A Substation Cloud Data Processing Method for Constructing 3D Models

Abstract

The present application relates to the technical field of digital modeling of power facilities, and particularly relates to a substation point cloud data processing method for constructing a three-dimensional model. The method comprises the following steps: obtaining substation point cloud data; performing convolution filtering on the substation point cloud data to obtain convolution filtering data; performing structure change feature extraction according to the convolution filtering data and the substation point cloud data to obtain structure change feature data; performing component attention screening on the substation point cloud data according to the structure change feature data to obtain point cloud component pseudo-labeled data; performing component clustering on the point cloud component pseudo-labeled data to obtain point cloud component clustering data; performing pose calibration on the substation point cloud data according to the point cloud component clustering data to obtain pose calibration data; and performing topological structure modeling according to the pose calibration data to obtain a three-dimensional model of the substation. The present application can realize high-precision structure extraction of substation point cloud data, and improve the construction efficiency and topological expression accuracy of the three-dimensional model.

Description

Technical Field

[0001] This invention relates to the field of digital modeling technology for power facilities, and in particular to a cloud data processing method for substations used to construct three-dimensional models. Background Technology

[0002] Digital twin 3D model construction refers to the process of reconstructing the geometric structure, topological relationships, and semantic attributes of a physical entity (such as a substation) in a virtual environment using 3D modeling technology, based on spatial data collected from the physical entity. This creates a virtual mirror model that is highly consistent with the physical system. This model not only possesses the appearance and spatial layout corresponding to the real system but also carries multi-dimensional information such as equipment status and operation and maintenance logic. It enables perception mapping, dynamic synchronization, and simulation prediction of physical objects, forming the fundamental core of a digital twin system. In the power sector, it supports functions such as visual management, fault simulation, and remote inspection, making it one of the key supporting technologies in smart grid construction. Currently, with the continuous improvement of the digitalization level of power infrastructure, substations, as critical nodes in the power system, urgently need to achieve high-precision, high-consistency, and high-scalability digital twin 3D model construction. Traditional modeling methods mainly rely on manual annotation and CAD drawing replication, which is not only inefficient and costly but also lacks effective correction and abstraction capabilities when facing real-world errors (such as equipment offset, component occlusion, and posture distortion). Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a substation cloud data processing method for constructing three-dimensional models, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a substation cloud data processing method for constructing a three-dimensional model, the method comprising:

[0005] S1. Obtain substation cloud data; perform convolution filtering on the substation cloud data to obtain convolution-filtered data;

[0006] S2. Extract structural change features from convolutional filtered data and substation cloud data to obtain structural change feature data; perform component attention filtering on substation cloud data based on structural change feature data to obtain point cloud component pseudo-label data; perform component clustering on point cloud component pseudo-label data to obtain point cloud component clustering data.

[0007] S3. Perform attitude calibration on the substation point cloud data based on the point cloud component clustering data to obtain attitude calibration data;

[0008] S4. Based on the attitude calibration data, perform topology modeling to obtain a three-dimensional model of the substation.

[0009] This invention effectively enhances the perception of key geometric structures and boundary changes in substation point cloud data by combining convolutional filtering with structural change feature extraction, thereby improving the accuracy of component selection. Through component attention filtering and pseudo-annotation mechanisms, high-confidence equipment component identification can be achieved without manual annotation, reducing labor costs. Combining component clustering and attitude calibration steps spatially regularizes the equipment structure in the original point cloud, providing a unified geometric basis for topology modeling. By integrating spatial location, functional connectivity, and semantic rules into a topology modeling method, accurate and reliable 3D models of substations are generated.

[0010] Optionally, the convolutional filtering includes:

[0011] The normal vector covariance tensor diagram of the substation cloud data is calculated to obtain tensor diagram data;

[0012] Perform spectral domain point cloud response transformation on the tensor graph data to obtain point cloud spectral data;

[0013] Multi-kernel convolution processing is performed on point cloud spectral data to obtain convolutionally filtered data. The multi-kernel convolution processing includes spherical kernel convolution processing, planar kernel convolution processing, and cylindrical kernel convolution processing.

[0014] This invention, by introducing a normal vector covariance tensor diagram for calculation, can fully explore the changing trends and normal distribution differences of local geometric structures in point clouds, achieving sensitive capture of boundary features, corner structures, and heterogeneous surfaces. Through spectral domain point cloud response transformation, point cloud data can be mapped from the spatial domain to the frequency domain, enhancing the response capability to high-frequency structures (such as sharp edges and small protrusions). Multi-kernel convolution processing, combining the directionality and morphological adaptability of spherical, planar, and cylindrical kernel convolutions, is adapted to extract structural features of point-like, planar, and linear components, respectively, thereby achieving highly robust geometric perception of different types of power equipment components.

[0015] Optionally, the structural change feature extraction includes:

[0016] Structural change data is obtained by extracting structural changes from convolutional filtered data and substation cloud data.

[0017] The structural jump index is calculated based on the structural change data to obtain the structural jump index data.

[0018] Structural fluctuation projection analysis was performed on substation site cloud data based on structural change data and structural jump index data to obtain structural change characteristic data.

[0019] This invention, by fusing convolutionally filtered data with raw point cloud data, can fully extract local geometric perturbation information from the surface of equipment components, identifying areas that may have abrupt edge changes, geometric breaks, or irregular structures. The introduction of a structural jump index allows the system to quantitatively measure the continuity of connectivity and the degree of morphological abrupt changes in the point cloud, thereby effectively detecting equipment seams, transitions, and incomplete components. Combined with structural wave projection analysis, it can directionally enhance structural change trends across multiple scales, making the structural change characteristics not only sensitive to local responses but also directional selectivity and scale adaptability.

[0020] Optionally, the component attention filtering includes:

[0021] Point-level attention data is obtained by performing point-level attention calculation based on structural change feature data.

[0022] A priori masking enhancement process is performed on the substation point cloud data based on the point-level attention data to obtain the point cloud enhanced data.

[0023] Local aggregation and enhancement are performed on the point cloud augmentation data to obtain locally aggregated data;

[0024] Pseudo-labels are generated based on local aggregated data to obtain pseudo-label data for point cloud components.

[0025] This invention introduces a point-level attention mechanism driven by structural change features to achieve response weighting for key structural regions in substation cloud data. This allows the system to automatically focus on areas with abrupt edge changes, morphological anomalies, or sensitive connections, thereby improving the robustness of structural identification. Prior masking enhancement based on point-level attention data significantly enhances the feature saliency of potential component regions and effectively suppresses interference from background noise and unstructured areas. Local aggregation enhancement spatially improves structural coherence and regional consistency, ensuring that the generated pseudo-labels possess spatial integrity and semantic reliability. This invention achieves accurate pseudo-label generation through self-supervised attention-guided processing without requiring manual annotation, providing high-confidence input data and exhibiting good engineering scalability and structural awareness intelligence.

[0026] Optionally, the component clustering includes:

[0027] Feature extraction is performed on the pseudo-labeled data of point cloud components to obtain pseudo-labeled feature data;

[0028] Density clustering is performed on the pseudo-labeled feature data to obtain feature clustering data;

[0029] Structural feature data is obtained by extracting structural features from the feature clustering data;

[0030] Based on the structural feature data, a device component structure nesting graph is constructed from the feature clustering data to obtain the number of point cloud component clusters.

[0031] This invention extracts multi-dimensional features from pseudo-annotated point cloud component data, integrating multi-source information such as geometric morphology, spatial density, and topological distribution to characterize the structural differences and similarities between components. A density clustering algorithm is used to aggregate the pseudo-annotated feature data, which is independent of the preset number of categories and possesses good adaptability, making it suitable for scenarios in substations with complex component types and tightly interwoven structures. Combining structural features extracted from the feature clustering results, such as connectivity, compactness, and orientation consistency, helps improve the accuracy of cluster boundaries and the completeness of component identification. The device component nesting graph constructed through structural features effectively represents the spatial subordination and hierarchical dependencies between components, providing structured input for topology modeling. This invention not only improves the structural rationality of the clustering results but also enhances the system's ability to understand the spatial organization of complex power components, exhibiting good engineering adaptability and model generalization.

[0032] Optionally, the pseudo-labeled feature data includes first pseudo-labeled feature data and second pseudo-labeled feature data, and feature extraction includes:

[0033] Geometric features are extracted from the pseudo-annotated data of point cloud components to obtain the first pseudo-annotated feature data;

[0034] Local connectivity features are extracted from the pseudo-annotated point cloud component data to obtain the second pseudo-annotated feature data.

[0035] This invention comprehensively enhances the structural representation capability of pseudo-annotated point cloud component data by jointly modeling geometric features and local connectivity features. The geometric feature extraction stage obtains fine-grained morphological information such as curvature distribution, normal direction, and edge variations of the components, facilitating the identification of equipment boundaries or protruding components with distinct structural outlines. Meanwhile, local connectivity feature extraction constructs a nearest-neighbor connection graph, analyzing the connectivity, structural consistency, and skeleton path tension between point clouds, effectively revealing the spatial continuity and potential fracture characteristics of components. By fusing the first and second pseudo-annotated feature data, not only is the discriminative power of the clustering model enhanced, but the sensitivity to identifying weakly connected and small components is also improved.

