Tower crane outer surface defect inspection method based on unmanned aerial vehicle

Abstract

The invention discloses a tower crane outer surface defect inspection method based on an unmanned aerial vehicle, and the method comprises the steps: 1, collecting the three-dimensional point cloud data of the outer surface of a tower crane, and building a complete point cloud model under a unified coordinate system; step 2, constructing and preprocessing a three-dimensional model, performing filtering, registration and surface reconstruction on the point cloud data obtained in the step 1, generating a tower crane triangular surface model, and dividing a key inspection area based on model characteristics; 3, space division and cube modeling are carried out, space division is carried out on the triangular surface model generated in the step 2, sub-blocks containing the tower crane structure are defined as obstacle cubes through triangle-sub-block intersection detection, and virtual safety cubes are generated on the peripheries of the obstacle cubes; 4, generating a viewpoint and a sight line, and generating the viewpoint by adopting dynamic parameter clustering in the virtual security cube; and step 5, generating an unmanned aerial vehicle path connected with all viewpoints, and performing unmanned aerial vehicle inspection according to the path.

Description

Technical Field

[0001] This invention relates to the field of engineering machinery inspection and UAV path planning, and in particular to a method for inspecting defects on the outer surface of a tower crane based on a UAV. Background Technology

[0002] Traditional tower crane inspections are centered on "manual supervision + simple tool assistance," and are mainly divided into two categories: First, direct manual inspection, where workers wearing safety equipment climb ladders or baskets to a height of 30-100 meters to visually inspect for defects such as weld cracks and missing bolts. They use tools such as wrenches and measuring tapes to determine the degree of bolt looseness and the size of defects, and finally record the findings using paper forms or mobile phone photos. Second, fixed sensor-assisted inspection, where low-resolution cameras and adhesive temperature sensors are installed at fixed locations such as the middle of the tower crane and the hoisting mechanism. Workers periodically check the data and verify anomalies, but the sensor coverage is limited and the results need to be interpreted manually.

[0003] This type of inspection method is extremely unsafe. High-altitude operations rely on ladders or suspended platforms, which carries risks of falls from heights (data from the Ministry of Housing and Urban-Rural Development shows that such accidents accounted for 18% of all high-altitude fall accidents in construction from 2021 to 2023), falling objects (risk of falling tools), and environmental risks (impact from strong winds and thunderstorms). The fatality rate for these accidents exceeds 90%. Efficiency is also severely lacking. A complete inspection of a single QTZ63 tower crane takes 4-6 hours, and inspection cycles for multi-tower crane projects can reach 2-3 days, failing to meet the demands of intensive operations in modern projects. Data processing also takes 2-3 hours, further extending the process.

[0004] The drone tower crane inspection is based on "high-precision sensing + intelligent algorithms". By equipping it with LiDAR, high-definition cameras, infrared thermal imagers and other equipment, combined with path planning algorithms, it can realize fully automatic detection of the tower crane's outer surface - eliminating the need for manual high-altitude operations. Ground personnel can complete the inspection through remote control and terminal. It can also rely on 3D modeling and AI recognition technology to realize quantitative analysis of defects, completely changing the traditional inspection operation logic.

[0005] Its safety has achieved a qualitative leap, requiring no personnel to enter the high-altitude area throughout the entire process, fundamentally eliminating risks such as falls from heights and being struck by objects, with an accident rate approaching zero. Furthermore, the drone is less affected by the environment, operating stably even in wind speeds of 5-8 m / s and light rain, avoiding the interruptions caused by inclement weather in traditional inspections. Efficiency is significantly improved; an inspection of a single QTZ63 tower crane takes only 30-40 minutes, which is 1 / 8 to 1 / 10 of the traditional method. For multi-tower crane projects, multiple machines can work collaboratively, reducing the inspection cycle to several hours. Data is automatically processed to generate reports, eliminating manual data entry and improving overall process efficiency by 8-10 times. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for inspecting defects on the outer surface of tower cranes based on unmanned aerial vehicles (UAVs) to address the shortcomings of the existing technology.

[0007] To address the aforementioned technical problems, this invention discloses a method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs):

[0008] Step 1: Install high-precision lidar equipment on the ground around the tower crane to collect three-dimensional point cloud data of the outer surface of the tower crane and establish a complete point cloud model under a unified coordinate system.

[0009] Step 2: 3D model construction and preprocessing. The point cloud data obtained in Step 1 is filtered, registered and surface reconstructed to generate a high-precision triangular face model of the tower crane. Based on the model features, key inspection areas such as the tower body, boom and bolt connection are divided.

[0010] Step 3: Spatial partitioning and cube modeling. The triangular face model generated in Step 2 is partitioned using an octree algorithm. The octree depth is set to 4. The sub-block containing the tower crane structure is defined as an obstacle cube by triangle-sub-block intersection detection. A virtual safety cube (drone flight area) is generated around the obstacle cube.

[0011] Step 4: Viewpoint and line-of-sight generation. Within the safety cube of Step 3, based on the structural features of the tower crane, namely "modular tower body and segmented boom", dynamic parameter clustering is used to generate viewpoints.

[0012] Step 5: Inspection path planning. An inspection path generation algorithm is used to generate a drone path connecting all viewpoints in Step 4, and the drone inspection is carried out according to the path.

[0013] The acquisition method described in step 1 includes a combination of rotational scanning and multi-station stitching; the lidar device is a 128-line solid-state lidar with the following scanning parameters: horizontal field of view 120°, vertical field of view 30°, scanning frequency 10Hz, ranging range 3-50m, and multi-station stitching using the ICP algorithm based on common control points with a stitching error ≤3mm.

[0014] The specific implementation of multi-station splicing in step 1 is as follows: 3-5 lidar stations are evenly arranged around the tower crane, and the scanning time of each station is ≥2 minutes. Coordinate transformation between stations is achieved by placing more than 3 spherical targets with a diameter of 5cm in the scanning area. The target positioning accuracy is ≤±3mm.

