Distribution tower material recognition method based on spatial stereo segmentation and point cloud group structure

Abstract

The application discloses a material recognition method for distribution network towers based on spatial three-dimensional segmentation and point cloud group structure, which comprises the following steps: collecting three-dimensional point cloud data of distribution network towers and environment and pre-processing, and constructing a material feature database; calculating the spatial distance between points and constructing an undirected point cloud graph; selecting seed points according to the density, distribution or rules in the graph, and corresponding material categories of various seed points; performing random walk clustering with the seed points as the starting points, and completing spatial segmentation according to the principle of maximum probability; constructing an improved three-dimensional Gaussian mixture model, training parameters, and generating fine-grained clustering labels; combining the clustering labels and the feature database, and completing type calibration, quantity statistics and spatial positioning; and outputting a structured result for material inventory, inspection and management. The application realizes high-precision recognition and spatial positioning of the materials of the distribution network towers by fusing random walk clustering and the improved three-dimensional Gaussian mixture model.

Description

Technical Field

[0001] This invention relates to the field of intelligent power distribution network material identification technology, and in particular to a method for identifying power distribution network tower materials based on spatial three-dimensional segmentation and point cloud group structure. Background Technology

[0002] With the continuous improvement of the power industry's intelligence and automation levels, the efficient management and accurate identification of distribution network poles and their associated materials have become crucial aspects of smart distribution network construction. Traditional methods for identifying, counting, and inspecting distribution network pole materials mainly rely on manual inspections or two-dimensional image-based methods. These methods are limited by issues such as high human subjectivity, poor environmental adaptability, and insufficient spatial information acquisition, making it difficult to achieve large-scale, accurate, and efficient asset management. The rapid development of 3D point cloud technology provides a new approach for spatial modeling and intelligent identification of power facilities. 3D point cloud data acquired through devices such as lidar and structured light can fully reflect the spatial distribution and geometric shape of poles and surrounding materials, becoming an important data foundation for intelligent distribution network material identification.

[0003] However, existing 3D point cloud material recognition methods generally suffer from several technical bottlenecks. On the one hand, in complex power distribution network scenarios, where materials are diverse, densely distributed, and overlap, traditional methods based on single clustering, simple segmentation, or rule matching struggle to effectively separate various types of materials, easily leading to misclassification, omissions, or blurred boundaries. On the other hand, existing algorithms such as Gaussian mixture models and clustering segmentation mostly focus only on point cloud spatial coordinates, failing to fully integrate spatial topology, local attribute differences, and multi-scale hierarchical information, resulting in insufficient recognition capabilities for heterogeneous materials, fine-grained structures, and spatial hierarchies. Most methods lack adaptability and innovation in model structure, making it difficult to balance global and local, macroscopic and microscopic feature modeling, thus limiting recognition accuracy and engineering application.

[0004] Therefore, how to provide a material identification method for power distribution towers based on spatial three-dimensional segmentation and point cloud group structure is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for identifying materials on power distribution network towers based on spatial 3D segmentation and point cloud group structures. This invention comprehensively utilizes technologies such as 3D point cloud data processing, spatial segmentation, random walk clustering, and a multi-structure improved 3D Gaussian mixture model. It details how to achieve automated, high-precision identification and spatial positioning of power distribution network towers and their associated materials through spatial modeling, fine-grained clustering, and intelligent comparison of standardized point cloud data. This method boasts advantages such as full utilization of spatial information, high identification accuracy, strong adaptability to complex scenarios, and a high degree of automation, significantly improving the efficiency and intelligence level of power distribution network asset inventory, intelligent inspection, and operation and maintenance management.

[0006] The method for material identification of power distribution towers based on spatial three-dimensional segmentation and point cloud group structure according to embodiments of the present invention includes:

[0007] Collect three-dimensional point cloud data of power distribution towers and their surrounding environment, preprocess the three-dimensional point cloud data to obtain a standardized point cloud dataset, and construct a power distribution network material feature database.

[0008] Based on a normalized point cloud dataset, the spatial distance between each point and other points in its neighborhood is calculated to construct an undirected point cloud graph. Each point is treated as a node in the undirected point cloud graph, and the neighborhood relationship is used as the edge. The edge weight is defined using the spatial distance.

[0009] Based on the spatial density, geometric distribution, or manually set rules of points in the undirected point cloud, a set of seed points is selected as the starting nodes for random walk clustering, and each type of sub-point corresponds to a different material category.

[0010] Starting from the seed point, perform random walk clustering on the undirected point cloud map, calculate the random walk probability from each unclassified point to the seed points of each category, and assign each point to the category with the highest probability value to obtain the spatial segmentation result;

[0011] For each spatial segmentation region in the spatial segmentation result, a corresponding improved 3D Gaussian mixture model is established, the parameters of the 3D Gaussian mixture model are trained, and fine-grained clustering labels are output.

[0012] Based on fine-grained clustering labels and combined with the distribution network material feature database, the point cloud data of each spatial segmentation region is classified, counted, and spatially located to obtain material category, quantity, and spatial location information.

[0013] The system outputs information on material categories, quantities, and spatial locations in a structured format for inventory, inspection, and asset management of power distribution tower materials.

[0014] Optionally, the three-dimensional point cloud data specifically includes the spatial coordinate information, reflection intensity information, and color information of the power distribution tower and surrounding materials in three-dimensional space.

[0015] Optionally, the preprocessing of the 3D point cloud data specifically includes noise removal, downsampling, point cloud registration, and coordinate normalization.

[0016] Optionally, the construction of the distribution network material feature database specifically refers to combining manual annotation results and historical material information to extract geometric features, analyze spatial distribution, and organize attribute information for different distribution network materials. This involves summarizing and classifying the typical three-dimensional shapes, size parameters, spatial structural features, and related attributes of various materials to form a distribution network material feature database that can be used for material identification and comparison.

