A Smart Modeling Method Based on Point Cloud Data

Abstract

The present application relates to the technical field of point cloud processing, and more particularly to an intelligent modeling method based on point cloud data. The method comprises the following steps: scanning the inside of a substation to obtain visual field overlap map data; constructing a point cloud density distribution field for the visual field overlap map data to obtain point cloud density distribution field data; extracting a blocked area from the point cloud density distribution field data to obtain blocked area data; calculating the structure connectivity of the point cloud density distribution field data based on the blocked area data to obtain blocked area connectivity data; mapping the hollow residual of the blocked area connectivity data to obtain hollow residual data; performing graph neural network inference on the hollow residual data to obtain category inference data; fusing the physical rules based on the category inference data to obtain point cloud completion data; and constructing a point cloud model to obtain the point cloud model. The present application improves the spatial coverage integrity and blocked area recognition accuracy of point cloud data in a blocked environment, thereby improving the accuracy of point cloud modeling.

Description

Technical Field

[0001] This invention relates to the field of point cloud processing technology, and in particular to an intelligent modeling method based on point cloud data. Background Technology

[0002] With the development of laser scanning and 3D sensing technologies, point cloud data has been widely applied in fields such as industrial scene modeling and digital equipment management. In substation scenarios within the power industry, due to the dense distribution and diverse structures of equipment, conventional techniques often use scanners fixed at a single location / path, scanning the scene only from a single direction or limited angles, or the scanning viewpoint covers only a limited number of times during the scan. In existing technologies, when faced with gaps in point cloud data acquisition, missing areas are often filled in using rule-based template matching or simple point cloud density-based methods, lacking in-depth analysis of occluded areas. This results in poor performance in subsequent point cloud model construction, easily leading to problems such as incorrect completion or misclassification of equipment categories. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes an intelligent modeling method based on point cloud data, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides an intelligent modeling method based on point cloud data, the method comprising:

[0005] S1. Perform high-density multi-angle scanning of the substation interior to obtain view overlap map data; construct a point cloud density distribution field from the view overlap map data to obtain point cloud density distribution field data.

[0006] S2. Extract occlusion areas from the point cloud density distribution field data to obtain occlusion area data; calculate the structural connectivity of the point cloud density distribution field data based on the occlusion area data to obtain occlusion area connected data; perform hole residual mapping on the occlusion area connected data to obtain hole residual data.

[0007] S3. Perform graph neural network inference on the vacuous residual data to obtain category inference data;

[0008] S4. Perform physical rule fusion based on category reasoning data to obtain point cloud completion data; construct a point cloud model based on the point cloud completion data to obtain the point cloud model.

[0009] This invention constructs a view overlap map through high-density multi-angle scanning and combines it with a density distribution field to achieve fine-grained modeling of the complex internal space of substations, overcoming the local data sparsity problem caused by conventional single-view scanning. Through occlusion area extraction and structural connectivity calculation, it can accurately distinguish void areas caused by spatial structural occlusion and scanning blind spots, improving the accuracy of void identification. The use of void residual mapping to construct a local residual feature map helps reveal the potential geometric trends of hidden void areas, enhancing the targeting of model completion. In the graph neural network inference stage, category-specific inference is achieved by combining residual map features, avoiding the category confusion problem in traditional completion. Through physical rule fusion of the completion process, based on equipment process specifications and spatial logic constraints, a highly consistent completion structure is dynamically generated, ensuring the consistency of the generated model in geometric accuracy and electrical safety specifications.

[0010] Optionally, the high-density multi-angle scanning includes:

[0011] The scanner is controlled to acquire scanning data from multiple angles using a preset scanning posture and scanning movement path;

[0012] The scanned data is calibrated to obtain calibration data;

[0013] The view overlap relationship is calculated from the calibration data to obtain view overlap map data.

[0014] This invention controls the scanner to perform high-density multi-angle acquisition according to a preset scanning posture and movement path. This invention can effectively overcome the blind spot problem caused by the dense equipment, complex structure and limited space inside the substation, and significantly improve the spatial coverage of the scan data. By calibrating the posture of the acquired scan data, the problem of coordinate error accumulation and posture drift in the multi-view stitching process is solved, ensuring the spatial consistency between data from different scanning rounds. By calculating the view overlap relationship, view overlap map data is generated, so that the scanning redundancy and occlusion risk of each spatial region are explicitly expressed.

[0015] Optionally, the point cloud density distribution field construction includes:

[0016] Spatial raster data is obtained by dividing the view overlap map data into spatial raster data;

[0017] Point cloud mapping is performed on spatial raster data based on view overlap map data to obtain point cloud mapping data;

[0018] The point cloud mapping data is obtained by performing view overlap weighting correction on the point cloud mapping data;

[0019] The density gradient field is calculated based on the point cloud mapping data to obtain the point cloud density distribution field data.

[0020] This invention establishes a unified spatial indexing system by dividing the view overlap map data into spatial grids, effectively supporting the localized processing of large-scale, high-density point cloud data and avoiding storage redundancy and computational bottlenecks in traditional global processing. Point cloud data is mapped according to the spatial grid, forming a one-to-one correspondence between spatial locations and point cloud data, improving the spatial distribution representation of point cloud density. Through view overlap weighted correction, multi-angle redundant scanning information is fully integrated, dynamically balancing the uneven point density caused by differences in scanning perspectives and the presence of occlusions, enhancing the spatial consistency and density representation of point cloud data. Through density gradient field calculation, the density change trends of different spatial regions are accurately extracted, significantly improving the accuracy of occlusion area extraction and hole region identification.

[0021] Optionally, the extraction of the occlusion area includes:

[0022] Scanning field simulation calculations are performed on point cloud density distribution field data to obtain scanned field data;

[0023] Based on the point cloud density distribution field data and the scanned visible field data, the visible field contrast residual is calculated to obtain the visible field contrast data;

[0024] Topological occlusion chain analysis was performed on the visible field comparison data to obtain local occlusion chain map data;

[0025] The shading potential is calculated based on the local shading chain map data to obtain the shading potential data;

[0026] Cross-temporal occlusion processing is performed based on occlusion potential data to obtain occlusion area data.

[0027] This invention utilizes point cloud density distribution data to perform scanning visibility simulation calculations, enabling the establishment of a theoretically reachable space model for multi-view scanners. This accurately characterizes the theoretical visibility of various areas within the complex environment of a substation, significantly enhancing the initial identification capability of obstructed areas. Based on residual comparison analysis between actual observation data and the theoretical visibility field, it effectively distinguishes abnormal void areas caused by equipment obstruction, dead zones, or data gaps, improving the accuracy of obstructed area extraction. Through topological obstruction chain analysis, it extracts the causal chain relationship of obstruction in local space, gaining a deeper understanding of the obstruction causes in structures such as multi-layered equipment stacking and cross-obstruction, providing more physically meaningful obstruction features for the completion process. In the obstruction potential calculation stage, by fusing multi-dimensional features such as chain tension, foldback characteristics, and penetration, it refines the credibility judgment of obstructed areas, reducing the risk of false or missed completion. Combined with cross-temporal obstruction processing, it dynamically identifies and eliminates instantaneous false obstructions during the scanning process, improving the stability and reliability of obstructed area data.

[0028] Optionally, the shielding potential calculation includes:

[0029] The tension distribution of the masking chain is calculated based on the local masking chain map data, and the tension distribution data of the masking chain is obtained.

[0030] The shielding chain reversal coefficient is calculated based on the shielding chain tension distribution data to obtain the shielding chain reversal coefficient data;

[0031] The temporal fluctuation residuals are calculated based on the local shading chain map data to obtain the temporal fluctuation residual data;

[0032] The field of view penetration coefficient is calculated based on the temporal fluctuation residual data to obtain the field of view penetration coefficient data;

[0033] Based on the shading chain return coefficient data and the field of view penetration coefficient data, the potential segmentation coherence mapping of the local shading chain map data is performed to obtain the shading potential data.