[0036] Optionally, the geometric feature extraction includes:

[0037] Local covariance is calculated on the pseudo-annotated data of point cloud components to obtain local covariance data;

[0038] Curvature spectrum data is obtained by extracting curvature spectrum data from local orthometric data;

[0039] Multi-scale point cloud structural fluctuation analysis was performed on the pseudo-annotated data of point cloud components based on curvature spectrum data to obtain structural fluctuation spectrum data.

[0040] The first pseudo-labeled feature data is obtained by performing a density projection along the principal axis based on the structural wave spectrum data.

[0041] This invention effectively captures the spatial distribution characteristics of point clouds within their local neighborhoods by introducing local covariance calculation, enhancing the ability to perceive differences in the microstructure of components. By extracting curvature spectrum data through eigenvalue analysis of the covariance matrix, the morphological change trend of the equipment surface can be quantitatively described, identifying key geometric features such as abrupt edge changes and curved surface corners. Combined with multi-scale structural fluctuation analysis, the stability and morphological perturbations of the point cloud structure can be detected at different scales, improving the ability to identify components of different sizes and complex geometric details. Through density projection along the principal axis, local structural features are mapped to the principal axis coordinate space, making linear extended structures (such as cables and grounding wires) and planar devices (such as terminals and covers) exhibit more obvious distribution differences in the feature space.

[0042] Optionally, the local connectivity feature extraction includes:

[0043] The pseudo-labeled point cloud component data is processed using a nearest neighbor graph to obtain nearest neighbor graph data.

[0044] Local subgraph connectivity analysis is performed based on the nearest neighbor connection graph data to obtain local subgraph connectivity data;

[0045] Pseudo-connectivity supplementation is performed based on the local subgraph connectivity data to obtain subgraph connectivity data;

[0046] Subgraph skeleton path extraction is performed based on subgraph connectivity data to obtain subgraph skeleton path data;

[0047] The connection tension data is obtained by calculating the connection tension based on the subgraph skeleton path data;

[0048] The pseudo-connectivity is calculated based on the connection tension data to obtain the second pseudo-label feature data.

[0049] This invention constructs a nearest-neighbor connectivity graph to model the spatial structure of pseudo-annotated point cloud component data, revealing potential connections between points and effectively capturing structural continuity features. Through local subgraph connectivity analysis, broken regions or multiple separated structures within pseudo-annotated components can be identified, providing a basis for connection completion. A pseudo-connectivity supplementation strategy can compensate for sparse connections while maintaining structural rationality, enhancing the overall connectivity of the pseudo-annotated region. Based on subgraph skeleton path extraction and connection tension calculation, the stability and continuity of the main structural paths within the point set can be quantified, identifying potential deformation, stretching, or twisting locations. By generating second pseudo-annotated feature data through pseudo-connectivity calculation, the system possesses the ability to intelligently identify and compensate for connection anomalies and structurally incomplete regions during structural recognition.

[0050] Optionally, S3 includes:

[0051] Based on the point cloud component clustering data, the substation point cloud data is processed by the main direction of the components to obtain the main direction data of the components.

[0052] The attitude offset is calculated based on the main direction data of the component to obtain the attitude offset data;

[0053] The attitude collinearity residual map is processed based on the attitude offset data to obtain the attitude collinearity residual map data;

[0054] The attitude calibration data is obtained by performing pseudo-annotation expectation optimization on the substation site cloud data based on the attitude collinearity residual map data.

[0055] This invention effectively extracts the spatial orientation information of each point cloud cluster component through component principal orientation processing, solving the problem of difficult alignment in scenarios with inconsistent attitudes and chaotic orientations in traditional point clouds. By calculating attitude offset data, the system can quantify the offset between each component and the global reference direction, thereby identifying components with rotational errors or orientation distortions. Through attitude collinearity residual map processing, the principal orientation differences between multiple related components are uniformly modeled and residual optimized to ensure that components that should be arranged collinearly or parallel in space (such as busbars and circuit breaker strings) maintain overall structural consistency. Based on the residual map results, pseudo-annotation expectation optimization can be performed to correct components with large attitude deviations by combining semantic prior constraints, improving the spatial regularity of the equipment in the global coordinate system.

[0056] Optionally, S4 includes:

[0057] Based on the attitude calibration data, a spatial relationship diagram of device components is constructed to obtain component relationship diagram data;

[0058] The component structure data is obtained by reconstructing the structure based on the component relationship diagram data;

[0059] The component structure data is connected and assembled to obtain a three-dimensional model of the substation.

[0060] This invention constructs a spatial relationship diagram of equipment components from point cloud data after attitude calibration. This accurately reconstructs the spatial dependencies and adjacency relationships between components in a substation, avoiding connection errors caused by attitude errors or component misalignment. The establishment of the component relationship diagram is not only based on spatial geometric relationships but also incorporates constraints such as orientation consistency and interface compatibility, thereby improving the accuracy of structural modeling and scene adaptability. During structural reconstruction, the system can perform geometric completion and contour fitting based on the spatial configuration of each component in the relationship diagram, achieving breakpoint connection and component contour closure, enhancing the integrity and visualization of the overall model. Through connection and assembly, structural registration and logical assembly between components are achieved, generating a high-quality 3D substation model with realistic spatial distribution, reasonable topological connectivity, and semantic labeling information.

[0061] The purpose of this invention is to effectively enhance the edge preservation and noise suppression capabilities of key geometric structures in point clouds during the initial stage by introducing convolutional filtering operations. Based on the fusion of convolutionally filtered data and the original point cloud, structural change features are extracted, capturing not only local morphological changes such as curvature abrupt changes and normal fluctuations, but also dynamically identifying potential equipment boundaries and component segmentation lines. A component attention screening mechanism driven by structural changes enables automatic focusing on high-response regions. Combined with mask enhancement and local aggregation strategies, high-confidence pseudo-labels are generated, providing prior guidance for component clustering. During component clustering, geometric and connectivity features are fused, and a nested structural graph is constructed to effectively distinguish various equipment components and identify their hierarchical relationships. In the attitude calibration stage, based on principal direction extraction, collinear residual mapping, and pseudo-label expectation optimization, the orientation and position of equipment components in three-dimensional space are unified, providing a consistent coordinate system for topology modeling. Through spatial connectivity graph construction and semantically driven topology assembly, a high-precision three-dimensional substation model with complete geometric structure, connection logic, and semantic information is obtained. Attached Figure Description

[0062] 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:

[0063] Figure 1 A flowchart illustrating the steps of a substation cloud data processing method for constructing a three-dimensional model is shown in one embodiment.

[0064] Figure 2 A flowchart illustrating the steps of a convolutional filtering method according to an embodiment is shown.

[0065] Figure 3 A flowchart illustrating the steps of a structural change feature extraction method according to an embodiment is shown.

[0066] Figure 4 A flowchart illustrating the steps of an attitude calibration method according to one embodiment is shown.

[0067] Figure 5 A flowchart illustrating the steps of a topology modeling method according to an embodiment is shown.

[0068] 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

[0069] 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.

[0070] 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.

[0071] 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.

[0072] Please see Figures 1 to 5 This application provides a substation cloud data processing method for constructing a three-dimensional model, the method comprising:

[0073] S1. Obtain substation cloud data; perform convolution filtering on the substation cloud data to obtain convolution-filtered data;

[0074] Specifically, the system uses fixed LiDAR equipment (such as a static 3D laser scanning system mounted on a tripod) or mobile point cloud scanning platforms (such as handheld or vehicle-mounted ground laser scanners) to perform spatial scanning of the substation site and acquire raw point cloud data. The acquired data adopts common point cloud formats such as .pcd or .las, and the reference coordinate system is a local geographic coordinate system. Preliminary noise reduction processing is performed on the raw point cloud data, including but not limited to isolated point removal based on point density thresholds or statistical filters. In the filtered point cloud, the system constructs a fixed-size local neighborhood for each point and uses the k-nearest neighbor search algorithm to obtain its k nearest neighbors (e.g., k=20), forming the set of neighbors for that point. The system performs local feature enhancement on the point cloud data based on a structure-preserving point cloud convolution kernel function. For a point, its convolution response is defined as the weighted sum of the feature values ​​of all points in its neighborhood, and the convolution calculation formula is: ,in For point The convolution response value, The coordinates of the current center point. The coordinates of adjacent points For point The neighborhood set, For point-to-point The convolution weights, Adjacent points The input feature values. The weight function is defined as: ,in For point-to-point The convolution weights, It is an exponentially decaying function. The coordinates of the current center point. The coordinates of adjacent points For distance attenuation control parameters, The cosine of the angle between the normals. For point and The angle between the normal vectors is calculated using the above method to obtain the convolution response value of each point, thereby generating a set of point cloud convolution feature data that reflects the characteristics of local structural changes.

[0075] S2. Extract structural change features from convolutional filtered data and substation cloud data to obtain structural change feature data; perform component attention filtering on substation cloud data based on structural change feature data to obtain point cloud component pseudo-label data; perform component clustering on point cloud component pseudo-label data to obtain point cloud component clustering data.

