Unmanned aerial vehicle laser radar point cloud tower head deformation detection method and system

Abstract

The invention discloses an unmanned aerial vehicle laser radar point cloud tower head deformation detection method and system, and the method comprises the steps: obtaining an original tower point cloud through an unmanned aerial vehicle laser radar, carrying out the preprocessing, obtaining a purification tower point cloud, and extracting a purification tower head point cloud based on the purification tower point cloud; then feature extraction and enhancement are carried out on the purification tower head point cloud to obtain a tower head enhancement feature vector; determining the type of the tower head based on the tower head enhanced feature vector, obtaining a standard tower head point cloud, and carrying out model matching on the purification tower head point cloud and the standard tower head point cloud to obtain a matching matrix; based on the matching matrix, performing deformation area positioning on the purification tower head point cloud to obtain a final deformation area; in the final deformation area, the corresponding area difference between the purification tower head point cloud and the standard tower head point cloud is calculated, and a final deformation quantification result is obtained; and finally, calculating a comprehensive risk value based on the final deformation quantification result, and dividing risk levels according to the comprehensive risk value to realize tower head deformation detection.

Description

Technical Field

[0001] This invention belongs to the field of laser point cloud monitoring, early warning and risk prevention technology, and relates to a method and system for detecting tower head deformation, specifically a method and system for detecting tower head deformation of UAV lidar point cloud. Background Technology

[0002] As smart grid operation and maintenance develops towards automation and high precision, UAV lidar, with its advantages of non-contact measurement, wide coverage, and complete three-dimensional information, is gradually becoming the core technology for detecting tower head deformation. The tower head, as the key area at the top of the tower that integrates crossarms, insulators, and hanging points, is subjected to conductor tension and wind loads for a long time, and is prone to deformation such as tilting and denting, which directly affects the power grid transmission safety.

[0003] Traditional methods widely employ single-stream feature extraction based on raw point clouds. This method directly uses the raw tower head point cloud as input, performs nonlinear transformation on the 3D coordinates of each point through a multilayer perceptron to capture local positional correlation information, and then aggregates the global feature vector covering the entire tower head through global max pooling or average pooling operations. Finally, this feature vector is used for subsequent tower head type determination, standard model matching, and other steps to provide basic feature support for deformation detection.

[0004] However, traditional methods can only construct features based on the location information of the original points and cannot simultaneously capture the structural information of the tower head (such as the hierarchical structure at the connection between the crossarm and the tower body, and the spatial distribution pattern of the insulators). As a result, the extracted features lack effective characterization of the structural correlation between the complex components of the tower head, making it difficult to accurately distinguish the structural differences between different types of tower heads. Ultimately, this creates hidden dangers for the subsequent accurate determination of tower head types and matching with standard models, limiting the accuracy and reliability of the overall deformation detection process. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for detecting deformation of the tower head of a drone's lidar point cloud.

[0006] The technical solution adopted in this invention is: a method for detecting the deformation of the tower head of a UAV lidar point cloud, comprising the following steps: Step 1: Obtain the original pole point cloud using UAV lidar and preprocess it to obtain the purified pole point cloud. Extract the purified tower head point cloud based on the purified pole point cloud. Step 2: Perform feature extraction and enhancement on the point cloud of the purification tower head to obtain the tower head enhanced feature vector; Step 3: Determine the tower head type based on the tower head enhancement feature vector and obtain the standard tower head point cloud. Perform model matching between the purification tower head point cloud and the standard tower head point cloud to obtain the matching matrix. Step 4: Based on the matching matrix, locate the deformation region of the point cloud at the head of the purification tower to obtain the final deformation region; Step 5: Within the final deformation region, calculate the difference between the corresponding regions of the purification tower head point cloud and the standard tower head point cloud to obtain the final deformation result; Step 6: Calculate the comprehensive risk value based on the final deformation results, classify the risk level according to the comprehensive risk value, and realize the tower head deformation detection.

[0007] Preferably, in step 1, the original tower point cloud is processed. ,in, Let N be the 3D coordinates of the k-th point cloud, and N be the total number of point clouds. Principal Component Analysis (PCA) is used to calculate the coordinates of each point cloud. normal vector And optimize normal vector consistency through anisotropic smoothing:

[0008] in, The optimized normal vector, for of - Neighborhood point set, for of -The j-th reference point in the neighborhood, for The center of gravity of the neighborhood, Distance is the influencing factor. Angle threshold; The point cloud after optimizing the normal vectors is voxelized according to the voxel size to generate a voxel set. M is the number of voxels, for each voxel The point closest to the voxel center within the cluster is selected as the cluster seed point. ; Seed point Define a similarity measurement function centered on [the target]. Voxels are clustered into supervoxels using K-means clustering. ,in, Let be the number of supervoxels; the formula for calculating the similarity measure function is:

[0009] in, These are the weighting coefficients. , Voxels , The center of mass, The standard deviation of spatial distance. , Voxels , The initial average normal vector, It is a voxel The initial average normal vector transpose; For adjacent supervoxels , The vector connected by its centroid = - and the cosine of the angle between the vector and the connecting vector. Determine the concavity / convexity relationship and eliminate non-target hypervoxels; among which... The vector connecting the superbody's core and the superbody. For supravoxels The transpose of the average normal vector, For supravoxels The cosine of the angle between the normal vector and the connecting vector. For supravoxels The cosine of the angle between the normal vector and the connecting vector; if the connection is concave, supervoxels with a point cloud density lower than 10 points / voxel are marked as redundant noise and removed; if the connection is convex, the supervoxels are retained; thus, a preliminarily cleaned tower point cloud is obtained. ; Based on point cloud local density and normal vector dispersion , The initial purification of the tower cloud The noise is categorized into three levels: high-intensity noise, low-intensity noise, and high-noise noise. For noise of different intensities, a distance-angle determination factor is introduced to improve the traditional bilateral filter. The filtering formula is as follows: ,in, These are the coordinates of the filtered point. The improved bilateral filter factor is calculated and dynamically adjusted in three levels: The formula for calculating the bilateral filter factor for high-intensity noise suppression is:

[0010] in, This is the improved bilateral filter factor. Distance weights For normal projection weights, For intensity weight, This is the transpose of the optimized normal vector; Finally, the point cloud of the purification tower was obtained. ; Based on point cloud of purification poles To extract the point cloud of the purified tower head, firstly, the boundary elevation between the tower head and the tower body is determined through elevation distribution analysis. Points in the purified tower point cloud with elevations higher than the tower shoulder elevation threshold are initially classified as candidate tower head point clouds, while the main tower body point cloud with elevations lower than the tower shoulder elevation threshold is simultaneously removed. Then, for each point in the candidate tower head point cloud, the number of neighboring points within a preset horizontal search radius is counted, and the point cloud density per unit area is calculated, removing scattered debris points with insufficient density. Finally, the dispersion of the normal vector within the neighborhood of each point in the candidate tower head point cloud is calculated, retaining points with dispersion below the threshold and removing debris points with excessive dispersion. The final output is the purified tower head point cloud. .

[0011] Preferably, in step 2, the point cloud at the head of the purification tower is used to extract features from the multi-source flow feature set; the specific implementation process is as follows: Firstly, regarding the cloud point at the purification tower head. The original points in the dataset are used to directly extract location features using a Multilayer Perceptron (MLP). The feature extraction formula is as follows:

[0012] in, Let be the eigenvector of the t-th original point. for of Neighborhood point set, This is the default value; for Average offset from neighboring points This is the normalized distance factor; Through the above steps, the original point source flow feature set is obtained. ; Then, the tower head space is partitioned according to voxel size to generate a voxel set. K represents the number of voxels. A non-empty voxel is defined as a voxel containing at least one cleansing point. For a set of non-empty voxels, a feature pyramid is constructed using four layers of three-dimensional sparse convolution. The feature extraction formula is as follows:

[0013] in, Let be the multi-scale feature vector of the m-th non-empty voxel. for The number of points inside, For density statistics of points within voxels, () represents a 4-layer sparse convolution stacked module; Through the above steps, the voxel source flow feature set is obtained. The voxel features are then mapped back to the original point dimension using bilinear interpolation to obtain the voxel features of each cleaned point. ; Finally, the purification tower head cloud. Middle primitive Cartesian coordinates Convert to cylindrical coordinates Rasterize the cylindrical coordinates to generate a perspective view. The contextual features are extracted using a 6-layer 2D CNN. The output feature map of the CNN is then mapped back to the original point dimension using bilinear interpolation to obtain the perspective features of each cleaned-up point. Through the above steps, the source flow feature set of the perspective view is obtained. .

