Power transmission tower intelligent modeling method and system based on three-dimensional point cloud data

Abstract

The invention relates to the technical field of three-dimensional modeling, in particular to a transmission tower intelligent modeling method and system based on three-dimensional point cloud data. The method comprises the following steps: obtaining area scanning data, and carrying out tower area identification according to the area scanning data to obtain tower area data; carrying out tower extraction on the tower region data to obtain tower segmentation data; carrying out improved component semantic recognition on the tower segmentation data to obtain tower component data; performing structure topological relation processing on the tower member data to obtain tower structure topological graph data; and constructing a power transmission tower model according to the tower structure topological graph data to obtain the power transmission tower model. According to the method, the automation degree and the structure recognition precision of power transmission tower modeling are improved, and the spatial distribution and connection relation of various components can be accurately restored.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and in particular to an intelligent modeling method and system for power transmission towers based on 3D point cloud data. Background Technology

[0002] In the current digital operation and maintenance system of smart grids and transmission lines, transmission towers, as critical infrastructure, rely heavily on digital modeling and visualization of their structural state for support of operation and maintenance decisions. Traditional tower modeling methods largely depend on manual surveying, reconstruction of structural drawings, or matching of rule templates, resulting in low automation, poor adaptability to non-standard components, low modeling efficiency, and missing model semantics. In recent years, with the application of technologies such as LiDAR and UAV remote sensing, 3D point clouds have become an important means of acquiring spatial structural data of towers. However, how to reconstruct a 3D model with clear structural semantics and well-defined connections from massive amounts of unstructured point clouds has become a challenge. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an intelligent modeling method and system for power transmission towers based on three-dimensional point cloud data, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a method for intelligent modeling of power transmission towers based on 3D point cloud data, the method comprising:

[0005] S1. Obtain area scanning data, identify tower areas based on area scanning data to obtain tower area data; extract towers from tower area data to obtain tower segmentation data.

[0006] S2. Perform improved component semantic recognition on the tower segmentation data to obtain tower component data;

[0007] S3. Perform structural topology processing on the tower component data to obtain tower structure topology data;

[0008] S4. Construct the transmission tower model based on the tower structure topology data to obtain the transmission tower model.

[0009] In this invention, an improved component semantic recognition mechanism is introduced in step S2, which not only enables fine-grained identification of key internal components of the tower (such as the tower body, crossarms, and insulator attachment points), but also enhances the robustness of identifying irregular or damaged components. Step S3 constructs a component connection map through structural topological relationship processing, clarifying the physical connections and hierarchical relationships between components, effectively supporting the structural constraints of subsequent model construction. Compared to existing methods that only perform geometric fitting or template matching, step S4 combines topological structure to reconstruct the 3D model, which can dynamically adapt to various tower types and structural deformation scenarios, ensuring that the modeling results have higher engineering interpretability and visualization interaction value. Therefore, this method not only improves the accuracy and efficiency of tower model reconstruction, but also significantly enhances the model's adaptability to complex scenarios (such as mountain slopes and tilted towers).

[0010] Optionally, S1 includes:

[0011] Acquire area scan data;

[0012] Preliminary filtered data is obtained by filtering based on height thresholds using regional scan data.

[0013] The vertical structure density of the initial screening data is detected to obtain secondary screening data.

[0014] Deep learning segmentation is performed on the secondary filtered data to obtain tower area data;

[0015] Target geometric data is obtained by fitting the target geometry based on the tower area data;

[0016] Structural clustering is performed on the target geometric data to obtain tower segmentation data.

[0017] This invention employs height threshold screening to effectively exclude low-lying non-structural elements such as ground and vegetation, improving data processing efficiency. Subsequently, vertical structure density detection further highlights the continuity and characteristic distribution of towers in the vertical direction, enhancing the accuracy of target area localization in complex backgrounds. Combined with deep learning segmentation technology, intelligent identification of tower areas is achieved, adapting to various terrains and occlusion scenarios. Then, representative morphological features are extracted through target geometric fitting, and structural clustering technology is used to effectively separate towers from non-tower components at the spatial structure level, outputting tower segmentation data with complete component boundaries.

[0018] Optionally, S2 includes:

[0019] Using pre-defined tower structure spatial data, spatial constraint clustering is performed on the tower segmentation data to obtain initial component data;

[0020] An improved semantic network recognition method is used to identify the initial component data to obtain the tower component data.

[0021] This invention uses spatially constrained clustering of tower segmentation data based on pre-defined tower structure spatial data. This fully leverages the prior geometric characteristics of the tower (such as symmetry, hierarchy, and spatial arrangement rules) to perform preliminary segmentation of components based on their positional relevance, avoiding common missegmentation problems in traditional point cloud clustering. The subsequently introduced improved semantic network recognition model integrates local geometric features with component contextual information, improving the accuracy of component recognition (such as crossarms, insulator attachment points, and reinforcement supports), making it suitable for component analysis of heterogeneous tower structures.

[0022] Optionally, the spatially constrained clustering includes:

[0023] The tower segmentation data is divided into point cloud spatial blocks to obtain point cloud spatial segmentation data;

[0024] Using the preset tower structure spatial data, the point cloud spatial division data is matched with component spatial constraint rules to obtain point cloud label data;

[0025] Clustering of similar point regions based on point cloud label data yields initial component data.

[0026] This invention effectively reduces the complexity of global point cloud processing by dividing the tower segmentation data into point cloud spatial blocks, while preserving local geometric and spatial relationship features. Subsequently, spatial constraint matching is performed on the divided units using pre-set tower structure spatial data (such as typical height ranges, relative positions, and support relationships of each component). This embeds the topological prior knowledge of the components into the clustering process, avoiding the problems of improper handling of boundaries, irregularities, or connection parts in traditional unsupervised clustering. Clustering similar point regions based on label data yields initial component data with semantic similarity and structural integrity, providing accurate candidate regions for deep semantic recognition and geometric modeling.

[0027] Optionally, the improved semantic network recognition is performed using a pole and tower component recognition model, the construction steps of which include:

[0028] Obtain historical initial component data;

[0029] Graph neural network processing is applied to the historical initial component data to obtain component graph feature data;

[0030] Local connectivity dependency feature extraction and global structural attention feature extraction are performed on the component graph feature data to obtain local connectivity dependency feature data and global structural attention feature data, respectively.

[0031] Feature fusion is performed on local connectivity-dependent feature data and global structural attention feature data to obtain feature fused data;

[0032] Elevation datum surface is extracted from the historical initial component data to obtain elevation datum surface data;

[0033] Relative elevation features are extracted from elevation datum data to obtain elevation feature data;

[0034] Elevation-guided feature fusion is performed on elevation feature data and feature fusion data to obtain elevation feature fusion data.

[0035] The tower structure hierarchical feature data is obtained by performing tower structure hierarchical feature characterization on the elevation feature fusion data.

[0036] Based on the structural hierarchical feature data of the tower, the semantic relationship regularization term of the components is processed to obtain the training data, and the constraint optimization is performed based on the training data to obtain the tower component recognition model.

[0037] This invention utilizes historical initial component data to construct a component graph structure and extracts local connectivity dependencies and global structural attention features between point clouds through a graph neural network, effectively capturing the spatial correlation within components and the structural context information across components. By extracting elevation datum surfaces and modeling relative elevation features, vertical distribution characteristics are introduced as auxiliary variables to enhance the model's sensitivity to the vertical distribution patterns of the tower structure. After fusing elevation features with multi-scale semantic features, the hierarchical features of the tower structure are further extracted, achieving a layered representation of components in different structural layers (such as the tower base, tower body, and tower head). Constraint optimization of the model through component semantic relationship regularization not only improves the clarity of semantic boundaries but also significantly suppresses misidentification and cross-class interference.

[0038] Optionally, the local connectivity dependency feature extraction includes:

[0039] Spatial hierarchical encoding is performed on the component diagram feature data to obtain the component diagram structural data;

[0040] The adjacent node data is obtained by filtering the structural data of the component diagram using a connection mask.

[0041] By applying component status guidance to adjacent node data, node aggregation data is obtained;

[0042] Based on the node aggregation data, semantic guidance is performed on the component connection edges to obtain the edge semantic connection data;

[0043] Component relation residual enhancement is performed based on edge semantic connection data to obtain local connection dependency feature data.

[0044] This invention employs spatial hierarchical encoding of component graph feature data, explicitly encoding the spatial location and semantic hierarchy of components within the structural hierarchy. This helps distinguish the connection methods and functional attributes between components at different levels. Redundant connections are eliminated through connection masking, retaining only adjacent nodes with physical or logical significance, improving the targeting and robustness of feature aggregation. Guided by component state, the features of first-order adjacent nodes are dynamically aggregated, ensuring that feature updates consider not only spatial adjacency but also the structural state information of the nodes themselves. Subsequently, edge semantic guidance is applied based on the node aggregation results, assigning semantic labels (such as support, suspension, connection) to connecting edges, enhancing the interpretability of the topological relationships between components. Finally, a component relationship residual enhancement mechanism introduces a difference enhancement channel, retaining the most sensitive change signals in the connection relationships, thereby generating locally connected dependency feature data with significant discriminative power.