[0017] Optionally, based on the normalized point cloud dataset, the spatial distance between each point and other points in its neighborhood is calculated to construct an undirected point cloud graph. Each point is treated as a node in the undirected point cloud graph, and the neighborhood relationship is used as the edge. The edge weight is defined using the spatial distance, including:

[0018] For each point in the normalized point cloud dataset, the set of neighboring points for each point is determined using the radius neighborhood algorithm;

[0019] For each point and other points in its neighborhood, calculate the Euclidean distance between the three-dimensional coordinates. The Euclidean distance is the square root of the sum of the squares of the distance differences along the three-dimensional coordinate axes, reflecting the actual distance between the two points in three-dimensional space.

[0020] Each point in the normalized point cloud dataset is used as a node in the undirected point cloud graph, and each pair of points that satisfy the neighborhood relationship is used as an edge in the undirected point cloud graph. Based on the calculated Euclidean distance, an edge weight is assigned to each edge by setting a distance decay rule. The closer the edge is, the larger the weight, and the farther the edge is, the smaller the weight. This completes the construction of all nodes, edges and edge weights in the undirected point cloud graph.

[0021] Optionally, based on the spatial density, geometric distribution, or manually defined rules of the points in the undirected point cloud, a set of seed points is selected as the starting nodes for random walk clustering. The various sub-points correspond to different material categories, including:

[0022] Based on the undirected point cloud graph, the spatial density is calculated for each node, where the spatial density is the number of neighboring nodes contained within a preset radius centered on the node.

[0023] All nodes are sorted from high to low according to spatial density, and nodes with spatial density greater than a set threshold are initially selected as candidate nodes.

[0024] Spatial uniformity screening is performed on all candidate nodes;

[0025] Based on manually set rules or historical recognition experience, seed points selected for spatial uniformity are labeled with corresponding material categories by combining spatial location and point cloud features. Each material category contains at least one seed point, and the seed points corresponding to each material category are used as the starting nodes for random walk clustering.

[0026] Optionally, obtaining the spatial segmentation result includes:

[0027] Starting with the seed points corresponding to each material category, we initialize category labels for each node on the undirected point cloud map. Initially, only the seed points are assigned category labels, and other nodes are in an unclassified state.

[0028] On an undirected point cloud graph, for each unclassified node, the random walk probability to each category seed point is calculated based on the transition probability defined by the edge weight.

[0029] For each unclassified node, the category label is set to the category with the highest probability of random walk, thus completing the initial spatial segmentation of the current round;

[0030] The preliminary spatial segmentation results are evaluated by self-supervised feedback. Based on the local consistency of the node's category, neighborhood confidence, and boundary uncertainty, nodes with ambiguous boundaries and uncertain classifications are automatically identified and temporarily marked as pseudo-label seed points.

[0031] The pseudo-label seed points are added to the new seed set. Combined with the original seed points, random walk clustering is performed again on the undirected point cloud graph. The random walk probability to each category seed point is recalculated for all unclassified nodes and pseudo-label seed points, and the category labels are updated.

[0032] After the initial spatial segmentation results are corrected by self-supervised feedback, pseudo-labeled seed points are added to a new seed set and combined with the original seed points. Random walk clustering is then performed again on the undirected point cloud map. The class assignment process is iterated until the class labels of all nodes remain stable or the preset maximum number of iterations is reached. Finally, the spatial segmentation results are output.

[0033] Optionally, the output fine-grained clustering labels include:

[0034] For each spatial segmentation region in the spatial segmentation result, point cloud data within the spatial segmentation region is extracted. The point cloud data includes three-dimensional spatial coordinate information and serves as input data for the improved three-dimensional Gaussian mixture model.

[0035] The input point cloud data is fed into the topological Gaussian module, the heterogeneous covariance module, and the hierarchical fusion module, respectively. The topological Gaussian module includes a topological feature extraction unit and a feature fusion unit. The heterogeneous covariance module includes a covariance type discrimination unit and a covariance parameter generation unit. The hierarchical fusion module includes a multi-scale initialization unit and a hierarchical probability fusion unit. The three modules serve as substructures of the improved 3D Gaussian mixture model.

[0036] In the topological Gaussian module, the topological feature extraction unit takes point cloud data and radius neighborhood connection information as input, calculates the local topological features of each point, and the feature fusion unit fuses the local topological features with the spatial coordinate information of the points to output a topological fusion feature set.

[0037] In the heterogeneous covariance module, the covariance type discrimination unit takes the topology fusion feature set as input and combines the spatial characteristics of the point cloud distribution to determine the covariance type corresponding to each Gaussian component. The covariance parameter generation unit automatically generates covariance parameters based on the covariance type and related point set, and outputs the covariance type and parameter set of each Gaussian component.

[0038] In the hierarchical fusion module, the multi-scale initialization unit takes the covariance type, parameter set, and point cloud data as input to initialize Gaussian components at different spatial scales. The hierarchical probability fusion unit fuses the component parameters and attribution probabilities at each spatial scale and outputs a multi-scale Gaussian component parameter set.

[0039] Using the multi-scale Gaussian component parameter set as initial parameters, combined with topological fusion features and covariance parameters, within the overall framework of the three-dimensional Gaussian mixture model, the expectation-maximization algorithm is used to iteratively optimize the mean, covariance, and weight of each component, and outputs the final converged Gaussian component parameter set.

[0040] Based on the final converged set of Gaussian component parameters, the probability of belonging to each point cloud data point in the spatial segmentation region under all Gaussian components is calculated. Fine-grained clustering labels are assigned according to the principle of maximum probability. The fine-grained clustering labels and corresponding Gaussian component parameters of each point in the spatial segmentation region are output for fine differentiation and category labeling of materials.

[0041] Optionally, obtaining the material category, quantity, and spatial location information includes:

[0042] Based on the output fine-grained clustering labels, the point cloud data within the spatial segmentation region are grouped according to the clustering labels, with each group corresponding to a Gaussian component or fine-grained category.

[0043] For each fine-grained category, extract the spatial coordinate features and distribution features of the point cloud, as well as the standard material geometric features, spatial structure features, and attribute information from the power distribution network material feature database, and construct the point cloud feature vector and the database feature vector;

[0044] Calculate the similarity between the point cloud feature vector and the database feature vector. The similarity is based on cosine similarity, and the one with the highest similarity is selected as the basis for category matching.