[0034] This invention calculates the tension distribution of the occlusion chain in local occlusion chain maps, quantifying the spatial tension characteristics between different occlusion chain segments. This accurately identifies high-stress occlusion chain segments formed by equipment structures in complex spaces, improving the accuracy of occlusion causal relationship modeling. By calculating the occlusion chain return coefficient, it identifies multi-layered return phenomena in the occlusion path, enhancing the ability to identify highly complex occlusion areas and reducing potential calculation errors caused by misjudgment of the occlusion chain. Combined with scanning time-series fluctuation residual analysis, it dynamically captures the stability characteristics of the occlusion chain's evolution over time, significantly improving the ability to distinguish between false and true occlusion. The calculation of the field-of-view penetration coefficient reflects the scan permeability of local spatial regions, improving the overall spatial connectivity expression of the occlusion chain. Through potential segment coherence mapping, it achieves dynamic optimization of the potential continuity and spatial consistency between occlusion chain segments, ensuring that the generated occlusion potential data possesses good local coherence and global consistency.

[0035] Optionally, the structural connectivity calculation includes:

[0036] Heteroscedastic density features are extracted from the point cloud density distribution field data based on the occlusion area data to obtain heteroscedastic density feature data.

[0037] Restricted adjacency graphs are constructed based on heteroscedasticity density feature data to obtain restricted adjacency graph data;

[0038] Guided connectivity propagation is performed on restricted adjacency graph data to obtain occluded region connectivity data.

[0039] This invention extracts heteroscedastic density features from the point cloud density distribution field corresponding to the occluded area data, effectively characterizing the density variation characteristics of different spatial localities within the occluded area, highlighting the density differences between device edges, overlapping occluded areas, and background areas, and enhancing the discernibility of occluded area boundaries. By constructing a constrained adjacency graph using heteroscedastic density features, it fully integrates spatial adjacency relationships and density gradient characteristics, dynamically suppressing high-noise, low-reliability adjacency relationships, enhancing the boundary constraint effect in connectivity calculation, and avoiding false connectivity. Furthermore, guided connectivity propagation is employed, under the physical topological constraints of the occlusion chain, to enhance connectivity within the occluded area along a reliable path, accurately separating effective occlusion structures in locally complex occlusion areas, and significantly improving the spatial consistency and physical rationality of connectivity representation within the occluded area.

[0040] Optionally, the cavity residual mapping includes:

[0041] The occlusion region connectivity potential field is generated from the occlusion region connectivity data to obtain the occlusion region connectivity potential field data;

[0042] Local density missing kernel mapping is performed on the connected potential field data of the shading region to obtain local density missing kernel data;

[0043] Hollow residual path backtracking is performed based on locally missing density kernel data. Hollow backtracking data is obtained.

[0044] A global void residual field is constructed based on void backtracking data to obtain void residual data.

[0045] This invention generates a connected potential field for occluded regions by processing connected data of occluded regions. This quantifies the potential energy distribution of different connected regions in the spatial dimension, enhances the structural guidance characteristics within the occluded regions, and effectively solves the problem of lack of spatial gradient expression in connected regions in traditional methods. Based on the connected potential field data, local density missing kernels are extracted, which can keenly capture areas with abnormally low point cloud density within the occluded region, accurately locate potential hole center regions, and reduce the risk of hole misjudgment caused by global density equalization. Combined with hole residual path backtracking, the hole formation chain is backtracked layer by layer along the high-confidence connected paths in the connected potential field of the occluded region, improving the accuracy of hole boundary extraction under occluded structures. Through the construction of a global hole residual field, local density missing information and path backtracking data are effectively integrated to form a residual field with good spatial continuity.

[0046] Optionally, S3 includes:

[0047] The residual plot structure is constructed from the vacant residual data to obtain the residual plot structure data;

[0048] Injecting prior category labels into the residual graph structure data yields graph-labeled data.

[0049] Category-specific graph convolutional inference is performed on the graph-labeled data to obtain category inference data.

[0050] This invention constructs a residual graph structure from the void residual data, enabling the establishment of a high-dimensional feature relationship network within the void region at the spatial topology level. This fully explores the intrinsic correlation between the residual features of each local void, enhancing the graph structure's ability to express complex void morphologies and avoiding the category ambiguity problem encountered by traditional point cloud-based local feature processing methods in large-scale occlusion areas. By injecting prior category label information, semantic guidance can be given to the nodes in the graph structure based on substation equipment process specifications and existing category knowledge, giving the graph convolutional inference process stronger category discrimination capabilities and improving the recognition accuracy between different equipment categories. Through category-specific graph convolutional inference, convolutional kernels and convolutional paths specifically designed for different equipment categories are used, effectively enhancing the targeting and robustness of category feature extraction and avoiding the problem of insufficient adaptability of traditional single graph convolutional models to multi-category occlusion areas.

[0051] Optionally, S4 includes:

[0052] The physical constraint rule set is dynamically loaded based on the category reasoning data to obtain the physical constraint data.

[0053] Based on the physical constraint data, rule-driven completion candidate generation is performed to obtain completion candidate data;

[0054] The point cloud completion data is obtained by performing multi-rule collaborative optimization and completion based on the candidate completion data.

[0055] This invention dynamically loads the corresponding category's physical constraint rule set based on category inference data, enabling automatic matching of refined process specifications and spatial structural constraints for different equipment categories (such as busbars, disconnectors, and surge arresters) in substations. This avoids the problem of decreased completion accuracy in heterogeneous equipment scenarios using traditional fixed rule models. By using physical constraint data to drive the completion candidate generation process, it can accurately generate completion candidate segments with physical rationality and spatial consistency based on multi-dimensional rules such as structural geometric dimensions, electrical safety distances, and layout symmetry, significantly reducing the risk of morphological distortion caused by a lack of process constraints. Through multi-rule collaborative optimization completion, it integrates the coupling relationships between multiple physical constraints and uses a multi-objective optimization method to balance the consistency of equipment functional logic and spatial coherence, effectively improving the performance of point cloud completion data in terms of geometric accuracy, structural consistency, and electrical specification compliance.

[0056] Optionally, this application also provides an intelligent modeling system based on point cloud data, used to execute the intelligent modeling method based on point cloud data as described above, the intelligent modeling system based on point cloud data comprising:

[0057] The multi-view high-density point cloud field construction module is used to perform high-density multi-angle scanning of the interior of the substation to obtain view overlap map data; and to construct the point cloud density distribution field from the view overlap map data to obtain point cloud density distribution field data.

[0058] The occlusion region connectivity residual mapping module is used to extract occlusion regions from point cloud density distribution field data to obtain occlusion region data; calculate the structural connectivity of point cloud density distribution field data based on the occlusion region data to obtain occlusion region connectivity data; and perform hole residual mapping on the occlusion region connectivity data to obtain hole residual data.

[0059] The category-specific graph inference module is used to perform graph neural network inference on the punctuous residual data to obtain category inference data.

[0060] The physical rule-driven completion modeling module is used to fuse physical rules based on category reasoning data to obtain point cloud completion data; and to construct a point cloud model based on the point cloud completion data to obtain the point cloud model.

[0061] The purpose of this invention is to effectively improve the point cloud coverage integrity of complex spatial structures within substations by combining high-density multi-angle scanning with view overlap map construction, significantly reducing data sparsity areas caused by scanning blind spots and stacked obstructions; by constructing a point cloud density distribution field, it accurately expresses the changes in point cloud density gradient within a spatial region, enhancing the ability to identify potential obstruction areas; in the obstruction area extraction and structural connectivity calculation stages, topological occlusion chain analysis and connectivity potential field optimization are adopted to achieve accurate modeling of complex equipment superimposed occlusion relationships, improving the spatial coherence and physical rationality of cavity residual area identification; through graph neural network inference driven by residual map structure, combined with category prior and specific convolution strategy, the accuracy of equipment category discrimination is enhanced, avoiding the multi-category misclassification problem in traditional methods; in the physical rule fusion stage, based on dynamic rule set loading and multi-rule collaborative optimization, it is ensured that the completion process fully meets equipment process standards and spatial structure logic, effectively improving the geometric consistency and electrical safety specification compliance of the completion model. 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 an intelligent modeling method based on point cloud data according to one embodiment is shown.