[0014] Preferably, in step 2, features are first extracted from the point cloud of the purification tower head using a multi-flow feature aggregation method to obtain a multi-source flow feature set; then, a channel attention mechanism is introduced to dynamically allocate the feature weights of each source flow, and the fusion formula is as follows:

[0015]

[0016] in, Let be the fused feature vector of the t-th point. , and Let be the channel attention weights for the t-th original point, voxel, and perspective feature. () represents the attention weight calculation network. () represents global average pooling. Let be the eigenvector of the t-th original point. Let be the voxel characteristic of the t-th purification point. The perspective view features of the t-th purification point; The above steps yield the fused feature set. ={ }; Next, based on the tower head structure, the purification tower head point cloud is... Divided into M non-overlapping RoI regions For each Calculate the percentage of non-empty voxels:

[0017] in, for The number of non-empty voxels within, for The corresponding total number of voxels, The value represents the proportion of non-empty voxels. A larger value indicates a denser point cloud within the RoI, while a smaller value indicates a sparser point cloud. Construct a 4-layer voxel feature pyramid, based on The feature layer is dynamically selected for aggregation, and then feature sampling is optimized through deformable RoI pooling, as shown in the following formula:

[0018] in, The pyramid optimization feature vector for the m-th RoI. To fuse features, the feature mapping of the i-th layer of the pyramid. and The sparsity threshold, () represents the deformable RoI pooling operation; Finally, the pyramid optimization features of all RoIs are concatenated and then normalized through a fully connected layer to obtain the pyramid head enhancement feature vector. .

[0019] Preferably, in step 3, the enhanced feature vector of the tower head is... Feature channel weights are dynamically allocated through a channel attention mechanism, and the selection formula is as follows:

[0020] in, These are the core feature vectors after filtering. () is the core feature selection network. () represents the global max pooling operation; For standard CAD models of mainstream tower head types, generate corresponding core feature vectors according to the feature extraction process in step 2, and build a feature library. Calculate the core feature vector of the measured tower head. With each standard feature in the feature library cosine similarity If the maximum similarity is greater than the preset similarity threshold, the measured tower head type is determined to be the standard type corresponding to the maximum similarity; if both are less than the preset similarity threshold, it is determined to be a non-standard type. Based on the determined tower head type, retrieve the corresponding standard CAD model, convert it into a standard tower head point cloud, and calculate the purification tower head point cloud. The characteristic centroid is obtained by averaging the coordinates of all points at the tower head, and the standard tower head point cloud is calculated. The characteristic centroid is used to translate the standard tower head point cloud along three-dimensional directions, so that... and Coincident, translation vector The standard tower head point cloud after translation is recorded as Based on core features Extract the principal direction of the tower head, calculate the angle between the principal direction of the standard model and the measured principal direction of the tower head, and obtain the rotation matrix. Based on rotation matrix Achieve a rotational transformation to align the principal directions of both components; Based on core features Locate the key feature points at the end points of the crossarm and the connection points of the insulator, and calculate the length of the crossarm. Tower height ,right Perform the same key feature point localization and calculate the corresponding standard dimensions. , Calculate the scaling factor ,Will Multiply the coordinates of all points by Standard tower head point cloud obtained .

[0021] Preferably, in step 3, the purification tower head point cloud and the standard tower head point cloud are model-matched by the iterative nearest point algorithm based on KD tree to accelerate neighborhood search, and a matching matrix is ​​obtained. The implementation process includes: Using standard tower head cloud point Based on the three-dimensional coordinates, the point cloud space is recursively divided using the x, y, and z axes as alternating dividing axes to generate a multi-layer binary tree structure for the point cloud of the purification tower head. Each point in Quickly locate its position using a Kd-tree binary tree traversal. The algorithm for iterative nearest point selection is improved by introducing feature similarity weights, using the k nearest neighbors in the given set. The objective function is as follows:

[0022] in, For rotation matrix, It is a translation vector. for exist The nearest neighbor is obtained by searching a Kd-tree. For feature similarity weights, To purify each point in the cloud of points at the top of the tower, The total number of valid points in the purified tower head point cloud; Rotation matrix Initialize to Translation vector Initialize to The optimal solution for each iteration is obtained by using Singular Value Decomposition (SVD). and Update the point cloud coordinates of the purification tower head, re-search for nearest neighbors using a Kd-tree, and repeat the optimization process. Set two convergence conditions; iteration stops when either condition is met. After convergence, the final output is the matching matrix. .

[0023] Preferably, in step 4, based on the matching matrix, a density clustering algorithm is used to locate the deformation region of the point cloud at the head of the purification tower to obtain the final deformation region. The implementation process includes: First, based on the matching matrix, each point in the point cloud of the purification tower head is analyzed. Perform coordinate transformation to obtain the aligned point cloud of the purification tower head in the standard coordinate system; After alignment, the root mean square error of the point cloud matching must be calculated to satisfy RMS < threshold A. If it is not satisfied, return to step 3 and re-execute the precise matching. Then, radius filtering is used to remove isolated noise points remaining after alignment, while retaining continuous structural points; voxel lattice downsampling is used to reduce the number of point clouds, resulting in an aligned point cloud. ; For Alignment Point Clouds Each point in Search for its nearest neighbor and calculate the average distance from each point to its nearest neighbor. All Arrange the data in ascending order to generate KNN curves. The inflection points of the KNN curves correspond to abrupt changes in point cloud density; the density difference between deformed and undeformed regions will form inflection points on the curve. The first derivative of the curve is then calculated. Set a derivative threshold; when the first derivative exceeds the derivative threshold, the corresponding point is the inflection point. Let the average distance corresponding to the inflection point be denoted as . ; right Each point in If its The number of points in the neighborhood is greater than or equal to the neighborhood threshold of the core point. If a point is identified as a core point, all points within its ε-neighborhood are associated with it. Starting from each directly associated point, all unassociated qualified points within its ε-neighborhood (points indirectly adjacent to the core point with acceptable density) are then associated. This process continues until no new qualified points can be associated. All associated points then form a complete cluster, generating multiple clusters. Points not included in any cluster are identified as noise points and are directly removed. Calculate the minimum bounding box area for each cluster. Remove Small clusters smaller than the preset value; calculate the distance between the cluster center and the corresponding region center in the standard model. ,like If the value is less than the preset value, it is determined to be a non-deformation difference and is discarded. The remaining clusters after filtering are the final deformable regions. Record the point set of each deformed region and the range of spatial coordinates.

[0024] Preferably, in step 5, the difference between the corresponding regions of the purification tower head point cloud and the standard tower head point cloud is calculated using the improved Hausdorff distance algorithm within the final deformation region to obtain the final deformation result. The specific implementation process is as follows: Based on the final deformation region Spatial coordinate range Cloud point at the standard tower head Extract the standard point set of the corresponding region ; With standard point set The minimum bounding box is used as the initial space, which is recursively divided into 8 sub-cubes according to the hierarchy until the number of points contained in each leaf node is less than or equal to the threshold B, thus generating an octree. ; By improving the Hausdorff distance calculation to determine the spatial difference between the deformed region and the standard region, three core quantitative indicators are defined, including average deformation. Maximum deformation and tilt angle It adapts to different types of deformation; among them, For deformation region The number of points, For all points in the deformation region arrive The nearest point distance; define the maximum deformation. ,Pick The maximum value; define the tilt angle. The degree of quantification of tilt, among which, Let be the normal vector of the plane fitted to the point set of the deformed region. The normal vector of the plane fitted to the point set of the standard region; Apply constraint correction to quantitative indicators: if a certain point >N If it is an outlier, it will be removed and recalculated. and tilt angle The deformation is limited to the range [0, M]. If the deformation exceeds the range, it is recorded as M and marked as severe deformation. N and M are preset values. The final deformation result is obtained, which includes a set of quantitative indicators for each deformation region. And the corresponding spatial coordinates.

[0025] Preferably, in step 6, the average deformation is taken into account. Maximum deformation Inclination angle The three types of quantitative indicators are normalized to obtain a normalized indicator set. ; The weighted summation method is introduced to calculate the overall risk value. The formula for calculating the overall risk value is as follows:

[0026] in, This represents the overall risk value for the k-th deformation region; the larger the value, the higher the risk. This represents the structural sensitivity weight for the region. , and The weights are the indicator weights.

[0027] The technical solution adopted by the system of the present invention is: a UAV lidar point cloud tower head deformation detection system, comprising: One or more processors; A storage device is provided for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the UAV lidar point cloud tower head deformation detection method.