[0045] Optionally, S3 includes:

[0046] Component node diagrams are generated from the tower component data to obtain component diagram data;

[0047] The component diagram data is processed to generate candidate connection relationships between component pairs, thus obtaining the component connection diagram data.

[0048] Based on the component connection diagram data, the topological relationships are classified and labeled to obtain the tower structure topology diagram data.

[0049] This invention maps tower component data into a graph structure, constructing a component node graph. This not only preserves the geometric and semantic features of each component but also provides a unified structured framework for subsequent connection relationship modeling. Based on rules such as spatial adjacency, relative direction, and elevation distribution, candidate physical connections between node pairs are generated, effectively covering complex relationships commonly found in tower structures, such as rigid connections, suspended connections, and multi-component intersections. Subsequently, graph classification and edge classification algorithms are used to topologically label these candidate connections, clarifying the type and hierarchical affiliation of each connection, thus achieving complete modeling of the tower structure from geometric components to topological semantics.

[0050] Optionally, the generation of candidate connection relationships by the component includes:

[0051] Enumerate the component node pairs in the component diagram data to obtain the component node pair data;

[0052] Spatial proximity filtering is performed on the component node pair data to obtain filtered node pair data;

[0053] Based on the node pairs, the filtered data is matched by component type to obtain the node pair combination data;

[0054] Geometric angle matching calculation and type combination prior weight calculation are performed on the node-pair combination data to obtain geometric angle matching data and type combination prior data, respectively.

[0055] Based on the geometric angle matching data and the prior data of type combination, the connection edges of the node-pair combination data are filtered to obtain the connection edge data;

[0056] The component connection diagram data is generated by generating a component connection diagram based on the connection edge data and component nodes.

[0057] This invention enumerates node pairs in the component diagram data to comprehensively cover all possible component combinations, providing a complete candidate set for connection relationship modeling. Subsequently, a spatial proximity filtering mechanism is employed to eliminate node pairs that are too far apart or lack physical connection possibility, significantly reducing redundant computation and improving processing efficiency. Building upon this, component type combination matching rules are introduced to semantically screen the rationality of physical connections between different types of components (such as tower columns and crossarms, crossarms and insulator attachment points), ensuring that the connection logic conforms to engineering reality. Next, geometric angle matching calculations are used to accurately analyze the spatial directional relationships between components, and the connection probability is evaluated by combining the prior weights of component type combinations, further enhancing the interpretability and discriminativeness of connection prediction. A multi-condition fusion strategy is used to filter connection edges for node pair combinations, generating a component connection diagram and constructing a topological relationship map that highly conforms to the actual structural layout.

[0058] Optionally, S4 includes:

[0059] The component geometry is generated based on the tower structure topology data to obtain the component geometry data;

[0060] Structural connection-driven assembly is performed based on component geometry data to obtain component assembly data;

[0061] The component assembly data is fitted with component parameters to obtain component fitting data;

[0062] Based on the component fitting data, a semantic 3D model is constructed to obtain the transmission tower model.

[0063] This invention generates component geometry based on tower structure topology data, giving each component an independent and adjustable geometric representation, facilitating flexible assembly and shape adjustment. Next, the topological connection relationships drive the component assembly process, ensuring a high degree of consistency between the model's assembly sequence and the component connection logic, suitable for tower structures with multiple intersecting or cross-layer connections. Subsequently, a component parameter fitting stage is introduced, using real point cloud data to finely calibrate the scale, shape, and other parameters of each component geometry, improving the model's performance in dimensional accuracy and structural fidelity. Combining structural semantic tags with the fitting results generates a complete semantic 3D model with excellent visualization, structural analysis, and engineering expansion capabilities.

[0064] Optionally, this application also provides an intelligent modeling system for transmission towers based on 3D point cloud data, used to execute the intelligent modeling method for transmission towers based on 3D point cloud data as described above, wherein the intelligent modeling system for transmission towers based on 3D point cloud data includes:

[0065] The point cloud region segmentation module is used to acquire region scanning data, identify tower regions based on the region scanning data to obtain tower region data, and extract towers from the tower region data to obtain tower segmentation data.

[0066] The component semantic recognition module is used to perform improved component semantic recognition on the pole segmentation data to obtain pole component data;

[0067] The topology parsing module is used to process the structural topology relationships of tower component data to obtain tower structure topology data.

[0068] The 3D modeling and structural reconstruction module is used to construct transmission tower models based on tower structure topology data, thus obtaining transmission tower models.

[0069] The purpose of this invention is to accurately extract tower regions through elevation threshold screening and vertical structural density detection, and to achieve fine-grained separation of various components (such as tower columns, crossarms, and hanging points) by combining deep learning segmentation networks and tower space rules. Subsequently, a graph neural network is introduced to construct a component topology graph, integrating geometric relationships and type priors to complete edge identification and structural semantic annotation. In the geometric modeling stage, component assembly is driven by the topological structure, and a point cloud residual fitting mechanism is introduced to optimize component parameters, ensuring that the generated 3D model maintains a high degree of consistency in geometric accuracy and connection logic. The constructed semantic 3D model not only possesses a complete structural hierarchy and semantic attributes, but also supports standardized export for digital twin platforms. Attached Figure Description

[0070] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0071] Figure 1 A flowchart illustrating the steps of an intelligent modeling method for power transmission towers based on 3D point cloud data is shown in one embodiment.

[0072] Figure 2 A flowchart illustrating the steps of a point cloud region segmentation method according to an embodiment is shown.

[0073] Figure 3 A flowchart illustrating the steps of a topology analysis method according to an embodiment is shown.

[0074] Figure 4 A flowchart illustrating the steps of a three-dimensional modeling and structural reconstruction method according to an embodiment is shown.

[0075] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0076] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0077] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0078] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0079] Please see Figures 1 to 4 This application provides a method for intelligent modeling of power transmission towers based on 3D point cloud data, the method comprising:

[0080] S1. Obtain area scanning data, identify tower areas based on area scanning data to obtain tower area data; extract towers from tower area data to obtain tower segmentation data.

[0081] In one embodiment, the system uses a drone or airborne platform equipped with LiDAR to inspect the power transmission line area and acquire raw 3D point cloud data covering the towers and their surrounding environment. The collected data can be saved as a standard point cloud format file, such as .las or .pcd. Height threshold filtering is performed. The system sets a ground height threshold based on the actual scene, for example, setting data points with a Z-axis coordinate higher than 8 meters as initial targets to eliminate low-elevation point cloud data such as ground, shrubs, and low-lying structures, thus obtaining a preliminary filtered dataset. Vertical structure density detection is then performed. The system divides the preliminary filtered point cloud into fixed-size grid windows (e.g., 1 meter × 1 meter) in the XY plane and calculates the point distribution density and height hierarchy continuity in the Z-direction of each grid. If a grid has a continuous and high-density point layer structure in the vertical direction, it is determined that it may contain vertical components and is marked as a candidate tower area, forming secondary filtered data. Deep learning semantic segmentation processing is performed on the secondary filtered data. The system calls a pre-trained 3D point cloud segmentation network model (such as PointNet++ or KPConv) to classify the point cloud data point by point, outputting point-level semantic labels such as "tower," "cable," and "background." Only points identified as "towers" are retained to form the tower region data. After obtaining the tower region, target geometry fitting is performed. The system uses the Random Sample Consensus Algorithm (RANSAC) to perform basic geometry fitting on the local point cloud within the region, including typical component shapes such as cylinders and cuboids, to extract preliminary structural features and spatial distribution information, forming target geometric data. Structural clustering is then performed. Based on the point cloud geometric consistency and spatial continuity rules, the system uses an improved DBSCAN algorithm that combines point density thresholds and fitting residual control to perform component-level clustering of the point cloud. During clustering, not only the Euclidean distance between points is considered, but geometric fitting error is also introduced as a clustering weight adjustment factor, thereby effectively distinguishing different structural components. The output result is tower segmentation data with clear component boundaries.

[0082] S2. Perform improved component semantic recognition on the tower segmentation data to obtain tower component data;

[0083] In one embodiment, spatially constrained clustering is performed. The system spatially divides the tower segmentation data obtained in step S1 into blocks along the Z-axis, uniformly dividing the point cloud data into multiple vertical blocks from bottom to top, for example, dividing each block into a sub-block every 5 meters. Subsequently, based on a preset tower structure spatial template (such as a "three-segment tower structure model"), the tower space is divided into the tower base segment, tower column segment, and tower head crossarm segment, and the typical distribution range of each segment in the vertical space is defined. For example, the tower base is usually located at the bottom 0–8 meters, the tower column is concentrated in the middle 8–25 meter area, and the tower head and crossarm are mainly distributed in the top 25 meters and above. Based on the center elevation value and point density distribution characteristics of the sub-blocks, the system determines the structural segment to which it belongs and assigns preliminary component labels, thereby forming an initial set of component sub-blocks with spatial semantics. Semantic recognition driven by a graph neural network is then performed. The system uses the aforementioned spatial sub-blocks as graph nodes to construct a component graph structure G = (V, E), where V represents the set of component sub-block nodes and E represents the set of edges between components presumed to be spatially adjacent or physically connected. For each graph node, its local spatial geometric features are extracted, including but not limited to voxel size, fitting residual, point density, and morphological fitting type (e.g., columnar, plate-like). Simultaneously, the system further extracts contextual structural features between nodes, such as relative elevation difference, spatial connection direction, and edge similarity. These features are embedded into a graph neural network, and a graph attention mechanism is used to train the nodes for information aggregation and category recognition, outputting semantic category labels for each component node, such as "main tower column," "crossarm," "insulator hanging point," and "auxiliary support." The system performs elevation feature fusion processing on the component nodes. Based on the elevation datum of the tower area, the relative elevation of each component is extracted, and its structural hierarchical position in the overall tower structure (e.g., bottom, middle, top) is determined according to its vertical distribution. The system uses this elevation level information as a location encoding feature and fuses it with the aforementioned graph embedding features to enhance the model's ability to distinguish between upper and lower structures of a component. This helps reduce semantic ambiguity in cases where components have similar shapes but are located at different levels. The system outputs a dataset of tower components with clear semantic labels and spatial boundaries.

