, and the occlusion utility term is calculated based on the number of obstructions to obtain a third utility component, and performing a weighted combination of the first utility component, the second utility component, and the third utility component to obtain a comprehensive utility value; A spherical coordinate system detection model is established based on the three-dimensional spatial coordinates, the position of the detection platform is mapped into position parameters in the spherical coordinate system, a value range of the position parameters is determined according to the maneuverability of the detection platform, and a constraint condition set including a pitch angle constraint, an azimuth angle constraint, and a detection distance constraint is established; Adopting an adaptive particle swarm optimization algorithm to solve the multi-objective observation utility function, randomly generating an initial particle swarm within the constraint condition set, each particle carries a set of position parameters, and calculating the comprehensive utility value corresponding to each particle; Determining an individual optimal position and a global optimal position according to the comprehensive utility value, calculating a velocity update of the particle swarm based on an adaptive inertia weight, wherein the adaptive inertia weight decays as the number of iterations increases, and superimposing the velocity update with the current position to obtain an updated particle position; Determine whether the updated particle position meets the constraint condition set, perform boundary constraint processing on the particle position that exceeds the constraint range, remap the processed particle position to the spherical coordinate system detection model, repeat the iteration until the preset termination condition is met, and extract the position parameters corresponding to the global optimal position as the optimal detection perspective.

[0073] For example, first, obtain the 3D spatial coordinate information and surface normal vector information of the target detection point. A 3D laser scanner can be used to acquire 3D point cloud data of the target scene. Then, a point cloud processing algorithm is used to extract the 3D coordinates (x, y, z) and surface normal vectors (nx, ny, nz) of the target detection point. Assume that the 3D coordinates of a target detection point are (10, 20, 5) and the surface normal vector is (0, 0, 1).

[0074] Next, a ray tracing model is built based on the acquired 3D spatial coordinates to calculate the number of obstructions in the detection line of sight. A ray is constructed, starting from the detection platform's position and ending at the target detection point. The number of intersecting objects is determined by determining where this ray intersects with other objects in the scene. This is the number of obstructions. Assuming the detection platform is at (0, 0, 0), the target detection point is at (10, 20, 5), and the ray intersects two obstacles, the number of obstructions is 2.

[0075] Then, a multi-objective observation utility function is constructed, which contains three utility terms: visibility utility term, resolution utility term, and occlusion utility term.

[0076] The visibility utility term is calculated based on the cosine of the angle between the observation line vector and the surface normal vector. The observation line vector is the vector pointing from the detection platform to the target detection point. The larger the cosine of the angle, the more perpendicular the observation angle is to the target surface, and the better the visibility. Assuming the observation line vector is (1, 2, 0.5) and the surface normal vector is (0, 0, 1), the cosine of the angle is 0.447.

[0077] The resolution utility term is calculated as the ratio of the detection distance to the preset distance threshold. The detection distance is the distance between the detection platform and the target detection point. The preset distance threshold is set based on actual needs. The smaller the detection distance, the higher the resolution. Assuming the detection distance is 22.36 and the preset distance threshold is 30, the ratio is 0.745.

[0078] The occlusion utility term is calculated based on the number of occluders. The greater the number of occluders, the higher the degree of occlusion and the lower the utility value. For example, if there are 2 occluders, the occlusion utility value is 0.5 (assuming the maximum number of occluders is 4, the occlusion utility value = 1 - number of occluders / 4).

[0079] The three utility items are weighted together to obtain a comprehensive utility value. Assuming the visibility utility item has a weight of 0.4, the resolution utility item has a weight of 0.3, and the occlusion utility item has a weight of 0.3, the comprehensive utility value is 0.4*0.447+0.3*0.745+0.3*0.5=0.562.

[0080] Next, a spherical coordinate detection model is established based on the three-dimensional spatial coordinates. The position of the detection platform is mapped to position parameters in the spherical coordinate system, including azimuth, pitch angle, and detection distance. The value range of these position parameters is determined based on the maneuverability of the detection platform. Assume that the azimuth range is 0 to 360 degrees, the pitch angle range is -30 to 30 degrees, and the detection distance range is 10 to 50 meters. Then, a constraint set is established, including pitch angle constraints, azimuth constraints, and detection distance constraints.