[0028] Compared with the prior art, the beneficial effects of the present invention include: (1) The multi-flow feature aggregation method of the present invention uses original points, voxels, and perspective. Figure 3 Parallel extraction of source and stream preserves positional accuracy, structural integrity, and global context, overcoming the shortcomings of traditional single-stream feature information. It is precisely adapted to the scenario of UAV LiDAR collecting point clouds on tower heads, providing more comprehensive feature support for subsequent detection.

[0029] (2) Based on the multi-stream feature aggregation method, this invention introduces dynamic weighting of point cloud sparsity and deformable voxel RoI pooling, which can specifically solve the problem of sparse point cloud in distant areas and dense point cloud in close areas of the tower head. Dynamic weighting enhances the feature contribution of sparse areas, and deformable RoI pooling adaptively focuses on non-empty voxel dense areas, avoiding the sparse area information being masked in multi-stream feature fusion, further improving feature quality and adapting to the detection requirements of complex tower head structures.

[0030] (3) The present invention improves the Hausdorff distance. Based on the traditional Hausdorff distance, it combines the characteristics of tower head point cloud and introduces octree layering and local search optimization. This not only solves the problems of low calculation efficiency and poor adaptability to large-scale point clouds in the traditional method, but also improves the distance calculation accuracy of near-distance deformation area. It can more quickly and accurately quantify the degree of deformation such as tower head indentation and tilt, and perfectly adapts to the actual needs of tower operation and maintenance to obtain deformation data efficiently and accurately. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a hierarchical diagram of the steps for obtaining the matching matrix in the method of this embodiment of the invention; Figure 3 This is a step hierarchy diagram of the method for obtaining the comprehensive risk value in an embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Please see Figure 1 , Figure 2 and Figure 3 This embodiment provides a method for detecting the deformation of the tower head of a drone's lidar point cloud, which includes the following steps: Step 1: Obtain the original pole point cloud using UAV lidar and preprocess it to obtain the purified pole point cloud. Extract the purified tower head point cloud based on the purified pole point cloud. In one implementation, the original pole point cloud is acquired using UAV lidar and preprocessed. First, large-scale redundant noise generated by non-measurement targets (such as surrounding trees, crossing cables, and ground debris) in the original pole point cloud is targeted and removed using a supervoxel segmentation-based method. The specific implementation process is as follows: Point cloud of the original tower ,in, Let N be the 3D coordinates of the k-th point, and N be the total number of point clouds. Principal component analysis (PCA) is used to calculate the coordinates of each point. normal vector And optimize normal vector consistency through anisotropic smoothing:

[0035] in, The optimized normal vector, for of - Neighborhood point set, for of -The j-th reference point in the neighborhood, for The center of gravity of the neighborhood, Distance is the influencing factor. The angle threshold controls the weights of normal vector similarity.

[0036] In this embodiment, Dynamically adjusted based on point cloud density, greater than 80 points / square meter. When the value is 20, and it falls between 50 and 80... When the value is 25, it is less than 50. The value is 30. for The center of gravity of the neighborhood, The distance influence factor has a value of 0.08. The angle threshold is set to 25 degrees. normal vector transpose, for of - The normal vector of the neighboring points.

[0037] about The core basis for determining the distance influence factor is the density characteristics of the tower point cloud, the core requirements for normal vector optimization, and verification through multiple sets of comparative experiments. In this embodiment, the design density of the tower point cloud is 50~80 points / square meter, and 0.08 can accurately cover the average distance range between the target point and its effective neighboring points. The core objective of anisotropic smoothing is to optimize the consistency of normal vectors, but it cannot be over-smoothed to the point of losing structural features. The distance weight assigns an effective weight to "neighboring points within 0.08m", and the weight of "points outside 0.15m" is reduced to below 0.1. This can both converge the normal vectors of the same structural surface (improving consistency) and distinguish the structural edges from stray points (preserving details). Secondly, the optimal value is determined through multiple sets of controlled variable experiments. When the value is 0.08, the normal vector dispersion is reduced to 8°, and the edge details of structures such as tower head crossarms and insulators are not blurred, which is the optimal balance point between consistency and detail.

[0038] about Although it is a fixed value, it directly reflects the "normal vector similarity weight and angle threshold". "Relationship", through " "The design logic of weight decay can be quickly understood, and subsequent optimizations..." The value (e.g., adjusted to 30°) only needs to be modified. The angle values ​​do not require restructuring the formula, making it more adaptable.

[0039] The point cloud after optimizing the normal vectors is voxelized according to the voxel size (the voxel size is set according to the accuracy requirements of the key structure of the tower head to ensure that details such as crossarms and insulator connection points can be covered by voxels), generating a voxel set. M is the number of voxels, for each voxel The point closest to the voxel center within the cluster is selected as the cluster seed point. This ensures that seed points are evenly distributed on the surface of the target structure.

[0040] Seed point Define a similarity measurement function centered on [the target]. Voxels are clustered into supervoxels using K-means clustering. , ( (where the number of supervoxels is 1), the formula for calculating the similarity measure function is:

[0041] in, This is the weighting coefficient, which is set to 0.6 in this embodiment. , Voxels , The center of mass, The standard deviation of spatial distance is 0.3m in this embodiment. , Voxels , The initial average normal vector is obtained by averaging the optimized normal vectors of the voxel in-matrix points. In this embodiment, the value is 30°. It is a voxel The initial average normal vector The transpose of .

[0042] For adjacent supervoxels , The concavity / convexity relationship is determined by the angle between the line connecting their centroids and the normal vector, and non-target hypervoxels are eliminated: Define the vector connecting the superbody's core. = - Calculate the cosine of the angle between the normal vector and the connecting vector:

[0043] in, The vector connecting the superbody's core and the superbody. For supravoxels The transpose of the average normal vector, For supravoxels The cosine of the angle between the normal vector and the connecting vector. For supravoxels The cosine of the angle between the normal vector and the connecting vector.

[0044] like - <-0.2 (determined as "concave connection", indicating that the gap between supervoxels is not a target, such as the gap between trees and towers), then supervoxels with a point cloud density lower than 10 points / voxel are marked as redundant noise and removed; if - If the value is >0.1 (indicating a "convex connection", meaning a continuous tower structure), then the hypervoid is retained.

[0045] This step yields a preliminary purified point cloud from the towers. ,against For residual small-scale noise (such as scanning glitch points and local outliers), a noise classification-improved bilateral filtering strategy is adopted, and the filtering parameters are dynamically adjusted according to the noise intensity. The specific implementation steps are as follows: Based on the local density of point clouds and the dispersion of normal vectors, The noise level is divided into three levels: High-intensity noise: point of - Neighborhood ( Dynamically adjusted based on point cloud density, greater than 80 points / square meter. When the value is 20, and it falls between 50 and 80... When the value is 25, it is less than 50. Average distance (value 30) >0.15, k=15 in this embodiment; normal vector dispersion >45 degrees (e.g., isolated scanning burrs); Medium-intensity noise: 0.08< <0.15, and 30 degrees < <45 degrees (e.g., slight outlier points at the edge of the tower head); low-intensity noise: <0.08, and <30 degrees (e.g., slight local texture noise); where, normal vector dispersion , The normal vector is used to quantify the consistency of the neighborhood normal vectors.

[0046] For noise of different intensities, a distance-angle decision factor is introduced to improve the traditional bilateral filter. The filter formula is as follows: ,in, These are the coordinates of the filtered point. The improved bilateral filter factor is calculated and dynamically adjusted in three levels.

[0047] The formula for calculating the bilateral filter factor for high-intensity noise suppression is:

[0048] in, This is the improved bilateral filter factor. Distance weights For normal projection weights, Intensity weights are used to reduce the impact of high-intensity noise on the filtering results. This is the transpose of the optimized normal vector; It should be noted that the formula for calculating the distance weight is: ,in, A value of 0.1 is used to enhance the suppression of high-intensity noise. The formula for calculating the normal projection weight is as follows: In this embodiment The value is 0.05.

[0049] For medium-intensity noise suppression, adjust the parameters as follows: =0.15、 =0.08, =0.6, balancing noise suppression and detail preservation, low-intensity noise suppression introduces an angle decision factor. Avoid smoothing the details of the tower head edges: ,in, for and The angle between the normal vectors, the bilateral filter factor for low-intensity noise suppression. Among them, the standard deviation of low-intensity noise distance weights =0.2, standard deviation of the projection weight of the low-intensity noise normal. =0.12, low-intensity weight =0.9, It is a dimensionless minimum constant, taking the value of This is to avoid the formula becoming meaningless when the denominator is 0.

[0050] The above two steps produce the point cloud of the purification tower. .