[0076] Specifically, in the process of extracting structural change features, the system constructs a local neighborhood covariance matrix for each point. This matrix is ​​calculated based on the three-dimensional coordinates of the point and its k nearest neighbors. Eigenvalue decomposition is then performed on the covariance matrix to obtain three eigenvalues ​​arranged in ascending order. , representing the local spatial extensibility of a point along the principal axis, secondary axis, and normal direction, respectively. The system uses these to calculate the local curvature index of the point, defined as... This index reflects the degree of surface abrupt change at a point; a higher value indicates that it is located in a region of drastic structural change, such as a corner or edge. The curvature value and normal vector change at the system's connection point construct the structural change response index. (Calculation point) The magnitude of the vector difference between the normal vector and the average normal vector of its neighborhood is expressed as: ,in For point The normal vector, It is the average of the normals of its neighborhood. The structural change index is defined as follows: ,in This is the structural change response index. Let be the local curvature value of the point. The normal variation weighting coefficient is used. For point The normal vector, The average of its neighborhood normals is used. The system uses the Sigmoid activation function to map the above structural change exponent to point-level attention weights between 0 and 1, i.e. ,in To represent the Sigmoid function, For point The intensity of attention. Based on attention weights, the system sets an empirical threshold. The system selects all points with attention intensity greater than a threshold as the initial salient point set to construct an attention mask region. For each point within the mask region, the system performs a spherical neighborhood dilation operation with an dilation radius of 0.2 meters, merging overlapping dilated regions into several local clusters. Each cluster is considered a potential component candidate region. During the pseudo-label generation stage, the system assigns a unique pseudo-label ID to each cluster. Label initialization is based on the geometric centroid location of the cluster, using a spatial grid or clustering index for preliminary classification. Simultaneously, the system calculates the average attention value of all points within the cluster as the label confidence score.

[0077] S3. Perform attitude calibration on the substation point cloud data based on the point cloud component clustering data to obtain attitude calibration data;

[0078] Specifically, during attitude calibration, the system extracts geometric features from each clustered pseudo-labeled component cluster. These features include: a normal consistency index (the average cosine similarity between the pairwise normal vectors of all points within the cluster); a point density index (calculating the number of neighboring points within a unit radius sphere centered on each point in the cluster and averaging the results); and a compactness index (mapping the component point set to a unit voxel grid and calculating the proportion of voxels occupied by at least one point relative to the total number of voxels). The system sequentially constructs a three-dimensional feature vector from these three features and inputs it into a density-based clustering algorithm (DBSCAN, or its hierarchical variant HDBSCAN) for secondary clustering. The minimum sample size for the clustering algorithm is set to 10, and the distance threshold is set to 0.3 meters. The clustering results output the category label of each component point and its corresponding point set. For each component point set in the clustering results, the system uses principal component analysis (PCA) to extract its principal axis direction. This principal axis direction is the eigenvector corresponding to the largest eigenvalue of the point set covariance matrix, representing the main spatial extension direction of the component. Based on prior domain knowledge, such as the vertical arrangement of cable components and the horizontal extension of busbar components in a substation, the system pre-determines the target alignment direction for each type of component. For each principal axis vector, the system calculates the corresponding three-dimensional rotation matrix to ensure that the principal axis vector aligns with the target alignment direction after rotation. This rotation matrix is ​​obtained by minimizing vector differences. During attitude transformation, the system also calculates the geometric centroid position of each clustered component point set and performs translation operations accordingly to maintain relative spatial structural consistency after directional rotation. Through rigid body transformation operations consisting of rotation and translation, the standardized calibration of component attitudes in a unified reference coordinate system is achieved.

[0079] S4. Based on the attitude calibration data, perform topology modeling to obtain a three-dimensional model of the substation.

[0080] Specifically, in the topology modeling process, the system constructs an initial candidate topology graph based on the attitude-calibrated components. Each independent point cloud component cluster is used as a node in the graph, and the existence of structural connections is determined according to the following rules: if the Euclidean distance between any two components is less than a preset threshold (e.g., 0.5 meters), and the angle between their principal axes is less than a set threshold (e.g., 20 degrees), a candidate edge is established between them, representing a potential physical connection. The system filters the legality of the candidate edges according to a preset functional connection rule base. This rule base records the valid connection types between typical equipment components in a substation, such as allowing a "busbar" component to connect to a "surge arrester," and allowing a "circuit breaker" to connect to a "support," while connection relationships that do not conform to the rules (such as component pairs that lack physical or functional reachability) are eliminated. The verified legal connection edges are used to construct the final component topology graph. The system encodes the above topology relationships into a structural connection relationship table, preferably organized using JavaScript Object Notation (JSON) format, where each record contains fields such as component identifier, connection type, spatial location information, and connection edge attributes. In terms of 3D structural model construction, the system performs morphological restoration processing on the point cloud data of each component cluster. Specifically, it can use Poisson surface reconstruction to obtain a continuous and smooth surface representation, or use α-shape envelope extraction to extract boundary configurations, to adapt to point cloud inputs with different densities or completeness. For component pairs with confirmed connections, the system interpolates and generates bridging structures based on the spatial docking region between them, such as connecting pipes, connecting lines, or surface patches. The system outputs two types of model data: one is a complete 3D point cloud model of the substation based on components and bridging structures, which can be in PLY, OBJ, or other mainstream 3D model file formats; the other is a topology diagram recording the connection relationships between components, preferably in GraphML or JSON format.

[0081] Optionally, the convolutional filtering includes:

[0082] S11. Calculate the normal vector covariance tensor diagram of the substation cloud data to obtain tensor diagram data;

[0083] Specifically, for any point in the point cloud dataset Construct a spherical neighborhood with this point as its center and a radius of 0.2 meters, denoted as . , This is a subset of the point cloud set P. For the neighborhood point set... Calculate its three-dimensional spatial covariance tensor. Define the average location (centroid) of points in the neighborhood as... The covariance matrix is ​​defined as: ,in Let covariance matrix be the variance matrix. For point The neighborhood point set, The spatial coordinates of the neighboring points For point The neighborhood point set, Let the coordinates be the centroid coordinates of the neighborhood. This is a transpose operation. Eigenvalue decomposition is performed on the covariance matrix to obtain three non-negative eigenvalues. , and Assume it satisfies Each eigenvalue corresponds to an eigenvector, where corresponding unit eigenvector This represents the main direction of that neighborhood. Each point... The tensor feature information, i.e., the set of eigenvalues ​​{ } and the principal direction vector It is stored as a structured element in the tensor graph data.

[0084] S12. Perform spectral domain point cloud response transformation based on tensor graph data to obtain point cloud spectral data;

[0085] Specifically, for any point in the point cloud dataset Within its neighborhood, from the unit sphere M direction vectors are uniformly sampled. For each sampled direction, a local direction response function is defined, such as... ,in For point In direction The local response function value on, For point Neighborhood any point in, For point The neighborhood point set, It is an exponential function. The coordinates of the center point, This is the bandwidth control parameter for the Gaussian weighting function. For the first Each sampling direction vector For point The corresponding principal direction unit vector, for and The inner product of the above direction functions. The frequency domain is mapped to the spherical frequency domain through spherical harmonic transformation, expanded using spherical harmonic basis functions, and low-order frequency terms (such as spherical harmonic order l∈[0,4]) are retained to form a spectral description vector.

[0086] S13. Perform multi-kernel convolution processing on the point cloud spectrum data to obtain convolutionally filtered data. The multi-kernel convolution processing includes spherical kernel convolution processing, planar kernel convolution processing, and cylindrical kernel convolution processing.

[0087] Specifically, the system processes the point cloud spectral data using different convolution kernels to obtain convolutionally filtered data. Spherical kernel convolution is used for uniform spherical regions, such as the top of a surge arrester or the spherical end of an insulating support. The convolution weight is defined as follows: for any point within the center point and its neighborhood, the weight value is composed of the square of the Euclidean distance between the two points, and is attenuated by an exponential function. The weight function is expressed as: ,in The weights are spherical kernel convolution weights. It is a natural exponential function. The three-dimensional coordinates of the center point The three-dimensional coordinates of the neighboring points Here is the bandwidth parameter for the spherical kernel. The spherical kernel convolution response is calculated by summing the weighted spectral values ​​of the neighborhood: ,in For point The spherical convolution response value, The three-dimensional coordinates of the neighboring points For point The neighborhood point set, The weights are spherical kernel convolution weights. For point The spectral eigenvectors.

[0088] Planar kernel convolution processing is used for grounding layers, platforms, or other locally approximate planar regions in substations. This convolution method uses the consistency of the orthogonal direction of the normal vector as the weighting criterion; that is, for each point, the system calculates its normal vector direction and projects the vector difference between neighboring points onto that normal vector. The convolution weight function is calculated exponentially based on this projected distance. ,in For planar kernel convolution weights, It is a natural exponential function. The three-dimensional coordinates of the neighboring points The three-dimensional coordinates of the center point For point The unit normal vector, This is the parameter for suppressing planar projection.