[0015] Specifically, 3-5 lidar stations (using 128-line solid-state lidar, 120° horizontal field of view, 30° vertical field of view, 10Hz scanning frequency, and a ranging range of 3-50m) are evenly deployed within a 3-50m radius around the tower crane. Each station is equipped with a 360° rotating gimbal (rotation accuracy 0.1°). Four 5cm diameter spherical reflective targets are placed near the tower crane (as registration references, with a positioning error ≤3mm). Each station performs a 360° rotating scan with a scanning angle interval of 0.5°, and a single station scan time of 3 minutes to acquire point cloud data of the tower crane's outer surface (point cloud density 120 points / ㎡). For easily missed areas such as the top of the tower crane and the far end of the boom, two additional scans are performed. A global coordinate system is established with the tower crane's slewing center as the origin (X-axis pointing vertically upwards, Y-axis pointing to the front end of the boom). Multi-station point cloud registration is achieved through target matching (using an improved ICP algorithm, registration error ≤3mm), and the data is stitched together to form a complete point cloud model (number of points ≥10 million).

[0016] Existing methods for collecting tower crane point cloud data mostly employ "static scanning of unit stations," which cannot cover the entire tower crane body and blind spots at high and low ends. This method ensures a 360° coverage around the tower crane, eliminating scanning blind spots.

[0017] The filtering process in step 2 employs a combined algorithm of voxel filtering and statistical filtering: the voxel filtering uses a directional noise removal module to process noise sources specific to tower cranes, including point cloud jitter caused by boom vibration and interference points from on-site hoisting equipment. The voxel grid size is set to 0.02m × 0.02m × 0.02m to remove redundant points. The statistical filtering is set to K=8 neighborhood and standard deviation factor=1.5 to remove outlier noise points.

[0018] After voxel filtering, temporal consistency verification of jitter point clouds is performed: by comparing multiple frames of point clouds, instantaneous abnormal points caused by tower crane jitter are eliminated.

[0019] After statistical filtering, spatial clustering and removal of equipment interference points are performed: the point cloud clustering features of the hoisting equipment on site are identified and filtered.

[0020] Specifically, the raw point cloud data of the tower crane's outer surface collected from multiple stations in step 1 is imported into the processing software, a unified coordinate system is established, and the data is stitched together to form a complete point cloud. A voxel grid downsampling algorithm is used to simplify the stitched point cloud, setting the voxel grid size to 0.02m × 0.02m × 0.02m, dividing the point cloud space into uniform voxels. The directional noise removal module performs multi-frame comparison filtering on the point cloud (non-static noise) caused by instantaneous shaking due to wind or mechanical vibration of the tower crane boom. For point clouds at the same voxel location, the frequency of occurrence in 10 sub-point clouds is counted: if the occurrence frequency is ≤3 times, it is determined as an "instantaneous shaking point" and removed; if the occurrence frequency is >3 times, it is determined as a "stable point" and retained. For point clouds that have undergone time-series verification, an outlier removal algorithm based on neighborhood statistics is used, setting the neighborhood parameter K=8, to remove outlier noise points with an average distance > global mean + 1.5 x standard deviation. Euclidean clustering algorithm was used to cluster the statistically filtered point cloud, and the linearity and sphericity features of each cluster were extracted. If a cluster satisfies linearity ≥ 3.0 or sphericity ≥ 0.6, and the distance from the cluster center to the tower crane surface is > 0.8m, it is identified as a "hoisting equipment interference cluster" and is removed entirely.

[0021] The surface reconstruction in step 2 uses the Poisson reconstruction algorithm, generating a triangular model with ≥500,000 triangular faces and a model accuracy ≤±3mm;

[0022] Specifically, the Poisson reconstruction algorithm is used to convert the point cloud into a triangular face model (≥500,000 triangular faces). MeshLab software is then used to repair the model and fill in any gaps in the triangular faces. Subsequently, the key areas of the tower crane are divided into primary, secondary, and tertiary areas, namely the standard section connection welds of the tower body, the lower chord welds of the jib, the bolted connections, the main limbs of the tower body, the web members of the jib, and the remaining non-load-bearing outer surface areas.

[0023] The triangle-sub-block intersection detection in step 3 includes: determining whether the vertex of the triangle face is inside the sub-cube, or whether any edge of the triangle face intersects with the 6 faces of the sub-block (sub-cube). If either condition is met, it is determined to be an obstacle sub-block.

[0024] Specifically, in step 3, a cubic bounding box surrounding the tower crane is first constructed. Then, an octree algorithm is used to recursively partition the initial space four times (depth = 4), resulting in 4096 small cubes (voxels). These small cubes are then further divided into obstacle cubes and safety cubes. Obstacle cubes: If a small cube contains at least one tower crane triangle, it is marked as an obstacle cube, and its spatial coordinates and dimensions are recorded. Safety cubes are obtained by extending the obstacle cubes in 12 directions by 0.3-0.5m, ensuring a minimum safe distance of ≥0.3m between the drone and the tower crane structure.

[0025] Step 4 uses dynamic parameter clustering to generate viewpoints, specifically as follows:

[0026] Step 4-1: Divide the tower crane into a vertical tower body module and a horizontal boom module. Divide the vertical tower body module according to the standard section height, and divide the horizontal boom module into three sections: root, middle and far end.

[0027] Step 4-2: The clustering radius of the vertical modules of the tower body is half the length of the standard section, and the clustering density is the height of one standard section. Each cluster center corresponds to one viewpoint. The fixed clustering radius of the horizontal modules of the crane boom is 0.5 meters, and the clustering density is 1.5 times that of the tower body modules.

[0028] Step 4-3: Based on the geometric curvature characteristics of the lidar, divide the region into high curvature and low curvature areas. The number of clusters in the high curvature area is twice that in the low curvature area. Sample each region at different densities. Based on the sample points obtained from the sampling, perform modular dynamic clustering of the tower crane. The viewpoint position is offset outward from the cluster center along the normal direction of the triangular face by 1-1.2m (the safe distance between the UAV and the tower crane surface). The line of sight direction accuracy is ≤±3° to ensure coverage of all triangular faces within the cluster area.

[0029] Step 4-4: Conduct a preliminary assessment of the path feasibility (turning angle, flight distance) and defect coverage of the generated viewpoints. Based on the assessment results, fine-tune the viewpoint positions or add new viewpoints to ensure that they meet the needs of subsequent path optimization. Verify the line-of-sight reliability of all generated viewpoints.