[0064] Figure 2 A flowchart illustrating the steps of a high-density multi-angle scanning method according to an embodiment is shown.

[0065] Figure 3A flowchart illustrating the steps of an occlusion region extraction method according to an embodiment is shown.

[0066] Figure 4 A flowchart illustrating the steps of a category-specific graph reasoning method according to an embodiment is shown.

[0067] Figure 5 A flowchart illustrating the steps of a physical rule-driven completion modeling method according to one 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 an intelligent modeling method based on point cloud data, the method comprising:

[0073] S1. Perform high-density multi-angle scanning of the substation interior to obtain view overlap map data; construct a point cloud density distribution field from the view overlap map data to obtain point cloud density distribution field data.

[0074] In one embodiment, a laser scanner (robotic arm type or trajectory-guided type) is controlled to perform multi-angle high-density scanning operations according to a preset set of scanning posture parameters, including horizontal rotation angle, pitch angle, and roll angle, and in conjunction with scanning path planning (covering equipment passages, structural blind spots, etc. within the substation scenario). During the scanning process, the multi-view redundancy of any single point in space is ensured to be no less than 3. The redundancy is defined as the number of views (with a posture difference greater than 15°) effectively covered by a single spatial grid cell being ≥ 3. Scanning data from each viewpoint is collected, and the spatial position coordinates and posture angles of the scanner are recorded in real time. An initial reference frame (the first frame of scan data) is selected as the origin of the global coordinate system, and the scan data of each frame are sequentially matched between frames in chronological order. For each pair of scan frames, a nearest-point matching strategy is employed to calculate the set of corresponding point pairs between the source and target frames. Based on minimizing the mean square error of the Euclidean distance between matched point pairs, the rotation matrix R and translation vector T are iteratively optimized to solve for the rigid transformation matrix [R|T] from the source to the target frame. The pose relationship is then updated, and joint pose calibration is performed on the multi-view scan data to generate a global point cloud dataset. The global point cloud dataset is divided into fixed spatial grids, with grid cell sizes selected from 1 / 10 to 1 / 5 of the device's detail scale, ranging from 2cm to 5cm. The number of points within each spatial grid cell is statistically analyzed to generate point cloud density distribution field data. Simultaneously, the number of times each grid cell is covered by different scanning views is calculated and standardized into a view coverage redundancy coefficient.

[0075] S2. Extract occlusion areas from the point cloud density distribution field data to obtain occlusion area data; calculate the structural connectivity of the point cloud density distribution field data based on the occlusion area data to obtain occlusion area connected data; perform hole residual mapping on the occlusion area connected data to obtain hole residual data.

[0076] In one embodiment, based on a theoretical visible field simulation model (generated jointly by the scanner's view cone projection parameters and the substation equipment spatial layout model), the theoretical visible field accessibility coefficient for each grid cell in the space is calculated. The actual point cloud density in the point cloud density distribution field is compared with the theoretically expected density to calculate the residual rate. ,in For the first The residual rate of each spatial raster cell This refers to the raster index number of the spatial raster cell in the X direction. This refers to the raster index number of the spatial raster cell in the Y direction. This refers to the grid index number of the spatial grid cell in the Z direction. This represents the number of point clouds actually observed within the grid. This represents the theoretically expected number of point clouds for the raster. This is determined by setting a threshold (e.g., ...). >0.6) Classify the residual graph, marking high residual regions as suspected occlusion areas. Construct a restricted adjacency graph based on the suspected occlusion areas, where graph nodes correspond to high residual raster cells, and edge weights are calculated as follows: ,in For the node To the node The weight of adjacent edges, This is the grid spatial distance weighting coefficient, with a value of 1.0. Indicates grid spatial distance, This is the weighting coefficient for the difference in the number of point clouds, with a value of 0.5. For nodes The corresponding number of actual observed point clouds within the grid. For nodes The corresponding number of actual observed point clouds within the grid. This is the adjacency consistency weight coefficient, with a value of 0.5. This represents the adjacency consistency metric. The adjacency graph extracts occluded connected regions through guided connectivity propagation. Seed nodes are prioritized for propagation, selecting nodes with the highest residual rates (top 10%). Propagation paths primarily extend along device boundaries and occlusion chains to avoid accidental diffusion to unoccluded areas. After extracting the connected occluded regions, gradient backtracking is performed on the residual kernel regions within the occluded areas via residual path backtracking. Combined with the spatial order of the connected chains, a continuous residual field is generated. Path backtracking stops when the path gradient drops below 10% or the cumulative path length exceeds the physical scale of the device (e.g., 1.5m). The grid step size of the generated void residual field is controlled within ≤5cm.

[0077] S3. Perform graph neural network inference on the vacuous residual data to obtain category inference data;

[0078] In one embodiment, based on the void residual field data, the center points of high residual grids are extracted as graph nodes in the graph neural network structure. The selection criterion for the high residual grids is the residual value. ≥0.6. The edge weights of the graph structure form an edge weight matrix. Prior category label data originates from equipment design models (such as CAD / BIM models) or manually labeled known equipment category regions, and is injected into some graph nodes as a known label set. Unlabeled nodes serve as target inference nodes. The graph neural network model uses Graph Convolutional Network (GCN), Graph Attention Network (GAT), or a GIN structure, with 3–5 layers. Each GIN convolutional unit in the structure contains adjacent node feature aggregation, employing a Sum aggregation strategy. ,in for, For the first A multilayer perceptron (a fully connected network with 2-3 layers of ReLU activation). For learnable parameters, For nodes In the Layer feature representation, For nodes The adjacent node index, For nodes The set of adjacent nodes, For nodes In the The graph neural network (Graph NNN) uses layer-level feature representation. BatchNorm regularization is employed to balance the training process; node-level Dropout is used to prevent overfitting; convolutional kernels are designed with class-specific parameter configurations, and differentiated convolutional weight combinations are used for different device categories to enhance class discrimination. Input features include the residual value, spatial coordinates, adjacency degree (node ​​degree), and mean residual value of the neighborhood. During the inference phase, the Graph NNN outputs a class prediction vector for each graph node. After softmax normalization, nodes with a class prediction confidence score ≥ 0.8 are classified as high-confidence class nodes.

[0079] S4. Perform physical rule fusion based on category reasoning data to obtain point cloud completion data; construct a point cloud model based on the point cloud completion data to obtain the point cloud model.

[0080] In one embodiment, based on the category reasoning results, a physical rule set corresponding to the equipment category is dynamically loaded. For example, for busbar equipment, straight structure rules and bracket connection point constraint rules are loaded; for disconnector equipment, rotary shaft motion rules and closing angle restriction rules are loaded. The physical rule set includes the nominal geometric dimension range of the equipment (nominal dimension ± 5% allowable deviation), spatial relationship constraints (such as minimum safety distance ≥ 300mm), electrical safety specification requirements (such as distance to ground ≥ 1.5m), and structural connection specifications specific to the equipment category. A parametric completion template combined with a local void residual field-guided fitting method is used to generate preliminary completion candidate segments. The segment morphology is dynamically adjusted according to the distribution of high missing areas in the residual field and the equipment category rules. During the completion process, a multi-rule collaborative optimization method is used, such as: Objective 1: geometric fitting error ≤ 3mm, defined as the fitting error between the completed segment and the residual high missing area point cloud; Objective 2: physical rule compliance ≥ 95%, i.e., the matching ratio between the morphological parameters of the completed equipment segment and the physical rule items; Objective 3: spatial continuity between the completed segment and the adjacent original point cloud ≥ 0.9. Initialize the completion parameter space (e.g., completion length, attitude angle, connection position offset); generate a candidate solution set within the parameter space; for each candidate solution, calculate the above three target index values; use Pareto sorting to filter the current non-dominated solution set; through iterative search and fine-tuning, converge to the target convergence condition, with the target improvement rate <1% for 5 consecutive iterations, or reaching the maximum number of iterations = 100. For the current optimal solution, use the L-BFGS algorithm for gradient optimization to refine the morphological parameters of the completed fragment and improve the target convergence effect. Integrate the optimized completed fragment into the overall point cloud model, recalculate the global structural consistency index, including device structure connectivity, device alignment accuracy (controlled ≤5m), and spatial consistency score between the completed region and the original point cloud, to form a complete point cloud model.