[0089] Cylindrical kernel convolution processing is used in substation scenarios for slender cylindrical structures such as cables and pipes. This convolution kernel identifies the major axis direction of a point by constructing a local principal axis direction vector at that point (which can be determined by the eigenvector corresponding to the largest eigenvalue in the tensor graph). For neighboring points, the system calculates their relative position vector, then removes the projection of this vector onto the principal axis direction, extracting the radial deviation component. The convolution weight function depends on this radial distance and is weighted and suppressed using an exponential function. ,in The weights are the convolution weights of the cylinder kernel. It is a natural exponential function. The three-dimensional coordinates of the neighboring points The three-dimensional coordinates of the center point For point The local principal axis direction vector, The bandwidth parameter, which controls the radial attenuation, suppresses convolution interference from axially repeating structures, thereby improving the recognition accuracy of the outer contour of columnar structures.

[0090] Optionally, the structural change feature extraction includes:

[0091] S21. Extract structural changes from the convolutional filter data and the substation cloud data to obtain structural change data.

[0092] Specifically, the system uses the convolutionally filtered data after multi-kernel convolution processing as input to perform structural disturbance feature analysis on each point in the original point cloud of the substation. The original point cloud dataset is denoted as... , where each point Includes its three-dimensional spatial coordinates and normal vector And the tensor features generated by the preprocessing step. After multi-kernel convolution processing, the system generates a corresponding convolution response value for each point, denoted as . For each point The system calculates the degree of difference between its convolutional response and that of its neighboring points, defining it as the gradient change rate index. A set of its k nearest neighbors is constructed. For all neighboring points The mean absolute difference is calculated as follows: ,in The convolution response difference rate, For point The number of neighboring points, The first in the neighborhood One point, For point The neighborhood set, For point The convolutional filter response value, For point The convolutional filter response value. System calculation points. Normal jump rate, based on point Its neighboring points The mean cosine difference of the angle between the normal vectors is calculated, and the expression is as follows: ,in Normal perturbation rate, For point The number of neighboring points, The first in the neighborhood One point, For point The neighborhood set, For trigonometric functions, For point The unit normal vector, For point The unit normal vector, This is the dot product of two unit normal vectors. The system performs structural perturbation calculations, which are expressed as follows: ,in This is the structural disturbance index. These are the convolution perturbation weight coefficients. The convolution response difference rate, This is the normal perturbation weighting coefficient. This represents the normal perturbation rate. The system outputs the structural perturbation index corresponding to each point, forming a structural change heatmap data covering the entire point cloud.

[0093] S22. Calculate the structural jump index based on the structural change data to obtain the structural jump index data;

[0094] Specifically, for each point in the substation site cloud, the system assesses the degree of structural difference between it and its local neighborhood based on the generated structural disturbance score. The system constructs a nearest neighbor set for each point and defines a connectivity change rate to characterize the structural disturbance difference between the point and its neighborhood. The calculation expression is as follows: ,in The rate of change of connectivity, For point The number of neighboring points, For the neighboring region The three-dimensional coordinates of the points For point The neighborhood set, For point The structural disturbance intensity value, For point The structural disturbance intensity value. The rate of change of the system with respect to all points. After normalization, the index is mapped to the standard interval [0,1] using the minimum-maximum normalization formula, resulting in the standardized structural jump index. The system uses a preset jump threshold. (For example, set to 0.6) to make a judgment, when the structural jump index at a certain point At this point, the point is considered a potential structural break or discontinuous connection edge. The system outputs a structural jump index layer, with each point in the layer accompanied by its corresponding jump index value.

[0095] S23. Based on the structural change data and structural jump index data, perform structural fluctuation projection analysis on the substation cloud data to obtain structural change characteristic data.

[0096] Specifically, based on structural change data and the structural jump index, the system constructs a local fluctuation analysis model for each point in the substation site cloud. The system performs principal component analysis (PCA) on the neighborhood point set of each point, calculates its neighborhood covariance matrix, and extracts the eigenvector corresponding to the largest eigenvalue as the principal axis direction of the local region at that point. The system constructs a directional projection function for the local structural disturbance based on this principal axis direction. For each point in the neighborhood, the system calculates the weighted projection of its disturbance value onto the principal direction, as shown in the following formula: For each point... The projected contribution is defined as the product of the structural disturbance value and the cosine weight of the inter-point direction vector along the principal axis. The system performs a weighted sum of all projected contributions to obtain the structural disturbance component along the principal direction: ,in The main direction of the disturbance component. The three-dimensional coordinates of the points in the neighborhood. For point The neighborhood point set, For point The structural disturbance value, The three-dimensional coordinates of the points in the neighborhood. The three-dimensional coordinates of the current analysis point. For point Local principal axis direction, This represents the distance between points. The system calculates the residual disturbance term of the disturbance in the plane orthogonal to the principal direction. This can be approximated by subtracting the principal direction disturbance from the total disturbance. Calculate the structural wave directionality. , which represents the degree of concentration of structural disturbances in the principal direction, is defined as follows: ,in For structural wave direction rate, The main direction of the disturbance component. These represent the residual perturbation components in orthogonal directions. The calculation results are used to determine whether local perturbations are mainly concentrated along the principal axes. Values ​​closer to 1 indicate more linearly concentrated perturbations, while values ​​closer to 0 indicate uniform or randomly oriented perturbations. The system obtains structural change characteristic data. .

[0097] Optionally, the component attention filtering includes:

[0098] Point-level attention data is obtained by performing point-level attention calculation based on structural change feature data.

[0099] Specifically, each point has a three-dimensional structural change feature vector. Attention score is calculated as ,in Point-level attention weights, It is the sigmoid activation function. For structural residual weighting coefficients, The structural change residual index of the point. These are the weighting coefficients for the jump response. For the boundary jump response of a point, This is the normal fluctuation weighting coefficient. Normal volatility This is the bias term. Output This serves as the point-level attention weight. The attention score for each point is obtained, forming a point-level attention layer.

[0100] A priori masking enhancement process is performed on the substation point cloud data based on the point-level attention data to obtain the point cloud enhanced data.

[0101] Specifically, the system constructs a semantic mask layer based on the attention score of each point. An attention threshold is set. (Values ​​range from 0.6 to 0.8). For each point, if its attention weight is greater than or equal to the threshold... If a point is active, it is marked as a masked active point (mask value 1); otherwise, it is marked as an inactive point (mask value 0). The system performs a spatial dilation operation based on a spherical neighborhood on all masked active points. A neighborhood sphere with a radius of 0.25 meters is constructed centered on each masked point, and all points within this neighborhood not covered by the initial mask are marked as enhanced points. This process effectively prevents the omission of component identification information due to local point cloud sparsity, scan occlusion, or structural breaks. The system then performs a weighted boosting process on the attention scores of these enhanced points. The original attention weights are increased by an amplitude. (amplitude) The values ​​are between 0.1 and 0.2, and the results are truncated to ensure that their maximum value does not exceed 1. Therefore, the enhanced attention weights are expressed as follows: .

[0102] Local aggregation and enhancement are performed on the point cloud augmentation data to obtain locally aggregated data;

[0103] Specifically, the system performs local clustering on point cloud data with high attention weights within a three-dimensional space to identify candidate regions that may belong to the same structural component, thereby enhancing the expressive power of component-level semantic structures in the point cloud data. The system constructs a point connectivity graph. The system then filters all enhanced attention weights from the augmented point cloud data. Points with a distance greater than 0.5 are considered candidate structural points. Connection edges between candidate points are constructed based on Euclidean spatial distance. When the spatial distance between any two points does not exceed the preset maximum connection distance (set to 0.3 meters), the two points are considered to have spatial adjacency in structure, and an undirected edge is established between them. All edges constitute a connection graph between points. Connected subgraphs are extracted from the connection graph as initial component clusters. Several connected regions are identified from the above connection graph using graph traversal methods (e.g., Breadth-First Search or Depth-First Search). Each connected subgraph represents a group of interconnected high-attention points, which the system uses as potential component candidate clusters. Local consistency filtering is performed on the initial component clusters. For each connected subgraph, the system calculates the normal vector consistency index and point density index for all points in the cluster. Normal vector consistency is evaluated by calculating the standard deviation of the normal vector. If the standard deviation of the normal vector in a cluster exceeds a preset threshold, the region is considered structurally inconsistent. The density index is measured by the number of points per unit volume. If it is less than the minimum density threshold, the cluster is considered to lack structural support. The system will discard any component cluster that does not meet any of the above conditions.

[0104] Pseudo-labels are generated based on local aggregated data to obtain pseudo-label data for point cloud components.

[0105] Specifically, the system treats each local cluster that passes the structural consistency screening as a potential candidate unit for structural components, constructs corresponding pseudo-label units and assigns them identification numbers, and assigns a corresponding confidence score to each unit based on attention information. Label numbers and attributes are initialized. For each local cluster (i.e., the connected subgraph extracted and filtered through the connection graph in step S16), the system assigns a unique identification number, such as cluster_001, cluster_002, etc. The system records the indices of all point sets contained in the component cluster and calculates its spatial geometric centroid. The system assigns a pseudo-label confidence value. For example, the system calculates the average attention score of the overall structural cluster region based on the attention enhancement value of each point in each component cluster, and uses this as the confidence score of the pseudo-label unit. Suppose a local cluster contains several points, and its attention enhancement value is... The confidence level of the pseudo-labeled unit is defined as the average of the attention enhancement values ​​of all points, i.e. ,in The confidence score for pseudo-labeled units. For the points contained in this cluster, Index the points in the point cloud. For aggregated cluster point set, This represents the attention enhancement value for each point. A higher confidence score indicates that the region has high consistency and saliency in its spatial structure. The output point cloud component pseudo-annotation data includes the identifier of each pseudo-annotated unit, the index of its point set, the geometric centroid coordinates, and its confidence score.