[0030] The calculation of the geometric curvature characteristics of the lidar in step 4-3 is as follows: by calculating the angle between the normal vectors of the current triangle and the adjacent triangles, the average value of the angle is taken as the "average normal vector angle" of the triangle; at the same time, the average side length of the adjacent triangles is calculated as the "average side length", and finally the curvature data of each triangle is obtained by using the formula "curvature value = average normal vector angle / average side length".

[0031] Step 4 also includes viewpoint validity verification. For a single triangular mesh to be visible to the viewpoint, the following four conditions must be met: (1) The triangular mesh must be within the camera's field of view, determined by the camera's field of view parameters; (2) The distance from the triangular mesh to the camera must be within the camera's visible distance, determined by the camera's depth of view parameters; (3) The angle between the triangle normal vector and the viewpoint direction should be less than a given threshold; (4) There should be no other solid structures obstructing the viewpoint. This can be calculated using a ray-triangle intersection test algorithm.

[0032] After completing the spatial division and "obstacle-safety" binary cube modeling, it is necessary to generate inspection viewpoints covering all key areas of the tower crane's outer surface within the safety cube, and ensure that the line of sight of each viewpoint can accurately focus on the detection target. This step is the core link to achieve defect-free detection and high-precision identification. The specific process revolves around "modular curvature calculation - regional sample screening - tower crane modular dynamic clustering - viewpoint optimization - validity verification", forming a complete viewpoint generation closed loop.

[0033] Specifically, the first step is to calculate the curvature of the modular triangular faces, which is crucial for distinguishing between high-risk and low-risk defect areas in each module. Using a discrete curvature estimation method, the tower crane is divided into standard tower section modules, boom root modules, boom middle modules, and boom far-end modules. For each triangular face within the model generated in step 2, its adjacent triangular faces are searched one by one (10 adjacent faces between the tower body and boom root modules, and 15 adjacent faces between the boom middle and far-end modules). The angle between the current triangular face and the normal vectors of its adjacent triangular faces is calculated, and the average angle is taken as the "mean normal angle" of that triangular face. Simultaneously, the average side length of adjacent triangular faces is calculated as the "average side length." Finally, the curvature data for each triangular face is obtained using the formula "curvature value = mean normal angle / average side length". Based on the curvature values ​​and module characteristics, each area is classified. For the tower body standard section module and the far end module of the crane boom, the high curvature areas with a curvature >0.05mm⁻¹ mainly include weld edges, bolt head edges, and structural corners. These areas are high-incidence areas for defects such as weld cracks and loose bolts. For the root and middle modules of the crane boom, the threshold for judging high curvature areas is adjusted to >0.04mm⁻¹, because the stress in this area is relatively mild, and the defect risk is slightly lower than that in the tower body and far end. Through modular curvature calculation, the high-risk areas that need to be focused on in each module can be accurately identified, providing a quantitative basis for subsequent clustering and point placement, and avoiding redundancy in inspection resources.

[0034] Next, the sample preparation stage for modular dynamic clustering of the tower crane proceeds. Differentiated sampling is applied to different modules and risk areas. For the tower body standard section module, high curvature areas are sampled at a grid density of 0.4m × 0.4m, and low curvature areas at 0.6m × 0.6m. For the far end of the boom module, high curvature areas are sampled at 0.3m × 0.3m, and low curvature areas at 0.5m × 0.5m. For the boom root and middle areas, high curvature areas are sampled at 0.5m × 0.5m, and low curvature areas at 0.8m × 0.8m. Each sample point must include the three-dimensional coordinates and normal vector of the triangular face. Invalid sample points located inside the tower crane, in the ground background, or outside the safety cube are strictly eliminated. The final valid samples focus only on key external surface areas such as the tower body, boom, and bolted connections. Furthermore, each module's samples are clustered independently to ensure a high degree of match between the results and actual inspection requirements.

[0035] Subsequently, a modular dynamic clustering algorithm for tower cranes was implemented, adapting key parameters based on the structural complexity of each module to balance cluster discriminative power and stability. For the standard tower section module, the fuzzy coefficient m=1.6, the cluster radius is half the length of the standard section, the maximum number of iterations is 40, and the convergence threshold is 0.01m. For the distal boom module, the fuzzy coefficient m=1.8, the fixed cluster radius is 0.5m, the maximum number of iterations is 40, and the convergence threshold is 0.01m. For the root and middle boom modules, the fuzzy coefficient m=1.7, the maximum number of iterations is 40, and the convergence threshold is 0.01m. In terms of the number of clusters, the high curvature region of the tower body and distal boom modules has twice the number of clusters as the low curvature region, while the high curvature region of the root and middle boom modules has 1.5 times the number of clusters as the low curvature region. During the iteration process, the initial cluster centers of the tower body and the root module of the crane boom are randomly selected from the valid samples. The initial cluster centers of the middle and far end modules of the crane boom are preferentially selected from the "high confidence zone of defects" (geometric high curvature + temperature anomaly area). Then, the membership degree of each sample point to each cluster center is calculated, and the cluster centers are updated according to the membership degree so that the center gradually approaches the geometric core of the region. These two steps are repeated until the change in the distance between the cluster centers of two adjacent iterations is less than the convergence threshold, and the offset of the cluster centers of adjacent modules is ≤0.3m. At this time, the output cluster centers accurately represent the geometric features of the corresponding module region, and each cluster center serves as the core position of the potential viewpoint.

[0036] Finally, viewpoint optimization and validity verification were performed to ensure the safety and accuracy of the generated viewpoints: 1. Verify viewpoint field of view coverage; the triangular mesh must be within the camera's field of view. 2. Distance verification; the distance from the triangular mesh to the viewpoint must be within the camera's visible range. 3. Normal vector angle verification; the angle between the triangular normal vector and the viewpoint direction must be <60°. 4. Obstruction verification; a ray-triangle intersection test algorithm was used to ensure no physical obstruction between the triangular facets and the viewpoint. The cluster center of the tower body standard section module was offset outward by 1.2m along the normal vector, the far end module of the crane boom was offset by 1.0m, and the root and middle modules of the crane boom were offset by 1.2m. The deviation between the line of sight of the calibration viewpoint and the triangular facet normal vector of all modules was ≤±3° to avoid image distortion. Redundant points with a spacing of less than 0.6m were deleted from the tower body module, the far end module of the crane boom with a spacing of less than 0.4m, and the root and middle modules of the crane boom with a spacing of less than 0.6m. The high curvature area coverage of the tower body and the far end module of the crane arm is ≥99%, and the low curvature area coverage is ≥95%. The high curvature area coverage of the crane arm root and middle module is ≥98%, and the low curvature area coverage is ≥95%. The overlap rate of adjacent viewpoint observations is ≥20%. Finally, effective inspection viewpoints with no blind spots and no redundancy are generated and adapted to the structure of each module, providing a high-quality set of viewpoints for subsequent path planning.