[0081] An adaptive particle swarm algorithm is used to find the optimal detection angle. 100 initial particles are randomly generated within the constraints, each containing a set of position parameters. The overall utility value corresponding to each particle is calculated, and the individual historical optimal position and the swarm's global optimal position are recorded. An adaptive inertia weight mechanism is introduced, with the weight initially set to 0.9 and decreasing linearly to 0.4 with the number of iterations. This ensures both global search capability in the early stages and local convergence performance in the later stages. A velocity update is calculated based on the particle's current velocity, individual optimal position, and global optimal position, and then superimposed on the current position to obtain the new position.

[0082] Boundary constraints are applied to particle positions that exceed the constraints after the update. Periodic boundary conditions are used to handle azimuth angle violations, while rebound boundary conditions are used to handle pitch angle and detection range violations. The processed positions are remapped to the spherical coordinate system, and the iterative optimization process continues. When the improvement in the comprehensive utility value is less than 0.001 after 50 consecutive iterations, the algorithm is considered converged, and the global optimal position is extracted as the optimal detection angle.

[0083] For example, if the azimuth angle corresponding to the global optimal position is 120 degrees, the pitch angle is 15 degrees, and the detection distance is 25 meters, then the optimal detection viewing angle is (120, 15, 25).

[0084] In this embodiment, by comprehensively considering visibility, resolution, and occlusion utility to construct a multi-target observation utility function, and optimizing the solution under constraints, accurate target detection of drones and unmanned vehicles in complex environments can be achieved. This method improves the rationality of the detection perspective and the observation efficiency, ensuring the comprehensive acquisition of target surface features. The ray tracing model is used to effectively evaluate the impact of occlusion and improve the robustness of the observation model. The detection model constructed based on the spherical coordinate system optimizes the spatial positioning capability of the detection platform, and the adaptive particle swarm optimization algorithm enhances the global search capability and local convergence speed of the solution process. The overall solution improves the adaptability and flexibility of the detection system and can be widely used in monitoring and navigation tasks in dynamic environments.

[0085] Figure 2 FIG. 1 is a structural diagram of a multi-platform collaborative unmanned inspection task planning system for wind power scenarios according to an embodiment of the present invention. Figure 2 As shown, the system includes: The first unit is used to obtain wind farm point cloud data based on a dual-band laser scanner, obtain wind farm image data using multi-view oblique photogrammetry, align the wind farm image data according to the wind farm point cloud data, generate a three-dimensional scene model including wind turbine location information and target detection point information, divide the inspection area into multiple inspection blocks using an adaptive grid division method based on the three-dimensional scene model, construct an accessibility path network including terrain relief coefficient and equipment occlusion coefficient, calculate the inspection difficulty coefficient of each inspection block according to the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; The second unit is configured to deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, obtain environmental parameter data and device status data, extract feature vectors from the environmental parameter data and device status data using tensor decomposition and feature mapping methods, establish an inspection constraint model in combination with the three-dimensional scene model, and generate a trunk inspection path and a branch inspection path on the accessibility path network based on the inspection constraint model using a hierarchical progressive path planning algorithm; The third unit is used to build a multi-platform relative positioning network based on ultra-wideband radio waves, obtain the relative position coordinates of drones and unmanned vehicles on the main inspection path and branch inspection path in real time, establish a distributed formation control model based on the relative position coordinates, use the distributed formation control model to calculate the optimal detection angle based on the target detection point information, generate a collaborative detection trajectory for drones and unmanned vehicles, configure binocular vision cameras and millimeter wave radars along the collaborative detection trajectory, collect device surface image data and internal feature data; input the device surface image data and internal feature data into the edge computing node, identify the device status characteristics through a deep learning algorithm, and generate a device health score; dynamically optimize and adjust the drone inspection path and the unmanned vehicle inspection path based on the device health score and the inspection constraint model.