[0045] Each fine-grained category is matched with the most similar material category in the database to identify the material type of the point cloud category. The number of point cloud data and the centroid coordinates or boundary range in three-dimensional space are recorded as the basis for quantity statistics and spatial positioning.

[0046] By summarizing the material type, quantity, and spatial location information of all fine-grained categories, a structured result of the material type, quantity, and spatial location information within the spatially segmented region is obtained.

[0047] Optionally, the step of outputting material category, quantity, and spatial location information in a structured result format for inventory, inspection, and asset management of power distribution tower materials specifically refers to organizing the name, corresponding quantity, three-dimensional spatial coordinates or spatial range, category label, and related attribute information of each type of material that has been automatically identified into a standardized data format, and outputting it as a table, database record, or interface data for direct use by automatic inventory, inspection scheduling, and asset management.

[0048] The beneficial effects of this invention are:

[0049] This invention improves the automated identification and management of power distribution towers and their materials in complex scenarios by integrating spatial 3D segmentation, random walk clustering, and a structurally innovative 3D Gaussian mixture model. Compared to existing technologies that rely solely on single spatial clustering or traditional Gaussian mixture models, this invention not only fully explores the spatial topological relationships and multi-scale structural features of 3D point clouds, but also achieves collaborative identification of local details and overall distribution of materials through modular innovation. This effectively solves problems such as difficulty in distinguishing dense, overlapping, and heterogeneous materials, blurred boundaries, and insufficient spatial hierarchy.

[0050] The method of this invention enables high-precision, structured output of the category, quantity, and spatial location of power distribution towers and their associated materials without extensive manual intervention, thereby improving the automation and accuracy of asset inventory and inspection.

[0051] This invention has good adaptability and engineering scalability, and can provide a strong data foundation and technical support for the operation and maintenance, asset management and intelligent decision-making of smart distribution networks. It has significant practical application value and promotion prospects. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 The flowchart shows the distribution network tower material identification method based on spatial three-dimensional segmentation and point cloud group structure proposed in this invention.

[0054] Figure 2 This is a schematic diagram of the three-dimensional Gaussian mixture model of the material identification method for power distribution towers based on spatial three-dimensional segmentation and point cloud group structure proposed in this invention.

[0055] Figure 3 This is a schematic diagram of the point cloud segmentation and recognition results of the power distribution tower material identification method based on spatial three-dimensional segmentation and point cloud group structure proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1 , Figure 2 and Figure 3 A method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure includes:

[0058] Collect three-dimensional point cloud data of power distribution towers and their surrounding environment, preprocess the three-dimensional point cloud data to obtain a standardized point cloud dataset, and construct a power distribution network material feature database.

[0059] Based on a normalized point cloud dataset, the spatial distance between each point and other points in its neighborhood is calculated to construct an undirected point cloud graph. Each point is treated as a node in the undirected point cloud graph, and the neighborhood relationship is used as the edge. The edge weight is defined using the spatial distance.

[0060] Based on the spatial density, geometric distribution, or manually set rules of points in the undirected point cloud, a set of seed points is selected as the starting nodes for random walk clustering, and each type of sub-point corresponds to a different material category.

[0061] Starting from the seed point, perform random walk clustering on the undirected point cloud map, calculate the random walk probability from each unclassified point to the seed points of each category, and assign each point to the category with the highest probability value to obtain the spatial segmentation result;

[0062] For each spatial segmentation region in the spatial segmentation result, a corresponding improved 3D Gaussian mixture model is established, the parameters of the 3D Gaussian mixture model are trained, and fine-grained clustering labels are output.

[0063] Based on fine-grained clustering labels and combined with the distribution network material feature database, the point cloud data of each spatial segmentation region is classified, counted, and spatially located to obtain material category, quantity, and spatial location information.

[0064] The system outputs information on material categories, quantities, and spatial locations in a structured format for inventory, inspection, and asset management of power distribution tower materials.

[0065] In this embodiment, the three-dimensional point cloud data specifically includes the spatial coordinate information, reflection intensity information, and color information of the power distribution tower and surrounding materials in three-dimensional space.

[0066] In this embodiment, the preprocessing of the three-dimensional point cloud data specifically includes noise removal, downsampling, point cloud registration, and coordinate normalization.

[0067] In this embodiment, the construction of the distribution network material feature database specifically refers to combining manual annotation results and historical material information to extract geometric features, analyze spatial distribution, and organize attribute information for different distribution network materials. The typical three-dimensional shapes, size parameters, spatial structural features, and related attributes of various materials are summarized and classified to form a distribution network material feature database that can be used for material identification and comparison.

[0068] In this embodiment, the step of calculating the spatial distance between each point and other points in its neighborhood based on a normalized point cloud dataset, constructing an undirected point cloud graph, treating each point as a node in the undirected point cloud graph, using neighborhood relationships as edges, and defining edge weights using spatial distance includes:

[0069] For each point in the normalized point cloud dataset, the set of neighboring points for each point is determined using the radius neighborhood algorithm, specifically as follows:

[0070] Set a predetermined neighborhood radius threshold for each point;

[0071] For each point, iterate through all other points in the normalized point cloud dataset and calculate the Euclidean distance between the point and the other points in three-dimensional space.

[0072] All points whose Euclidean distance to a given point is not greater than the neighborhood radius threshold are added to the set of the given point's neighborhood points.

[0073] For each point and other points in its neighborhood, calculate the Euclidean distance between the three-dimensional coordinates. The Euclidean distance is the square root of the sum of the squares of the distance differences along the three-dimensional coordinate axes, reflecting the actual distance between the two points in three-dimensional space.

[0074] Each point in the normalized point cloud dataset is used as a node in the undirected point cloud graph, and each pair of points that satisfy the neighborhood relationship is used as an edge in the undirected point cloud graph. Based on the calculated Euclidean distance, an edge weight is assigned to each edge by setting a distance decay rule. The closer the edge is, the larger the weight, and the farther the edge is, the smaller the weight. This completes the construction of all nodes, edges and edge weights in the undirected point cloud graph.