[0084] S3. Perform structural topology processing on the tower component data to obtain tower structure topology data;

[0085] In one embodiment, a component node graph is generated. The system treats each component that has undergone semantic recognition in the previous step as a node in the graph structure, with each node corresponding to a component instance. For each component node, the system extracts the following as node features: the three-dimensional coordinates (X, Y, Z) of the center point, the geometric shape type of the component's outer body (e.g., cylinder, cuboid, etc.), the size range (e.g., length, width, height), and the identified component category label (e.g., "tower column," "crossarm," "hanging point," etc.). After construction, an initial set of component graph nodes is formed to describe the spatial component distribution in the tower structure. Candidate connection relationships for component pairs are then generated. The system traverses all node pairs (pairwise combinations of components) in the component graph and calculates the Euclidean distance between components. If the spatial distance between components is less than a preset connection threshold (e.g., 3 meters), the node pair is retained as a candidate connection relationship. For each retained pair of components, the angle between their principal axis directions is calculated (obtained through the principal component directions of the component point cloud). Combining the prior connection relationship rules of component types, a set of type combination weight matrices is defined (either manually input or by calling a preset weight library). For example, a high connection probability weight is assigned to the tower column and crossarm (strong connection), while a medium weight is assigned to the tower column and the hanging point (weak connection), and crossarms are assigned low weight or no connection. The system calculates a connection confidence score for each candidate component pair based on the geometric angle consistency and type combination weights. If the connection confidence score is higher than a set threshold (e.g., 0.6), the component pair is determined to have a structural connection relationship, and a connection edge is added to the component diagram to form a preliminary component connection diagram. The system then performs topological structural relationship classification and labeling. Based on the spatial orientation features of the connection edge, component type combinations, and relative elevation differences, the system uses decision rules to semantically classify the connection relationships. Specifically, when the center height of component A is significantly higher than that of component B, and the component type is a "crossarm-tower column" combination, it is determined to be a "support connection"; if the component is suspended on the side of another component, it is determined to be a "suspended connection"; if the components are at the same level and placed parallel to each other, they are marked as "parallel connection". The labeling results are appended to the attributes of the connection edges to form a structural semantic topological diagram. The system outputs a topology diagram of the tower structure, which includes node semantics, connection edge types, and spatial structural relationships.

[0086] S4. Construct the transmission tower model based on the tower structure topology data to obtain the transmission tower model.

[0087] In one embodiment, component geometry generation is performed. The system generates an initial geometric model by calling a predefined standard geometric template library based on the component category corresponding to each node in the component topology diagram. The template library contains various typical component shapes; for example, tower columns use multi-segmented cylindrical structures, crossarms use flat cuboids or T-shaped beams, and hanging points and auxiliary components can use spheres, short columns, etc. The initial parameters of the template are provided by component semantic annotations, such as type and size range. Simultaneously, the key parameters of the geometric template are adaptively adjusted based on the previous point cloud fitting results, including component height, cross-sectional radius, and direction vector, to ensure a high degree of consistency between the geometric model and the actual point cloud structure. Structural connection-driven assembly is then performed. Based on the connection edge information in the tower structure topology diagram, the system performs assembly operations sequentially according to structural dependencies. The assembly sequence typically follows a logical order from bottom to top and from main components to auxiliary components; that is, first assembling the tower base and tower columns, then assembling the crossarms, and finally installing hanging points and auxiliary supports, etc. During assembly, the system introduces component connection interface rules to ensure that the connection methods between components conform to the actual structural characteristics. For example, the top of the tower column needs to be aligned with and embedded in the center connection interface of the crossarm, and the hanging point needs to be attached to the reserved attachment position on the side of the crossarm. Interface types include face-to-face, point-to-face, and edge-to-edge, and the system automatically identifies and matches the interface shape based on geometric features. Parameter fitting and error optimization are then performed. After assembly, the system performs spatial overlap analysis between the initially constructed 3D structure and the original point cloud data. Specifically, residual distance is calculated for each component model and its corresponding point cloud subset, and the position, orientation, and size parameters of the component are fine-tuned using the least squares method as the optimization objective, thereby reducing fitting errors. This optimization process updates the spatial transformation matrix of each component and can introduce connection stiffness or relative offset tolerance models, enabling the assembled structure to maintain physical connection logic while achieving higher point cloud fitting accuracy. Semantic 3D model construction is then completed. The system adds corresponding semantic label information to each component, such as "main tower column," "crossarm," "support," or "hanging point," and establishes parent-child connection relationships between components for hierarchical management. The generated transmission tower model not only possesses a complete three-dimensional geometric shape, but also includes structural semantics and spatial connection topology information. The model can be exported to various industry-standard formats, such as the industrial basic model format .IFC, the three-dimensional general format .OBJ, or a BIM platform-specific format that supports structural identification and operation and maintenance management.

[0088] Optionally, S1 includes:

[0089] S11. Acquire area scan data;

[0090] In one embodiment, a drone equipped with a lidar system (such as Livox or Velodyne) is used to perform a panoramic scan of the area surrounding the power transmission line, collecting three-dimensional point cloud data in .las or .pcd format. The flight path spacing is set to no more than 30m, the altitude to be between 80 and 120 meters, and the scanning resolution to be no less than 20 points / m².

[0091] S12. Perform height threshold filtering based on the area scan data to obtain preliminary filtered data;

[0092] In one embodiment, the system performs height-dimension filtering on the acquired raw 3D point cloud data. Each point cloud sample contains its position coordinates in 3D space, namely, horizontal coordinates, vertical coordinates, and vertical height. The system sets a minimum height threshold, for example, eight meters, as a filtering condition. During execution, all point cloud data with height values ​​below this threshold will be removed to effectively exclude low-lying, unstructured point clouds such as terrain features, shrubs, temporary facilities, and vehicles.

[0093] S13. Perform vertical structure density detection on the preliminary screening data to obtain secondary screening data;

[0094] In one embodiment, the system projects the initially filtered point cloud data onto a horizontal coordinate plane and divides it into two-dimensional grid cells at fixed intervals, for example, each grid cell has a side length of one meter. For each grid cell, the system sorts the point cloud data within it according to its vertical height value and divides it into multiple consecutive slice layers in the height direction with a fixed step size (e.g., 0.5 meters). The system counts the number of consecutive slice layers with point cloud distribution and calculates the average point density of the grid in three-dimensional space. If a grid has three or more consecutive effective height slice layers and the average point density is not lower than a set threshold (e.g., thirty points per cubic meter), it is determined to be a potential candidate region for vertical structures. The system summarizes the point cloud data corresponding to all grid cells that meet the conditions to form a secondary filtering dataset.

[0095] S14. Perform deep learning segmentation on the secondary filtered data to obtain tower area data;

[0096] In one embodiment, the system performs point-by-point semantic recognition processing on the 3D point cloud data obtained from secondary filtering using a deep learning-based point cloud semantic segmentation model. The model used is a 3D structure-aware network pre-trained on a standard point cloud dataset, such as a neural network model with local feature aggregation capabilities. The model's input features include the 3D spatial coordinates (horizontal, vertical, and height coordinates) of each point, the corresponding normal vector information (reflecting the directionality of the surface where the point is located), and local curvature indices (representing the geometric shape change trend at the point). During classification, the model maps each point to its corresponding semantic category. Preset labels include "tower structure," "power transmission cable," "surface vegetation," and "background area," etc. The system filters and retains all point cloud samples labeled "tower structure" and uses them as tower area data.

[0097] S15. Perform target geometry fitting based on tower area data to obtain target geometry data;

[0098] In one embodiment, the system performs vertical layering processing on the extracted 3D point cloud data of the tower area. Specifically, the overall point cloud data is segmented along the Z-axis at fixed height intervals (e.g., every five meters) to construct a layer-by-layer point cloud data structure. At each layer, the system performs geometric structure fitting on the local point cloud, such as randomly selecting point pairs or sets, constructing candidate geometric models, validating in-model points, repeating sampling, and scoring the model to obtain the parameter results of the geometric structure fitting. For the main tower area, it preferentially fits an axisymmetric cylinder; while for transverse component areas such as crossarms, it attempts to fit a cuboid or a plane. Each successfully fitted geometry records its key structural parameters, including the geometric center position, principal axis direction vector, geometric dimensions (e.g., height, radius, or side length), and fitting confidence (indicating the degree of matching between the fitted model and the original point cloud). The system summarizes all qualified fitted structures into a target geometric data set.