[0051] The tower head is the core area at the top of the tower, including crossarms and insulators. Its elevation is significantly higher than that of the main tower body. The boundary elevation between the tower head and the tower body is determined through elevation distribution analysis: First, the point cloud of the purification tower is analyzed. The elevation data of all points were statistically analyzed to generate an elevation histogram. The top three local maxima with the highest frequency were selected, with the highest value corresponding to the top of the tower and the second highest value corresponding to the area near the tower shoulder (the junction between the tower head and the tower body). A fixed threshold of 3 meters was set (based on the structural parameters of mainstream 110kV-500kV towers, the elevation difference between the tower shoulder and the top of the tower is usually 2-4 meters, hence the use of 3 meters). The elevation corresponding to the second highest value was subtracted by 3 meters to obtain the tower shoulder elevation threshold. Points in the purified point cloud with elevations higher than this tower shoulder elevation threshold were initially classified as candidate tower head point clouds, while the main body point clouds with elevations lower than this threshold were simultaneously removed.

[0052] The candidate tower head point cloud may contain a small number of non-target points, which need to be further filtered using the unique structural features of the tower head: First, horizontal distribution density screening is performed: For each point in the candidate tower head point cloud, in this embodiment, a horizontal search radius of 0.5 meters is used to count the number of neighboring points within this radius and calculate the point cloud density per unit area. The point cloud density in the core structure area of ​​the tower head is usually not less than 30 points / square meter, while the point cloud density in the debris area is generally less than 10 points / square meter. Based on this, scattered debris points with insufficient density are eliminated.

[0053] Normal vector consistency screening: Calculate the dispersion (standard deviation) of the normal vector in the neighborhood of each point in the candidate tower head point cloud. In this embodiment, the dispersion threshold is set to 25 degrees. Points with dispersion below this threshold (tower head structure points with consistent normal vectors) are retained, and debris points with excessive dispersion are removed.

[0054] Final output: Purification Tower Head Cloud .

[0055] Step 2: Extract and enhance features from the point cloud of the purification tower head using the multi-stream feature aggregation method and the deformable variogram feature pyramid method to obtain the tower head enhanced feature vector; In one implementation, the purification tower head point cloud is subjected to feature extraction and enhancement using a multi-stream feature aggregation method and a deformable variogram feature pyramid method to obtain a tower head enhanced feature vector. The specific implementation process is as follows: against The original points in the dataset are used to directly extract fine-grained positional features using a multilayer perceptron (MLP), avoiding information loss caused by voxelization or projection. The feature extraction formula is as follows:

[0056] in, Let be the eigenvector of the t-th original point. for of Neighborhood point set, in this embodiment The value is 25. for The average offset from neighboring points captures local location correlation. This is the normalized distance factor. To purify the clouds at the top of the tower The three-dimensional coordinates of the t-th original point. For Qt in The s-th point in the set of k' neighborhood points.

[0057] It should be noted that the formula for calculating the normalized distance factor is: ,in, As the center of gravity of the tower head, by The coordinates of all points are averaged. In this embodiment, the MLP consists of 3 fully connected layers (input dimension 6, hidden layer dimension 128, output dimension 64). The activation function is ReLU to avoid gradient vanishing and to achieve non-linear mapping of position information.

[0058] This step yields the original point source flow feature set. It fully preserves the precise location and local correlation information of the tower head point.

[0059] Will Voxelization, through three-dimensional sparse convolution, extracts multi-scale structural features to adapt to the hierarchical structure of tower head crossarm-insulator-tower shoulder. The specific implementation is as follows: The tower head space is divided according to voxel size (the voxel size is set based on the minimum structural size of the tower head to ensure that structural details are not obscured by voxelization), generating a voxel set. (K is the number of voxels), a non-empty voxel is defined as a voxel containing at least one cleansing point. For the set of non-empty voxels, a feature pyramid is constructed using four layers of 3D sparse convolution. In this embodiment, the output feature dimensions of each convolution layer are 32, 64, 128, and 256, respectively, and the kernel size is [missing information]. With a step size of 2, the feature extraction formula is as follows:

[0060] in, Let be the multi-scale feature vector of the m-th non-empty voxel. for The number of points inside, For density statistics of voxel in-points, in this embodiment It is a 4-layer sparse convolution stacked module that only performs calculations on non-empty voxels.

[0061] This represents the summation of the three-dimensional coordinates of all points within a voxel and the calculation of the average voxel coordinates. This average coordinate can be regarded as the "center position" of the voxel, providing a spatial position reference for the voxel for sparse convolution and ensuring that the convolution can capture the hierarchical spatial relationship of the tower structure. voxels The total number of purification points contained within is used as a "density feature" of the voxel to supplement the input sparse convolution, thereby enhancing the expression of density differences among different voxels. and Common input 4-layer 3D sparse convolution .

[0062] This step yields the voxel source flow feature set. The voxel features are then mapped back to the original point dimension using bilinear interpolation to obtain the voxel features of each cleaned point. .

[0063] Will Projecting the image onto a cylindrical coordinate system generates a perspective view. A 2D CNN is then used to extract global contextual features, compensating for the lack of global correlation between the original points and voxel features. The specific implementation is as follows: Purify the tower head cloud The three-dimensional coordinates of the t-th original point Convert to cylindrical coordinates The projection formula is as follows:

[0064] in, The center of gravity of the tower head's horizontal projection. The horizontal projection radius is This is the horizontal azimuth angle.

[0065] The cylindrical coordinates are rasterized (in this embodiment, the raster size is 0.3 degrees). 0.05m, Direction covers all angles, (Direction covers the tower head elevation range), generate perspective view. (In this embodiment, H=1200, W=800, adapted to the tower head size); In this embodiment, contextual features are extracted using a 6-layer 2D CNN (Conv2d). The CNN output feature map is mapped back to the original point dimension through bilinear interpolation to obtain the perspective features of each purification point. Through this step, the source flow feature set of the perspective view is obtained. This supplements the global context association information of the tower head.

[0066] To address the information redundancy and correlation differences among the three source stream features, a channel attention mechanism is introduced to dynamically allocate the weights of each source stream feature, highlighting core features (such as the positional features of the original points and the structural features of voxels) and suppressing redundant information. The fusion formula is as follows:

[0067]

[0068] in, Let be the fused feature vector of the t-th point. , and Let be the channel attention weights for the t-th original point, voxel, and perspective feature. () represents the attention weight calculation network. () represents global average pooling. Let be the eigenvector of the t-th original point. Let be the voxel characteristic of the t-th purification point. The perspective view features of the t-th purification point; It should be noted that the input dimension of the attention weight calculation network is the feature dimension of each source stream. In this embodiment, the hidden layer dimension is 1 / 4 of the input dimension, and the output dimension is 1. and The computational logic and Consistent.

[0069] This step yields the fusion feature set. .

[0070] To address the differences in point cloud sparsity across different regions of the tower top (e.g., dense point clouds in the crossbeam area captured at close range, and sparse point clouds at the top of the tower captured at a distance), a deformable variogram feature pyramid is constructed. Feature layers are dynamically selected for aggregation to enhance feature representation in sparse regions while avoiding feature redundancy in dense regions. The specific implementation is as follows: Based on the prior knowledge of the tower head structure (such as crossarms, tower shoulders, and top), The region is divided into M non-overlapping RoI regions (M is dynamically adjusted according to the tower head type; in this embodiment, M=5 for double crossarm tower heads and M=3 for single crossarm tower heads). Each RoI contains a subset of point clouds with continuous structures.

[0071] For each Calculate the percentage of non-empty voxels:

[0072] in, for The number of non-empty voxels within, for The corresponding total number of voxels, The value represents the proportion of non-empty voxels. A larger value indicates a denser point cloud within the RoI, while a smaller value indicates a sparser point cloud.

[0073] In this embodiment, a four-layer voxel feature pyramid is constructed (the feature layer resolutions are 1 / 1, 1 / 2, 1 / 4, and 1 / 8 of the original resolution, corresponding to features at different scales). The feature layer is dynamically selected for aggregation, and then feature sampling is optimized through deformable RoI pooling, as shown in the following formula:

[0074] in, The pyramid optimization feature vector for the m-th RoI. To fuse features in the feature mapping of the 4th layer pyramid, and In this embodiment, the sparsity threshold is used. =0.6, =0.3, obtained based on a large number of tower head point cloud statistics. This is a pooling operation for deformable RoIs.

[0075] It should be noted that dense RoI (in this embodiment) >0.6) Only high-level features are needed to meet the expression requirements, sparse RoI (in this embodiment) <0.3) Multi-layer feature fusion is required to supplement information.