[0075] This invention constructs an undirected point cloud graph using a radius-based neighborhood algorithm based on a standardized point cloud dataset. This not only achieves refined modeling of spatial structures but also improves the accuracy of material segmentation and clustering. Unlike traditional point cloud connection methods that rely solely on global distance or fixed adjacency relationships, this method dynamically filters closely related point pairs in the actual space by setting a predetermined neighborhood radius for each point. This ensures that neighborhood connections reflect the true physical structure of objects while filtering out irrelevant and redundant connections, enhancing the spatial expressive power of the point cloud graph. By employing Euclidean distance and distance decay rules to assign weights to edges, the strength of spatial relationships is quantified, allowing clustering and subsequent segmentation to fully utilize local geometric features. This invention not only introduces spatial constraints in the selection of nodes and edges but also provides a solid data foundation and a controllable structural framework for spatial segmentation, clustering, and material identification through a flexible edge weight allocation mechanism. This improves the accuracy and robustness of material boundary identification and spatial topology modeling in complex environments, effectively supporting the automated intelligent identification of multiple categories of materials in point cloud scenarios.

[0076] In this embodiment, a set of seed points is selected as the starting nodes for random walk clustering based on the spatial density, geometric distribution, or manually set rules of the points in the undirected point cloud. The various sub-points correspond to different material categories, including:

[0077] Based on the undirected point cloud graph, the spatial density is calculated for each node, where the spatial density is the number of neighboring nodes contained within a preset radius centered on the node.

[0078] All nodes are sorted from high to low according to spatial density, and nodes with spatial density greater than a set threshold are initially selected as candidate nodes.

[0079] Spatial uniformity screening is performed on all candidate nodes, specifically as follows:

[0080] The candidate nodes are traversed sequentially. The first candidate node is used as the seed point. Then, only candidate nodes whose three-dimensional Euclidean distance from the selected seed point is greater than the preset minimum distance are retained until all candidate nodes are traversed. This ensures that the selected seed points are evenly distributed in space and are spaced apart from each other by more than the minimum distance.

[0081] Based on manually set rules or historical identification experience, seed points selected for spatial uniformity are labeled with their corresponding material categories by combining their spatial location and point cloud features. Each material category contains at least one seed point. The seed points corresponding to each material category are used as the starting nodes for random walk clustering. Manually set rules refer to pre-defining the typical distribution location, shape characteristics, or size thresholds of different material categories in space according to the actual distribution specifications or industry standards of distribution network tower materials. For example, insulators are usually located at both ends of crossarms, and guy wires are commonly found near the tower base. Experience helps determine the category to which the seed point belongs. Historical identification experience refers to the knowledge of the correlation between material categories and spatial features and point cloud attributes accumulated based on the manual annotation results of a large amount of previous point cloud data or verified identification cases. For example, the statistical laws of the spatial density, geometric features, and reflection intensity features of a certain type of hardware are known. Combined with the distribution patterns of similar materials in historical data, the newly identified seed points are labeled with their categories.

[0082] In this embodiment, obtaining the spatial segmentation result includes:

[0083] Starting with the seed points corresponding to each material category, we initialize category labels for each node on the undirected point cloud map. Initially, only the seed points are assigned category labels, and other nodes are in an unclassified state.

[0084] On an undirected point cloud graph, for each unclassified node, the random walk probability to each category seed point is calculated based on the transition probability defined by the edge weight. The random walk probability is the probability distribution of reaching each category seed point from the unclassified node along the path with the largest edge weight. Specifically, the calculation of the random walk probability to each category seed point is as follows:

[0085] Based on the edge weights between nodes in the undirected point cloud graph, the edge weights of each pair of adjacent nodes are normalized to obtain the transition probability matrix between nodes. The probability of each node transitioning to a neighboring node is equal to the ratio of the edge weight to the sum of all outgoing edge weights.

[0086] Assign labels to all seed points according to their corresponding categories, initialize the random walk probability of seed points to 1, and initialize the random walk probability of other unclassified nodes to 0;

[0087] For each unclassified node, based on the transition probability matrix and the probability distribution of known nodes, the cumulative probability of each node to each category seed point is continuously updated through iteration until convergence or the preset number of iterations is reached, and finally the probability distribution of each unclassified node to each category seed point is obtained.

[0088] For each unclassified node, the category label is set to the category with the highest probability of random walk, thus completing the initial spatial segmentation of the current round;

[0089] The preliminary spatial segmentation results are evaluated using a self-supervised feedback mechanism. Based on the local consistency of a node's category, neighborhood confidence, and boundary uncertainty, nodes with ambiguous boundaries and uncertain classifications are automatically identified and temporarily labeled as pseudo-label seed points. The local consistency of a node's category refers to the degree to which a node belongs to the same category as most nodes in its spatial neighborhood. If a node's category label is consistent with that of most nodes in its neighborhood, the node is considered to have high local consistency; conversely, if there are many nodes of different categories in the neighborhood, the node's local consistency is low. The neighborhood confidence refers to the proportion of nodes whose category labels are consistent with those of nodes in their neighborhood, used to measure the reliability of the node's category determination. The boundary uncertainty refers to the fact that the probability distribution of a node's category is relatively close or changes frequently in the spatial segmentation, making it difficult to clearly determine its category, reflecting the degree of uncertainty of a node being located in the boundary region of different categories.

[0090] The pseudo-label seed points are added to the new seed set. Combined with the original seed points, random walk clustering is performed again on the undirected point cloud graph. The random walk probability to each category seed point is recalculated for all unclassified nodes and pseudo-label seed points, and the category labels are updated.

[0091] After the initial spatial segmentation results are corrected by self-supervised feedback, pseudo-labeled seed points are added to a new seed set and combined with the original seed points. Random walk clustering is then performed again on the undirected point cloud map. The class assignment process is iterated until the class labels of all nodes remain stable or the preset maximum number of iterations is reached. Finally, the spatial segmentation results are output.