[0106] Optionally, the component clustering includes:

[0107] Feature extraction is performed on the pseudo-labeled data of point cloud components to obtain pseudo-labeled feature data;

[0108] Specifically, for each pseudo-labeled block, the system extracts the following feature dimensions, including the geometric center coordinates, and calculates the mean position of the point set in the three-dimensional coordinate system, represented as follows: , representing the corresponding X, Y, and Z coordinate values. The size vector represents the boundary range of the statistical point set along the three coordinate axes. , , This is used to characterize the length, width, and height dimensions of the pseudo-labeled block. The shape compactness index is based on the eigenvalue sequence extracted by principal component analysis (PCA). , This represents the spatial extent of the point cloud in the main direction. The spatial extent of the point cloud in the secondary principal direction. To calculate the spatial extent of the point cloud in the minimum direction, calculate the ratio of the maximum principal axis to the minimum principal axis. This is used to measure whether the geometric distribution of the block exhibits a slender or flattened trend. Normal consistency is calculated by determining the standard deviation of the angle between the normal vectors of all points within the pseudo-annotation block, denoted as . Point density index is calculated by counting the number of points within a unit sphere (e.g., with a radius of 0.2 meters) and dividing by the volume of the sphere to obtain an estimate of the local point density. .

[0109] Density clustering is performed on the pseudo-labeled feature data to obtain feature clustering data;

[0110] Specifically, for each pseudo-annotated sub-block, the system has extracted its multi-dimensional feature vector, including geometric center, size vector, point density, shape compactness index, connectivity index, and the aforementioned feature data. For any two pseudo-annotated component sub-blocks i and j, based on their normalized feature vectors... and Calculate its distance in the feature space. The distance function used is weighted Euclidean distance: ,in For sub-blocks and Weighted distance in feature space, Indexed by feature dimensions, This refers to the weight of that dimension; for example, a higher weight can be assigned to the geometric dimension (e.g., ...). , These are the weight values ​​for the size dimension. (Weight values ​​for the point density dimension) For the first The sub-block in the Values ​​in each feature dimension For the first The sub-block in the The system calculates the values ​​across several feature dimensions. For each pseudo-labeled sub-block, it counts the number of neighboring samples within a specified radius to construct a local neighborhood sample set. If the number of neighbors for a sub-block is not less than a preset minimum neighbor threshold (e.g., 3-5), it is marked as a core sub-block; otherwise, it is considered a boundary or isolated sub-block. Starting from any core sub-block, the system merges all mutually reachable core and boundary sub-blocks in its neighborhood into the same cluster. By repeating this process, the system can progressively identify all clusters with density connectivity, while sub-blocks not participating in any cluster are individually marked as noise or anomalous structures—these are mislabeled, isolated components, or structurally broken samples. The system assigns a cluster number to each pseudo-labeled sub-block, forming feature clustering data.

[0111] Structural feature data is obtained by extracting structural features from the feature clustering data;

[0112] Specifically, for each cluster Structural features are extracted from the following four dimensions, such as bounding box collinearity feature extraction, specifically for each sub-block within a cluster. Principal component analysis (PCA) is performed to extract the first principal axis direction, and the variance of the included angles between the principal axes of all sub-blocks is calculated. Smaller included angles and lower variances indicate a uniform orientation of components, such as cable layouts or parallel busbar arrangements. Hierarchical distribution features are extracted by projecting the geometric centers of all sub-blocks along the vertical direction (Z-axis) or the principal axis direction, and counting the number of layers where the component center points are distributed in that direction to characterize the layering of components, such as multi-layered structures like circuit breakers and brackets. The adjacency matrix is ​​calculated for any pair of sub-blocks. Calculate the Euclidean distance between their nearest points. If this distance is less than the adjacency threshold, then... (e.g., 0.1 meters), then in the adjacency matrix The markers indicate direct adjacency or physical contact; this information helps identify spatial relationships such as supports, connections, and suspensions. Geometric inclusion relationships are identified, specifically for pairs of sub-blocks. Determine its bounding box Does it fully include? If true, it can be recorded as a geometric nesting relationship, such as the shell and internal components.

[0113] Based on the structural feature data, a device component structure nesting graph is constructed from the feature clustering data to obtain the number of point cloud component clusters.

[0114] Specifically, each pseudo-annotated sub-block is treated as a node in the graph, and each node is bound to its corresponding geometric attributes, such as center position, bounding box volume, principal axis direction, and class probability (generated by the aforementioned pseudo-annotation). Edges are established between node pairs, specifically including geometric inclusion edges. The smallest bounding box completely contains the child blocks. The bounding box is then used to add directed edges to the graph. This indicates that B is a nested component of A. Physical adjacency edges are defined if the distance between the nearest points of any two sub-blocks is less than a set adjacency threshold (e.g., ...). If the distance is 1 meter, then add an undirected edge. This indicates that the two components have a physical contact or adjacency relationship. Node attributes include geometric volume, principal axis direction vector, and pseudo-label category distribution probability; edge attributes include relative direction angle (principal axis angle), geometric inclusion ratio (inclusion strength), contact area ratio, etc., enabling quantitative characterization of the tightness and hierarchy of component relationships. A weakly connected subgraph extraction algorithm is performed on the constructed nested structure graph, i.e., after removing directional information from the graph, all connected regions are searched; each weakly connected subgraph is considered a complete device component unit; each subgraph is assigned a unique number as a component cluster ID. A component ID label is attached to each point in the point cloud to identify its component; the generated component structure graph can be exported as graph structure data (such as GraphML or JSON format); it also supports visualization of the spatial layout of each component and the point cloud clustering results.

[0115] Optionally, the pseudo-labeled feature data includes first pseudo-labeled feature data and second pseudo-labeled feature data, and feature extraction includes:

[0116] Geometric features are extracted from the pseudo-annotated data of point cloud components to obtain the first pseudo-annotated feature data;

[0117] Specifically, geometric feature extraction is performed on the pseudo-annotated data of point cloud components, including point set extraction for each pseudo-annotated sub-block. Construct a geometric descriptor subset. This includes extracting the geometric center coordinates, calculating the dimension vector, extracting the principal axis directions, calculating the vector shape ratio, and estimating the point density. The extracted geometric center coordinates form a computation point set. The geometric center coordinates of the point set are defined as the average of all three-dimensional coordinates of that point set, i.e. ,in The coordinates of the geometric center are For the geometric center at Coordinate values ​​in the three axes, For the geometric center at Coordinate values ​​in the three axes, For the geometric center at Coordinate values ​​in the three axes, For the first The number of points in each pseudo-labeled sub-block. For point set The first in One point, For the first The point set of a pseudo-labeled sub-block. for exist Coordinate values ​​in the three axes, for exist Coordinate values ​​in the three axes, for exist The coordinate values ​​along the three axes. The dimension vector is calculated by taking the difference between the maximum and minimum values ​​of the point set along each coordinate axis, thus forming the dimension vector: ,in For the size vector in Components in the axial direction For the size vector in Components in the axial direction For the size vector in Components in the axial direction for The maximum component in the axial direction, for Minimum component in the axial direction, for The maximum component in the axial direction, for Minimum component in the axial direction, for The maximum component in the axial direction, for The minimum component along the axis represents the extent of the point cloud bounding box along each axis. The principal axis directions are extracted as point sets. Principal component analysis (PCA) is performed to obtain the eigenvectors of the covariance matrix, where the principal axis direction is the eigenvector corresponding to the largest eigenvalue, representing the dominant distribution direction of the point cloud in 3D space. The vector shape ratio is calculated by extracting the three eigenvalues ​​from the PCA decomposition. ,in The largest principal component eigenvalue, The second largest principal component eigenvalue. Calculate the ratio relationship of the smallest principal component eigenvalue. To determine whether the point set structure is elongated, flat, or blocky, if It is slender; if If the three are approximately equal, it is a flat type; if they are approximately equal, it is a block type. Point density estimation is based on a set of points. Includes There are points, and the bounding box volume is... Then the point density is defined as: ,in For point density, The number of points within the point set. The bounding box volume of the sub-block represents the number of points per unit volume, reflecting the density of the point cloud distribution.

[0118] Local connectivity features are extracted from the pseudo-annotated point cloud component data to obtain the second pseudo-annotated feature data.

[0119] Specifically, a nearest neighbor connection graph is constructed for all pseudo-labeled sub-blocks. The geometric center of each sub-block is treated as a node in the graph. If the distance between the nearest points of any two sub-blocks is less than a preset connection threshold (e.g., 0.5 meters), an undirected edge is created in the graph to represent their adjacency. Based on the above graph structure, for each graph node (i.e., pseudo-labeled sub-block), the following structural graph metrics are extracted: node degree, average edge length, clustering coefficient (the ratio of the actual number of triangular closures formed between the current node's neighbors to the theoretical maximum possible number of closures), betweenness centrality, and pseudo-connectivity metrics (i.e., whether the current sub-block has redundant connected paths in its topology (e.g., whether there are alternative edges that bypass the current node and still connect to adjacent nodes)).