[0081] Optionally, the high-density multi-angle scanning includes:

[0082] S11. Control the scanner to collect data from multiple angles using a preset scanning posture and scanning movement path;

[0083] In one embodiment, the laser scanner is controlled to perform multi-angle scanning operations according to a preset scanning posture parameter set. The scanning posture parameters include pitch angle, rotation angle, and roll angle. The range and step angle of each angle parameter set are optimized based on the equipment layout and occlusion characteristics. The scanning posture set is layered according to the equipment height levels within the substation, divided into three height levels: 0-1m, 1-2.5m, and 2.5-5m. Corresponding scanning angle combinations are planned for each level to improve the integrity of spatial data coverage at different heights. The scanning path planning adopts a strategy combining spiral detours and horizontal passageways, automatically generated based on the substation CAD / BIM plan and equipment obstacle avoidance model. The moving speed is controlled within ≤0.2m / s, and the scanner's single-frame scanning step distance is controlled to be 1 / 10 of the equipment's characteristic scale, within the range of 2-5cm. During the scanning process, it is ensured that any unit grid in the space is effectively covered by at least 3 different scanning perspectives (perspective difference greater than 15°), with a field-of-view redundancy ≥3. The collected data includes point cloud spatial coordinates, reflection intensity, scanning timestamp, and scanner real-time pose parameters, which include the scanner's spatial position coordinates and attitude angles.

[0084] S12. Perform pose calibration on the scanned data to obtain calibration data;

[0085] In one embodiment, a global optimization registration algorithm is used to perform pose calibration on the scanned data. During the corresponding point matching process, a local covariance matrix is ​​constructed for each matching point pair, and the target is jointly optimized and minimized. ,in The objective function value for global registration error. For the first The covariance matrix corresponding to each point pair For reference point cloud The spatial coordinates of the points Points after transformation The coordinates, soon By transforming the matrix Coordinates mapped to the reference coordinate system For the first point in the cloud to be registered The spatial coordinates of the points This is the pose transformation matrix, which includes rotation and translation. For the first The inverse of the covariance matrix of each point. The registration process uses the global coordinate system corresponding to the BIM design model as the reference coordinate system, mapping all scan frames uniformly to the global reference coordinate system. The initial reference frame is set as the starting position frame of the first round of scanning, serving as the benchmark for global registration. During the registration process, residual optimization is performed on the point cloud of each scan frame and the current global model, with the optimization objective being to control the root mean square error of the residuals to no more than 3mm. To address potential dynamic error sources during scanning (such as micro-vibration of the scanning platform, support offset, platform jitter, etc.), a time window moving average filter is introduced during pose estimation, with the moving window length set to 10-15 frames.

[0086] S13. Calculate the view overlap relationship of the calibration data to obtain view overlap map data.

[0087] In one embodiment, the scanning space is spatially divided according to a global reference coordinate system into fixed-size cubic grid cells, each grid cell having a side length of 5 cm. For each spatial grid cell, the number of times the grid is effectively covered by different scanning angles is counted. Different scanning angles are defined as scanning posture differences (pitch angle or rotation angle) greater than 15° to ensure sufficient discriminative power for these angle differences. The field-of-view redundancy of each grid cell is calculated using the following formula: ,in This represents the view redundancy coefficient of the raster cell. This represents the number of different scanning angles that are effectively covered by the grid cell. This represents the maximum number of times all rasters in the current scanning scene are scanned and covered, representing the expected maximum redundancy. ≥3. The output view overlap map data includes the spatial coordinates of the center point of each grid, the corresponding view redundancy, and a list of scan frame IDs that participated in the scanning of that grid.

[0088] Optionally, the point cloud density distribution field construction includes:

[0089] Spatial raster data is obtained by dividing the view overlap map data into spatial raster data;

[0090] In one embodiment, the overall scanning space of the substation is divided into fixed-size three-dimensional cubic grid units based on a globally unified coordinate system (either the BIM design coordinate system or a registered global scanning coordinate system). The selection rule for the grid side length is determined according to the detailed dimensions of typical equipment, within the range of 2cm to 5cm. =5cm, corresponding to 1 / 10 to 1 / 5 of the typical detail scale of the device. The spatial information of each grid cell includes the coordinates of the grid center point. The grid number and the corresponding spatial bounding box range, wherein the bounding box range is defined as follows: It is used to define the boundary position of the grid in three-dimensional space.

[0091] Point cloud mapping is performed on spatial raster data based on view overlap map data to obtain point cloud mapping data;

[0092] In one embodiment, the globally calibrated point cloud data is mapped to the aforementioned spatial grid cells. The mapping criterion is that the spatial coordinates of each point in the point cloud fall within the bounding box of the corresponding grid cell, thus determining that the point belongs to that grid cell. For each point, its corresponding grid cell is located and the point is added to that grid cell. For each grid cell, the following data is accumulated and statistically analyzed: point cloud count, representing the total number of points belonging to that grid cell; a coverage view ID list, recording which different scanning views (i.e., scan frame IDs or view sequence numbers) covered the grid cell during the scanning process; and coverage per view cell, defined as the cumulative number of scans performed by each scanning view cell within the grid cell.

[0093] The point cloud mapping data is obtained by performing view overlap weighting correction on the point cloud mapping data;

[0094] In one embodiment, for point cloud mapping data, the view redundancy of each grid cell is calculated, and the formula for calculating the view redundancy of each grid cell is as follows: ,in The redundancy coefficient is the field of view. This indicates the number of independent viewing angles that are effectively covered by the grid from different scanning angles (viewpoint orientation difference ≥15°). This represents the maximum number of independent viewpoints covered by all rasters in the current scanned scene. Based on view redundancy, the rasters are divided into regions with different redundancy levels: high redundancy regions are defined as... ≥0.7, low redundancy region is defined as <0.3. To balance the point cloud sparsity bias caused by low-view redundancy areas (corresponding to occlusion or blind spots), a weighted correction is applied to the point cloud count within the low-redundancy areas. The correction formula is as follows: ,in The corrected number of raster cell point clouds. This represents the actual number of point clouds observed within the original raster cells. The redundancy coefficient is the field of view. =0.01 is a stabilizing term to prevent division by zero. Output the weighted corrected point cloud mapping data, recording the corrected point cloud count for each raster.

[0095] The density gradient field is calculated based on the point cloud mapping data to obtain the point cloud density distribution field data.

[0096] In one embodiment, based on the weighted and corrected point cloud mapping data, the local density gradient of each spatial grid is calculated. The local density gradient is defined as the corrected point cloud count of that grid minus the count of its neighboring grids. The mean, calculated using the formula is: ,in For the first The local density gradient value of each spatial raster cell. (X-direction index, Index in the Y direction, For Z-direction index, This represents the mean of the number of points in the corrected point cloud. For Centered on the target, this represents the number of point clouds within a 6- or 18-neighborhood of spatial raster cells. Neighboring raster cells are selected based on a local neighborhood difference strategy, using either 6-neighborhood (i.e., adjacent raster cells in the X / Y / Z directions) or 18-neighborhood (including diagonally adjacent raster cells) for calculation. The specific neighborhood range is dynamically adjusted based on the device's spatial complexity. This local density gradient calculation effectively highlights areas of abnormal point cloud density gradients and enhances the edges of occluded areas and local occlusion features. The output point cloud density gradient field data table includes the raster number, the corrected number of point clouds, and the corresponding local density gradient.