[0092] This invention introduces a random walk-based spatial segmentation and self-supervised feedback mechanism into undirected point cloud maps, achieving automatic segmentation and boundary optimization of material categories, thus improving the accuracy and intelligence of point cloud segmentation. Unlike traditional spatial segmentation methods, this method not only utilizes edge weight normalization and transition probability matrices to efficiently simulate information transmission between points, enabling category labels to propagate orderly within the spatial structure, but also enhances the modeling capability for complex spatial relationships through the cumulative iteration of random walk probabilities. Employing a triple criterion of local consistency, neighborhood confidence, and boundary uncertainty, it automatically filters nodes with ambiguous classification boundaries and uncertain category determinations, forming pseudo-label seed points and dynamically expanding the clustering starting point, achieving self-correction and refinement of segmentation labels. The self-supervised feedback correction mechanism effectively solves the misclassification and omission problems that easily occur in traditional methods in scenarios with dense multi-material, overlapping boundaries, and weak categories, improving the sensitivity and robustness of the segmentation results to boundaries and details. The overall process possesses good adaptability and automation capabilities, providing more reliable basic data support for material identification and spatial analysis, and has significant engineering practical value and technological innovation significance.

[0093] In this embodiment, the output fine-grained clustering labels include:

[0094] For each spatial segmentation region in the spatial segmentation result, point cloud data within the spatial segmentation region is extracted. The point cloud data includes three-dimensional spatial coordinate information and serves as input data for the improved three-dimensional Gaussian mixture model.

[0095] The input point cloud data is fed into the topological Gaussian module, the heterogeneous covariance module, and the hierarchical fusion module, respectively. The topological Gaussian module includes a topological feature extraction unit and a feature fusion unit. The heterogeneous covariance module includes a covariance type discrimination unit and a covariance parameter generation unit. The hierarchical fusion module includes a multi-scale initialization unit and a hierarchical probability fusion unit. The three modules serve as substructures of the improved 3D Gaussian mixture model.

[0096] In the topological Gaussian module, the topological feature extraction unit takes point cloud data and radius neighborhood connectivity information as input to calculate the local topological features of each point. The feature fusion unit fuses the local topological features with the spatial coordinate information of the points and outputs a topological fusion feature set. The calculation of the local topological features of each point specifically involves:

[0097] A local subgraph is constructed based on the radius neighborhood of each point. First-order and higher-order spectral features are calculated using the graph Laplacian operator to capture the local structural complexity and connectivity of points in space, thereby obtaining a basic topological description of the point cloud.

[0098] Based on the spectral characteristics, principal component analysis is used to obtain the main direction of the neighborhood. By making small perturbations to the positions of neighborhood points, the changes in the main direction and structural characteristics before and after the perturbation are recorded to measure the stability and sensitivity of the local topology under spatial changes.

[0099] Based on the extraction of the spectral features, main direction and perturbation stability features of the multi-order neighborhood graph, the distribution of each type of point in the neighborhood of each point is statistically analyzed at multiple neighborhood scales, and the information entropy is calculated to obtain the diversity and redundancy of the spatial structure around the node at different scales. This reflects the uniformity of the distribution of the node at different spatial levels and the local complexity, and enhances the ability to identify complex structures and boundary regions.

[0100] In the heterogeneous covariance module, the covariance type discrimination unit takes the topological fusion feature set as input and, combined with the spatial characteristics of the point cloud distribution, determines the covariance type corresponding to each Gaussian component. The covariance parameter generation unit automatically generates covariance parameters based on the covariance type and related point sets, outputting the covariance type and parameter set of each Gaussian component. The determination of the covariance type corresponding to each Gaussian component based on the spatial characteristics of the point cloud distribution refers to selecting the most suitable covariance structure type to describe the component distribution characteristics based on the spatial distribution morphology, density, and directional characteristics of the point cloud within each component. This includes diagonal, full matrix, and sparse types. The automatic generation of covariance parameters by the covariance parameter generation unit based on the covariance type and related point sets is specifically as follows:

[0101] For the identified covariance type, determine the specific parameter elements that need to be calculated. For the diagonal type, only the variance of each coordinate axis is calculated. For the full matrix type, the covariance between all coordinate axes is calculated. For the sparse type, the preset key parameters are calculated.

[0102] For the relevant point set assigned to the Gaussian component, calculate the distance distribution from all points to the component center, and then extract the required spatial feature values ​​to provide basic data for the calculation of covariance parameters;

[0103] Based on the covariance type and spatial characteristic values, the covariance parameters are automatically generated using the corresponding statistical methods, and the generated parameter set is output as a spatial distribution description of the Gaussian components.

[0104] In the hierarchical fusion module, the multi-scale initialization unit takes the covariance type, parameter set, and point cloud data as input to initialize Gaussian components at different spatial scales. The hierarchical probability fusion unit fuses the component parameters and assignment probabilities at each spatial scale and outputs a multi-scale Gaussian component parameter set. The hierarchical probability fusion unit fuses the component parameters and assignment probabilities at each spatial scale by weighting and integrating the Gaussian component parameters obtained at different spatial scales with their respective point assignment probabilities to generate a Gaussian component parameter set and assignment probability distribution that comprehensively reflects the multi-scale structural characteristics.

[0105] Using the multi-scale Gaussian component parameter set as initial parameters, combined with topological fusion features and covariance parameters, within the overall framework of the three-dimensional Gaussian mixture model, the expectation-maximization algorithm is used to iteratively optimize the mean, covariance, and weight of each component, and outputs the final converged Gaussian component parameter set.

[0106] Based on the final converged set of Gaussian component parameters, for each point cloud data point within the spatial segmentation region, the probability of belonging under all Gaussian components is calculated. Fine-grained clustering labels are assigned according to the maximum probability principle, and the fine-grained clustering labels and corresponding Gaussian component parameters for each point within the spatial segmentation region are output. This is used for fine-grained material differentiation and category labeling. Specifically, the calculation of the probability of belonging under all Gaussian components is as follows:

[0107] Based on the spatial distance between each point and the center of each Gaussian component, and combined with the spatial distribution parameters of the components, the relative similarity of a point to each Gaussian component is evaluated.