[0099] S16. Perform structural clustering on the target geometric data to obtain tower segmentation data.

[0100] In one embodiment, the system performs component-level clustering based on the aforementioned target geometric data set, according to the spatial proximity and morphological similarity between geometric segments. Spatial proximity can be measured by the Euclidean distance between geometric segments, while morphological similarity is determined based on their principal direction vector, size parameters, and geometric shape features. The system constructs a spatial relationship network based on a graph structure, where each geometric segment is treated as a node in the graph structure. When the spatial distance between any two geometric segments is less than a set threshold (e.g., one meter) or the angle between their direction vectors is less than a specified angle (e.g., ten degrees), a connection edge is established between them, indicating that they are related. Based on this graph structure, the system employs an improved density-based spatial clustering algorithm (DBSCAN). Unlike the traditional DBSCAN, which is based on "point-to-point distance," the system assigns a weighted value to the connection edge between each pair of segments in the graph structure. The weight consists of two items: spatial distance weight (e.g., the reciprocal of the Euclidean distance) and geometric fitting residual weight (e.g., the error of fitting to a cylinder or plane; the smaller the error, the higher the weight). Synthetic weights are used to determine density reachability between segments, replacing traditional pure distance thresholds. During clustering, a fixed radius is no longer used; instead, the clustering neighborhood is adaptively adjusted based on component orientation consistency and local residual characteristics. For example, if components within a region have consistent orientation height, the system will automatically relax distance restrictions to aggregate more components. The system introduces a structural confidence index, combining fitting error, orientation variance, spatial distribution, and other parameters to determine core segments. For areas with prominent density changes at structural edges, a boundary optimization mechanism based on structural continuity detection is introduced to prevent structural segments from being misclassified as outliers due to critical distances. After initial clustering, geometric fitting verification is performed on each cluster unit; if the overall residual is greater than a threshold or the orientation dispersion is high, the cluster is split. The point cloud data corresponding to all clustering results are labeled and recombined to form a component-level semantic grouping dataset.

[0101] Optionally, S2 includes:

[0102] Using pre-defined tower structure spatial data, spatial constraint clustering is performed on the tower segmentation data to obtain initial component data;

[0103] In one embodiment, the tower segmentation data obtained in step S1 is first subjected to spatial layering along the Z-axis. The Z-axis is a vertical spatial axis, and the layering height is set to four meters per layer, with the specific height adjustable according to typical tower configuration characteristics. For any point in the point cloud data, its vertical coordinate value is subtracted from the Z-coordinate of the lowest point in the current dataset, divided by the preset layering height, and rounded down to determine the spatial layer to which the point belongs. The system loads a set of preset tower structure spatial templates to represent the component composition pattern of typical towers, such as a "three-segment structure": tower base, tower column, and tower head. The spatial distribution range of each part has been predefined, for example, the tower base corresponds to the lower 20% of the tower height, the tower column to the middle 60%, and the tower head to the upper 20%. The system matches the layered point cloud data with the spatial range of each structural segment according to its height, thereby assigning corresponding initial component labels to each point. Local clustering operations are performed within each spatial layer to achieve preliminary aggregation of components. This clustering method starts with the point with the highest point cloud density in each layer as the seed point, and recursively searches for its neighboring points within a spatial radius (0.5 meters). If a neighboring point satisfies the minimum density and geometric similarity (similarity greater than 0.85), it is added to the current cluster and continues to expand outwards until it can no longer grow. If the number of points after clustering is less than 100, they are discarded. After the clusters are formed, the system analyzes the geometric attributes of each cluster object, specifically including its length-width-height ratio and the fitting error of its spatial shape. When the geometric shape of a cluster can be approximately fitted to a regular component (such as a cylinder or square prism) and the fitting error does not exceed 0.1 meters, the cluster is retained as a candidate component. The system generates an initial set of components with good spatial coherence and geometric consistency, denoted as the initial component set. Each item in this component set represents a structural component that has been successfully identified and satisfies the geometric constraints.

[0104] An improved semantic network recognition method is used to identify the initial component data to obtain the tower component data.

[0105] In one embodiment, a component graph structure is constructed. Each component in the initial component set is considered a node in the graph, and an undirected graph structure is constructed. Nodes represent individual initial component objects. Edges in the graph are established based on the spatial connections between components; for example, components are considered connected if the Euclidean distance between them is less than 1 meter or if they are at the same spatial level. Furthermore, the system records candidate pairs of component types; for example, a common connection is between a main tower column and a crossbeam. The system extracts node and edge features from the graph structure. Node features include the component's geometric parameters (such as size, shape fitting residuals), point cloud density features, the component's Z-axis height position, fitting type (e.g., approximating a cylinder or cuboid), and the component's centroid spatial position. Edge features include the spatial orientation angle between two components, the Euclidean distance, and their connection category. After feature construction, the system uses an improved semantic graph neural network for inference and recognition. In this embodiment, a graph attention network model is used to encode the component graph. During model training and inference, not only are the features of each node considered, but also the semantic relationship information between its adjacent components is introduced. The model introduces two auxiliary pieces of information in each round of graph information propagation: one is the relative elevation embedding, which expresses the relative position of nodes in the vertical direction; the other is the component type prior weight matrix, which reflects the high-confidence connection relationship between specific component pairs. For example, the connection between the main tower column and the crossarm is a common pattern, and is given a higher association weight. 0.9, Hanging point: 0.8 0.3, 0.6, 0.1, 0.5, other combinations without explicit connections: 0.0~0.2 (set according to experience and sample statistics). The model is trained using supervised learning, with the goal of enabling each component node to output a corresponding semantic label. Semantic categories include, but are not limited to, main tower column, crossarm, support, insulator hanging point, and other component types. After initial identification, the system performs post-processing steps. Specifically, this includes secondary aggregation of component regions with inconsistent labels but similar shapes and spatial proximity; if obvious semantic conflicts are found (e.g., a component located at the top of the tower is misidentified as a tower base), logical correction is performed based on the component's elevation information and global identification probability. Through the above post-processing process, a tower component dataset with accurate label assignment and clear spatial boundaries is obtained. This dataset is used as output and is denoted as the component set after structural identification.

[0106] Optionally, the spatially constrained clustering includes:

[0107] The tower segmentation data is divided into point cloud spatial blocks to obtain point cloud spatial segmentation data;

[0108] In one embodiment, the system performs spatial block partitioning on the point cloud dataset segmented from the output tower data to facilitate subsequent component identification and clustering operations. This point cloud dataset is denoted as P, where each point contains its three-dimensional coordinate information in the form (x...). i ,y i ,z i The system performs voxelization on the point cloud data in three-dimensional space, dividing the entire space into several cubic voxel units of uniform size. Specifically, the resolution of the spatial grid is set to 0.5 meters in the horizontal direction (X and Y axes) and 1 meter in the vertical direction (Z axis). This resolution can be adjusted appropriately according to the geometry of the tower and the density of the point cloud. For any point in the point cloud, its voxel unit is obtained by calculating the distance of that point relative to the minimum boundary coordinate in each dimension, dividing by the voxel size, and rounding down. Specifically, the minimum values ​​of the X, Y, and Z dimensions in the point cloud dataset are determined, denoted as x. min y min With z min The system calculates the offset of a point relative to its starting position, divides it by the resolution per dimension, and rounds down to determine the voxel index to which the point belongs. The system then performs voxel mapping on all points and counts the number of points, point density (i.e., the number of points per unit volume), and spatial distribution range of points (e.g., bounding box size) within each voxel. Through these operations, a set of spatially partitioned data units, denoted as a voxel set, is generated, with each voxel containing its statistical attributes.

[0109] Using the preset tower structure spatial data, the point cloud spatial division data is matched with component spatial constraint rules to obtain point cloud label data;

[0110] In one embodiment, the typical tower structure spatial data includes spatial layout rules for different component areas. The height range rule distinguishes the vertical distribution range of each functional component, specifically including: tower base area: height between 0 and 8 meters; main tower body area: height between 8 and 25 meters; crossarm area: height between 25 and 30 meters; and top auxiliary area: height above 30 meters. Furthermore, the preset component distribution pattern includes: main tower columns should be continuously distributed along the vertical direction (i.e., the Z-axis) and have an approximately columnar geometric shape; crossarms should have horizontal extension characteristics and a distinct lateral protrusion; insulator hanging points should be located on both sides of the crossarm, symmetrically distributed, and with relatively high local point density. During the matching process, the system processes each spatial voxel unit. For the center point of each voxel, its three-dimensional coordinates are represented as (x... i ,y i ,z iThe system determines the height range corresponding to the vertical coordinate value of a point, thereby identifying the possible component region to which it belongs. Further discrimination is performed based on the geometric features of the point cloud contained within the voxel. Specific judgment indicators include the spatial orientation distribution of points, point density distribution, and statistical measures of principal components of directions. For example, if the principal direction distribution of the point cloud within a voxel is concentrated in a certain direction, and the variance of the principal component in that direction is below a threshold (e.g., below 0.1), it indicates that the point cloud morphology within that voxel exhibits a columnar structure. If the above geometric morphological features match the rules of a certain component region, the system assigns corresponding candidate labels to the points contained in that voxel. The system outputs a set of point clouds with component candidate labels, where each point is associated with a semantic label. Label types include, but are not limited to, "tower column candidate," "crossarm candidate," "hanging point candidate," and "other."