[0097] Optionally, the extraction of the occlusion area includes:

[0098] S21. Perform scanning field simulation calculations on the point cloud density distribution field data to obtain scanning field data;

[0099] In one embodiment, known frustum parameters of the scanner, including horizontal and vertical viewing angles and an upper limit of scanning distance, are utilized. These parameters are determined by scanner calibration data or device specification parameters. For each frame of scan data, an attitude transformation is performed based on the scanner's real-time pose parameters (spatial position coordinates and attitude angles) to map the scanner's frustum model to the global coordinate system, and a theoretical scan field of view model is established through simulation. The overall space is divided into fixed-size three-dimensional grid cells with grid side lengths... =5cm. For each grid cell, determine whether its center point is within the view frustum spatial envelope of the current scan frame, as a criterion for visibility. The determination result is represented by a Boolean variable; if the grid center is within the view frustum's reachable range, it is marked as visible. (Visible); if not within reach, mark as (Invisible). Output a visibility data table for each raster, including the raster number and its corresponding visibility tag.

[0100] S22. Calculate the visual field contrast residual based on the point cloud density distribution field data and the scanned visual field data to obtain the visual field contrast data.

[0101] In one embodiment, for the established point cloud density distribution field data and scanned view area data, for each raster cell within the view area, the corresponding view area contrast residual is calculated. The formula for calculating residuals is: ,in Represents a grid The number of point clouds after internal correction. This represents the reference mean point cloud density for this category of equipment within a normal area. The reference mean is determined through expert experience rules or equipment design data. Based on the calculation results, a residual judgment threshold is set. If… If the value is greater than 0.6, the raster is marked as a potential occlusion raster. A view area contrast residual map data table is generated, including the raster number and its corresponding residual value.

[0102] S23. Perform topological occlusion chain analysis on the visible field comparison data to obtain local occlusion chain map data;

[0103] In one embodiment, for the visible field contrast residual data, grid cells with residual values ​​higher than 0.6 are used as graph nodes in the graph structure to establish a topological adjacency graph. In the adjacency graph, graph nodes correspond to high residual grid cells, and the formula for calculating the edge weights between nodes is as follows: ,in For graph nodes to Edge weights, This is the spatial distance weighting coefficient, with a value of 1.0. For graph nodes and The distance between the center points of the corresponding grid cells This is the residual difference weighting coefficient, with a value of 0.5. For graph nodes The residual rate corresponding to the grid cell, For graph nodes The residual rate corresponds to the grid cell. Using this adjacency graph, the top 5% of nodes with the highest residual rates are selected as initial traversal seed nodes, sorted, and used as search starting points in sequence. In each search round, starting from the seed node, the path is expanded preferentially along the direction with the lowest adjacent edge weight, constructing chain paths and prioritizing highly connected and residual-consistent paths. During traversal, a cumulative residual consistency threshold is used; if the path residual consistency fluctuation is >20%, the current path expansion is terminated, and the next candidate path is used. In the DFS search, unvisited nodes are traversed recursively in depth-first order to construct deep chains. In the BFS search, neighboring nodes are searched preferentially to construct wide-coverage chains. After the search, short chain segments with a length of less than 3 grid cells are removed to avoid interference from short chain artifacts on the chain structure representation. The output locally occluded chain map data includes the chain ID, the corresponding node sequence (i.e., the grid number sequence), the total residual value of the chain (i.e., the cumulative residual values ​​of each node on the chain), and the chain length.

[0104] S24. Calculate the shading potential based on the local shading chain map data to obtain the shading potential data;

[0105] In one embodiment, for local shading chain map data, the following shading characteristic indices are calculated for each shading chain: tension distribution. Defined as the mean gradient of the residual values ​​along the path of the shading chain, reflecting the trend of residual changes within the shading region; return coefficient. Defined as the proportion of the number of turnback segments in the shading chain to the total number of chain segments; visual penetration. , defined as the average visibility coefficient within the shading chain region, represents the overall visibility of the chain space region. For the above characteristic indicators, a normalized nonlinear combination formula is used to calculate the shading potential score. The formula is ,in , This is the weighting coefficient for the tension index, with a value of 0.4. This is the return factor weighting coefficient, with a value of 0.3. The visual penetration weight coefficient is 0.3; the output occlusion potential data table includes the chain number (chain ID) and the corresponding occlusion potential score.

[0106] S25. Perform cross-temporal occlusion processing based on occlusion potential data to obtain occlusion area data.

[0107] In one embodiment, cross-temporal occlusion consistency processing is performed on the occlusion potential data acquired through multiple rounds of scanning to distinguish between stable and dynamic occlusion chains. For occlusion chains within adjacent time periods, if the chain number (chain ID) is the same or the spatial overlap of chain nodes is ≥80%, it is considered as the continuous performance of the same chain in different time periods. The change in occlusion potential score of the corresponding chain in different time periods is calculated. The calculation formula is: ,in Time period The shielding potential score of the inner chain and the range of change in the shielding potential score. Time period The shielding potential score of the inner chain, For the first Time-series index For the first A time-series index. If... If the value is less than 0.1, it is determined to be a stable occlusion chain, and all grid cells covered by the corresponding chain are marked as occlusion areas. If the occlusion value is greater than 0.3, it is determined to be a dynamic occlusion or pseudo-occlusion chain, and the chain is removed and not included in the occlusion area data. For values ​​0.1 ≤ Chains with a value ≤0.3, such as busbars and rigid support equipment, ≤0.2 is preferred for retention, such as disconnectors and flexible connectors. Values ​​greater than 0.2 require manual confirmation and are not directly retained. Alternatively, based on expert configuration, the system can configure a sensitivity threshold (e.g., 0.15). If the value is less than the threshold, the occlusion chain is retained; otherwise, it is discarded. The output occlusion area data includes: spatial raster number, corresponding occlusion marker (occluded / unoccluded), and chain ID information.

[0108] Optionally, the shielding potential calculation includes:

[0109] The tension distribution of the masking chain is calculated based on the local masking chain map data, and the tension distribution data of the masking chain is obtained.

[0110] In one embodiment, the tension distribution characteristics of each shading chain are calculated for local shading chain map data. Each shading chain contains a sequence of nodes arranged in spatial order. Each node Corresponding to the center coordinates of the spatial raster and associated with the residual values, This is the node order item, with a value of 1... For adjacent node segments within the chain, calculate the residual gradient. The formula is: ,in The residual gradient is the sum of the adjacent node segments within the chain. For nodes The visible contrast residual value corresponding to the spatial raster. For nodes The visible contrast residual value corresponding to the spatial raster. =1,2,..., -1. Calculate the tension value of the corresponding segment. , defined as the residual gradient divided by the distance between nodes: ,in For nodes and The spatial Euclidean distance between them. The tension distribution is calculated for the entire shielding chain using the following formula: ,in For all within this shielding chain The maximum value obtained The output shielding chain tension distribution data includes the chain number (chain ID) and the tension sequence.

[0111] The shielding chain reversal coefficient is calculated based on the shielding chain tension distribution data to obtain the shielding chain reversal coefficient data;

[0112] In one embodiment, the corresponding shielding chain reversal coefficient is calculated based on the shielding chain tension distribution data. This is applied to the spatially ordered sequence of nodes within the shielding chain. Calculate the path vector between adjacent nodes: ,in For nodes to Spatial position difference vector between them =1,2,..., -1. Calculate the overall principal direction vector of the chain. Defined as a vector of all path segments The mean vector reflects the trend of the chain's main space. For each path vector... Calculate its relationship with the principal direction vector. The included angle The calculation formula is: ,in The unit is degrees. If... If the angle is greater than 90°, the segment is considered a turnaround segment, indicating a local path reversal. When determining whether a path segment is a turnaround segment, in addition to considering the included angle... >90° is used as the primary criterion, while auxiliary criteria are also considered, such as those for tension distribution. The local extreme point, if There is a clear jump, such as (i.e., adjacent within the shielding chain) The difference is ≥0.5, and the corresponding The boundary section is then identified as a turnaround section. Turnaround coefficient Defined as: , the total number of chain segments = -1. Output a table of occlusion chain turnaround coefficients, including chain number (chain ID) and corresponding turnaround coefficient, as an important indicator of the space complexity and occlusion instability of the occlusion chain.