[0037] In the UAV viewpoint generation stage of tower crane external surface defect inspection, traditional viewpoint generation methods have core technical flaws, making it difficult to meet the inspection requirements of high precision and high safety. Traditional methods only focus on the spatial location coverage of the viewpoint, without verifying the observation effectiveness and flight safety of the viewpoint: some viewpoints cannot effectively collect defect data due to tower crane structure obstruction, camera field of view outside the limit, or excessive line of sight, forming "invalid viewpoints"; some viewpoints are too close to the tower crane structure (e.g., less than 0.3m), posing a collision risk during UAV flight, and traditional methods lack a systematic effectiveness verification and safe distance guarantee mechanism. Tower crane defects (such as weld cracks and missing bolts) are mostly concentrated in high curvature areas (weld edges, bolt heads, structural corners), while the defect incidence rate in low curvature areas (such as flat steel plates of the tower body and non-load-bearing sides of the boom) is extremely low. Traditional methods only place points according to spatial location without considering the correlation between geometric curvature and defect distribution. This results in insufficient viewpoint density in high-curvature defect-prone areas and redundant viewpoints in low-curvature, low-risk areas, which increases both inspection time and the probability of missed defects. This paper proposes a viewpoint generation strategy: Based on the 3D point cloud data of the tower crane acquired by LiDAR, the curvature value of each triangular face is calculated. By obtaining the mean angle between the normal vectors of the current triangular face and its adjacent triangular faces, and combining it with the average side length of the adjacent triangular faces, the formula "curvature value = mean angle between normal vectors / average side length" is used to identify "high curvature areas = high-defect areas". Secondly, a differentiated clustering density is set according to the curvature value: the number of clusters for high curvature areas in the tower body and the far end of the boom is twice that of low curvature areas, and the number of clusters for high curvature areas in the root / middle of the boom is 1.5 times that of low curvature areas, ensuring dense coverage of key areas such as welds and bolts. Finally, a quantitative coverage target is set: the coverage rate of high curvature areas is ≥98%-99% (99% for the tower body / far end), the coverage rate of low curvature areas is ≥95%, and the overlap rate of adjacent viewpoints is ≥20%, achieving "precise focusing on high-risk areas and no redundancy in low-risk areas", solving the problems of "uneven coverage and high missed detection rate" in traditional methods. The problem was addressed by performing four layers of effective verification on the generated viewpoints: first, field-of-view verification to ensure the target triangular mesh was within the UAV camera's field of view; second, distance verification to ensure the distance from the triangular mesh to the viewpoint was within the 3-50m visibility range of the LiDAR; third, normal vector angle verification to ensure the angle between the triangular face normal vector and the viewpoint direction was <60° to avoid image distortion; and fourth, unobstructed verification using a "ray-triangle intersection test" to ensure there was no tower crane structure obstructing the viewpoint and the target triangular face, thus completely eliminating invalid viewpoints.

[0038] The method of using an octree-based hierarchical obstacle cube configuration addresses potential collisions during UAV flight. By incorporating sharp-turn penalties into the basic functions, the number of sharp turns by the UAV is reduced, resulting in a smoother flight path. Furthermore, the method of segmenting tower cranes and partitioning them into high-curvature and low-curvature zones ensures that the generated viewpoints are more closely aligned with the tower crane's structure.

[0039] By generating viewpoints within safe areas, this method addresses several issues: 1) it focuses solely on model coverage without considering the safety of UAVs during flight; 2) it solves the problem of insufficient viewpoint coverage and high missed detection rates in high-defect areas; and 3) it generates fewer viewpoints in low-risk areas, reducing viewpoint redundancy and inspection risks.

[0040] The specific implementation steps of the path planning algorithm in step 5 are as follows:

[0041] Fitness is calculated using an objective function. Individuals achieve iterative optimization by moving towards their own historical best position and the group's global best position. The path satisfies the following conditions: it does not intersect with obstacle cubes, it stays within safe cubes throughout the entire path, and it has the shortest total path length.

[0042] The objective function incorporates collision penalties and sharp turn penalties. The fitness function (i.e., the objective function) of the algorithm is: ;

[0043] in: A single line represents the path length. This is the collision penalty coefficient. This represents the number of collisions on a single path. This is the penalty coefficient for sharp turns. This represents the number of sharp turns made by the drone on a single path.

[0044] The specific steps include:

[0045] 1. Initialization: Set the number of individuals to 50, the maximum number of iterations to 100, and the individuals to be ordered path sequences containing all viewpoints;

[0046] 2. Fitness Calculation: Calculate the fitness value for each individual based on the fitness function described above;

[0047] 3. Individual Learning Stage: Individuals update their paths based on historical best practices through "path segment replacement," with the movement magnitude weighted according to the individual's experience. Positive correlation;

[0048] 4. Group Collaboration Phase: The group moves towards the globally optimal position through "path segment fusion," with the movement range determined by the group's experience weight. Positive correlation;

[0049] 5. Convergence judgment: Output the optimal path when the maximum number of iterations is reached.

[0050] Specifically, step 5 employs the algorithm proposed in this paper for coverage path planning. The core of this algorithm is mutual learning between individuals and the group through four steps: "initial modeling - fitness calculation - iterative optimization - result output," to generate an inspection path that covers all valid viewpoints, satisfies safety constraints, and has the lowest total cost. Essentially, it solves the "traveling salesman problem with multiple constraints."

[0051] The first step is the construction and initialization of the path optimization model. First, the optimization objective and constraints are clearly defined: the objective function is set as "total path length + number of collisions per path x collision penalty + number of sharp turns per path x sharp turn penalty". During the initialization phase, the effective viewpoint coordinates generated by the previous modular dynamic clustering of the tower crane are mapped to "unexplored" locations. The number of individuals and the maximum number of iterations are set. Simultaneously, an initial path is randomly assigned to each individual (i.e., the viewpoint access order is randomly arranged), and the "individual optimal path" for each individual and the "global optimal path" for the entire group are recorded.