[0086] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0087] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0088] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-platform collaborative unmanned inspection task planning method for wind power scenarios, characterized by: include: Acquire wind farm point cloud data using a dual-band laser scanner, acquire wind farm image data using multi-view oblique photogrammetry, align the wind farm image data based on the wind farm point cloud data, generate a three-dimensional scene model containing wind turbine location information and target detection point information, use an adaptive grid division method based on the three-dimensional scene model to divide the inspection area into multiple inspection blocks, construct an accessibility path network that includes terrain relief coefficient and equipment occlusion coefficient, calculate the inspection difficulty coefficient of each inspection block based on the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; Deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, and obtain environmental parameter data and device status data; use tensor decomposition and feature mapping methods to extract feature vectors from the environmental parameter data and device status data; establish an inspection constraint model based on the three-dimensional scene model; and generate a trunk inspection path and a branch inspection path on the accessibility path network based on the inspection constraint model using a hierarchical progressive path planning algorithm; A multi-platform relative positioning network is constructed based on ultra-wideband radio waves. The relative position coordinates of drones and unmanned vehicles are obtained in real time on the main inspection path and branch inspection paths. A distributed formation control model is established based on the relative position coordinates. The distributed formation control model is used to calculate the optimal detection angle based on the target detection point information, and a collaborative detection trajectory for the drone and unmanned vehicle is generated. Binocular vision cameras and millimeter-wave radars are deployed along the collaborative detection trajectory to collect device surface image data and internal feature data. Inputting the device surface image data and internal feature data into the edge computing node, identifying the device status features through a deep learning algorithm, and generating a device health score; The UAV inspection path and the unmanned vehicle inspection path are dynamically optimized and adjusted according to the equipment health score and the inspection constraint model.

2. The method according to claim 1, characterized in that Acquiring wind farm point cloud data based on a dual-band laser scanner, acquiring wind farm image data using multi-view oblique photogrammetry, registering the wind farm image data based on the wind farm point cloud data, and generating a three-dimensional scene model containing wind turbine location information and target detection point information includes: Acquire wind farm point cloud data based on a dual-band laser scanner, acquire terrain point cloud data using a near-infrared band scanning unit of the dual-band laser scanner, and acquire wind turbine surface point cloud data using a mid-infrared band scanning unit of the dual-band laser scanner. Dynamically adjust the scanning angle according to the terrain complexity of the target area to generate an adaptive density scanning strategy. Perform noise removal and classification processing on the terrain point cloud data and the wind turbine surface point cloud data according to the adaptive density scanning strategy to obtain preprocessed point cloud data. A five-lens tilted camera array is used to acquire multi-view image data of a wind farm. The five-lens tilted camera array includes one vertical view sensor for acquiring vertical view images and four tilt view sensors for acquiring tilt view images. Multiple control points are arranged around the wind turbine and the three-dimensional coordinates of the multiple control points are acquired. The multi-view image data is subjected to distortion correction and color balancing to obtain corrected multi-view image data. Extract image feature points from the corrected multi-view image data and establish a correspondence relationship between the same-name points, and perform bundle adjustment calculation based on the three-dimensional coordinates of multiple control points to obtain camera attitude parameters; extract terrain edge feature points and device contour feature points from the pre-processed point cloud data, and establish a spatial registration relationship between the terrain edge feature points and the device contour feature points and the extracted image feature points; According to the camera posture parameters, the terrain point cloud in the preprocessed point cloud data is roughly aligned with the vertical perspective image to obtain an initial spatial transformation matrix, and the wind turbine surface point cloud is further aligned with the oblique perspective image to obtain an optimized spatial transformation matrix; according to the optimized spatial transformation matrix, the preprocessed point cloud data and the corrected multi-perspective image data are spatially interpolated and texture mapped to generate a three-dimensional scene model, wherein the wind turbine tower coordinates and the nacelle orientation are used as the wind turbine position information, and the blade leading edge position and the hub connection position are marked as target detection point information.