[0113] The temporal fluctuation residuals are calculated based on the local shading chain map data to obtain the temporal fluctuation residual data;

[0114] In one embodiment, for the local occlusion chain map data, scanning time-series fluctuation residuals are calculated to measure the stability of the occlusion state of the occlusion chain during multiple scanning rounds. For the results of multiple scanning rounds, statistics are compiled for each scanning time slice. In (corresponding scan round or time period number), the occlusion chain node residual value For the same node Calculate the standard deviation of its time-slice residual sequence to obtain the temporal residual fluctuation of the nodes. The calculation formula is: ,in Indicates standard deviation, This represents all time slice sequences involved in the computation. To determine the total number of scan rounds, calculate the mean temporal fluctuation of the entire masking chain. Defined as all nodes within the chain The mean, calculated using the formula is: Output a time series fluctuation residual data table, including the chain number (chain ID) and the corresponding average time series fluctuation of the entire chain.

[0115] The field of view penetration coefficient is calculated based on the temporal fluctuation residual data to obtain the field of view penetration coefficient data;

[0116] In one embodiment, during the calculation of the field of view penetration coefficient, the temporal fluctuation residual from the previous stage is incorporated. Data, for High volatility chain segment of 0.3, reducing The adoption weight (e.g., multiplied by a decay factor of 0.8) is used to suppress the impact of dynamically unstable chain segments. The impact of the calculation is assessed by determining the corresponding view penetration coefficient for local occlusion chain map data. This calculation is performed for each node in the occlusion chain. Check the corresponding field of view visibility data. The The value is determined by the aforementioned scanned visible area simulation phase, where 1 indicates that the node is within the visible area, and 0 indicates that it is invisible. The total number of visible nodes within the occlusion chain is then counted. and the total number of chain nodes (i.e., the total number of nodes in the chain) ). Calculate the visual penetration coefficient. The calculation formula is: , This value describes the average visibility level of the corresponding spatial area along the entire occlusion chain; a lower value indicates a higher degree of overall occlusion. The output is a table of view penetration coefficients, including the chain number (chain ID) and its corresponding view penetration coefficient.

[0117] Based on the shading chain return coefficient data and the field of view penetration coefficient data, the potential segmentation coherence mapping of the local shading chain map data is performed to obtain the shading potential data.

[0118] In one embodiment, for local shading chain map data, combined with return coefficient data and field-of-view penetration coefficient data, potential segment coherence mapping is performed to calculate shading potential data. This is based on the shading chain tension distribution. The curve employs local extremum detection methods (such as first derivative zero-crossing detection or sliding window extremum determination) to divide the shading chain into several potential segments, with the segment boundaries corresponding to... Local maxima or minima are found within the potential range. For each potential segment, the segment potential characteristic value is calculated. The calculation formula is: ,in The chain reversal coefficient, This is the return factor weighting factor, with a value of 0.5. The chain's field of view penetration coefficient. This is the visual penetration weighting coefficient, with a value of 0.5. Potential coherence mapping is performed; for adjacent potential segments, if the potential difference between segments... Adjacent segments are merged and smoothed to form a potential continuity region, improving the continuity and stability of the spatial potential representation of the occlusion chain. The output occlusion potential data includes the chain number (chain ID) and the corresponding potential segment feature sequence.

[0119] Optionally, the structural connectivity calculation includes:

[0120] Heteroscedastic density features are extracted from the point cloud density distribution field data based on the occlusion area data to obtain heteroscedastic density feature data.

[0121] In one embodiment, heteroscedastic density features are extracted from the point cloud density distribution field data for occlusion area data to enhance the spatial representation capability of the occlusion boundary. This is achieved using occlusion residuals. The corresponding grid cells of the region are used as initial candidate occlusion areas. For each grid cell within the occlusion area... Count the number of corrected point clouds in the neighboring raster cells, and calculate the average number of point clouds in the range of 6, 18, or 26 neighbors. and variance Calculate the corresponding heteroscedasticity ratio. The calculation formula is: , This reflects the degree of change in point cloud density in a local area. A higher value indicates a more pronounced abrupt change in local density and a greater likelihood of occlusion boundaries forming. A heteroscedasticity threshold is set. The value ranges from 0.5 to 1.0 and can be flexibly adjusted according to different device scenarios. If > The grid regions identified as having strong density transitions are characterized as potential occlusion boundary regions. A heteroscedastic density feature data table is output, including the grid number and corresponding heteroscedasticity index.

[0122] Restricted adjacency graphs are constructed based on heteroscedasticity density feature data to obtain restricted adjacency graph data;

[0123] In one embodiment, a restricted adjacency graph is constructed for heteroscedasticity density feature data. The heteroscedasticity density feature value is used as the node attribute to construct the graph structure. Graph nodes correspond to spatial grid cells, and node numbers are uniquely determined by grid spatial coordinates. Node attributes are used for graph edge constraints and edge weight calculation. The graph edge connection rule is to connect only grid cells within 6-neighborhoods and 18-neighborhoods, and the heteroscedasticity difference between adjacent grid cells is considered to be within the range of 6-neighborhoods and 18-neighborhoods. satisfy ,in For grid cells The corresponding heteroscedasticity density eigenvalue is defined as the ratio of the local neighborhood density variance to the mean. To and Adjacent grid The heteroscedasticity density eigenvalues, This is the threshold for heteroscedasticity differences. ≤ Edge connections are allowed at times. The value is 0.2. Edge weight. Defined as ,in =0.01 is a stable term to prevent division by zero. This is a weighting factor with a value of 0.1. This represents the spatial Euclidean distance between adjacent nodes. It prevents erroneous connections at high-density abrupt boundary conditions, strengthens the restricted propagation characteristics of the graph structure, improves the spatial accuracy of connectivity analysis, and ensures that graph connectivity better reflects the characteristics of real occlusion structures.

[0124] Guided connectivity propagation is performed on restricted adjacency graph data to obtain occluded region connectivity data.

[0125] In one embodiment, based on the full-field heteroscedasticity density characteristics Distribution, sorting of grid cells, and selection The top 10% of high heteroscedasticity grids are used as initial propagation seed nodes. Propagation is prioritized along directions with higher edge weights, and edges with higher weights are used as the search priority queue, dynamically updating the search path. The propagation termination condition is set as follows: (1) Cumulative propagation path length. (1) Reaching the physical scale range of the occlusion area, with a value of 0.5m-1.5m; (2) During the continuous 3-step propagation process, the edge weight A weight decrease exceeding 50% is considered path decay, preventing false propagation into unoccluded areas. For each completed propagation path, the corresponding path number (chain ID, generated by sorting seed nodes) and the propagation path node sequence are recorded. By removing duplicate nodes and integrating paths, a complete connectivity data structure for occluded areas is formed. The output data includes the connectivity chain number (connectivity chain ID), the corresponding node sequence, and a connectivity strength score. The connectivity strength score can be calculated based on the weighted average of the paths.