[0111] Clustering of similar point regions based on point cloud label data yields initial component data.

[0112] In one embodiment, the system divides point cloud sets with the same label into several subsets. Taking the label "crossarm candidate" as an example, the system groups all point cloud points marked as "crossarm candidate" into a subset, denoted as the candidate set for that category. Each subset is entered into the clustering process as an independent input. During the clustering process, the system uses a morphologically constrained region growing algorithm. A set of seed points is selected from the current subset as the starting point for clustering, prioritizing random selection from regions with higher point density. Based on a threshold or after random sorting, the top few clusters with higher density are selected for random selection. The spatial proximity growth radius is set to 0.3 meters, meaning that if the distance between a point and the current cluster boundary is less than this radius, it is considered to have growth potential. The clustering condition also requires directional consistency. The system performs principal component analysis on the local orientation of the point cloud. If the angle between the point to be clustered and the principal orientation of the current clustering region is less than 15 degrees, it can be determined that the orientation is consistent, thus satisfying the clustering condition. For each successful cluster growth, the system continuously adds points that meet the conditions until the growth conditions are no longer met or a specified minimum number of cluster points is reached. In this embodiment, the minimum number of points is set to 80. After completing the initial clustering, the system performs geometric verification on each clustering result. This is mainly done by fitting standard geometric shapes (such as cylinders, planes, cuboids, etc.) to evaluate the structural rationality of the clustered regions. If the residual error generated by the clustered region during the fitting process is less than 0.1 meters, the clustered component is considered to have good geometric consistency and is retained; otherwise, it is considered an abnormal structure and is removed. Through the combined strategy of the above-mentioned region growth and geometric verification, the system forms a set of initial component units, each unit representing a local structural component with a consistent label and shape. The set of all initial component units is denoted as the initial component set.

[0113] Optionally, the improved semantic network recognition is performed using a pole and tower component recognition model, the construction steps of which include:

[0114] Obtain historical initial component data;

[0115] In one embodiment, the 3D point cloud samples are tower structure data processed with high-precision modeling and manual component annotation. Each sample contains its spatial point cloud information and component semantic labels. The system extracts and structures the initial component regions in each sample. The extracted content includes spatial geometric data, such as the center point coordinates, geometric dimensions (e.g., length, width, and height), and fitted shape type (e.g., cylinder, cuboid, or planar solid) of each component; and category label information, i.e., the structural type to which the component belongs, including but not limited to "tower column," "crossarm," "hanging point," and "attached component." The system organizes the above information into component sample pairs, with each component consisting of its spatial geometric description and semantic label. The final training dataset can be represented as a set of pairings between components and labels, with each sample corresponding to a "component-category" mapping relationship.

[0116] Graph neural network processing is applied to the historical initial component data to obtain component graph feature data;

[0117] In one embodiment, the system first constructs a graph structure based on historical component data, denoted as a graph model. This graph structure consists of a set of nodes and a set of edges, where nodes represent individual component objects; edges represent spatial connections between components, with the connection condition being that the spatial distance between components is less than a preset distance threshold, which is set to 1.5 meters in this embodiment. After the graph structure is constructed, the system extracts attribute features for each component node as input features. Node features mainly include the following: the component's geometric dimensions, including its length, width, and height; the residual value of the component's shape fitting; the component's principal direction vector; and the component's point cloud density features. The system uses a graph neural network to encode the graph structure. In this embodiment, a graph convolutional network or a graph attention network can be used as the modeling method. Through iterative computation of a multi-layer graph neural network, the system aggregates the semantic and structural information of adjacent components in each round of propagation, thereby updating the semantic expression of each node. The system generates a fixed-dimensional graph embedding feature vector for each component node in the graph, serving as a compressed expression of the spatial relationships and semantic information contained in the component within the graph structure.

[0118] Local connectivity dependency feature extraction and global structural attention feature extraction are performed on the component graph feature data to obtain local connectivity dependency feature data and global structural attention feature data, respectively.

[0119] In one embodiment, for each node in the graph structure, a set N of its first-order neighbors is constructed.i Only edges with a connection confidence score higher than a threshold (e.g., 0.6) are retained; the geometric state and connection angle features of neighboring nodes are aggregated; and the weighted aggregation features of nodes are calculated through a state-guided mechanism. in For local feature representation (local connectivity depends on feature data), α ij The attention weights are calculated from the geometric direction angle and connection type, where j is the node index and N is the number of nodes. i For the first-order neighbor set, MLP(x) j The output is from a multilayer perceptron. Global structural attention feature extraction involves constructing a fully connected graph (establishing edges between all components); the global context representation is computed using the Transformer self-attention mechanism. Q i =W q h i K j =W k h j V j =W v h j This yields a global structural dependency representation for each node. This represents the global feature, where j is the node index and N is the total number of components. For attention weights (softmax output), For query vector transpose, K j V is the key vector. j W is a value vector. q To query the matrix weights, h i W is a node embedding representation. k h represents the weights of the key matrix. j W is an embedded representation for another node. v The values ​​are the weights of the matrix.

[0120] Feature fusion is performed on local connectivity-dependent feature data and global structural attention feature data to obtain feature fused data;

[0121] In one embodiment, features are fused for each node: in For feature fusion data, ReLU is the activation function, and W1 is the fusion weight matrix for locally connected features (pre-set based on prior experience). W1 represents the local connectivity dependency features, and W2 represents the fusion weight matrix of the global structural attention features (pre-set based on prior experience). 'b' represents the global structural attention feature, and 'b' represents the bias term; thus, a multi-scale structural semantic fusion representation at the component level is obtained.

[0122] Elevation datum surface is extracted from the historical initial component data to obtain elevation datum surface data;

[0123] In one embodiment, for each tower structure in the point cloud sample, its bottom region is identified. In this embodiment, the bottom region can be defined as the lowest 5% of the total tower height, representing the part of the structure closest to the ground. Within this bottom region, the system calculates the mean Z-axis coordinate of all points and sets this average Z-value as the elevation reference value for the sample, denoted as the elevation reference surface. The system performs plane fitting on the point cloud of the bottom region (specifically, based on the input data, the model parameters are set according to preset parameters; a subset is randomly sampled from the input data; a candidate plane model is calculated on the sampled subset according to the model parameters to obtain a candidate plane model; interior point detection is performed on the candidate plane model based on the unsampled input data and supplemented into the model to obtain a supplementary plane model; this process is repeated iteratively to obtain the plane containing the most interior points as the fitting reference surface for the region), thereby fitting a more stable set of local reference planes and eliminating Z-axis drift problems caused by tilt, scanning errors, or changes in device posture. After the reference surface is extracted, the system performs Z-axis coordinate normalization processing on all components in the entire sample based on this elevation reference value. The specific operation involves subtracting the corresponding reference Z value from the Z-axis coordinate of all points in each component, thereby aligning the elevation standards of components in different samples.

[0124] Relative elevation features are extracted from elevation datum data to obtain elevation feature data;

[0125] In one embodiment, the difference between the Z-coordinate of the center point of each component and the reference plane is calculated: z relative =z center -Z base , where z relative z is the relative elevation value. center Z is the Z-coordinate of the center point of the component. base The relative height is divided into discrete elevation levels (such as tower base section, tower column section, and tower head section) or retained as a continuous variable, based on the elevation datum value; this value is added as an additional feature to the input of each node.

[0126] Elevation-guided feature fusion is performed on elevation feature data and feature fusion data to obtain elevation feature fusion data.

[0127] In one embodiment, the system designs and introduces an elevation modulation fusion module as a substructure for feature processing. This module receives two inputs: the relative elevation features of the component nodes, i.e., the relative Z-axis height calculated in the aforementioned steps; and the existing fusion feature vectors of the component nodes, which contain structural features aggregated from multiple sources such as local connectivity and global context. The fusion module maps the relative elevation information and the original structural features to the same feature space by introducing linear transformation operations on these two feature branches respectively. Specifically, a set of learnable linear weights is applied to the relative elevation features, and another set of linear transformations is applied to the node fusion features. The weighted results of the two are then added and input into a nonlinear activation function (such as ReLU) to obtain the elevation enhancement features. Through the above elevation modulation mechanism, the system can adjust and distinguish the semantic representation of the same type of component at different elevation segments, so that even if they have similar geometric shapes or local structures, the model can still assign different representation weights to them according to their elevation positions, thereby significantly improving the modeling ability of hierarchical relationships and component distribution patterns in tower structures.

[0128] The tower structure hierarchical feature data is obtained by performing tower structure hierarchical feature characterization on the elevation feature fusion data.