3. The method according to claim 1, characterized in that Based on the three-dimensional scene model, the inspection area is divided into blocks using an adaptive grid division method to obtain multiple inspection blocks. An accessibility path network including the terrain relief coefficient and the equipment occlusion coefficient is constructed. The inspection difficulty coefficient of each inspection block is calculated according to the target detection point information. The generation of drone inspection areas and unmanned vehicle inspection areas includes: Extract terrain elevation data and wind turbine distribution data from the three-dimensional scene model, divide the inspection area into an initial uniform grid, and calculate the elevation gradient value and equipment density value within the initial uniform grid; recursively divide the initial uniform grid based on the elevation gradient value and equipment density value, and subdivide the grid according to a quadtree structure in areas where the elevation gradient value is greater than a preset gradient area and the equipment density value is greater than a preset density threshold, to obtain multiple inspection blocks; Extracting terrain characteristic parameters based on the inspection block, inputting the terrain characteristic parameters into a pre-trained terrain assessment model to obtain a terrain relief coefficient; extracting spatial geometric information of the wind turbine from the three-dimensional scene model, calculating the degree of occlusion of the inspection block by the wind turbine using a ray casting method to obtain a device occlusion coefficient, and constructing an accessibility path network based on the terrain relief coefficient and the device occlusion coefficient; Extract the 3D coordinates and detection azimuth of the target detection point from the 3D scene model, determine the detection distance range and viewing angle range of each target detection point according to the equipment inspection specification, calculate the detection coverage rate of each inspection block using viewpoint accessibility analysis, and calculate the inspection difficulty coefficient based on the detection coverage rate, accessibility path network, and environmental complexity; Based on the deep learning model, features of each inspection block are extracted to obtain the inspection block features. The inspection difficulty coefficient, terrain undulation coefficient and inspection block features are input into a pre-trained regional classification model to obtain the optimal allocation strategy for the inspection blocks. According to the optimal allocation strategy, the inspection blocks are divided into drone inspection areas and unmanned vehicle inspection areas.

4. The method according to claim 3, characterized in that Extracting terrain characteristic parameters based on the inspection block, inputting the terrain characteristic parameters into a pre-trained terrain assessment model to obtain a terrain relief coefficient; extracting spatial geometric information of the wind turbine from the three-dimensional scene model, calculating the degree of occlusion of the inspection block by the wind turbine using a ray casting method to obtain a device occlusion coefficient; constructing an accessibility path network based on the terrain relief coefficient and the device occlusion coefficient includes: An adaptive grid partitioning method is used to divide the inspection area into grid cells. The elevation data of each grid cell is calculated based on the digital elevation model. The elevation gradient between adjacent grid cells is extracted based on the elevation data to obtain the slope change rate. The elevation standard deviation within a specified radius is calculated to obtain the terrain roughness. The terrain position index is obtained based on the relative height difference between the central grid and the surrounding grids. The slope change rate, terrain roughness, and terrain position index are combined into a feature vector, the manually annotated terrain difficulty level is used as a training label to construct a terrain assessment model, a cross-validation method is used to determine the kernel function parameters and penalty factor of the terrain assessment model, the feature vector of each grid cell is input into the terrain assessment model, and a terrain relief coefficient is output; Extract the spatial geometric information of the wind turbine from the three-dimensional scene model and construct a spherical coordinate system at the center of each grid cell; generate a ray direction vector based on the preset azimuth sampling interval and pitch angle sampling interval; use the bounding box acceleration algorithm to calculate the number of intersections between the ray direction vector and the wind turbine, and use the ratio of the number of intersections to the total number of rays as the device occlusion coefficient; Calculating a first weight coefficient of the terrain relief coefficient and a second weight coefficient of the device occlusion coefficient based on the motion characteristics of the UAV, and adding the product of the terrain relief coefficient and the first weight coefficient and the product of the device occlusion coefficient and the second weight coefficient to obtain a comprehensive cost coefficient of the grid unit; An eight-connected undirected graph structure is constructed with the center points of grid cells as nodes. The weighted average of the comprehensive cost coefficients of adjacent grid cells is calculated as the edge cost. The edge cost is used as the input of the minimum spanning tree algorithm. The set of edges that constitute the minimum spanning tree is retained, and the remaining edges are deleted to obtain the final reachability path network.