[0126] Optionally, the cavity residual mapping includes:

[0127] The occlusion region connectivity potential field is generated from the occlusion region connectivity data to obtain the occlusion region connectivity potential field data;

[0128] In one embodiment, occluded region connectivity data is generated for the occluded region connectivity data. The input occluded region connectivity data includes the connectivity chain number (connectivity chain ID) and the corresponding set of nodes. Each connectivity chain is considered an independent potential source, and the central node of the connectivity chain is defined as either the node in the middle of the chain or the node with the highest connectivity strength score. The initial value of the potential field is set to... =1.0 (potential field strength at the center point). For each grid node. The potential field value is calculated using a distance-attenuated potential field model, and the calculation formula is as follows: ,in for The minimum propagation path distance to the center node of the corresponding connected chain (cumulative distance along the graph path). The potential field attenuation coefficient has a value range of 100%. =0.89, which can be flexibly adjusted according to the space dimensions and obstruction characteristics of the equipment. For example, when facing the busbar corridor / cable channel, the value is 0.95, and when facing the high-voltage disconnect switch / outdoor overhead equipment area, the value is 0.8. The larger the value, the faster the potential field decays, reflecting the spatial decay characteristics of shading intensity. For spatial grid nodes covered by multiple connected chains, the potential field superposition principle is adopted, and the potential field is calculated for each chain according to its location. Superimpose the values ​​and take the maximum potential field value. This serves as the final potential field value for each grid node. The output table of potential field data for the occluded region includes the spatial grid number and its corresponding potential field value.

[0129] Local density missing kernel mapping is performed on the connected potential field data of the shading region to obtain local density missing kernel data;

[0130] In one embodiment, for each spatial grid cell, the statistically corrected point cloud density is calculated within a 6-neighbor or 18-neighbor range. mean and standard deviation The neighborhood range can be dynamically selected based on the device's spatial complexity. Calculate the missing kernel value for this grid. The calculation formula is: ,in This serves as a global or corresponding device category reference mean, derived from the reference mean density within the normal region. It is differentiated based on device type, reflecting the typical density level of that device category under unobstructed conditions. To enhance the representation of missing values ​​in high-potential regions, potential field values ​​are used for weighted balancing, and the weighted missing value is calculated. The calculation formula is: ,in For the corresponding grid The connectivity potential value originates from the aforementioned stage of generating the connectivity potential in the occluded region. The threshold for determining the missing kernel is set to... It can be flexibly adjusted according to different equipment scenarios. Regions exceeding the threshold are prioritized as key missing regions. A local missing kernel mapping data table is output, including the raster number and the missing kernel value after corresponding weights.

[0131] Hollow residual path backtracking is performed based on locally missing density kernel data. Hollow backtracking data is obtained.

[0132] In one embodiment, a region with a high degree of kernel deletion is selected ( > The raster cells are used as seed nodes for path search, and the seed node number is uniquely determined by its spatial location within the raster. An adjacency raster graph is constructed, with adjacency relationships based on 6-neighbor or 18-neighbor spatial connections, and graph edges do not cross occlusion boundaries. A breadth-first search (BFS) queue is initialized, employing a priority queue structure, with priority based on the missing kernel weight gradients in adjacent raster cells. ,in The missing kernel weight gradient between the current node and its neighboring nodes. This represents the missing kernel weight value of the current search node's raster. For adjacent nodes (i.e., adjacent grids) The missing kernel weight value is corresponding to the missing kernel. Expansion is preferentially carried out along the gradient descent direction of the missing kernel, with the gradient descent direction determined by preferentially expanding to adjacent cells. The direction of decrease. During the path search process, the path backtracking score is calculated. The calculation formula is: ,in This represents the cumulative number of nodes along the path, or, depending on the scenario, the cumulative spatial path length. The search stops when the currently expanded node is missing a core. , =0.2; Cumulative score for the current path threshold =0.3; or the cumulative number of nodes in the path exceeds the set maximum length L_max, where L_max = 10-20 nodes. Short paths (e.g., path with <3 nodes) are filtered out, or path backtracking is scored. Low-trust path, threshold Set to 0.3. Output a hole backtracking data table, including path number, path node sequence, and path backtracking score.

[0133] A global void residual field is constructed based on void backtracking data to obtain void residual data.

[0134] In one embodiment, a global void residual field is constructed for void backtracking data. The global void residual field is initialized. Defined in the full-space raster domain, initial value =0. For each hollow backtracking path The weights are accumulated along the path node sequence, and the accumulation formula is as follows: ,in In the global void residual field, the grid The initial cumulative residual value, initial value = 0. For the natural index term, Scoring for path backtracking Nodes within the path The cumulative spatial distance relative to the starting point of the path, The distance attenuation coefficient within the path ranges from 0.5 to 1.0. In areas where multiple paths overlap, the path contributions are accumulated, forming local high-value residual regions, highlighting the cumulative degree of spatial loss in the missing areas. After all paths have been traversed, the void residual field is processed using the following formula: ,in For the standardized global void residual field value, For the whole audience The maximum value, the standardized global void residual field value, is mapped to the [0,1] interval. The output is a global void residual field data table, including the spatial raster number and the corresponding standardized residual value.

[0135] Optionally, S3 includes:

[0136] S31. Construct a residual map structure from the vacant residual data to obtain residual map structure data;

[0137] In one embodiment, a residual graph structure is constructed for the cavitary residual data, serving as an important input graph structure for the graph neural network in the category reasoning stage. The input is global cavitary residual field data. ,in The residual value originates from the aforementioned void residual field construction stage. High residual region raster cells are selected based on the following criteria: ≥ threshold =0.5, which can be flexibly adjusted according to different device scenarios. Grid cells that meet the conditions are used as a graph node set. A graph node is defined as node ID= Node attributes = { coordinate }, where spatial coordinates A unified mapping to the global coordinate system is used. Graph adjacency is defined as adjacent nodes with high residuals within a 6-neighborhood or 18-neighborhood. ≥ Edges are established between () to form an edge set E. The formula for calculating edge weights is: ,in The edge weights represent the similarity between adjacent nodes with high residuals. It is a natural exponential function. This is the edge weight smoothing coefficient, with a value of 0.5. for The residual value, for The residual values ​​are calculated. The output residual graph structure data includes the graph node set, edge set, node attributes, and edge weights.

[0138] S32. Inject category prior labels into the residual graph structure data to obtain graph label data;

[0139] In one embodiment, prior category labels are injected into the residual graph structure data to generate graph label data, which serves as supervision information for the category-specific graph convolutional network. The sources of prior category labels include: (1) CAD / BIM data of known equipment types, matched by spatial location projection, with the matching area defined as grid nodes within a spatial overlap ≥ 80%; (2) historical high-confidence classification data, such as classification results with a confidence level higher than 0.9 in the previous modeling round, serving as sources of reliable labels. For nodes within the matching area, category labels are injected, assigning node category numbers: And label the corresponding confidence score. ), where score( )∈[0,1], set to ≥0.9, indicating manually confirmed or highly reliable historical results. For unlabeled nodes, assign a label( =None, confidence score )=0, serving as the learning target node (unlabeled node) during the training phase of the graph neural network. The output graph labeling data includes the graph node set, node category labels, node confidence, and adjacent edge set.

[0140] S33. Perform category-specific graph convolutional inference based on the graph labeling data to obtain category inference data.