[0076] The pyramid optimization features of all RoIs are concatenated and then normalized through a fully connected layer to obtain the pyramid head enhancement feature vector. .

[0077] Step 3: Determine the tower head type based on the tower head enhanced feature vector and obtain the standard tower head point cloud. Use the iterative nearest point algorithm based on KD tree to accelerate neighborhood search and perform model matching between the purified tower head point cloud and the standard tower head point cloud to obtain the matching matrix. In one implementation, for the tower head enhancement feature vector, feature channel weights are dynamically allocated through a channel attention mechanism to highlight core features related to the tower head type and suppress redundant features. The selection formula is as follows:

[0078] in, These are the core feature vectors after filtering. The network is used to select core features. This is a global max pooling operation that extracts the maximum value of each channel of the enhanced feature.

[0079] It should be noted that the core feature selection network in this embodiment consists of two fully connected layers with an input dimension of 1024, a hidden layer dimension of 256, and an output dimension of 1024. The activation function used is LeakyReLU to avoid the vanishing of negative gradients.

[0080] Based on core feature vectors The method determines the tower head type by comparing feature similarity, adapting to more tower head types and eliminating the need for manual annotation of the training set. The specific implementation process is as follows: For standard CAD models of mainstream tower head types (single crossarm, double crossarm, triple crossarm, and crossarm), corresponding core feature vectors are generated according to the feature extraction process in step 2, and a feature library is constructed. Calculate the core features of the measured tower head With each standard feature in the feature library Cosine similarity: The closer the value is to 1, the higher the feature matching degree between the measured tower head and the standard type. In this embodiment, the similarity threshold is set to 0.75. If the maximum similarity is greater than the similarity threshold, the measured tower head type is determined to be the standard type corresponding to the maximum similarity. If both are less than the threshold, it is determined to be a "non-standard type", the general standard model is called and it is marked that manual review is required later.

[0081] Based on the determined tower head type, the corresponding standard CAD model is retrieved and converted into a standard tower head point cloud. Dynamic adaptation is then achieved through coarse attitude adjustment and scale adaptation, making the standard tower head point cloud closely resemble the approximate attitude of the purified tower head point cloud. The specific implementation process is as follows: Computational purification tower head point cloud The characteristic centroid is obtained by averaging the coordinates of all points at the tower head, and the standard tower head point cloud is calculated. The characteristic centroid is used to translate the standard tower head point cloud along three-dimensional directions, so that... and Coincident, translation vector The standard tower head point cloud after translation is recorded as Based on core features Extract the main direction of the tower head (e.g., the horizontal extension direction of the crossarm), calculate the angle between the main direction of the standard model and the measured main direction of the tower head, and perform a rotation transformation to make their main directions consistent. The rotation matrix is ​​then used. It is calculated from the angle between the principal direction vectors.

[0082] Because the structural dimensions of the same type of tower head vary across different voltage levels, dimensional adaptation is necessary to ensure that the standard model matches the measured tower head dimensions. Based on core features Locate key feature points such as the crossarm endpoints and insulator connection points, and calculate the crossarm length. Tower height ,right Perform the same key feature point localization and calculate the corresponding standard dimensions. , Calculate the scaling factor ,Will Multiply the coordinates of all points by Obtain the standard tower head point cloud after scale adaptation. In this embodiment, the scale factor is limited to the range of [0.8, 1.2] (set according to the size difference range of the same type of tower head of mainstream towers). If it exceeds this range, it is judged as abnormal and marked for manual review to avoid excessive scale adaptation leading to matching distortion.

[0083] The point cloud of the purification tower head is model-matched with the point cloud of the standard tower head using an iterative nearest-point algorithm based on KD-tree to accelerate neighborhood search, and a matching matrix is ​​obtained. The specific implementation process is as follows: Using standard tower head cloud point Based on the three-dimensional coordinates, the point cloud space is recursively divided using the x, y, and z axes as alternating dividing axes to generate a multi-layer binary tree structure. In this embodiment, each leaf node contains no more than 30 points (based on the density of the tower head point cloud). This is used to purify the tower head point cloud. Each point in Quickly locate its position using a Kd-tree binary tree traversal. The k nearest neighbors (k=15 in this example) are selected to significantly improve matching efficiency. Traditional ICP only calculates matching error based on spatial distance and ignores feature correlation. The improved version introduces feature similarity weights, so that point pairs with high feature matching degree account for a higher proportion in the optimization. The objective function is as follows:

[0084] in, For rotation matrix, It is a translation vector. for exist The nearest neighbor is obtained by searching a Kd-tree. For feature similarity weights, To purify each point in the cloud of points at the top of the tower, The total number of valid points in the purified tower head point cloud; It should be noted that the formula for calculating feature similarity weights is as follows: ,in, for Enhanced feature vectors, for The standard feature vector has a higher feature matching degree and a higher weight.

[0085] Rotation matrix Initialized to the state after coarse attitude adjustment Translation vector Initialize to To ensure the initial attitude is close to the optimal solution, the optimal solution for each iteration is obtained through singular value decomposition (SVD). and Update the point cloud coordinates of the purification tower head, re-search for nearest neighbors using a Kd-tree, and repeat the optimization process. Set two convergence conditions; iteration stops when either condition is met: in this embodiment, the difference between the objective function values ​​of two adjacent iterations is less than 1. The maximum number of iterations is set to 50.

[0086] After iterative convergence, the final rigid body transformation parameters are output. and matching matrix (In this embodiment, a 4×4 homogeneous transformation matrix is ​​used.) This matching matrix can transform the point cloud at the head of the purification tower. Transform to the standard model coordinate system to achieve precise alignment between the two.

[0087] Step 4: Based on the matching matrix, a density clustering algorithm is used to locate the deformation region of the point cloud at the head of the purification tower to obtain the final deformation region; In one implementation, the point cloud of the purification tower head is first transformed to the standard model coordinate system to achieve spatial alignment between the two. The specific implementation is as follows: Based on the matching matrix, each point in the point cloud of the purification tower head... Perform a coordinate transformation to obtain the aligned point cloud of the purification tower head in the standard coordinate system. The transformation formula in this embodiment is as follows:

[0088] in, To align the three-dimensional coordinates of the measured points, This is the matching matrix.

[0089] It should be noted that the transformation logic is as follows: through the rotation matrix Correcting the attitude deviation of the purification tower head point cloud by translation vector Compensation for positional deviations to ensure no deformation and Space overlap.

[0090] After alignment, the root mean square error (RMS) of the point cloud matching is calculated. In this embodiment, RMS < 0.03 must be satisfied. If not satisfied, return to step 3 and re-execute precise matching to avoid the alignment error interfering with the deformation judgment.

[0091] To reduce computational complexity and avoid interference from residual noise in the matching process for deformation detection, Perform mild noise reduction and uniform sampling: Radius filtering (search radius 0.05m, neighbor threshold 5) is used to remove isolated noise points remaining after alignment, while retaining continuous structural points. Voxel lattice downsampling (voxel size 0.08m×0.08m×0.08m in this embodiment) is used to reduce the number of point clouds while preserving the details of the tower head structure, resulting in an aligned point cloud output. .

[0092] For Alignment Point Clouds With standard tower head cloud For the differential region (deformation region), the DBSCAN density clustering algorithm with dynamic calculation parameters is used to automatically locate the suspected deformation region, solving the problem of traditional DBSCAN relying on manual parameter tuning and poor adaptability.

[0093] By constructing a KNN curve of the point cloud, the core parameters (core radius) of DBSCAN are automatically calculated. Core point neighborhood threshold To adapt to local density differences in the point cloud at the tower head, the specific implementation process is as follows: right Each point in In this embodiment, the 20 nearest neighbors of each point are searched (based on the density of the preprocessed point cloud to ensure sufficient coverage of neighboring points to reflect the local density), and the average distance from each point to its 20 neighbors is calculated. All Arrange them in ascending order to generate KNN curves.

[0094] The inflection point of the KNN curve corresponds to the abrupt boundary of point cloud density. The density difference between deformed and undeformed regions will form a significant inflection point on the curve. This can be determined by calculating the first derivative of the curve. In this embodiment, a derivative threshold of 0.02 is set (based on the fluctuation range of the point cloud density at the tower head). When the first derivative is greater than the derivative threshold, the corresponding point is the inflection point, and the average distance corresponding to the inflection point is denoted as . .

[0095] Calculate DBSCAN parameters: Core radius in this embodiment Based on a 1.2x amplification of the inflection point distance, continuous points covering deformed regions are ensured, while avoiding excessive fusion of non-deformed regions in clustering; core point neighborhood threshold. =12, determined through testing 100 sets of tower head deformation data; the core radius is limited to the range of [0.05m, 0.2m], and the boundary value is taken if it exceeds the range.