[0120] Optionally, the geometric feature extraction includes:

[0121] Local covariance is calculated on the pseudo-annotated data of point cloud components to obtain local covariance data;

[0122] Specifically, for each point In its neighborhood of radius r (e.g., 0.05m) In the process, construct the covariance matrix. ,in For point The covariance matrix, For point The number of neighboring points, The coordinates of the neighboring points, For point The neighborhood set, The mean coordinates of the neighborhood center points; It is a local structure tensor, i.e., local covariance data.

[0123] Curvature spectrum data is obtained by extracting curvature spectrum data from local orthometric data;

[0124] Specifically, for each Perform eigenvalue decomposition: ,in It is a local covariance matrix. The eigenvector matrix, It is a diagonal matrix of eigenvalues. The largest eigenvalue, It is the second largest eigenvalue. To obtain the minimum eigenvalue, extract the following spectral curvature indices, such as linearity. flatness divergence Anisotropy curvature This forms curvature spectrum data.

[0125] Multi-scale point cloud structural fluctuation analysis was performed on the pseudo-annotated data of point cloud components based on curvature spectrum data to obtain structural fluctuation spectrum data.

[0126] Specifically, multiple sets of neighborhood scales are defined. The system sets a set of radius parameters to construct local neighborhoods at different scales, denoted as { }, where typical values ​​can be taken as follows =0.03 meters, =0.05 meters, =0.08 meters. The curvature spectrum extraction operation is repeated at each scale. For each point, the system performs covariance calculation and curvature spectrum construction processes at different neighborhood radii, thereby obtaining the corresponding curvature spectrum vector at each scale. The system calculates the degree of spectral variation between different scales. For each point, the system calculates the pairwise curvature spectrum variation amplitude under all scale combinations. For any two scales, the curvature spectrum difference is defined as: ,in For the first Points at radius and The curvature spectrum variation value below, For the first The neighborhood radius value used to extract the curvature spectrum at each scale, such as 0.03. For the first The neighborhood radius value used to extract the curvature spectrum at each scale, such as 0.05. For the first Points at radius The curvature spectrum vector below, For the first Points at radius The curvature spectrum vector is then used to construct the structural wave spectrum vector. The system then applies this to all of the above. Statistical analysis of the fluctuation values ​​is performed to extract the maximum curvature fluctuation amplitude, defined as the set of curvature spectrum differences with the largest number of scale combinations; and the average curvature fluctuation rate, defined as the average of the curvature spectrum differences across all scale combinations. The system combines the maximum curvature fluctuation amplitude and the average fluctuation rate at each point to form a two-dimensional structural fluctuation spectrum vector.

[0127] The first pseudo-labeled feature data is obtained by performing a density projection along the principal axis based on the structural wave spectrum data.

[0128] Specifically, principal component analysis (PCA) is performed on the entire pseudo-labeled point set. The system takes the current pseudo-labeled cell as the object and performs PCA on the 3D coordinate data of all points contained within it, extracting the principal direction basis of the point set. The PCA result outputs three mutually orthogonal unit vectors, which are the first principal direction vectors. (Direction of maximum variance), second principal direction and the third main direction (Minimum variance direction), all are three-dimensional vectors. The system calculates the projected response of the structural wave spectrum along each principal axis. For any point in the pseudo-labeled region, the system calculates its response projection along each principal direction based on its structural wave spectrum vector (containing both maximum curvature and average volatility components) and its normal vector. In the j-th principal direction... Above, the principal axis response value at that point is defined as ,in For the first The points are in the main direction Projected response value on The amplitude of the structural wave spectrum (such as the L2 norm or a certain weighted combination). For the normal vector and the first The cosine of the angle between the principal directions. Normal vector With the main direction The angle between them. The projection values ​​in the three directions are combined to form a density response vector. For each point, the system combines its response values ​​in the three principal directions into a set of three-dimensional vectors. The statistics (mean, variance, maximum value, etc.) of the density vector of the point set within the entire pseudo-labeled block are calculated to form the first pseudo-labeled feature data.

[0129] Optionally, the local connectivity feature extraction includes:

[0130] The pseudo-labeled point cloud component data is processed using a nearest neighbor graph to obtain nearest neighbor graph data.

[0131] Specifically, input the set of pseudo-labeled points For each point, find its K nearest neighbors. During graph construction, each point in the point set is... As a node in the graph, it constitutes the node set V of the graph; and at each node... Points in its neighborhood Undirected edges are established between V and E, forming an edge set E. The graph structure is represented as G=(V,E), where each edge... Point Its neighboring points The spatial connections between them. Each edge Assign weights ,in For the edge The connection weights represent the spatial similarity between point pairs. It is a natural exponential function. For the first The three-dimensional spatial coordinate vector of a point For the first The three-dimensional spatial coordinate vector of a point This is the bandwidth parameter of the Gaussian kernel function. The output is the constructed nearest neighbor connectivity graph data.

[0132] Local subgraph connectivity analysis is performed based on the nearest neighbor connection graph data to obtain local subgraph connectivity data;

[0133] Specifically, for each pseudo-annotation component ,from Extracting subgraphs ,right Weakly connected component analysis is performed to identify whether there are multiple unconnected independent regions within a subgraph. If two or more weakly connected components exist, it indicates that the pseudo-labeled structure exhibits pseudo-splitting. For each connected component, its connectivity indices (such as edge density, average degree, path coverage, etc.) are calculated, and the connectivity vector of each subgraph is output. Including the number of sub-blocks Average Connectivity radius coverage ratio .

[0134] Pseudo-connectivity supplementation is performed based on the local subgraph connectivity data to obtain subgraph connectivity data;

[0135] Specifically, for multiple sub-blocks within the same pseudo-annotation component Search for the nearest pair of points between the centers of all sub-blocks; if ( For sub-blocks The geometric center point coordinate vector in For sub-blocks The geometric center point coordinate vector in The connectivity distance threshold is used to determine whether there is a potential connection between two sub-blocks (in meters, e.g., 0.2 meters). If the two sub-blocks are in different connected components, then a supplementary edge is established. Add these edges to the atomic graph The subgraph is reconstructed and completed, and the enhanced connected graph with topological coherence is output, which is called subgraph connectivity data.

[0136] Subgraph skeleton path extraction is performed based on subgraph connectivity data to obtain subgraph skeleton path data;

[0137] Specifically, the system extracts representative geometric principal axes or paths from the graph and calculates subgraphs. The geometric centroid is used as the starting point for path extraction; a graph search algorithm (such as Dijkstra's algorithm or A) is used. Search algorithms in graphs The system performs a full-graph shortest path search. Among all paths, the path with the longest reach from the centroid is selected as the main skeleton path. This path typically crosses the component's most important structural direction. Based on the main path, the system explores local branches of the path, employing a strategy of prioritizing the main path and expanding secondary branches to gradually grow the path set. The system outputs the skeleton path set for this component. Each path This represents a path consisting of a continuous sequence of points, with values ​​of 1 and 2. This sequence of points represents the topological connection order of the nodes in the graph.

[0138] The connection tension data is obtained by calculating the connection tension based on the subgraph skeleton path data;

[0139] Specifically, for each edge in the skeleton path Calculate the tension index: ,in For the edge The connection tension value on, Points on the skeleton path The three-dimensional coordinate vector, Points on the skeleton path The three-dimensional coordinate vector, For path The average side length (i.e., average distance) of all sides. The tension penalty weight is 0.5. For point and The cosine of the angle between the normal vectors. For point The unit normal vector, For point The unit normal vector is obtained; the tension vector is constructed, and the connection tension data of each pseudo-annotated sub-block is output.

[0140] Specifically, the system can access a predefined library of power facility structure models, where each model contains a standard geometric topology and its labeled tension threshold range (already set based on material and components). It extracts structural feature vectors (such as principal axis direction, size ratio, and normal distribution) from the current pseudo-labeled sub-block; searches the model library for the most similar structural model (using Euclidean distance or embedding space similarity); obtains the defined tension baseline range; and if the current tension exceeds the reference range, triggers the tension correction suggestion module: indicating structural connection anomalies, or automatically adjusting the path connection order to optimize structural stability.

[0141] The pseudo-connectivity is calculated based on the connection tension data to obtain the second pseudo-label feature data.

[0142] Specifically, for each pseudo-labeled sub-block Combine the following structural indicators to construct a fused feature vector. Subgraph connectivity index ,in This indicates the number of weakly connected components within a sub-block. For local average degree, Indicates subgraph connectivity radius coverage; connectivity tension index , representing average tension, maximum tension, and tension variability, respectively; skeleton path index This represents the length of the longest skeleton path within a sub-block. After merging, a fused structural feature vector is constructed. If... >1 and or A significantly larger value indicates that the sub-blocks contain discontinuous joins or false connections; if (For example, set to 0.3) and An abnormally small size indicates a structurally fragmented or incompletely labeled area; if If the tension significantly exceeds the model reference tension threshold, it indicates mis-splicing or extreme structural stretching.