[0108] The relative similarity of each Gaussian component is weighted by the mixture weight to obtain the weighted assignment score of each point to each Gaussian component.

[0109] The weighted assignment scores of all Gaussian components are normalized to obtain the final assignment probability distribution of each point under all Gaussian components, and the fine-grained clustering label of the point is determined with the highest probability.

[0110] This invention addresses point cloud data within spatially segmented regions by integrating a topological Gaussian module, a heterogeneous covariance module, and a hierarchical fusion module, achieving innovative structural and functional breakthroughs in a 3D Gaussian mixture model. By introducing multi-dimensional topological features such as local graph spectrum, principal direction perturbation, and multi-scale information entropy, it achieves a refined depiction of the complex spatial structure of point clouds, enabling the model to fully perceive local connectivity, stability, and spatial diversity, effectively improving its ability to distinguish material boundaries, fine-grained structures, and overlapping scenes of multiple categories. The heterogeneous covariance module adaptively selects the covariance type based on the point cloud distribution pattern and automatically generates parameters based on different spatial features, enhancing the model's adaptability to spatial anisotropy and complex material shapes, avoiding the problems of traditional Gaussian models being susceptible to single structures and noise sensitivity. The hierarchical fusion module, through weighted fusion of multi-scale Gaussian components and probabilities, enables the model to simultaneously capture the local details and overall pattern of point cloud data, enhancing its clustering robustness in multi-level structures. The overall process supports full-parameter self-learning and structural self-adaptation, which improves the automation, intelligence and engineering scalability of material clustering and category labeling, and provides higher accuracy, stronger generalization ability and better data foundation for intelligent material identification and asset management in complex power distribution network environments.

[0111] In this embodiment, obtaining the material category, quantity, and spatial location information includes:

[0112] Based on the output fine-grained clustering labels, the point cloud data within the spatial segmentation region are grouped according to the clustering labels, with each group corresponding to a Gaussian component or fine-grained category.

[0113] For each fine-grained category, extract the spatial coordinate features and distribution features of the point cloud, as well as the standard material geometric features, spatial structure features, and attribute information from the power distribution network material feature database, and construct the point cloud feature vector and the database feature vector;

[0114] Calculate the similarity between the point cloud feature vector and the database feature vector. The similarity is based on cosine similarity, and the one with the highest similarity is selected as the basis for category matching.

[0115] Each fine-grained category is matched with the most similar material category in the database to identify the material type of the point cloud category. The number of point cloud data and the centroid coordinates or boundary range in three-dimensional space are recorded as the basis for quantity statistics and spatial positioning.

[0116] By summarizing the material type, quantity, and spatial location information of all fine-grained categories, a structured result of the material type, quantity, and spatial location information within the spatially segmented region is obtained.

[0117] In this embodiment, the step of outputting material category, quantity, and spatial location information in a structured result form for inventory, inspection, and asset management of power distribution tower materials specifically refers to organizing the name, corresponding quantity, three-dimensional spatial coordinates or spatial range, category label, and related attribute information of each type of material that has been automatically identified into a standardized data format, and outputting it as a table, database record, or interface data for direct use by automatic inventory, inspection scheduling, and asset management.

[0118] Example 1:

[0119] To verify the feasibility of this invention in practice, it was applied to a power distribution network tower inspection and automatic material identification project in a suburban area. This region has a dense distribution of power transmission and distribution lines, totaling 24 towers, and the surrounding environment is complex, including green belts, construction sites, and piles of construction debris. Previously, manual inspection methods not only required significant manpower and resources but were also susceptible to obstruction and stacking issues when counting and locating tower materials, leading to inaccurate statistics, material omissions, and unclear spatial distribution descriptions. This was especially true at night or in inclement weather conditions, where manual efficiency and accuracy were even lower. Addressing these industry challenges, this project is the first to apply this invention's power distribution network tower material identification method based on spatial three-dimensional segmentation and point cloud group structure on a large scale in this region, comprehensively improving the automation and intelligence level of power distribution network material management.

[0120] In practical applications, the on-site inspection team uses a drone system equipped with high-precision lidar to collect comprehensive 3D point cloud data of each tower and its surrounding 5-meter radius. Each tower generates approximately 430,000 to 530,000 point cloud data points per inspection, all uploaded to the cloud in real time. The system first performs noise removal, downsampling, registration, and coordinate unification on the raw point cloud data to generate a standardized point cloud dataset. Then, based on spatial neighborhood relationships, an undirected point cloud graph is constructed, the point cloud density distribution is automatically analyzed, representative seed points are extracted, and a multi-round self-supervised feedback mechanism guides a random walk clustering algorithm to complete preliminary spatial segmentation. For each segmented region, the system automatically constructs an innovative 3D Gaussian mixture model that integrates topological features, adaptive covariance, and multi-scale probability to perform fine-grained clustering of the point cloud materials. This model is then matched with standard templates in the power distribution network material feature database to automatically determine the material category, quantity, and spatial location.

[0121] Table 1. Comparison of the method of the present invention and manual inspection in material identification of power distribution towers.

[0122] Tower number Total amount of materials (labor) Total number of materials (in this invention) Recognition accuracy (human, %) Recognition accuracy (in this invention, %) Recognition time (manual, min) Recognition time (in this invention, min) Spatial positioning error (human, m) Spatial positioning error (in this invention, m) Tower number A01 188 187 98.4 99.5 24.3 2.7 0.42 0.13 A01 A02 170 171 97.6 99.4 23.1 2.3 0.48 0.15 A02 A03 193 193 100.0 100.0 25.7 2.6 0.36 0.11 A03 A06 182 179 94.0 98.3 21.9 2.2 0.53 0.17 A06 A08 183 182 97.3 99.5 22.8 2.5 0.46 0.12 A08 A09 176 174 92.6 97.9 24.8 2.4 0.55 0.18 A09 A11 189 189 100.0 100.0 23.5 2.6 0.38 0.12 A11 A12 187 187 100.0 100.0 23.3 2.3 0.41 0.10 A12 average 183.5 183.0 96.8 99.3 23.7 2.5 0.45 0.14 average

[0123] Based on the comparison data in Table 1, it can be clearly seen that the "material identification method for power distribution towers based on spatial three-dimensional segmentation and point cloud group structure" proposed in this invention is significantly better than the traditional manual inspection method in several key performance indicators.