[0129] In one embodiment, the system combines relative elevation information and identified component types to analyze the spatial dependencies of each component within the tower structure. For example, the tower column typically serves as the main structural member, with crossarms attached to the upper part of the tower column, while auxiliary components such as insulator attachment points are attached to the sides of the crossarms. The system establishes a component hierarchical tree structure, with the main tower column as the root node. The remaining components are connected level by level according to their spatial relative positions and structural dependencies, forming an ordered tree-like hierarchical relationship. The system generates path encoding information for each component node within this hierarchical tree, representing the structural path it traverses from the root node to the current node. For example, for a crossarm component, its path encoding can be represented as "root node → tower column → crossarm," reflecting the component's relative position and dependent path within the entire tower structure. The system vectorizes the path encoding, transforming it into a learnable structural position vector through an embedding mechanism. This position embedding represents the semantic position of the component within the overall hierarchical structure, supplementing the missing topological hierarchical information in the original geometric and elevation features. The system generates a set of hierarchical structure-aware embedding vectors for each component node, which serve as an expression of the hierarchical features of the tower structure.

[0130] Based on the structural hierarchical feature data of the tower, the semantic relationship regularization term of the components is processed to obtain the training data, and the constraint optimization is performed based on the training data to obtain the tower component recognition model.

[0131] In one embodiment, the system constructs a regularization loss term to constrain the similarity of semantic labels among component nodes with spatial proximity and structural hierarchical similarity. The system traverses all connecting edges in the component graph. For any pair of connected nodes in the graph, if their predicted semantic labels are the same, higher similarity in their hierarchical embedding features is encouraged. This similarity can be calculated using a similarity metric function between hierarchical embedding vectors (such as dot product, cosine similarity, etc.), thus constituting the structural regularization term. If two components are adjacent in actual space and also in close positions in the structural hierarchy, they are more likely to belong to the same or similar types of components, and therefore should be reinforced during model training. In overall training, the system uses a joint loss function comprising two parts for network optimization: the first part is a semantic classification loss, employing the standard cross-entropy loss function to measure the difference between the predicted semantics and the true labels of component nodes; the second part is the structural regularization loss constructed above, used to reinforce spatial and hierarchical consistency. L total The total loss of the model, The cross-entropy loss is used for semantic classification, and λ is the weight coefficient of the regularization term, set between 0.1 and 0.3. The loss term is the structure regularization term. These two loss terms are combined by a weighted sum to form the total loss function. The weight coefficient of the regularization term can be set as an adjustment parameter between 0.1 and 0.3. During model training, the system iteratively updates the network parameters (specifically, initializing the model parameters and their first and second moment estimates, as well as the global step size (learning rate); calculating the model output and loss value through forward propagation; calculating the gradient of the loss with respect to each parameter through backpropagation; calculating the update of the first moment estimate (momentum term) and the update of the second moment estimate (gradient squared); performing bias correction to update the parameters; repeating the iteration) to obtain an optimized tower component recognition model.

[0132] Optionally, the local connectivity dependency feature extraction includes:

[0133] Spatial hierarchical encoding is performed on the component diagram feature data to obtain the component diagram structural data;

[0134] In one embodiment, the input component graph feature dataset Each node corresponds to a component; calculate the relative elevation value of each component node: Based on the empirical division rules of tower structure (e.g., tower base area 0–8m, main tower body 8–25m, tower head 25m+), the component nodes are assigned discrete hierarchical codes L. i ∈{0,1,2}; embed the hierarchical encoding and the original component graph into h i By concatenating the data, a hierarchical perceptual feature vector is formed. Output component diagram structural data Where A is the initial adjacency matrix in the graph.

[0135] The adjacent node data is obtained by filtering the structural data of the component diagram using a connection mask.

[0136] In one embodiment, for edge A in the adjacency matrix A ij The adjacency mask M is obtained by performing masking according to the following two types of rules. ij ∈{0,1}: If the Euclidean distance d between components ij A connection greater than 2.5m is considered to have no actual connection; set M. ij =0; Calculate the angle θ between the normal directions of the components. ij If θ ij >45° indicates inconsistent directions; set M. ij =0; retain the node pairs (i,j) that satisfy the above conditions to form a first-order effective neighbor set. Output the set of first-order adjacent nodes corresponding to each component node.

[0137] By applying component status guidance to adjacent node data, node aggregation data is obtained;

[0138] In one embodiment, for each pair of nodes (i,j), the current feature vector of node i is concatenated with the feature vector of its neighboring node j, and input into the reset gate weight matrix. After activation by the Sigmoid function, the control gate r is output. ij This threshold controls the degree to which the target node selectively memorizes the features of its neighboring nodes. The features of node i are used to perform residual interaction with the features of its neighboring node j; specifically, the features of node i are compared with those of node r. ij After gating (i.e., product operation), the features of node j are fed into the nonlinear transformation module, passed through a two-layer feedforward network, and then a new adjacency feature representation is output using the hyperbolic tangent activation function, denoted as the guiding feature. The system calculates the weight coefficient of each adjacent component in the current node aggregation based on the spatial geometric information and directional relationship between two nodes. This includes two factors: the Euclidean distance between the nodes (closer distance, higher weight) and the cosine similarity between the principal axis direction vectors of the two nodes (higher directional consistency, higher weight). The weighted combination of these two factors, after normalization, serves as the attention weight α. ij Guide the characteristics of all adjacent nodes. According to the corresponding attention weight α ij The weighted summation, used as the state of the target node i, guides the aggregation result. This aggregation feature not only encodes the node's own structural state but also reflects its semantic association and structural dependency with surrounding components in the local structure.

[0139] Based on the node aggregation data, semantic guidance is performed on the component connection edges to obtain the edge semantic connection data;

[0140] In one embodiment, the system traverses all edge pairs (i,j) with connections in the component graph, where each edge represents a candidate spatial connection between two component nodes. For each edge, the state-guided aggregation features of its two endpoints are extracted and denoted as the aggregation features of node i. Aggregation features with node j The aggregated features are fed into the edge semantic classification module as input. Specifically, the system uses a lightweight feedforward neural network (e.g., a multilayer perceptron, MLP) to perform a nonlinear transformation on the feature pair and normalizes the output using a Softmax function to predict the semantic type probability distribution of the edge. The input is a concatenated representation of the node pair features, and the output is the normalized confidence probability value of the edge under various semantic relationships. In this invention, the semantic type of an edge can be defined as a finite set, denoted as ε = {support, hook, adjacency, unconnected}. Support connection indicates a bearing or vertical force transmission relationship between components (e.g., a tower column supporting a crossbeam); hook connection indicates a component attached to the surface of the main component structure (e.g., a hook point connecting a crossbeam); adjacency connection indicates two components arranged parallel in space but without physical connection; unconnected indicates that the current edge pair is not a valid structural relationship and should be discarded. The system uses a threshold to determine the maximum confidence value in the prediction results of each edge. If the maximum confidence is less than the set threshold (e.g., 0.6), the edge is considered to be in an "unconnected" state, marked as an invalid connection edge, and removed in subsequent processing. Conversely, if the maximum confidence is higher than the threshold, the edge is retained as a valid structural connection and classified into a specific semantic connection type.

[0141] Component relation residual enhancement is performed based on edge semantic connection data to obtain local connection dependency feature data.

[0142] In one embodiment, for each node i in the component graph, the system extracts the set of its adjacent nodes that are marked as having a valid connection relationship in the edge semantic classification, denoted as . This set only includes edges with explicit structural semantic meaning to node i, such as "supporting connection" or "hanging connection," while filtering out edges that do not have a physical coupling relationship, such as "adjacent" or "not connected." The system performs residual augmentation calculations on this node. Specifically, for node i, based on its original graph embedding feature vector, a residual representation of its semantically connected adjacent nodes is introduced. Each connected node... The enhancement contribution is driven by its state-guided features. With the target node feature h i The difference between them is used to construct the value, which is then multiplied by the confidence weight β of the corresponding edge. ij , where β ij∈[0,1] represents the confidence strength of edge (i,j) in the edge semantic classification as a valid connection. After weighted aggregation of all such residual information, it is added to the original node features to form structural semantic enhancement features, which means that the local connection dependency feature vector of the target node is equal to the sum of its original features and the residual enhancement terms of all semantically connected neighboring nodes.

[0143] Optionally, S3 includes:

[0144] S31. Generate component node diagrams from the tower component data to obtain component diagram data;

[0145] In one embodiment, the system constructs a graph structure model based on a tower component dataset. This dataset consists of multiple segmented tower components, each containing its three-dimensional spatial center coordinates (including horizontal, vertical, and height coordinates), geometric dimensional parameters (such as length, width, and height), and semantic component type labels, such as tower column, crossarm, or hanging point. During graph structure construction, the system defines each component as a node in the graph and assigns each node a set of attribute features, including its spatial location, structural dimensions, and semantic type. The component type is expressed using one-hot encoding. Initially, the graph structure contains only nodes and no arbitrary connecting edges, i.e., it is constructed as a borderless graph model with an empty edge set. The system simultaneously pre-calculates candidate spatial adjacency relationships between components, including the three-dimensional spatial distance between components and the degree of projection overlap in the vertical and horizontal directions.