[0096] DBSCAN clustering is performed based on dynamically calculated parameters to locate suspected deformation regions, and spurious clusters (such as local differences caused by matching errors) are eliminated through structural feature filtering. The specific implementation process is as follows: right Each point in If its The number of points in the neighborhood is greater than or equal to the neighborhood threshold of the core point. If a point is identified as a core point, all points within its ε-neighborhood are associated with it. Starting from each directly associated point, all unassociated qualified points within its ε-neighborhood (points indirectly adjacent to the core point with acceptable density) are then associated. This process continues until no new qualified points can be associated. All associated points then form a complete cluster, generating multiple clusters. Points not included in any cluster are identified as noise points and are directly removed. Calculate the minimum bounding box area for each cluster. In this embodiment, the area of ​​the tower head deformation region is typically ≥0.1㎡, and is excluded. Small clusters <0.1㎡; calculate the distance between the cluster center and the corresponding region center in the standard model. In this embodiment, if If the difference is less than 0.05m (less than the matching error threshold), it is considered a non-deformation difference and is discarded.

[0097] The remaining clusters after filtering are the final deformable regions. (K is the number of deformable regions), record the point set of each deformable region. and the range of spatial coordinates.

[0098] Step 5: Within the final deformation region, the improved Hausdorff distance algorithm is used to calculate the corresponding region differences between the purification tower head point cloud and the standard tower head point cloud to obtain the final deformation result; In one implementation, based on the final deformed region Spatial coordinate range Cloud point at the standard tower head Extract the standard point set of the corresponding region (Points whose coordinates fall within the deformation region in the standard model) to ensure spatial correspondence of the comparison objects.

[0099] The octree construction process is as follows: using the standard point set The minimum bounding box is used as the initial space, which is recursively divided into 8 sub-cubes according to the hierarchy until each leaf node contains ≤20 points (in this embodiment, the density is set according to the standard tower head point cloud), thus generating an octree. .

[0100] The logic of the local search is to find the nearest point to the standard point set for each deformed point. Utilizing the spatial continuity of the point cloud (the nearest point of a deformed point is usually adjacent to the nearest point of its neighboring deformed points), the starting point of the next search is set to the leaf node containing the nearest point of the previous point. The search prioritizes that node and its neighboring nodes. If a suitable node is found... If the point is at the minimum distance, the inner loop terminates directly, without needing to traverse all standard points.

[0101] The computational complexity of the traditional Hausdorff distance is O(N). (N is the number of deformation points, and M is the number of standard points). After improvement, using an octree and local search, the complexity is reduced to O(N). The computational efficiency is improved by more than 60% (logM).

[0102] By improving the Hausdorff distance calculation to determine the spatial difference between the deformed region and the standard region, three core quantitative indicators are defined, including average deformation. Maximum deformation and tilt angle It can adapt to different types of deformation (tilt, concavity, convexity). The average deformation is defined to reflect the overall degree of deformation. The formula for calculating the average deformation is:

[0103] in, For deformation region The number of points, For all points in the deformation region arrive The closest distance, It represents the average deformation.

[0104] Define the maximum deformation ,Pick The maximum value reflects the degree of deformation at the most severe location. For deformation areas with a clear principal direction, such as crossarms and tower tops, the tilt angle is defined to quantify the degree of tilt by fitting the angle between the principal plane of the deformation area and the principal plane of the standard area. The formula for calculating the tilt angle is:

[0105] in, Let be the normal vector of the plane fitted to the point set of the deformed region. The normal vector of the plane fitted to the standard region point set. The tilt angle.

[0106] To avoid interference from outliers, constraint corrections are applied to the quantitative indicators: In this embodiment, if a certain point... >3 If the value exceeds 3 times the average deformation, it is considered an outlier, removed, and the calculation is repeated. and tilt angle The range is limited to [0°, 10°] (in this embodiment, according to the tower safety operation standard, the tower head tilting more than 10° requires emergency handling). If it exceeds the range, it is recorded as 10° and marked as severe deformation.

[0107] The final deformation result is obtained, which includes a set of quantitative indicators for each deformation region. And the corresponding spatial coordinates.

[0108] To ensure the accuracy of deformation localization and quantification, further screening is conducted through morphological verification and matching error elimination to avoid interference from spurious deformations. The specific implementation is as follows: Calculate the aspect ratio (ratio of the longest side to the shortest side) of the deformed region. In this embodiment, the aspect ratio of the actual deformation of the tower head is usually ≤5. If the aspect ratio of the deformed region is greater than 5, it is judged as a slender false cluster and is removed. Calculate the ratio of the quantitative index to the matching error RMS. In this embodiment, if If the deformation is less than 2 (the deformation is less than twice the matching error), it is determined to be a false deformation caused by the matching error and is discarded.

[0109] The remaining deformation area and corresponding quantitative indicators after verification are the final deformation location and quantification results: including spatial coordinate range, point set data, average deformation, maximum deformation, tilt angle, and deformation type (determined based on quantitative indicators).

[0110] Step 6: Calculate the comprehensive risk value based on the final deformation results, classify the risk level according to the comprehensive risk value, and realize the tower head deformation detection.

[0111] In one implementation, the quantitative indicators output in step 5 (average deformation and maximum deformation in length units, and tilt angle in angle units) have different dimensions and significantly different numerical ranges, making it impossible to directly integrate and calculate the risk value. Therefore, they need to be normalized to the [0,1] interval to ensure that the indicators are comparable and can be integrated. The specific implementation is as follows: For the three core quantitative indicators, normalization functions based on engineering safety thresholds are designed to map the indicator values ​​to [0,1]. The normalization formulas for deformation variables (average deformation variable and maximum deformation variable) are as follows:

[0112] in, For the normalized shape variables, For the original shape variables, The minimum threshold for deformation is 0.02 in this embodiment. The maximum threshold value for deformation is 0.2 in this embodiment.

[0113] The formula for calculating the normalized tilt angle is:

[0114] in, The normalized tilt angle. This is the original tilt angle. The minimum threshold for the tilt angle is set to 0.5 degrees in this embodiment. The maximum threshold for the tilt angle is 5 degrees in this embodiment.

[0115] If the original value of an indicator is within the range of [lower threshold, upper threshold], the normalized result should fall within (0, 1); if it exceeds the threshold range, the result should be saturated to 0 or 1. If an anomaly occurs (such as a normalized value > 1 or < 0), return to step 5 to re-check the deformed result to avoid error propagation, and finally output the normalized indicator set. .

[0116] The impact of deformation on tower safety varies significantly depending on the structure of the tower head. Therefore, sensitivity weights need to be assigned to make the risk assessment more aligned with engineering realities. The specific implementation process is as follows: Based on the mechanical characteristics and operation and maintenance experience of the tower head structure, the tower head structure is divided into three sensitivity levels: high sensitivity structure (corresponding to the weight in this embodiment). =1), including insulator string connection points, crossarm end hanging points, and lightning protection wire fixing points at the top of the tower head. This type of structure directly bears the conductor tension or lightning protection wire load; slight deformation may lead to load imbalance. Medium-sensitivity structures (in this embodiment) =0.8): This includes the main body of the crossarm, the tower shoulder joint, and the ladder fixing point. This type of structure is the main load-bearing component of the tower head, and deformation exceeding a certain degree will affect the overall stability; low-sensitivity structure (in this embodiment) =0.6): This includes tower head decorative parts, non-load-bearing connecting rods, and guardrails. These structures do not participate in load transfer, and their deformation has a relatively small impact on safety. The weight value is determined in this embodiment through structural mechanics simulation analysis and statistical analysis of 50 historical operation and maintenance failure cases.

[0117] For each deformation region The structural type is determined by its spatial coordinate range: based on the standard model. Mark the spatial coordinate ranges of various structures, and mark the center coordinates of the deformation region. By comparing with the structural coordinate range library, the structural type is determined and the corresponding sensitivity weight is assigned. If the deformation region spans multiple structures, the maximum weight is taken to ensure that high-risk structures are not underestimated.

[0118] Taking into account the average deformation, maximum deformation, and tilt angle, a weighted summation is introduced to calculate the comprehensive risk value. The formula for calculating the comprehensive risk value is as follows:

[0119] in, This represents the overall risk value for the k-th deformation region; the larger the value, the higher the risk. This represents the structural sensitivity weight for the region. , and As the indicator weight, =0.4, =0.4, =0.2, determined by the analytic hierarchy process.