[0143] Specifically, tension feature space is mapped based on the connection tension data to obtain tension feature space data; skeleton graph embedding is performed based on the tension feature space data to obtain skeleton graph embedding data; cross-sub-block relative connectivity similarity is determined on the skeleton graph embedding data to obtain cross-sub-block relative connectivity data; and pseudo-connected graph is constructed based on the cross-sub-block relative connectivity data to obtain second pseudo-annotated feature data.

[0144] For each pseudo-labeled sub-block Construct a tension distribution layer in the skeleton path map, where This represents the location. The connection tension (such as the local tension gradient) is projected onto the local principal axis direction (provided by PCA) to form the principal direction tension trend spectrum. The tension spatial coupling coefficient is defined. , For the tensile strength variance, Let be the connection radius, representing the degree of coupling between tension fluctuations and the connection radius. The value is significantly too high, indicating the presence of pseudo-connections caused by tension distortion. (For sub-blocks) Perform local skeleton extraction to obtain a set of skeleton paths. Construct the skeleton path direction field, and define the angle between the path direction and the main direction. Statistical analysis was performed to calculate the skeleton orientation consistency index. ,in For sub-blocks The skeletal orientation consistency index. For sub-blocks The number of skeleton path segments in the data. This is an identifier for any path segment within the skeleton path segment. For path segment The angle between the direction vector and the local principal axis direction; if (e.g., 0.6) indicates a chaotic path structure and unnatural splicing. Calculate the path structure for any two adjacent sub-blocks. and Euclidean distance between the nearest edge pairs; for each connection tension vector and skeleton orientation features Execution similarity measure: tension similarity ,in For sub-blocks and Tension vector similarity between them The distance is the cosine of the angle between the tension vectors. For sub-blocks The average connection tension vector, For sub-blocks The average connection tension vector. Orientation similarity. ,in For sub-blocks and The similarity of the skeletons in the main direction between them For sub-blocks The principal direction unit vector, For sub-blocks The principal direction unit vector. If and , ( For sub-blocks and The minimum distance between the nearest edge pairs. The distance threshold represents a constant used to determine whether two sub-blocks are sufficiently close. For sub-blocks and Tension vector similarity between them As the tension similarity threshold, For sub-blocks and The similarity of the skeletons in the main direction between them If the threshold for directional consistency is used, then it is determined to be a region with high pseudo-connectivity. Construct a pseudo-connected graph. Nodes are child blocks Edges are constructed based on pseudo-connectivity scores (such as weighted similarity); the graph is partitioned using spectral clustering or Louvain methods to identify structural groups that may be miscut; soft merging is performed on sub-blocks within each group, and pseudo-label consistency labels are updated.

[0145] Optionally, S3 includes:

[0146] S31. Perform component main direction processing on the substation point cloud data based on the point cloud component clustering data to obtain component main direction data;

[0147] Specifically, principal orientation estimation is performed for each component. The system extracts components based on principal component analysis (PCA). The principal axis direction. For The geometric centroid of the component is obtained by averaging the three-dimensional coordinates of all points, denoted as . Calculate the coordinate offset vector of the point set relative to the centroid, and find the mean of the outer product to obtain the covariance matrix. The definition is as follows: ,in Let covariance matrix be the variance matrix. For the number of component points, For components The first in The three-dimensional coordinates of the points For the first A point cloud component (a clustered set of points). for The Middle The three-dimensional coordinates of the points For components The geometric centroid. For the covariance matrix. Eigenvalue decomposition yields three orthogonal eigenvectors. The corresponding eigenvalues ​​represent the spatial extent of the component along each principal direction. The eigenvectors with the largest eigenvalues ​​are selected by sorting the eigenvalues ​​from largest to smallest. As a component The main direction vector. The output is a set of main direction vectors for the components.

[0148] S32. Calculate the attitude offset based on the main direction data of the component to obtain the attitude offset data;

[0149] Specifically, the system uses a preset standard attitude direction as a reference to quantify and analyze the spatial offset between the principal orientation vector of each point cloud component and this reference direction, and constructs a rotation matrix for attitude correction or offset estimation. A standard attitude reference direction is defined. The standard orientation vector is taken from a principal axis direction of the world coordinate system (e.g., the vertical Z+ axis or Y+ axis) to represent the component orientation under ideal conditions. The system calculates the attitude angle. For the k-th component, its principal orientation vector is... The system calculates its relationship with the standard reference direction. The angle between The included angle is calculated by the ratio of the vector dot product to the magnitude, and its mathematical expression is as follows: ,in This represents the attitude offset angle of the current component's main direction relative to the standard direction. It is an inverse cosine function. For the first Components The principal direction vector, The standard attitude reference direction vector is used. The system calculates the rotation axis direction. The system calculates the rotation axis direction based on the vector cross product. Rotate to The required rotation axis direction is defined as the unit vector. ,Right now The system constructs an attitude rotation matrix. The system uses a rotation formula to rotate the current component to the reference attitude direction for any point. Its rotated coordinates The calculation is as follows: ,in The first after attitude rotation The three-dimensional coordinates of the points For components in the original point cloud The The three-dimensional coordinates of the points The attitude offset angle (in radians) represents the angle between the component's principal direction and the reference direction. Let be the unit vector in the direction of the attitude rotation axis, representing the direction from... Rotate to The desired axis. This rotation formula is constructed as a rotation matrix. This allows all component points to be transformed uniformly. Output attitude offset angle and rotation matrix The resulting attitude offset data.

[0150] S33. Perform attitude collinearity residual plot processing based on attitude offset data to obtain attitude collinearity residual plot data;

[0151] Specifically, the system constructs a graph structure with residual weights based on the main orientation information of all identified components. A collinearity residual graph is then constructed. This graph is denoted as... The graph node set V contains all components. Each component is a node; the graph edge set E represents the connection relationship between components, and edges are constructed based on spatial distance and facility topology. For any two components... and The system calculates the distance between their geometric centroids. If one of the following two conditions is met, then connect the edges in the graph. ,like Less than the set spatial connection threshold (unit: meters); Component and These belong to the same electrical facility module, such as substation areas divided according to preset electrical topology rules. The system is based on the principal direction vector of each component. and Calculate the collinear residual angle, which is used as the weight of the edge, as defined below: ,in The collinear residual angle represents a measure of the residual, indicating the angle between the principal directions. This is an inverse cosine function; output the included angle value (in radians). For components The principal direction vector, For components The principal direction vector. Weights The larger the value, the greater the difference in principal directions between the component pairs, the worse the collinearity, and the higher the degree of attitude offset. The output includes an attitude collinearity residual map. The graph structure encodes all component pairs that meet the connectivity conditions and their collinearity residual relationships; the attitude residual matrix ,in For the number of components, Representation Component and The collinear residual angle can be set to 0 or infinity at non-connected edges.

[0152] S34. Based on the attitude collinearity residual map data, perform pseudo-annotation expectation optimization on the substation cloud data to obtain attitude calibration data.

[0153] Specifically, the system analyzes the attitude consistency information among components in the collinear residual graph, adjusts the attitude directions of components with abnormal deviations to make them closer to the set of directions with high attitude consistency in adjacent structures, thereby optimizing the attitude of the overall pseudo-label. The system identifies component nodes with large attitude deviations. For any node in the attitude collinear residual graph... (Representing a component), the system uses collinear residual weights with its adjacent components. The system identifies whether the current node exhibits a significant deviation (e.g., the average residual exceeds a set threshold) and uses it as an optimization candidate. The system calculates the desired principal direction for each component requiring optimization. The system selects its set of adjacent nodes and constructs the expected principal direction vector of the component. This direction is traversed by the main direction vector of the adjacent node. The weighted average is obtained, where the weighting coefficients are inversely proportional to the residuals, i.e.: ,in Let be the expected principal direction vector of the component. For vector normalization operations, To be compatible with components Adjacent component order items, To be compatible with components The set of adjacent components (the set of adjacent nodes in the diagram). The attitude collinearity residual weights represent the component's weights. and The difference in angle (in radians) between the principal directions. for The principal direction vector. The system performs attitude correction. Based on the aforementioned desired principal direction... For reference to the target direction, the components The current principal direction vector Perform rotation optimization by updating the original principal direction vector using a weighted fusion strategy, as defined below: ,in For components The current principal direction vector, This is the fusion coefficient, with a value range of [0.5, 0.9]. For components The current principal direction vector, For components The expected principal direction vector is calculated by weighted averaging of the directions of adjacent components. The output is attitude calibration data, including the set of updated principal direction vectors for each component; and the set of spatial geometric coordinates of the overall point cloud components after attitude optimization.