5. The method according to claim 1, wherein Extracting feature vectors from the environmental parameter data and the equipment status data using tensor decomposition and feature mapping methods, establishing an inspection constraint model in combination with the three-dimensional scene model, and generating a trunk inspection path and a branch inspection path on the reachability path network using a hierarchical progressive path planning algorithm based on the inspection constraint model include: Environmental parameter data and device status data are constructed into a third-order tensor according to the sensor quantity dimension, feature dimension, and time step dimension, and the third-order tensor is reduced in dimension using a tensor decomposition method to obtain a core tensor and a modal factor matrix; a radial basis function feature mapping network is trained using the core tensor, and the network parameters of the radial basis function feature mapping network are iteratively optimized to obtain a feature vector by performing nonlinear feature mapping on the core tensor; Constructing a multidimensional constraint model based on the three-dimensional scene model, wherein the multidimensional constraint model includes spatial constraints, time constraints, energy consumption constraints, and safety constraints; calculating a dynamic weight coefficient using the feature vector, multiplying the dynamic weight coefficient with the corresponding constraint and superimposing them to obtain an inspection constraint model; A hierarchical path planning is performed on the accessibility path network. The characteristic vector is used to construct a sampling probability density function on the first layer of grid. Path sampling points are selected based on the sampling probability density function. The path extension step is determined according to the characteristic vector. A trunk inspection path is generated using a path optimization algorithm. The inspection constraint model is used as a path cost function on a sub-grid of the first layer of grid. The inspection point access order is determined based on the inspection point constraint parameters. A branch inspection path connecting each inspection point is generated. The connection between the trunk inspection path and the branch inspection path is smoothed using a spline interpolation method. Through iterative optimization, the smoothed connection is made to meet the kinematic constraint requirements in the spatial constraint.

6. The method according to claim 1, characterized in that A multi-platform relative positioning network is constructed based on ultra-wideband radio waves. The relative position coordinates of drones and unmanned vehicles are obtained in real time on the main inspection path and branch inspection path. A distributed formation control model is established based on the relative position coordinates. The distributed formation control model is used to calculate the optimal detection angle based on the target detection point information, and generate the collaborative detection trajectory of drones and unmanned vehicles, including: Establishing an ultra-wideband radio wave relative positioning network between multiple platforms, using a two-way ranging method to measure the round-trip time delay of radio wave signals, and calculating the relative distance between the UAV and the unmanned vehicle based on the round-trip time delay; Establishing a Kalman filter model to filter the relative distance to obtain a filtered relative distance measurement value, and using trilateration to calculate the relative position coordinates of the UAV and the unmanned vehicle in real time based on the filtered relative distance measurement value; Based on the relative position coordinates, a distributed formation control model is established, in which a desired formation matrix is set in the distributed formation control model, wherein column vectors correspond to the desired relative positions of the UAVs and the UGVs in the formation, and a distributed control law is constructed based on the consistency principle, wherein the distributed control law includes a position error term of the relative position coordinates and a preset trajectory tracking error term; Obtaining target detection point information, constructing an observation utility function based on the target detection point information, and optimizing and solving the observation utility function under pitch angle constraints and azimuth angle constraints to obtain the optimal detection viewing angles of the UAV and the unmanned vehicle; Based on the optimal detection perspective, a trajectory optimization objective function is established. The kinematic equations in the distributed formation control model are used as the dynamic constraints for trajectory optimization, and the expected formation matrix is used as the formation constraint. Based on the dynamic constraints and formation constraints, the trajectory optimization objective function is solved to obtain the expected trajectory that meets the optimal detection perspective. The expected trajectory is used as the preset trajectory, and the preset trajectory is dynamically corrected based on the distributed control law to generate a collaborative detection trajectory for the UAV and the unmanned vehicle.