[0120] The formula first integrates multi-dimensional risks through indicator weights, and then amplifies the risk contribution of key structures through structural sensitivity weights, ensuring that the risk assessment is both comprehensive and focused on the core.

[0121] Based on comprehensive risk value Based on the priority of engineering operation and maintenance, three risk levels are divided. The judgment rules in this embodiment are as follows: Minor risk (Level I): <0.3, included in the routine inspection plan, reviewed every 6 months, no immediate action required; Moderate risk (Level II): 0.3< <0.7, complete a special inspection within 3 months to assess the deformation trend and take reinforcement measures if necessary; Severe risk (Level III): If the value exceeds 0.7, immediately stop the machine for inspection and maintenance, limit the load on the tower, and take corrective or component replacement measures to avoid structural failure.

[0122] If there are multiple deformation areas on the same tower head and the spacing is ≤1.0m (considered as related deformations), the comprehensive risk value is the maximum value +0.1 (not exceeding 1.0), and the risk level is increased by one level; if historical monitoring data shows that the deformation of the tower head continues to increase (annual growth rate ≥0.05m), the risk level is increased by one level, and maintenance is prioritized.

[0123] The invention will be further illustrated below through specific experiments; To verify the effectiveness and practicality of this invention, actual power towers of three voltage levels (110kV, 220kV, and 500kV) were selected as research objects. A Jingwei M300 RTK UAV equipped with an L1 lidar was used for data acquisition. The flight altitude was controlled between 15-25m, and a non-repetitive scanning mode and triple-echo sampling (sampling frequency 160kHz / echo) were set. The data was output as LAS files in the UTM projection format of the National Geodetic Coordinate System 2000. A total of 20 sets of tower head point cloud data were acquired (including 5 sets under normal conditions, 6 sets with slight deformation, 5 sets with moderate deformation, and 4 sets with severe deformation). The point cloud density at the tower was ≥100 points / m², and the relative accuracy between the horizontal and vertical coordinates was ≤±5cm. The experimental hardware consisted of an Intel i9-14900K CPU, an NVIDIA RTX 4090 GPU, and 64GB of RAM. The software was based on Python + PCL. Implemented using the library and PyTorch framework, experimental results show that the present invention achieves a detection accuracy of 98.6% for deformations such as tower head crossarm offset and insulator tilt, with a deformation positioning error of ≤3cm. The recall rate for slight deformations (offset of 5-10cm) is 97.2%. Compared with the traditional ICP algorithm and K-means clustering method, the positioning error is reduced by 82.3% and the processing efficiency is improved by 35.7%. Moreover, it can still stably identify deformations in areas with vegetation cover and sparse point cloud (density of 50-80 points / ㎡), verifying the high accuracy and strong robustness of the method.

[0124] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0125] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for detecting the deformation of a UAV lidar point cloud tower head, characterized in that, The method comprises the following steps: Step 1: obtaining original tower point cloud by unmanned aerial vehicle laser radar and preprocessing to obtain purified tower point cloud, and extracting purified tower head point cloud based on the purified tower point cloud; Step 2: feature extraction and enhancement are performed on the purified tower head point cloud to obtain tower head enhanced feature vector; Step 3: determining the tower head type based on the tower head enhanced feature vector and obtaining standard tower head point cloud, and performing model matching on the purified tower head point cloud and the standard tower head point cloud to obtain a matching matrix; Step 4: based on the matching matrix, the deformation region of the purified tower head point cloud is located to obtain the final deformation region; Step 5: in the final deformation region, the difference between the purified tower head point cloud and the standard tower head point cloud in the corresponding region is calculated to obtain the final deformation quantization result; Step 6: based on the final deformation quantization result, a comprehensive risk value is calculated, and a risk level is divided according to the comprehensive risk value to realize tower head deformation detection. 2.The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method according to claim 1, characterized in that: In step 1, the original tower point cloud is processed. ,in, Let N be the 3D coordinates of the k-th point cloud, and N be the total number of point clouds. Principal Component Analysis (PCA) is used to calculate the coordinates of each point cloud. normal vector And optimize normal vector consistency through anisotropic smoothing: wherein, is the optimized normal vector, is is - the set of neighboring points, is is - the jth reference point within the neighborhood, is the center of gravity of the neighborhood, is the distance influence factor, is the angle threshold value; The point cloud after optimizing the normal vector is voxelized according to the voxel size, and a voxel set is generated , M is the number of voxels, for each voxel , the point closest to the center of the internal distance voxel is selected as the cluster seed point ; Defining a similarity measure function with a seed point Aggregating voxels into super-voxels by K-means clustering wherein, is the number of super-voxels; and the similarity measure function is calculated as: wherein, is a weight coefficient, , is a centroid of voxels , , is a spatial distance standard deviation, , is a centroid of voxels , , is an initial average normal vector of voxels is an initial average normal vector of voxels is a transpose of Adjacent super-voxels , , the vector of the line connecting their centroids = - , and the cosine value of the angle between the vector and the line connecting vector are used to determine the concave-convex relationship and remove non-target super-voxels; wherein, is the vector of the line connecting the centroids of super-voxels, is the transpose of the average normal vector of super-voxels , is the cosine value of the angle between the normal vector of super-voxels and the line connecting vector, is the cosine value of the angle between the normal vector of super-voxels and the line connecting vector; if it is determined to be concave connection, the super-voxels with point cloud density lower than 10 points per voxel are marked as redundant noise and removed; if it is determined to be convex connection, the super-voxels are retained; and a preliminary purified tower point cloud is obtained. Based on point cloud local density and normal vector dispersion , The noise in the preliminary purified pole tower point cloud is divided into three levels of high-intensity noise, high-intensity noise and low-intensity noise; for different intensity noises, a distance-angle judgment factor is introduced to improve the traditional bilateral filtering, and the filtering formula is as follows: , wherein is the coordinate of the filtered point, is the improved bilateral filtering factor, which is dynamically adjusted in three stages: The bilateral filtering factor calculation formula for high-intensity noise suppression is: wherein, is an improved bilateral filter factor, is a distance weight, is a normal projection weight, is an intensity weight, is a transpose of the optimized normal vector; Final cleaned pole tower point cloud ; Purification of tower point cloud based on net The extraction of the tower head point cloud is first determined by the elevation distribution analysis to determine the demarcation elevation of the tower head and the tower body. The points in the tower point cloud after purification with an elevation higher than the threshold value of the tower shoulder elevation are preliminarily classified as the candidate tower head point cloud, and the main body point cloud of the tower body with an elevation lower than the threshold value of the tower shoulder elevation is simultaneously removed. Then, for each point in the candidate tower head point cloud, the number of neighborhood points within a preset horizontal search radius is counted, the point cloud density within a unit area is calculated, and the scattered debris points with a density not meeting the standard are removed. Finally, the dispersion degree of the normal vector within the neighborhood of each point in the candidate tower head point cloud is calculated, the points with a dispersion degree lower than a threshold value are retained, and the debris points with a dispersion degree exceeding the threshold value are removed. The purified tower head point cloud is finally output . 3.The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method of claim 1, wherein: In step 2, the multi-stream feature aggregation method is used to extract features from the purified tower head point cloud to obtain a multi-source stream feature set; the specific implementation process is as follows: Firstly, the original points in the point cloud of the tower head are directly extracted by a multi-layer perception (MLP) to obtain position features, and the feature extraction formula is as follows: wherein, is a feature vector of the tthoriginal point, is is a neighborhood point set, is a preset value, is is an average offset from the neighborhood points, is a normalized distance factor; By the above steps, the original point source flow feature set is obtained ; Then the tower head space is divided by the voxel size to generate a voxel set , K is the number of voxels, and the non-empty voxel is defined as the voxel containing at least one purification point. For the non-empty voxel set, a feature pyramid is constructed by four layers of three-dimensional sparse convolution, and the feature extraction formula is as follows: wherein, is a multi-scale feature vector for the mth non-empty voxel, is is the number of points within is the density statistics of points within is a 4-layer sparse convolution stack module; Through the above steps, the voxel source flow feature set is obtained And the voxel features of each purified point are mapped back to the original point dimension through bilinear interpolation ; Finally, the purified tower head point cloud Convert to cylindrical coordinates Convert to cylindrical coordinates , and generate a perspective view by gridding the cylindrical coordinates , extract the context features by 6-layer 2D CNN, and map the output feature map of the CNN back to the original point dimension by bilinear interpolation to obtain the perspective view features of each purified point , and obtain the perspective view source flow feature set through the above steps .