[0052] Next, the iterative optimization cycle of "individual learning - group collaboration" begins. The core of the individual learning phase is optimizing the existing path based on its own experience: first, the fitness of each individual is calculated according to the objective function (the lower the value, the better the path); if the fitness of an individual's current path is better than its "individual optimal path," then the "individual optimal path" is updated. The group collaboration phase utilizes collective intelligence to avoid local optima: in each iteration, each individual adjusts its own path based on both the "individual optimal path" and the "global optimal path," with the movement weighted by the individual's experience. and group experience weight Control involves both drawing on one's own historical experience and learning from the best practices of the group; for example, when the group discovers that the path of "first covering the tower body and then extending to the boom" is better, multiple individuals will gradually move towards the access logic of that path.

[0053] Finally, convergence judgment and path post-processing are performed. When the change in the optimal fitness value in consecutive iterations is less than 1%, the algorithm is considered to have converged, and the viewpoint access order of the current optimal path is output; if it has not converged, the iteration continues until the maximum number of iterations of 100 is reached.

[0054] Beneficial effects:

[0055] 1. High modeling accuracy: Using ground-based lidar static scanning, the model accuracy reaches ±3mm, which is an order of magnitude higher than visual SLAM. After high-density supplementary scanning and defect enhancement algorithm processing, it can identify micro weld cracks ≥0.5mm in local areas.

[0056] 2. Highly targeted viewpoints: Modular dynamic clustering of tower cranes based on their structural characteristics doubles the viewpoint density in high curvature areas (high-defect areas), increases critical area coverage from 60% to 99%, and reduces redundant viewpoints by 40%.

[0057] 3. Excellent path optimization effect: The total path length optimized by the algorithm in this paper is shortened by 25-30% compared with the traditional TSP algorithm, the proportion of turning angles ≤90° reaches 80%, and the number of turns is reduced by 30.7% compared with the traditional TSP algorithm;

[0058] 4. High safety: The "obstacle-safety" cube-based detailed modeling ensures that the safe distance between the path and the tower crane structure remains stable at 0.3-1m, reducing the risk of collision by 90%;

[0059] 5. High degree of automation: The entire process from modeling to path generation is automated, reducing the preparation time for a single tower crane inspection from 4 hours to 0.5 hours, greatly improving inspection efficiency. Attached Figure Description

[0060] Figure 1 Generate a schematic diagram for the viewpoint and line of sight.

[0061] Figure 2 Plan routes for the path. Detailed Implementation

[0062] This embodiment describes the inspection of external surface defects of a tower crane based on unmanned aerial vehicles (UAVs). The following is the implementation process of this invention:

[0063] Step 1: Implementation Environment and Equipment Configuration: A 128-line solid-state LiDAR was selected, meeting the technical specifications of claim 2: horizontal field of view 120°, vertical field of view 30°, scanning frequency 10Hz, ranging range 3-50m, and ranging accuracy ±2cm; a 360° motorized rotating gimbal was used to ensure comprehensive scanning coverage of the tower crane's outer surface. A Leica TS60 total station was used, equipped with four 5cm diameter spherical reflective targets as common control points for multi-station LiDAR stitching. A DJI M300RTK4 rotary-wing drone was selected, equipped with an H20T gimbal, with a maximum payload of 5.5kg, a flight time of 40 minutes, RTK positioning accuracy ±2cm, and wind resistance level 7, meeting the stability requirements for high-altitude inspection.

[0064] Detailed implementation steps: Within 10m of the bottom of the tower crane, place four spherical reflective targets in an "equilateral triangle + center" layout. Use a total station to measure the three-dimensional coordinates of the targets (accuracy ±2mm) and record them as T1(x1,y1,z1), T2(x2,y2,z2), T3(x3,y3,z3), and T4(x4,y4,z4), which will serve as common control points for multi-station point cloud splicing. Start the lidar and rotating gimbal, and set the scanning parameters: horizontal angle 0°-360°, vertical angle -15°-+45°, angle interval 0.5°, and single-station scanning time 2.5 minutes (meeting the requirement of "≥2 minutes" in claim 8). For easily missed areas such as the top of the tower crane (50-60m) and the far end of the boom (50-55m), manually adjust the vertical angle of the gimbal to +30°-+45° for two supplementary scans to ensure that the point cloud density in high curvature areas (welds, bolts) is ≥150 points / ㎡. Import the original point cloud data from the four stations into CloudCompare, and achieve coarse registration through target coordinate matching. Calculate the transformation matrix between the point cloud and target coordinates at each station, and control the registration error to ≤5cm. Use the ICP algorithm based on common control points (claim 2), with the target coordinates as the reference, set the number of iterations to 500, the convergence threshold to 1e-6m, and the final stitching error to ≤3mm to generate a complete point cloud model in the global coordinate system. The global coordinate system is defined as follows: with the tower crane's rotation center as the origin, the X-axis is vertically upward along the tower body, the Y-axis points to the front end of the boom, and the Z-axis is perpendicular to the XY plane, forming a right-handed coordinate system.