[0154] Optionally, S4 includes:

[0155] S41. Construct a spatial relationship diagram of device components based on the attitude calibration data to obtain component relationship diagram data;

[0156] Specifically, the input is a set of pose-calibrated point cloud components, where each component already has a uniform pose orientation and includes principal axis direction vectors and bounding box boundary information. An undirected graph is then constructed. , where the node set Represents all components; edge set This indicates the possible physical or functional connections between components. For any pair of components... An edge is established if the following conditions are met, indicating that a connection is possible. The nearest point pair between the two components is less than a set adjacency distance threshold, such as 0.3 meters; and the angle between the main axes of the two components is less than a set maximum angle threshold, such as... The system sets prior structural combinations between specific types based on preset parameters. For example, if the component type is "transformer body," then components within 0.5 meters above it that are "bushing" type will be preferentially connected; if the component type is "busbar," then "switch" type components will be preferentially connected in its axial extension direction. These priors can be used to construct a device structure constraint table to restrict the generation conditions of edges in the diagram. Each edge can be assigned a connection score weight to express the strength of the connection probability, defined as follows: ,in For the edge The connection score weight, This is the distance factor weighting coefficient, with a value of 0.4. For components and The closest point-to-distance between them For the first A point cloud component, For the first A point cloud component, This is the adjacency distance threshold, used to determine whether two objects are adjacent. This is the weighting coefficient for the attitude angle factor, with a value of 0.3. The main axis direction consistency index (cosine of the included angle) is used; the larger the value, the more collinear the axis. For components and The angle between the principal axis direction vectors, This represents the structural prior factor weighting coefficient, with a value of 0.3. For components and The structural connection prior probability For component order items, This refers to the order of adjacent components. For components and Based on the pre-defined connection prior probabilities in the equipment knowledge base (e.g., the probability of a busbar connecting to a circuit breaker is 0.9, the probability of a transformer body connecting to a bushing is 0.95, and the probability of a bushing connecting to a conductor is 0.85), a component relationship diagram is constructed. ,in The edge weight matrix represents a quantitative expression of the connection possibilities between components.

[0157] S42. Reconstruct the structure based on the component relationship diagram data to obtain the component structure data;

[0158] Specifically, based on the edge information in the component relationship graph, a connection path is constructed between each pair of connected components, generating structural segments with practical engineering significance, and ensuring that the connection conforms to physical assembly logic and electrical engineering standards. Each edge in the graph represents a component. With components There are connections between components. Based on their spatial relationship and type information, the connection path type and geometry are determined. When the shortest distance between two components is less than the set connection threshold (e.g., 0.2 meters) and the main axis directions are basically the same (angle less than 10 degrees), the shortest connection segment between the two components can be used directly as the connection path. If the connection between the components requires electrical or mechanical plugging (e.g., transformer and bushing, busbar and switch), a standard connection geometry is generated according to the main direction, such as a cylindrical pipe or a rectangular busbar segment. If there is a significant deviation in the main axis directions of the two components (angle greater than the set threshold, such as 30 degrees), a bent connection component is introduced, such as an "elbow", "insulating bracket" or "adapter", to construct a reasonable connection path. Depending on the connection type, different interpolation methods are selected for structural modeling. For tubular connection components, Bezier curves or circular arc curves are used to interpolate the coordinate points at both ends, generating a smooth, continuous curvature pipe path. For plate connection components, triangular facet interpolation is performed using the bounding boxes of the two end components as constraints to generate a gradually transitioning plate. For structural voxelization, the above interpolated geometry is used for 3D mesh modeling to generate voxel or mesh models of the intermediate structural segments. Each connection segment includes the indices of the starting and ending components; the spatial length of the connection path (in meters); the direction vector of the connection segment (representing the spatial orientation between the starting and ending points); the label of the connection component type used (such as "straight segment", "elbow", "busbar", "pipe", "insulator", etc.); and the material properties or color code of the connection segment (encoded through preset parameters). The output is a set of component structural data, where each element represents a component connection segment and includes a description of its geometric shape and connection semantic information.

[0159] S43. Connect and assemble the component structure data to obtain the three-dimensional model of the substation.

[0160] Specifically, the input data includes a set of component point clouds and a set of connecting structure segments. The system maps all the point clouds and structure voxels to the same spatial coordinate system (such as the world coordinate system or the engineering coordinate system) to ensure that each part maintains the correct positional relationship in space. If some components have local coordinate systems, attitude transformation can be performed before fusion. For each connecting segment, based on its recorded start and end component indices, direction vectors, and connection length, it is embedded into the corresponding component. and Spatial connections between components are achieved between their boundaries. In the splicing area between components and connecting structural segments, point cloud redundancy or boundary overlap occurs. The system performs voxel overlap determination, dividing the splicing area into a unified voxel mesh. If connecting structural points and component points fall into the same voxel unit and the spatial distance is less than a set threshold (e.g., 2cm), the point closer to the center is retained, and redundant points are removed. Alternatively, stitching optimization can be performed. For discontinuous edge areas, geometric stitching strategies, such as sparse point set interpolation or boundary line topology reconstruction, can be implemented to improve visual and analytical continuity. The system performs 3D mesh reconstruction operations on the point cloud fusion results, including Poisson reconstruction and Marching Cubes algorithm. The generated mesh structure supports subsequent material mapping, mechanical simulation, or BIM system integration. The output 3D model data can be exported in various formats depending on the application scenario, including point cloud model formats: .pcd, .ply; mesh model formats: .obj, .stl; visualization / engineering structure formats: .gltf (suitable for web display), .ifc (suitable for BIM structured management).

[0161] 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.

[0162] 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 substation point cloud data processing method for constructing a three-dimensional model, characterized in that, The method comprises: S1, obtaining substation point cloud data; performing convolution filtering on the substation point cloud data to obtain convolution filtering data; S2, performing structure change feature extraction on the convolution filtering data and the substation point cloud data to obtain structure change feature data; performing component attention screening on the substation point cloud data according to the structure change feature data to obtain point cloud component pseudo-labeling data; performing feature extraction on the point cloud component pseudo-labeling data to obtain pseudo-labeling feature data; performing density clustering on the pseudo-labeling feature data to obtain feature clustering data; performing structure feature extraction on the feature clustering data to obtain structure feature data; and performing equipment component structure nested graph construction on the feature clustering data according to the structure feature data to obtain point cloud component clustering data; S3, performing pose calibration on the substation point cloud data according to the point cloud component clustering data to obtain pose calibration data; S4, performing topological structure modeling according to the pose calibration data to obtain a three-dimensional model of the substation; The pseudo-labeling feature data comprises first pseudo-labeling feature data and second pseudo-labeling feature data, and the feature extraction comprises: performing geometric feature extraction on the point cloud component pseudo-labeling data to obtain the first pseudo-labeling feature data; performing local connectivity feature extraction on the point cloud component pseudo-labeling data to obtain the second pseudo-labeling feature data; The geometric feature extraction comprises: performing local covariance calculation on the point cloud component pseudo-labeling data to obtain local covariance data; performing curvature spectrum extraction on the local covariance data to obtain curvature spectrum data; performing multi-scale point cloud structure fluctuation analysis on the point cloud component pseudo-labeling data according to the curvature spectrum data to obtain structure fluctuation spectrum data; performing principal axis direction density projection on the structure fluctuation spectrum data to obtain the first pseudo-labeling feature data.

2. The method of claim 1, wherein, The convolution filtering comprises: performing normal vector covariance tensor graph calculation on the substation point cloud data to obtain tensor graph data; performing frequency domain point cloud response transformation on the tensor graph data to obtain point cloud frequency spectrum data; performing multi-core convolution processing on the point cloud frequency spectrum data to obtain the convolution filtering data, wherein the multi-core convolution processing comprises spherical kernel convolution processing, planar kernel convolution processing, and cylindrical kernel convolution processing.

3. The method of claim 1, wherein, The structure change feature extraction comprises: performing structure change extraction on the convolution filtering data and the substation point cloud data to obtain structure change data; performing structure jump index calculation on the structure change data to obtain structure jump index data; performing structure fluctuation projection analysis on the substation point cloud data according to the structure change data and the structure jump index data to obtain the structure change feature data.

4. The method of claim 1, wherein, The component attention screening comprises: performing point-level attention calculation on the structure change feature data to obtain point-level attention data; performing prior mask enhancement processing on the substation point cloud data according to the point-level attention data to obtain point cloud enhanced data; performing local aggregation reinforcement on the point cloud enhanced data to obtain local aggregation data; performing pseudo-label generation on the local aggregation data to obtain the point cloud component pseudo-labeling data.

5. The method of claim 1, wherein, The local connectivity feature extraction comprises: performing neighbor connection graph processing on the point cloud component pseudo-labeling data to obtain neighbor connection graph data; According to the local subgraph connectivity data, local subgraph connectivity analysis is performed according to the near neighbor connection graph data, and local subgraph connectivity data is obtained; According to the local subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained; According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph skeleton path data, connection tension calculation is performed, and connection tension data is obtained; According to the connection tension data, pseudo connectivity is calculated, and second pseudo labeled feature data is obtained.

6. The method of claim 1, wherein S3 It includes: According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained; According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained.

7. The method of claim 1, wherein S4 It includes: According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained; According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained. It includes: According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained; According to the subgraph connectivity data, subgraph skeleton path extraction is performed, and subgraph skeleton path data is obtained; According to the subgraph connectivity data, pseudo connectivity is supplemented, and subgraph connectivity data is obtained.

Images