4. The unmanned aerial vehicle laser radar point cloud pylon tower head deformation detection method according to claim 1, characterized in that: In step 2, first, the multi-stream feature aggregation method is used to extract features from the purified tower head point cloud to obtain a multi-source stream feature set; then, a channel attention mechanism is introduced to dynamically allocate the weight of each source stream feature, and the fusion formula is as follows: wherein, is the fusion feature vector of the t-th point, , and is the channel attention weight of the t-th original point, voxel, perspective view feature, is the attention weight calculation network, is the global average pooling, is the feature vector of the t-th original point, is the voxel feature of the t-th purified point, is the perspective view feature of the t-th purified point. By the above steps, the fusion feature set is obtained ={ } Next, based on the tower head structure, the purification tower head point cloud is... Divided into M non-overlapping RoI regions For each Calculate the percentage of non-empty voxels: wherein, is the number of non-empty voxels within the RoI, is the total number of voxels corresponding to the RoI, is the non-empty voxel ratio, the larger the value, the more dense the point cloud within the RoI, the smaller the value, the more sparse. A 4-layer voxel feature pyramid is constructed, and the feature layers are aggregated according to The feature layers are aggregated dynamically, and the feature sampling is optimized through deformable RoI pooling, and the formula is as follows: wherein, is the pyramid optimized feature vector for the mth RoI, is the feature map of the fused feature at the i-th pyramid level, and is the sparsity threshold, is a deformable RoI pooling operation. Finally, pyramid optimization features of all RoI are spliced, and then normalized through 1 layer of fully connected layer to obtain tower head enhanced feature vector .

5. The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method of claim 1, wherein: In step 3, the tower head enhancement feature vector is obtained The feature channel weight is dynamically allocated through the channel attention mechanism, and the screening formula is as follows: wherein, is the screened core feature vector, is the core feature screening network, is a global max-pooling operation; The standard CAD model of the main stream pole tower head type is pre-extracted, and the corresponding core feature vector is generated according to the feature extraction process of step 2, and a feature library is constructed , the measured tower head core feature vector is calculated The cosine similarity of each standard feature in the feature library If the maximum similarity is greater than the preset similarity threshold, it is determined that the measured tower head type is the standard type corresponding to the maximum similarity; if they are all less than the preset similarity threshold, it is determined as a non-standard type; According to the determined tower head type, a corresponding standard CAD model is called, which is converted into a standard tower head point cloud, and a purified tower head point cloud is calculated The feature center of gravity is obtained by averaging the coordinates of all points of the tower head, and a standard tower head point cloud is calculated The feature center of gravity is obtained by averaging the coordinates of all points of the tower head, and a standard tower head point cloud is calculated The feature center of gravity is obtained by averaging the coordinates of all points of the tower head, and a standard tower head point cloud is calculated The translation vector is The translation vector is The translation vector is The main direction of the tower head is extracted based on the core feature The angle between the main direction of the standard model and the main direction of the measured tower head is calculated to obtain a rotation matrix The rotation matrix is used to realize a rotation transformation to make the main directions consistent Based on core features Locate the key feature points at the end points of the crossarm and the connection points of the insulator, and calculate the length of the crossarm. Tower height ,right Perform the same key feature point localization and calculate the corresponding standard dimensions. , Calculate the scaling factor ,Will Multiply the coordinates of all points by Standard tower head point cloud obtained .

6. The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method of claim 1, wherein: In step 3, the iterative closest point algorithm based on K-D tree accelerated neighborhood search is used to perform model matching on the purified tower head point cloud and the standard tower head point cloud to obtain a matching matrix; The implementation process is as follows: Based on the three-dimensional coordinates of the standard tower head point cloud , the x, y and z axes are alternately used as the segmentation axes, the point cloud space is recursively divided, and a multi-layer binary tree structure is generated. For each point in the purification tower head point cloud , , the k nearest neighbors in are quickly located through binary tree traversal of Kd-tree, and feature similarity weight is introduced to improve the iterative nearest point algorithm, and the objective function is as follows: wherein, is a rotation matrix, is a translation vector, is In the nearest neighbor point obtained by Kd-tree search in the middle, is a feature similarity weight, is each point in the purified tower head point cloud, is the total number of effective points in the purified tower head point cloud; Rotation matrix initialized to , translation vector initialized to , solve the optimal and for each iteration by singular value decomposition (SVD) , update the cleaned tower head point cloud coordinates, re-search the nearest neighbors by Kd-tree, repeat the optimization process, set two convergence conditions, stop iteration when one of them is met: after iteration convergence, the final output matching matrix 7. The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method of claim 1, wherein: In step 4, based on the matching matrix, a density clustering algorithm is used to locate the deformation region of the purified tower head point cloud to obtain the final deformation region; The implementation process is as follows: Firstly, based on the matching matrix, each point in the purified tower head point cloud is matched with the corresponding point in the standard tower head point cloud A coordinate transformation is performed to obtain the aligned purified tower head point cloud in the standard coordinate system; After alignment, the root mean square error of point cloud matching needs to satisfy RMS < threshold A, if not, return to step 3 to perform accurate matching again; Then the radius filter is used to remove the residual isolated noise points after alignment, and the continuous structure points are reserved; the voxel grid downsampling is used to reduce the number of point clouds, and the aligned point cloud is obtained ; For aligning point cloud Each point in the point cloud , search its nearest neighbor, calculate the average distance of each point to the nearest neighbor , all are sorted in ascending order, and the KNN curve is generated; the inflection point of the KNN curve corresponds to the mutation boundary of the point cloud density, and the density difference between the deformation area and the non-deformation area will form an inflection point on the curve. By calculating the first derivative of the curve , set the derivative, when the first derivative is greater than the derivative threshold, the corresponding point is the inflection point, and the average distance corresponding to the inflection point is recorded as ; right Each point in If its The number of points in the neighborhood is greater than or equal to the neighborhood threshold of the core point. If a point is identified as a core point, all points within its ε-neighborhood are associated with it. Starting from each directly associated point, all unassociated qualified points within its ε-neighborhood (points indirectly adjacent to the core point with acceptable density) are then associated. This process continues until no new qualified points can be associated. All associated points then form a complete cluster, generating multiple clusters. Points not included in any cluster are identified as noise points and are directly removed. Calculate the minimum bounding box area of each cluster , eliminate small clusters less than a preset value; calculate the distance between the cluster center and the corresponding region center of the standard model , if less than a preset value, it is determined as non-shape difference and eliminated; The cluster left after the screening is the final deformation region , record the point set of each deformation region and the range of spatial coordinates.

8. The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method of claim 1, wherein: In step 5, the improved Hausdorff distance algorithm is used to calculate the difference between the purified tower head point cloud and the standard tower head point cloud in the corresponding region to obtain the final deformation quantization result; The specific implementation process is as follows: Based on the final deformation region Spatial coordinate range Cloud point at the standard tower head Extract the standard point set of the corresponding region ; With the minimum bounding box of the standard point set as the initial space, it is recursively divided into 8 sub-cubes by level until the number of points contained in each leaf node is ≤ threshold B, generating an octree ; By improving the Hausdorff distance to calculate the spatial difference between the deformed region and the standard region, three core quantitative indicators are defined, including the average deformation , the maximum deformation and the tilt angle , which adapt to different types of deformation; wherein, is the number of points of the deformed region , and is the nearest point distance from all points of the deformed region to ; the maximum deformation is defined to take the maximum value of ; the tilt angle quantifies the tilt degree, wherein, is the normal vector of the plane fitted by the point set of the deformed region, and is the normal vector of the plane fitted by the point set of the standard region; Perform constraint correction on the quantitative indicators: if the point >N , it is determined as an abnormal value, which is removed and then recalculated and , the inclination angle is limited in the range of [0, M], and if it exceeds the range, it is recorded as M and marked as a serious deformation; wherein N and M are preset values. obtaining a final deformation quantification result, the final deformation quantification result including a set of quantification indices for each deformation region and corresponding spatial location coordinates.

9. The unmanned aerial vehicle LiDAR point cloud pole head deformation detection method according to any one of claims 1-8, characterized in that: In step 6, the average deformation is taken into account. Maximum deformation Inclination angle The three types of quantitative indicators are normalized to obtain a normalized indicator set. ; A weighted sum is introduced to calculate the comprehensive risk value, and the calculation formula of the comprehensive risk value is as follows: wherein, is the comprehensive risk value of the kth deformation region, the larger the value, the higher the risk, is the structure sensitivity weight of the region, , and is the index weight. 10.A unmanned aerial vehicle laser radar point cloud pole head deformation detection system, characterized in that, It includes: One or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the unmanned aerial vehicle laser radar point cloud tower head deformation detection method according to any one of claims 1 to 9.

Images