[0065] Step 2: Point Cloud Preprocessing and Surface Reconstruction. Based on the tower crane's outer surface point cloud data collected by the 128-line solid-state lidar in Step 1 and stitched using the ICP algorithm based on common control points (stitching error ≤ 3mm), a combined algorithm of "voxel filtering + statistical filtering + directional noise removal module" is used for filtering. Then, a high-precision triangular face model is generated using the Poisson reconstruction algorithm. The specific process is as follows: First, voxel filtering is performed. In CloudCompare, the voxel grid size is set to 0.02m × 0.02m × 0.02m. The stitched point cloud is downsampled, retaining one representative point in each voxel and removing more than 60% of redundant points. Next, "temporal consistency verification of jittery point cloud" is performed. Ten consecutive frames of point cloud are extracted. For each point in each frame, a neighboring point is searched in the adjacent three frames with a radius of 0.05m. Stable points that exist in ≥4 frames are retained, and points that only exist in 1~2 frames are removed. Instantaneous jitter points appearing in the frame (jitter point removal rate ≥90%); then statistical filtering is performed, setting K=8 neighborhood search, calculating the average distance between each point and its 8 neighbors, calculating the global mean (μ) and standard deviation (σ), and removing outlier noise points (such as construction cables, bird afterimages) with an average distance >μ+1.5×σ, ensuring an outlier noise removal rate ≥95%; then "spatial clustering removal of equipment interference points" is performed: Euclidean clustering algorithm is used to cluster the statistically filtered point cloud, extracting the linearity and sphericity features of each cluster. If the cluster satisfies linearity ≥3.0 or sphericity ≥0.6 and the cluster center distance from the tower crane surface is >0.8m, it is determined to be a "lifting equipment interference cluster" and is removed as a whole; finally, surface reconstruction is performed, estimating the normal vector (K=10 neighborhood) of the effective point cloud after filtering, and using the Poisson reconstruction algorithm in MeshLab, setting reconstruction depth = 12, sampling density = 1.5, and selecting 20 Verification of key structural points: number of model triangles = 620,000 (≥500,000), accuracy ≤ ±3mm.

[0066] Step 3: Define an initial cube space in MATLAB. Generate a cube that perfectly surrounds the tower crane, with the center of the tower crane triangular face model as the origin. This cube is the initial space. Recursively divide the initial space 4 times using the octree algorithm (depth = 4, claim 1), resulting in 4096 small cubes (voxels). Traverse all voxel units and use the "triangle-sub-block intersection detection" logic of claim 4: if a voxel contains ≥1 tower crane triangular face, or any vertex of a triangular face is within the voxel space, or any edge of a triangular face intersects with 6 faces of the voxel, it is determined to be an obstacle cube (marked in red), and its center coordinates (x, y, z) and dimensions are recorded. The safety cube is obtained by extending the obstacle cube 0.3-0.5m in 12 directions (±x, ±y, ±z and 6 diagonals) to ensure that the minimum safe distance between the UAV and the tower crane structure is ≥0.3m, which is defined as the UAV flight area.

[0067] Step 4: After completing the spatial division and "obstacle-safety" binary cube modeling, inspection viewpoints covering all key areas of the tower crane's outer surface need to be generated within the safety cube. It is also necessary to ensure that the line of sight of each viewpoint can accurately focus on the detection target. This step is the core link to achieve defect-free detection and high-precision identification. The specific process revolves around "curvature calculation - sample screening - tower crane modular dynamic clustering - viewpoint optimization - validity verification", forming a complete viewpoint generation closed loop.

[0068] First, the curvature of the triangular faces is calculated, which is the key basis for distinguishing between high-risk and low-defect areas. Using a discrete curvature estimation method, the tower crane is divided into four modules: standard tower section, boom root, boom middle, and boom distal. The tower section and boom root modules search for 10 adjacent triangular faces one by one, while the boom middle and distal modules search for 15 adjacent triangular faces. The angle between the current triangular face and the normal vectors of its adjacent triangular faces is calculated, and the average of these 10 angles is taken as the "mean normal angle" for that triangular face. Simultaneously, the side lengths of these 10 adjacent triangular faces are calculated, and their average is taken as the "average side length". Finally, the curvature data for each triangular face is obtained using the formula: "curvature value = mean normal angle / average side length". Based on the curvature value, the outer surface of the tower crane is divided into two types of regions: high curvature regions (curvature > 0.05 mm⁻¹), mainly including weld edges, bolt head edges, and structural corners between the tower body and the boom. These regions, due to abrupt geometric changes, are prone to stress concentration during tower crane loading, making them high-risk areas for defects such as weld cracks and loose bolts. Low curvature regions (curvature ≤ 0.05 mm⁻¹), mainly consisting of relatively simple structures such as flat steel plates of the tower body and non-load-bearing sides of the boom, have a lower probability of defect occurrence. Curvature calculations can accurately identify high-risk areas requiring focused attention, providing a quantitative basis for subsequent clustering and inspection, and avoiding redundant investment of inspection resources in low-risk areas.

[0069] Next, we proceed to the sample preparation stage for modular dynamic clustering of the tower crane. Based on the modular division of the tower crane (standard tower section module, boom root module, boom middle module, and boom distal module), a differentiated sample extraction strategy is implemented according to the structural characteristics and defect risk levels of different modules. For the standard tower section module, high curvature areas (welds, bolt edges) are sampled at a grid density of 0.4m × 0.4m, and low curvature areas (flat steel plates) are sampled at 0.6m × 0.6m. For the boom distal module, due to its slender structure and the ease with which defects are missed, the sampling density for high curvature areas is increased to 0.3m × 0.3m, and for low curvature areas, it is sampled at 0.5m × 0.5m. For the boom root / middle modules, high curvature areas are sampled at 0.5m × 0.5m, and low curvature areas at 0.8m × 0.8m. Each sample point must include the three-dimensional coordinates (x, y, z) and normal vector (nx, ny, nz) of the triangular face, and be screened independently by module: invalid points outside the internal structure of each module (such as the cross brace of the standard section and the web of the crane boom), the ground background, and the safety cube are removed to ensure that the tower body samples focus on the connection part of the standard section, and the crane boom samples focus on the luffing mechanism and the stress nodes, so as to achieve the clustering basis of "pure samples within the module and independent samples between modules".

[0070] Subsequently, a modular dynamic clustering algorithm for tower cranes is executed. The core principle is to adapt to structural differences in tower cranes using the aforementioned unified parameters, achieving a customized effect of "the higher the risk, the more precise the clustering." The fuzzy coefficient for the standard tower section module is m=1.6, with a maximum of 40 iterations and a convergence threshold ε=1e-5. The fuzzy coefficient for the far end of the boom module is m=1.8, with a maximum of 40 iterations and a convergence threshold ε=1e-6 (improving clustering accuracy for complex structures). The fuzzy coefficient for the root and middle boom modules is m=1.7, with a maximum of 40 iterations and a convergence threshold ε=1e-5 (adapting to medium-risk areas). The number of clusters in high-curvature areas of the tower body and far end of the boom modules is twice that of low-curvature areas, while the number of clusters in high-curvature areas of the root and middle boom modules is 1.5 times that of low-curvature areas. Optimization is performed during the iteration process. First, initial cluster centers are randomly selected from valid samples for the tower body and crane boom root modules. For the crane boom distal module, cluster centers are preferentially selected from the "high confidence zone of defects" (geometric high curvature + stress concentration area). Second, the offset of cluster centers of adjacent modules at the junction (such as the tower body and crane boom root) is ≤0.3m to avoid sharp turns in subsequent path planning. Finally, in addition to meeting the convergence threshold, the coverage rate of cluster centers in high curvature areas must be additionally verified to be ≥98%; otherwise, 5 more iterations are added. The final output cluster centers are stored according to module classification, with each center accurately corresponding to the typical regional features of its module, providing a module-level positioning benchmark for subsequent viewpoint generation.