7. The method according to claim 6, characterized in that Obtain target detection point information, construct an observation utility function based on the target detection point information, and optimize and solve the observation utility function under the conditions of pitch angle constraints and azimuth angle constraints to obtain the optimal detection angles of the UAV and the unmanned vehicle, including: Obtain the three-dimensional spatial coordinates and surface normal vector of the target detection point, establish a ray tracing model based on the three-dimensional spatial coordinates, and calculate the number of obstructions on the detection line of sight; Constructing a multi-target observation utility function, the multi-target observation utility function including a visibility utility term, a resolution utility term, and an occlusion utility term, wherein the visibility utility term is calculated based on the cosine value of the angle between the observation sight line vector and the surface normal vector to obtain a first utility component, the resolution utility term is calculated based on the ratio of the detection distance to a preset distance threshold to obtain a second utility component, and the occlusion utility term is calculated based on the number of obstructions to obtain a third utility component, and performing a weighted combination of the first utility component, the second utility component, and the third utility component to obtain a comprehensive utility value; A spherical coordinate system detection model is established based on the three-dimensional spatial coordinates, the position of the detection platform is mapped into position parameters in the spherical coordinate system, a value range of the position parameters is determined according to the maneuverability of the detection platform, and a constraint condition set including a pitch angle constraint, an azimuth angle constraint, and a detection distance constraint is established; Adopting an adaptive particle swarm optimization algorithm to solve the multi-objective observation utility function, randomly generating an initial particle swarm within the constraint condition set, each particle carries a set of position parameters, and calculating the comprehensive utility value corresponding to each particle; Determining an individual optimal position and a global optimal position according to the comprehensive utility value, calculating a velocity update of the particle swarm based on an adaptive inertia weight, wherein the adaptive inertia weight decays as the number of iterations increases, and superimposing the velocity update with the current position to obtain an updated particle position; Determine whether the updated particle position meets the constraint condition set, perform boundary constraint processing on the particle position that exceeds the constraint range, remap the processed particle position to the spherical coordinate system detection model, repeat the iteration until the preset termination condition is met, and extract the position parameters corresponding to the global optimal position as the optimal detection perspective.

8. A multi-platform collaborative unmanned inspection task planning system for wind power scenarios, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain wind farm point cloud data based on a dual-band laser scanner, obtain wind farm image data using multi-view oblique photogrammetry, align the wind farm image data according to the wind farm point cloud data, generate a three-dimensional scene model including wind turbine location information and target detection point information, divide the inspection area into multiple inspection blocks using an adaptive grid division method based on the three-dimensional scene model, construct an accessibility path network including terrain relief coefficient and equipment occlusion coefficient, calculate the inspection difficulty coefficient of each inspection block according to the target detection point information, and generate drone inspection areas and unmanned vehicle inspection areas; The second unit is configured to deploy environmental sensors and device status sensors at key nodes of the accessibility path network, collect environmental data and device status data, perform wavelet noise reduction and feature extraction, obtain environmental parameter data and device status data, extract feature vectors from the environmental parameter data and device status data using tensor decomposition and feature mapping methods, establish an inspection constraint model in combination with the three-dimensional scene model, and generate a trunk inspection path and a branch inspection path on the accessibility path network based on the inspection constraint model using a hierarchical progressive path planning algorithm; The third unit is used to build a multi-platform relative positioning network based on ultra-wideband radio waves, obtain the relative position coordinates of drones and unmanned vehicles on the main inspection path and branch inspection path in real time, establish a distributed formation control model based on the relative position coordinates, use the distributed formation control model to calculate the optimal detection angle based on the target detection point information, generate a collaborative detection trajectory for the drone and unmanned vehicle, deploy a binocular vision camera and millimeter wave radar along the collaborative detection trajectory, and collect device surface image data and internal feature data; Inputting the device surface image data and internal feature data into the edge computing node, identifying the device status features through a deep learning algorithm, and generating a device health score; The UAV inspection path and the unmanned vehicle inspection path are dynamically optimized and adjusted according to the equipment health score and the inspection constraint model.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images