[0146] S32. Generate candidate connection relationships for component pairs from the component diagram data to obtain component connection diagram data;

[0147] In one embodiment, for any pair of component nodes (v i ,v j (If i ≠ j) perform candidate generation operation; if the Euclidean distance d between components ij If the distance is less than 3m, it is listed as a candidate connection edge. Calculate the bounding box overlap rate of the two components. If there is overlap in the XY or YZ directions (overlap area ratio > 0.2), retain it; or calculate the directional similarity cos(θ) based on the angle between the axis alignment directions. ij If the value is greater than 0.85, proceed to the next step if any of the above conditions are met. The component combination must be reasonable, for example... (allow); Hanging point (permitted); (Not allowed); Use of predefined component combination rule matrix R type (t i ,t j )∈{0,1} to filter illegal combinations. Calculate the combination confidence score for each legal candidate edge: Where s ij The combined confidence score is given by α1, which is the distance weight term with a value of 0.4, and d ij Let be the distance between the i-th component and the j-th component, and α2 be the bounding box overlap ratio weight term with a value of 0.4. Let α be the bounding box overlap rate, and α3 be the directional similarity weight term with a value of 0.4, cos(θ) ij ) represents directional similarity, α4 is the weight term of the component combination rule matrix, with a value of 0.4, R type (t i ,t j ) is the component combination rule matrix, t i Let t be the component type corresponding to the i-th component. j For the component type corresponding to the j-th component; filter s ij >τ (e.g., 0.6) is used as a candidate edge. The output component connection graph data G′=(V,E′), where E′ is the set of edges with connection potential, and each edge contains two component IDs and their connection feature e. ij =(f i ,f j ,s ij ).

[0148] S33. Classify and label the topological relationships based on the component connection diagram data to obtain the tower structure topology diagram data.

[0149] In one embodiment, the system performs structural semantic classification and annotation of the connection relationships between components based on the component connection diagram, and constructs a tower structure topology diagram. The system extracts multi-dimensional structural features from each candidate connection edge in the diagram, specifically including spatial relative relationships such as Euclidean distance, vertical height difference, and horizontal projection distance between components; morphological matching relationships of components in size and direction; semantic type labels of the components at both ends; and the relative positions of components in the vertical direction, used to determine whether a support dependency relationship exists. The system uses a lightweight multilayer perceptron network model or a rule-based logic engine to classify the structural relationships of component edges. The semantic categories of edges include "support connection," such as the main load-bearing structural relationship between a tower column and a crossarm; "auxiliary connection," such as the suspension relationship between a crossarm and an insulator hanging point; "parallel adjacency," such as the relationship between multiple lateral components of a tower column at the same level; and "non-connection," indicating a redundant connection between two components with no actual structural association. The system uses type-prior connection probabilities as classification auxiliary constraints; for example, high-frequency structural connections such as "tower column-crossarm" will receive higher confidence weights. After classification, the system attaches semantic tags to the edge attributes of the original connectivity graph and removes all edges judged as "non-connected". The output tower structure topology graph consists of component nodes and valid connecting edges, with each edge containing a semantic identifier of the connection type.

[0150] Optionally, the generation of candidate connection relationships by the component includes:

[0151] Enumerate the component node pairs in the component diagram data to obtain the component node pair data;

[0152] In one embodiment, component drawing data is input. Each node v i Let V represent a component (the entire set is V, i.e., the set of nodes); perform all permutations and combinations of node pairs (excluding pairs of themselves): For each pair of nodes (v i ,v j Construct a node pair data structure to record the component ID; the coordinates of the two spatial center points of the node pair (x, y, z). i ,y i ,z i ),(x j ,y j ,z j ); geometric dimensions, type, and other attributes. Output a set of component node pairs.

[0153] Spatial proximity filtering is performed on the component node pair data to obtain filtered node pair data;

[0154] In one embodiment, for each component pair Calculate distance d ij =||c i -c j ||2;where d ij For the distance between components, c i Let c be the center point of the i-th component. j Let c be the center point of the j-th component. i and c j The center point of the two components is set; a distance threshold d is set. th =3.5m, if d ij >d th If the pair is not found, then discard it; determine whether the projections of the component's bounding box overlap in a certain direction (condition: XY plane IoU > 0.1); output the set of node pairs after spatial proximity filtering.

[0155] Based on the node pairs, the filtered data is matched by component type to obtain the node pair combination data;

[0156] In one embodiment, a component type combination rule matrix R is introduced. type (t i ,t j Define connectable component pairs, such as allow; allow; Not allowed; for each node pair If R type (t i ,t j If ) = 1, then retain it; output the set of node pairs with valid types.

[0157] Geometric angle matching calculation and type combination prior weight calculation are performed on the node-pair combination data to obtain geometric angle matching data and type combination prior data, respectively.

[0158] In one embodiment, the principal axis direction vector of the component is extracted. (Obtained through principal component analysis (PCA)); Calculate the included angle. Where θ ij Let be the included angle, and arccos be the inverse cosine function. Let be the principal axis direction vector of the i-th component. Let the principal axis direction vector of the j-th component be normalized to a matching score: in For the matching score (geometric angle matching), exp is an exponential function, and θ ij Let σ be the included angle, and σ be the matching degree control parameter. σ can be set to 15°, corresponding to a higher matching degree as the angle difference decreases. Prior weights for type combinations are calculated for each type combination (t). i ,t j ), set prior weights based on historical data statistics or engineering experience. like 0.9; 0.8; 0.2; Output geometric angle matching data; type combination prior data.

[0159] Based on the geometric angle matching data and the prior data of type combination, the connection edges of the node-pair combination data are filtered to obtain the connection edge data;

[0160] In one embodiment, for each component pair Calculate connectivity features: in For connectivity features, α is the weighting coefficient. Scoring is based on geometric angle matching. Let the prior weights for type combinations be α = 0.6; let the screening threshold be τ = 0.7, and retain... The connection pairs; each retained node pair is defined as a connection edge, along with its matching score and connection type candidates; the output connection edge dataset E = {e ij}, where each edge contains component pair indices and scores.

[0161] The component connection diagram data is generated by generating a component connection diagram based on the connection edge data and component nodes.

[0162] In one embodiment, the graph structure is initialized. Where V represents a component node; These are the connection edges generated after filtering; each edge... Add the following attributes: spatial vector (relative position of components); type combination (e.g., "tower column - crossarm"); matching score; output component connection diagram data structure.

[0163] Optionally, S4 includes:

[0164] S41. Generate component geometry based on the tower structure topology data to obtain component geometry data;

[0165] In one embodiment, the input data is a topology diagram of a tower structure, denoted as the structural diagram model. This topology diagram consists of a set of nodes and a set of edges. Each node represents a component unit, containing attributes such as component type, spatial center coordinates, and dimensional characteristics; edges represent the connection relationships between components. The system establishes template mapping rules for standard geometries based on component types, including but not limited to the following correspondences: tower columns are mapped to standard cylinders; crossarms are mapped to cuboids or slender beams; and hanging points or support components are mapped to spheres or small-sized connectors. For each node, the system instantiates the corresponding geometric template based on its included dimensional parameters (such as length, width, and height) to generate a standard component body. If node dimensional information is missing, the system can complete it in two ways: first, by filling in the missing information based on the historical average dimensional values ​​of similar components; second, by restoring the dimensional characteristics of the component through point cloud data fitting. To ensure the correct orientation of the component geometry in space, the system performs direction alignment processing. This process calculates the principal axis vector of the component based on the direction of its connecting edges in the topology diagram. The system aligns the default orientation of each geometry (i.e., the principal axis is the Z-axis) with the target spatial orientation, and transforms the geometry's orientation from the initial orientation to the target orientation using a rotation matrix to represent the component's orientation and connection relationships. The system outputs a component geometry dataset, where each component geometry contains three-dimensional attributes such as its spatial location (center coordinates), structural dimensions (length, width, height, or radius), and orientation information (principal axis vector).

[0166] S42. Perform structural connection-driven assembly based on component geometry data to obtain component assembly data;

[0167] In one embodiment, the attachment method and assembly process between components are determined based on the component connection relationships described by the edge set in the topology graph. For example, edge relationships can be represented as typical structural connection forms such as "tower column supporting crossarm" and "crossarm hanging point". Based on this, the system defines corresponding connection interface types for each connection relationship, including but not limited to face-to-face connection (the connection surfaces of the components fit together); point-to-face connection (the hanging point is attached to the crossarm surface through the connection point); edge-to-face connection (the edge of the bracket is connected to the structural surface), etc. The system determines the assembly sequence according to the structural hierarchy of the components, following a bottom-up assembly process, i.e., tower base, tower column, crossarm, and finally hanging point or auxiliary components. During the spatial alignment process, for each pair of components that need to be connected, the system obtains the position coordinates and normal vector information of its predefined connection surface. The spatial position and orientation of the sub-component are adjusted so that its connection surface coincides with the corresponding connection surface of the parent component in position and is aligned in orientation. The system sets a small tolerance distance (e.g., 0.01 meters) between the connection surfaces. During the assembly process, the system also performs collision detection and component rearrangement operations. If voxel overlap is detected between components during the docking process, the system determines geometric conflicts using methods such as axis-aligned bounding box (AABB) detection or Boolean intersection calculation. If an actual collision exists, the system fine-tunes the component positions, introducing rigid constraint optimization methods when necessary. While maintaining the original structural relationships of the connecting components, the system fine-tunes their positions and orientations to avoid structural conflicts and maintain overall topological consistency. The system outputs a set of assembled components, each correctly assembled according to the structural connection relationships defined in the topology diagram; the output result is the component assembly dataset.