[0071] Finally, modular viewpoint optimization and effectiveness verification are performed, combining the above unified parameters to ensure safety and accuracy. The cluster center of the standard tower section modules is offset outwards by 1.2m along the normal vector, the far end module of the crane boom is offset by 1.0m, and the root / middle module of the crane boom is offset by 1.2m. The deviation between the line of sight and the normal vector of all module viewpoints is ≤±3°. For the far end module of the crane boom, the parallelism between the line of sight and the crane boom axis needs additional verification. A ray obstruction rate >10% for the tower module is considered invalid, requiring a fine-tuning of 0.2-0.5m. A ray obstruction rate >8% for the far end module of the crane boom is considered invalid, allowing a fine-tuning of 0.3-0.6m along the direction perpendicular to the crane boom axis. A ray obstruction rate >10% for the root and middle modules of the crane boom is considered invalid, requiring a fine-tuning of 0.2-0.5m. Redundant points with a spacing <0.6m are deleted from the tower modules. High curvature area coverage is ≥99%, and low curvature area coverage is ≥95%. For the far-end module of the crane boom, remove redundant points with a spacing of <0.4m, ensuring coverage of ≥99% in high-curvature areas and ≥95% in low-curvature areas. For the root and middle modules of the crane boom, remove redundant points with a spacing of <0.6m, ensuring coverage of ≥98% in high-curvature areas and ≥95% in low-curvature areas. The overlap rate of adjacent viewpoints for all modules should be ≥20%.

[0072] Step 5: After obtaining the effective set of viewpoints, the optimal UAV path connecting all viewpoints needs to be generated using the algorithm in this paper. The core objective is to achieve the shortest path length, the lowest collision risk, and the smoothest turning while satisfying safety constraints (avoiding collisions with obstacle cubes), coverage constraints (accessing all viewpoints), and dynamic constraints (adapting to UAV flight characteristics). This step is crucial for transforming "static viewpoints" into "dynamically executable inspection tasks." The specific process revolves around "model building - iterative optimization - convergence judgment - path post-processing," fully leveraging the global optimization capabilities of the algorithm in this paper under complex constraints.

[0073] First, the path optimization model is constructed and initialized, which is the foundation for ensuring the correct direction of the algorithm. The optimization objective function is clearly defined as "total path length + number of collisions per path × collision penalty coefficient + number of turns per path × turn penalty coefficient". The total path length directly determines the inspection efficiency, while the number of collisions and the collision penalty coefficient, and the number of sharp turns and the turn penalty coefficient quantify the safety and stability risks. The more collisions and sharp turns, the larger the penalty coefficient, the higher the objective function value, and the worse the path. This function can simultaneously take into account efficiency, safety, and stability. The constraints are strictly adapted to actual needs: First, the "complete integer permutation constraint" requires that all valid viewpoints must be visited exactly once to avoid omissions or duplications; second, the "collision constraint" requires that the distance between any point on the path and the obstacle cube be ≥0.3m. This safe distance is determined by combining the UAV's fuselage size (0.5m) and flight error (±0.2m) to ensure that even with slight flight deviations, there will be no collision with the tower crane structure; third, the "dynamic constraint" requires that the angle between adjacent path segments be ≤90° to avoid flight instability caused by sharp turns of the UAV, while limiting the cruising speed to 2-3m / s and the turning speed to ≤1.5m / s to match the dynamic performance of the UAV. In the initialization phase, the coordinates of the previously generated effective viewpoints are mapped to "individual positions" in the path planning algorithm, with each individual corresponding to a potential path solution. The number of individuals is set to 50, and the maximum number of iterations is 100 to avoid the algorithm getting stuck in infinite iteration. Simultaneously, an initial path is randomly assigned to each individual (i.e., the viewpoint access order is randomly arranged), and the "individual optimal path" for each individual and the "global optimal path" for the entire group are recorded. The total length of the initial paths is usually quite long and contains many unreasonable intersections and backtrackings, requiring optimization through subsequent iterations. Individuals update their paths based on their own historical best experience, using the following formula:

[0074] ,

[0075] in, The path of the i-th individual in iteration t; : Indicates the historical best path for this individual; : Indicates local path switching; The individual learning factor is set to 0.3; This indicates the difference between the current path and the individual's optimal path, and is used to guide the adjustment of the path direction.

[0076] Next, the algorithm enters an iterative optimization loop of "individual learning - group collaboration," which is the core mechanism of this patented algorithm and effectively balances local optimization and global exploration. The individual learning phase focuses on optimizing existing paths based on individual experience: first, the fitness value of each individual is calculated according to the objective function; if the fitness of an individual's current path is better than its "individual optimal path," then the "individual optimal path" is updated. In each iteration, individuals adjust their paths by combining the "individual optimal path" and the "global optimal path"—referring to both their own historical experience and the best results from the group. For example, when the group discovers that the path "first covering the entire tower body, then extending along the crane arm from the root to the far end" is better, multiple individuals will gradually move towards the access logic of this path, further shortening the path length. This alternating "individual learning - group collaboration" mode can quickly optimize existing paths while also escaping local optima, ensuring that a globally optimal solution is ultimately obtained. The group collaboration phase represents the process by which individuals move towards the globally optimal solution, indicating that individuals learn the structural characteristics of excellent paths through shared information from the group, which can be represented as:

[0077] ,

[0078] in : Represents the current globally optimal path for the group; The group learning factor is set to 0.7.