[0124] Regarding recognition accuracy, the average accuracy of manual recognition is 96.8%, while the method of this invention reaches 99.3%, significantly improving recognition precision. For some towers, such as A03, A11, and A12, the recognition accuracy is 100% under both methods. However, for most other towers, such as A01, A06, and A09, this invention significantly outperforms manual methods, indicating that the algorithm possesses more stable material recognition capabilities and effectively reduces missed detections and misjudgments.

[0125] The difference in recognition time is even more striking: the average manual recognition time is 23.7 minutes, while the method of this invention takes only 2.5 minutes, representing an improvement in recognition efficiency of approximately 9.5 times. This advantage is particularly crucial for large-scale inspection tasks, as it can significantly reduce time costs and improve on-site operation and maintenance efficiency.

[0126] Regarding spatial positioning error, the average error of manual methods is 0.45 meters, while the present invention controls the error to within 0.14 meters, demonstrating a significant improvement in accuracy. The spatial recognition capability of this invention is particularly valuable in scenarios requiring high-precision installation, re-inspection, or inventory management.

[0127] This invention not only has high recognition and positioning accuracy, but also reduces labor and time costs. It is especially suitable for complex tower structures, environments with dense material loading, and multi-tower joint detection tasks, and can provide more reliable, fast and efficient technical support for intelligent operation and maintenance of power distribution networks.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure, characterized in that, include: Collect three-dimensional point cloud data of power distribution towers and their surrounding environment, preprocess the three-dimensional point cloud data to obtain a standardized point cloud dataset, and construct a power distribution network material feature database. Based on a normalized point cloud dataset, the spatial distance between each point and other points in its neighborhood is calculated to construct an undirected point cloud graph. Each point is treated as a node in the undirected point cloud graph, and the neighborhood relationship is used as the edge. The edge weight is defined using the spatial distance. Based on the spatial density, geometric distribution, or manually set rules of points in the undirected point cloud, a set of seed points is selected as the starting nodes for random walk clustering, and each type of sub-point corresponds to a different material category. Starting from the seed point, perform random walk clustering on the undirected point cloud map, calculate the random walk probability from each unclassified point to the seed points of each category, and assign each point to the category with the highest probability value to obtain the spatial segmentation result; For each spatial segmentation region in the spatial segmentation result, a corresponding improved 3D Gaussian mixture model is established, the parameters of the 3D Gaussian mixture model are trained, and fine-grained clustering labels are output. Based on fine-grained clustering labels and combined with the distribution network material feature database, the point cloud data of each spatial segmentation region is classified, counted, and spatially located to obtain material category, quantity, and spatial location information. The material category, quantity, and spatial location information are output in a structured result format for inventory, inspection, and asset management of power distribution tower materials. The output fine-grained clustering labels include: For each spatial segmentation region in the spatial segmentation result, point cloud data within the spatial segmentation region is extracted. The point cloud data includes three-dimensional spatial coordinate information and serves as input data for the improved three-dimensional Gaussian mixture model. The input point cloud data is fed into the topological Gaussian module, the heterogeneous covariance module, and the hierarchical fusion module, respectively. The topological Gaussian module includes a topological feature extraction unit and a feature fusion unit. The heterogeneous covariance module includes a covariance type discrimination unit and a covariance parameter generation unit. The hierarchical fusion module includes a multi-scale initialization unit and a hierarchical probability fusion unit. The three modules serve as substructures of the improved 3D Gaussian mixture model. In the topological Gaussian module, the topological feature extraction unit takes point cloud data and radius neighborhood connectivity information as input to calculate the local topological features of each point. The feature fusion unit fuses the local topological features with the spatial coordinate information of the points and outputs a topological fusion feature set. The calculation of the local topological features of each point specifically involves: A local subgraph is constructed based on the radius neighborhood of each point. First-order and higher-order spectral features are calculated using the graph Laplacian operator to capture the local structural complexity and connectivity of points in space, thereby obtaining a basic topological description of the point cloud. Based on the spectral characteristics, principal component analysis is used to obtain the main direction of the neighborhood. By making small perturbations to the positions of neighborhood points, the changes in the main direction and structural characteristics before and after the perturbation are recorded to measure the stability and sensitivity of the local topology under spatial changes. Based on the extraction of the spectral features, main direction and perturbation stability features of the multi-order neighborhood graph, the distribution of each type of point in the neighborhood of each point is statistically analyzed at multiple neighborhood scales, and the information entropy is calculated to obtain the diversity and redundancy of the spatial structure around the node at different scales. In the heterogeneous covariance module, the covariance type discrimination unit takes the topology fusion feature set as input and determines the covariance type corresponding to each Gaussian component based on the spatial characteristics of the point cloud distribution. The covariance parameter generation unit automatically generates covariance parameters based on the covariance type and related point set, and outputs the covariance type and parameter set of each Gaussian component. The determination of the covariance type corresponding to each Gaussian component based on the spatial characteristics of the point cloud distribution refers to selecting the most suitable covariance structure type to describe the distribution characteristics of the component based on the spatial distribution shape, density and directional characteristics of the point cloud in each component, including diagonal type, full matrix type and sparse type. In the hierarchical fusion module, the multi-scale initialization unit takes the covariance type, parameter set, and point cloud data as input to initialize Gaussian components at different spatial scales. The hierarchical probability fusion unit fuses the component parameters and assignment probabilities at each spatial scale and outputs a multi-scale Gaussian component parameter set. The hierarchical probability fusion unit fuses the component parameters and assignment probabilities at each spatial scale by weighting and integrating the Gaussian component parameters obtained at different spatial scales with their respective point assignment probabilities to generate a Gaussian component parameter set and assignment probability distribution that comprehensively reflects the multi-scale structural characteristics. Using the multi-scale Gaussian component parameter set as initial parameters, combined with topological fusion features and covariance parameters, within the overall framework of the three-dimensional Gaussian mixture model, the expectation-maximization algorithm is used to iteratively optimize the mean, covariance, and weight of each component, and outputs the final converged Gaussian component parameter set. Based on the final converged set of Gaussian component parameters, the probability of belonging to each point cloud data point in the spatial segmentation region under all Gaussian components is calculated. Fine-grained clustering labels are assigned according to the principle of maximum probability. The fine-grained clustering labels and corresponding Gaussian component parameters of each point in the spatial segmentation region are output for fine differentiation and category labeling of materials.

2. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The three-dimensional point cloud data specifically includes the spatial coordinates, reflection intensity, and color information of the power distribution tower and surrounding materials in three-dimensional space.

3. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The preprocessing of the 3D point cloud data specifically includes noise removal, downsampling, point cloud registration, and coordinate normalization.

4. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The construction of the distribution network material feature database specifically refers to combining manual annotation results and historical material information to extract geometric features, analyze spatial distribution, and organize attribute information for different distribution network materials. The typical three-dimensional shapes, size parameters, spatial structural features, and related attributes of various materials are summarized and classified to form a distribution network material feature database that can be used for material identification and comparison.

5. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, Based on the normalized point cloud dataset, the spatial distance between each point and other points in its neighborhood is calculated to construct an undirected point cloud graph. Each point is treated as a node in the undirected point cloud graph, and neighborhood relationships are used as edges. Edge weights are defined using spatial distance, including: For each point in the normalized point cloud dataset, the set of neighboring points for each point is determined using the radius neighborhood algorithm; For each point and other points in its neighborhood, calculate the Euclidean distance between the three-dimensional coordinates. The Euclidean distance is the square root of the sum of the squares of the distance differences along the three-dimensional coordinate axes, reflecting the actual distance between the two points in three-dimensional space. Each point in the normalized point cloud dataset is used as a node in the undirected point cloud graph, and each pair of points that satisfy the neighborhood relationship is used as an edge in the undirected point cloud graph. Based on the calculated Euclidean distance, an edge weight is assigned to each edge by setting a distance decay rule. The closer the edge is, the larger the weight, and the farther the edge is, the smaller the weight. This completes the construction of all nodes, edges and edge weights in the undirected point cloud graph.

6. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The process involves selecting a set of seed points as the starting nodes for random walk clustering based on the spatial density, geometric distribution, or manually defined rules of points in the undirected point cloud. Each seed point corresponds to a different material category, including: Based on the undirected point cloud graph, the spatial density is calculated for each node, where the spatial density is the number of neighboring nodes contained within a preset radius centered on the node. All nodes are sorted from high to low according to spatial density, and nodes with spatial density greater than a set threshold are initially selected as candidate nodes. Spatial uniformity screening is performed on all candidate nodes; Based on manually set rules or historical recognition experience, seed points selected for spatial uniformity are labeled with corresponding material categories by combining spatial location and point cloud features. Each material category contains at least one seed point, and the seed points corresponding to each material category are used as the starting nodes for random walk clustering.

7. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The obtained spatial segmentation result includes: Starting with the seed points corresponding to each material category, we initialize category labels for each node on the undirected point cloud map. Initially, only the seed points are assigned category labels, and other nodes are in an unclassified state. On an undirected point cloud graph, for each unclassified node, the random walk probability to each category seed point is calculated based on the transition probability defined by the edge weight. For each unclassified node, the category label is set to the category with the highest probability of random walk, thus completing the initial spatial segmentation of the current round; The preliminary spatial segmentation results are evaluated by self-supervised feedback. Based on the local consistency of the node's category, neighborhood confidence, and boundary uncertainty, nodes with ambiguous boundaries and uncertain classifications are automatically identified and temporarily marked as pseudo-label seed points. The pseudo-label seed points are added to the new seed set. Combined with the original seed points, random walk clustering is performed again on the undirected point cloud graph. The random walk probability to each category seed point is recalculated for all unclassified nodes and pseudo-label seed points, and the category labels are updated. After the initial spatial segmentation results are corrected by self-supervised feedback, pseudo-labeled seed points are added to a new seed set and combined with the original seed points. Random walk clustering is then performed again on the undirected point cloud map. The class assignment process is iterated until the class labels of all nodes remain stable or the preset maximum number of iterations is reached. Finally, the spatial segmentation results are output.

8. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The obtained material category, quantity, and spatial location information include: Based on the output fine-grained clustering labels, the point cloud data within the spatial segmentation region are grouped according to the clustering labels, with each group corresponding to a Gaussian component or fine-grained category. For each fine-grained category, extract the spatial coordinate features and distribution features of the point cloud, as well as the standard material geometric features, spatial structure features, and attribute information from the power distribution network material feature database, and construct the point cloud feature vector and the database feature vector; Calculate the similarity between the point cloud feature vector and the database feature vector. The similarity is based on cosine similarity, and the one with the highest similarity is selected as the basis for category matching. Each fine-grained category is matched with the most similar material category in the database to identify the material type of the point cloud category. The number of point cloud data and the centroid coordinates or boundary range in three-dimensional space are recorded as the basis for quantity statistics and spatial positioning. By summarizing the material type, quantity, and spatial location information of all fine-grained categories, a structured result of the material type, quantity, and spatial location information within the spatially segmented region is obtained.

9. The method for identifying materials in power distribution network towers based on spatial three-dimensional segmentation and point cloud group structure according to claim 1, characterized in that, The process of outputting material category, quantity, and spatial location information in a structured format for inventory, inspection, and asset management of power distribution tower materials specifically refers to organizing the name, corresponding quantity, three-dimensional spatial coordinates or spatial range, category label, and related attribute information of each type of material that has been automatically identified into a standardized data format, and outputting it as a table, database record, or interface data for direct use by automatic inventory, inspection scheduling, and asset management.

Images