[0168] S43. Fit component parameters to the component assembly data to obtain component fitting data;

[0169] In one embodiment, the system performs point cloud residual verification on the component assembly results. This process compares the geometry of each assembled component with the corresponding region in the original point cloud dataset. Specifically, for each component, a subset of local points cloud corresponding to its spatial location is extracted from the original tower point cloud data. The residual between the component surface and its corresponding point cloud is calculated, using the shortest distance from the point to the geometric surface as the residual evaluation index. The system can also evaluate the degree of overlap between the component's outer enclosure and the point cloud region, for example, using the intersection-over-union ratio (IoU) of the 3D bounding box as a geometric fit index. If the mean residual of the component fitting result exceeds a preset threshold (e.g., 0.1 meters), or the IoU is less than a set standard (e.g., less than 0.75), the component fitting quality is considered insufficient. During the fitting optimization process, the system uses a parameter minimization strategy to adjust the component's geometric parameters. Optional optimization methods include using the least squares method to perform global fitting optimization of the component parameters, with the goal of minimizing the sum of squared Euclidean distances between the component surface and the point cloud. That is, by adjusting the parameters of the component (such as position, orientation, and size), the distance residual between all sampling points and the component surface is minimized. Where min represents the minimization operation, θ i Let P be the parameter of the i-th component, and p be a point in the point cloud. i For the component, a subset of the point cloud. The system generates surface equations for the component geometry; alternatively, iterative optimization using gradient descent gradually updates parameters to reduce fitting error. Simultaneously, constraints are applied to connection points between components to maintain their structural connections and prevent disruption of the overall assembly topology. The system outputs a dataset of optimized component parameters. This fitted data contains updated 3D geometric parameters for each component, including its position coordinates, dimensions, and orientation.

[0170] S44. Construct a semantic 3D model based on the component fitting data to obtain the transmission tower model.

[0171] In one embodiment, the system binds semantic attributes to each component. This binding includes adding basic type tags, such as labeling components as predefined types like "tower column," "crossarm," or "hanging point"; adding structural hierarchy information, such as indicating whether a component is a main component or an auxiliary component, or adding connection dependency tags to describe the assembly sequence or structural affiliation between components. To support subsequent model management and parsing, the system establishes parent-child reference relationships between components, thereby supporting hierarchical management and semantic retrieval functions. During the model generation phase, the system integrates each component into a unified 3D semantic model based on its 3D geometric attributes, spatial location, and semantic tags. To ensure the model's compatibility and scalability across multiple platforms and application systems, the system supports exporting the model to various structured semantic model formats, including .IFC (for Building Information Modeling (BIM) systems); .GLTF / .OBJ+.JSON (for enhanced visual semantics); and a custom XML format (embedding component parameters, connection topology, semantic tags, etc.). Components are positioned using a unified world coordinate system.

[0172] Optionally, this application also provides an intelligent modeling system for transmission towers based on 3D point cloud data, used to execute the intelligent modeling method for transmission towers based on 3D point cloud data as described above, wherein the intelligent modeling system for transmission towers based on 3D point cloud data includes:

[0173] The point cloud region segmentation module is used to acquire region scanning data, identify tower regions based on the region scanning data to obtain tower region data, and extract towers from the tower region data to obtain tower segmentation data.

[0174] The component semantic recognition module is used to perform improved component semantic recognition on the pole segmentation data to obtain pole component data;

[0175] The topology parsing module is used to process the structural topology relationships of tower component data to obtain tower structure topology data.

[0176] The 3D modeling and structural reconstruction module is used to construct transmission tower models based on tower structure topology data, thus obtaining transmission tower models.

[0177] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0178] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for intelligent modeling of power transmission towers based on 3D point cloud data, characterized in that, The method includes: S1. Obtain area scanning data, identify tower areas based on area scanning data to obtain tower area data; extract towers from tower area data to obtain tower segmentation data. S2. Perform improved component semantic recognition on the tower segmentation data to obtain tower component data; S3. Perform structural topology processing on the tower component data to obtain tower structure topology data; S4. Construct the transmission tower model based on the tower structure topology data to obtain the transmission tower model.

2. The method according to claim 1, characterized in that, S1 includes: Acquire area scan data; Preliminary filtered data is obtained by filtering based on height thresholds using regional scan data. The vertical structure density of the initial screening data is detected to obtain secondary screening data. Deep learning segmentation is performed on the secondary filtered data to obtain tower area data; Target geometric data is obtained by fitting the target geometry based on the tower area data; Structural clustering is performed on the target geometric data to obtain tower segmentation data.

3. The method according to claim 1, characterized in that, S2 include: Using pre-defined tower structure spatial data, spatial constraint clustering is performed on the tower segmentation data to obtain initial component data; An improved semantic network recognition method is used to identify the initial component data to obtain the tower component data.

4. The method according to claim 3, characterized in that, The spatially constrained clustering includes: The tower segmentation data is divided into point cloud spatial blocks to obtain point cloud spatial segmentation data; Using the preset tower structure spatial data, the point cloud spatial division data is matched with component spatial constraint rules to obtain point cloud label data; Clustering of similar point regions based on point cloud label data yields initial component data.

5. The method according to claim 3, characterized in that, The improved semantic network recognition is performed using a pole and tower component recognition model. The construction steps of the pole and tower component recognition model include: Obtain historical initial component data; Graph neural network processing is applied to the historical initial component data to obtain component graph feature data; Local connectivity dependency feature extraction and global structural attention feature extraction are performed on the component graph feature data to obtain local connectivity dependency feature data and global structural attention feature data, respectively. Feature fusion is performed on local connectivity-dependent feature data and global structural attention feature data to obtain feature fused data; Elevation datum surface is extracted from the historical initial component data to obtain elevation datum surface data; Relative elevation features are extracted from elevation datum data to obtain elevation feature data; Elevation-guided feature fusion is performed on elevation feature data and feature fusion data to obtain elevation feature fusion data. The tower structure hierarchical feature data is obtained by performing tower structure hierarchical feature characterization on the elevation feature fusion data. Based on the structural hierarchical feature data of the tower, the semantic relationship regularization term of the components is processed to obtain the training data, and the constraint optimization is performed based on the training data to obtain the tower component recognition model.

6. The method according to claim 5, characterized in that, The local connectivity dependency feature extraction includes: Spatial hierarchical encoding is performed on the component diagram feature data to obtain the component diagram structural data; The adjacent node data is obtained by filtering the structural data of the component diagram using a connection mask. By applying component status guidance to adjacent node data, node aggregation data is obtained; Based on the node aggregation data, semantic guidance is performed on the component connection edges to obtain the edge semantic connection data; Component relation residual enhancement is performed based on edge semantic connection data to obtain local connection dependency feature data.

7. The method according to claim 1, characterized in that, S3 includes: Component node diagrams are generated from the tower component data to obtain component diagram data; The component diagram data is processed to generate candidate connection relationships between component pairs, thus obtaining the component connection diagram data. Based on the component connection diagram data, the topological relationships are classified and labeled to obtain the tower structure topology diagram data.

8. The method according to claim 6, characterized in that, The generation of candidate connection relationships by the component includes: Enumerate the component node pairs in the component diagram data to obtain the component node pair data; Spatial proximity filtering is performed on the component node pair data to obtain filtered node pair data; Based on the node pairs, the filtered data is matched by component type to obtain the node pair combination data; Geometric angle matching calculation and type combination prior weight calculation are performed on the node-pair combination data to obtain geometric angle matching data and type combination prior data, respectively. Based on the geometric angle matching data and the prior data of type combination, the connection edges of the node-pair combination data are filtered to obtain the connection edge data; The component connection diagram data is generated by generating a component connection diagram based on the connection edge data and component nodes.

9. The method according to claim 1, characterized in that, S4 includes: The component geometry is generated based on the tower structure topology data to obtain the component geometry data; Structural connection-driven assembly is performed based on component geometry data to obtain component assembly data; The component assembly data is fitted with component parameters to obtain component fitting data; Based on the component fitting data, a semantic 3D model is constructed to obtain the transmission tower model.

10. A smart modeling system for power transmission towers based on three-dimensional point cloud data, characterized in that, For executing the intelligent modeling method for transmission towers based on three-dimensional point cloud data as described in claim 1, the intelligent modeling system for transmission towers based on three-dimensional point cloud data includes: The point cloud region segmentation module is used to acquire region scanning data, identify tower regions based on the region scanning data to obtain tower region data, and extract towers from the tower region data to obtain tower segmentation data. The component semantic recognition module is used to perform improved component semantic recognition on the pole segmentation data to obtain pole component data; The topology parsing module is used to process the structural topology relationships of tower component data to obtain tower structure topology data. The 3D modeling and structural reconstruction module is used to construct transmission tower models based on tower structure topology data, thus obtaining transmission tower models.

Images