[0079] Finally, convergence judgment and path post-processing are performed to convert the algorithm output into flight commands executable by the UAV. The convergence judgment criterion is: when the optimal fitness value changes by less than 1% over 10 consecutive iterations, the algorithm is considered to have converged, at which point the path performance has stabilized and further iterations cannot significantly improve the optimization effect; if convergence has not occurred, iterations continue until the maximum number of iterations is reached. Finally, a collision detection program is used to perform secondary verification on all waypoints, calculating the distance between each waypoint and the nearest obstacle cube to ensure that the minimum distance is ≥0.3m and there is no risk of collision; the processed waypoint data is imported into the UAV ground control software to generate complete flight commands including position, velocity, and heading angle, and the UAV can automatically complete the inspection operation according to the commands, realizing an integrated closed loop of "viewpoint generation-path optimization-flight execution".

[0080] The path optimized by this algorithm not only shortens the total length by more than 22.7% compared to the traditional TSP algorithm, but also reduces the number of sharp turns by 30.7%. It also strictly meets all constraints, ensuring that the UAV can efficiently cover all effective viewpoints safely and smoothly, providing reliable flight support for subsequent defect data collection. Furthermore, the entire path planning process requires no manual intervention, exhibiting a high degree of automation; the path generation time for a single tower crane is only 30 minutes, significantly improving inspection preparation efficiency and fully demonstrating the practicality and superiority of this invention in engineering applications.

[0081] This invention provides a method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs). Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Step 1: Install lidar equipment on the ground around the tower crane to collect three-dimensional point cloud data of the outer surface of the tower crane and establish a complete point cloud model under a unified coordinate system. Step 2: 3D model construction and preprocessing. The point cloud data obtained in Step 1 is filtered, registered and surface reconstructed to generate a tower crane triangular face model. Key inspection areas are divided based on the model features. Step 3: Spatial partitioning and cube modeling. The triangular face model generated in Step 2 is spatially partitioned using the octree algorithm. The sub-blocks containing the tower crane structure are defined as obstacle cubes by triangle-sub-block intersection detection. A virtual safety cube is generated around the obstacle cube. Step 4: Viewpoint and line-of-sight generation. Within the virtual safety cube of Step 3, viewpoints are generated using dynamic parameter clustering based on the structural features of the tower crane. Step 5: Inspection path planning. An inspection path generation algorithm is used to generate a drone path connecting all viewpoints in Step 4, and the drone inspection is carried out according to the path.

2. The method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The acquisition method described in step 1 includes a combination of rotational scanning and multi-station stitching.

3. The method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The filtering process described in step 2 uses a combination algorithm of voxel filtering and statistical filtering: the voxel filtering uses a directional noise removal module to process the noise sources specific to the tower crane, including the point cloud jitter caused by the boom jitter and the interference points of the on-site hoisting equipment. After voxel filtering, the temporal consistency of the jitter point cloud is checked: by comparing multiple frames of point clouds, instantaneous abnormal points caused by the tower crane jitter are removed. After statistical filtering, spatial clustering and removal of equipment interference points are performed: the point cloud clustering features of on-site hoisting equipment are identified and filtered.

4. The method for inspecting defects on the outer surface of a tower crane based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The surface reconstruction in step 2 uses the Poisson reconstruction algorithm, and the number of triangles in the generated triangular model is ≥500,000.

5. The method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 4, characterized in that, The triangle-sub-block intersection detection in step 3 includes: determining whether the vertex of the triangle face is inside the sub-cube, or whether any edge of the triangle face intersects with the 6 faces of the sub-block. If any condition is met, it is determined to be an obstacle sub-block.

6. The method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 5, characterized in that, Step 4 uses dynamic parameter clustering to generate viewpoints, specifically as follows: Step 4-1: Divide the tower crane into a vertical tower body module and a horizontal boom module. Divide the vertical tower body module according to the standard section height, and divide the horizontal boom module into three sections: root, middle and far end. Step 4-2: The cluster radius of the tower body module is half the length of the standard section, the cluster density is the height of one standard section, and each cluster center corresponds to one viewpoint; the fixed cluster radius of the crane boom is R meters, and the cluster density is n times that of the tower body module. Step 4-3: Combine the geometric curvature characteristics of the lidar to divide the region into high curvature region and low curvature region; sample each region with different densities; perform modular dynamic clustering of the tower crane based on the sample points obtained from the sampling; the viewpoint position is the cluster center offset outward along the normal direction of the triangular face by m, and the line of sight direction accuracy is ≤k, to ensure that all triangular faces within the cluster area are covered. Step 4-4: Verify the line-of-sight reliability of all generated viewpoints.

7. A method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 6, characterized in that, The calculation of the geometric curvature characteristics of the lidar in step 4-3 is as follows: by calculating the angle between the normal vectors of the current triangle and the adjacent triangles, the average value of the angle is taken as the average angle of the normal vector of the triangle; at the same time, the average side length of the adjacent triangles is calculated as the average side length. Finally, the curvature data of each triangle is calculated by: curvature value = average angle of normal vector / average side length.

8. A method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 7, characterized in that, The reliability verification described in step 4-4 specifically includes the following four conditions for a single triangular mesh to be visible to the viewpoint: Condition 1: The triangular mesh must be within the camera's field of view, which is determined by the camera's field of view parameters; Condition 2: The distance from the triangular mesh to the camera must be within the camera's visible distance, which is determined by the camera's depth of view parameters; Condition 3: The angle between the triangle normal vector and the viewpoint direction should be less than a given threshold; Condition 4: There are no other solid structures obstructing the viewpoint between the triangular mesh and the viewpoint.

9. A method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 1, characterized in that, The specific implementation steps of the path planning algorithm in step 5 are as follows: Fitness is calculated using an objective function. Individuals achieve iterative optimization by moving towards their own historical best position and the group's global best position. The path satisfies the following conditions: it does not intersect with obstacle cubes, it stays within safe cubes throughout the entire path, and it has the shortest total path length.

10. A method for inspecting defects on the outer surface of a tower crane based on an unmanned aerial vehicle (UAV) according to claim 9, characterized in that, The objective function incorporates collision penalties and sharp turn penalties, and the fitness function of the algorithm is: ; in: A single line represents the path length. This is the collision penalty coefficient. This represents the number of collisions on a single path. This is the penalty coefficient for sharp turns. This represents the number of sharp turns made by the drone on a single path.

Images