[0141] In one embodiment, a category-specific graph convolutional neural network is used for category inference on graph-labeled data, outputting category inference data. The network architecture uses a graph convolutional neural network (GCN), a graph attention network (GAT), or a graph isomorphic network (GIN) as the backbone, employing a 3-5 layer GIN network structure. The network depth is dynamically adjusted based on the spatial complexity of the devices and the difficulty of category differentiation. For device categories with complex shapes or large local variations (such as disconnectors and composite switches), a 5-layer GIN network is used; for device categories with simple shapes and obvious linear continuity characteristics (such as busbars), a 3-4 layer GIN network is used. Each layer of the GIN network structure includes GINConv → BatchNorm → ReLU activation → Dropout (p=0.3) stacked layer by layer, with the overall network consisting of 3-5 layers. The calculation formula for the GINConv convolution kernel is as follows: ,in For the first Layer nodes Feature embedding, This is a multilayer perceptron independent of this layer (2 fully connected sub-networks, hidden dimensions = 64-128). For the first Layer nodes Feature embedding, For learnable parameters, the initial value is... =0, For nodes The set of adjacent nodes, For the first Layer nodes Feature embedding. Network input features (Extracted from node attributes in residual graph structure data) Includes the following features: ,in The residual value is ( ) represents the global coordinates of the node. This represents the degree of a node (the number of adjacent nodes). This represents the average residual value of adjacent nodes. The convolution strategy employs a combination of class-specific convolutional kernel weights and designs different feature enhancement paths for different classes using class templates. For example, busbar equipment uses convolutional kernels that enhance linear continuity features. ,in For the first After layer graph convolution, graph nodes The feature vector output results It is a multilayer perceptron. The structure consists of several fully connected layers (1 to 2 layers) and a ReLU activation function. This is the residual enhancement factor, with a value of 0.05. For the first After layer graph convolution, graph nodes The feature vector of the previous layer, For the adjacency graph node index. For graph nodes In the set of adjacent nodes, the set of linearly continuous adjacent nodes with a principal direction angle ≤ 20° is defined. represents the convolution weighting coefficients for linearly adjacent nodes, with a value of 1.0. For the first After layer graph convolution, the feature vector of the previous layer of adjacent node j, For graph nodes The set of adjacent nodes in a given set that has a principal direction angle > 20° is a non-linear set of adjacent nodes. The convolution weighting coefficients for nonlinear adjacent nodes are 0.2. For switch-type devices, enhanced local planar or articulated feature convolution kernels are used.

[0142]

[0143] ,

[0144] in For the first After layer graph convolution, graph nodes The feature vector output represents the feature representation of the current layer of the node. It is a multilayer perceptron. The structure consists of several fully connected layers (1 to 2 layers) and a ReLU activation function. This is the residual enhancement factor, with a value of 0.1. For the first After layer graph convolution, graph nodes The feature vector of the previous layer, For the adjacency graph node index. For graph nodes In the set of adjacent nodes, the set of adjacent nodes with a normal angle ≥ 60° with varying heights is mainly used to enhance the expression of hinge features. represents the convolution weighting coefficient for hinged adjacent nodes, with a value of 1.0. For the first After layer graph convolution, adjacent nodes The feature vector of the previous layer, For graph nodes In the set of adjacent nodes, except , The set of other adjacent nodes, The convolution weighting coefficients for the remaining adjacent nodes are set to 0.3. The loss function is designed as a weighted combination of supervised cross-entropy loss and graph smoothing regularization term. The overall loss function is: ,in This represents the overall loss value. To supervise the cross-entropy loss, it is derived from the labeled nodes in the graph labeling data. For graph adjacency consistency regularization, smoothing coefficient The value is set to 0.1. During the training phase, a semi-supervised graph convolution learning strategy is employed, utilizing labeled nodes for supervised training and employing a semi-supervised propagation mechanism for unlabeled nodes to enhance the overall generalization ability of the graph convolution. During the inference phase, the network learns from each graph node... Output class prediction vector P_class( ), P_class( ) is the class probability distribution vector P_class( after softmax) )∈[0,1], and determine the predicted class( ) and the corresponding confidence score ( Output category inference data, including node numbers. Predicted class( ), class probability distribution vector P_class( ), and category confidence score ( ).

[0145] Optionally, S4 includes:

[0146] S41. Dynamically load the physical constraint rule set based on the category reasoning data to obtain the physical constraint data;

[0147] In one embodiment, input category reasoning data, the data structure including { ,predicted_class( ),P_class( ), score( )}, where score( The score () originates from the aforementioned category-specific graph convolutional inference stage and represents the confidence level of the node's predicted category. Nodes with a confidence score ≥ 0.8 are classified into the high-confidence category region, meaning the confidence threshold is set to 0.8. This is based on the predicted class (predicted_class). The corresponding category ID (lass_ID) is used to dynamically load the physical rule set library `Rule_Set[Class_ID]`. `Rule_Set` is a predefined physical rule library, indexed by device category ID. Each device category corresponds to a set of physical rules, including the following rule items: device geometric dimension range (Length, Width, Height), allowing ±5% to 10% dimensional tolerance; structural connection constraint rules, such as the requirement for fulcrum connection constraints for disconnector devices and continuous straight segment connection rules for busbar devices; spatial positional constraints, maintaining a minimum safe distance of ≥300mm from adjacent devices; and electrical safety specifications, such as a safe ground height of ≥1.5m for devices. Through the above rule loading process, physical constraint data for the current category is generated. The data structure is `{Rule_Set[Class_ID], confidence map}`, where the confidence map records the spatial distribution of high-confidence category nodes.

[0148] S42. Based on the physical constraint data, rule-driven completion candidate generation is performed to obtain completion candidate data;

[0149] In one embodiment, for physical constraint data, within the high confidence category region, a corresponding parameterized device morphology template is generated according to the rule set defined in Rule_Set[Class_ID] to guide the generation of device completion fragments in the missing area. The parameterized device morphology template is defined as follows: busbar type equipment: straight segment template + bracket connection template; switch type equipment: rotating shaft structure template + closed plate structure template; surge arrester type equipment: columnar segment structure template. The preliminary generation method of the completion fragment includes the following steps: (1) taking the starting point of the high missing segment in the void residual region as the fitting starting point; (2) extending to the high residual region along the main axis of the rule or through the main direction fitted by principal component analysis (PCA) in the local high residual region; (3) during the fitting process, the length and morphology of the completion fragment are restricted by the rule items of Rule_Set (such as equipment length / size range, connection constraints, etc.) to ensure physical rationality. When each candidate completion fragment is generated, the completion score is calculated, and the calculation formula is: ,in To complete the rating, This is the consistency score weighting parameter, with a value of 0.35. The consistency score is calculated based on the fit between the morphology and the equipment rule model. The data is for continuous scoring with a weighting of 0.35. To complete the continuity score between the fragment and the adjacent known structures, The support weight data is set to 0.3. It provides local point cloud density support; outputs candidate completion data, including completion fragment number (completion fragment ID), corresponding morphological parameters, and completion score.

[0150] S43. Perform multi-rule collaborative optimization and completion based on the candidate completion data to obtain point cloud completion data.

[0151] In one embodiment, a multi-rule collaborative optimization method is used to optimize the parameters of the completed fragments and generate the final point cloud completed data for the candidate data to be filled in. The optimization process employs a multi-objective optimization algorithm, which can select a search strategy or the L-BFGS gradient optimization algorithm according to the scenario, and jointly optimizes the following three main objectives, including geometric fitting error. ,Require ≤3mm, defined as the average fitting error between the fitted shape of the device-completed segment and the existing structure or rule template; physical rule compliance, requiring Rule_Set physical constraint compliance ≥0.95, defined as the proportion of Rule_Set rule items satisfied; adjacent complete segment continuity C_c≥0.9, defined as the smoothness index of the connection between the completed segment and the adjacent original structure segment. A dynamic stopping condition is set for the optimization process: if the target improvement is <1% in 5 consecutive iterations, or the maximum iteration round max_iter=100 is reached, the optimization process terminates. max_iter can be flexibly adjusted according to the convergence rate. After optimization, the optimized scores are screened. Completed fragments with a value ≥0.85 ( Derived from the S42 completion scoring model, the complete fragments that meet the requirements are inserted into the overall point cloud model. The output point cloud completion data includes the complete fragment number (complete fragment ID), the optimized parameters (optimized parameters), and the insertion position.

[0152] Optionally, this application also provides an intelligent modeling system based on point cloud data, used to execute the intelligent modeling method based on point cloud data as described above, the intelligent modeling system based on point cloud data comprising:

[0153] The multi-view high-density point cloud field construction module is used to perform high-density multi-angle scanning of the interior of the substation to obtain view overlap map data; and to construct the point cloud density distribution field from the view overlap map data to obtain point cloud density distribution field data.

[0154] The occlusion region connectivity residual mapping module is used to extract occlusion regions from point cloud density distribution field data to obtain occlusion region data; calculate the structural connectivity of point cloud density distribution field data based on the occlusion region data to obtain occlusion region connectivity data; and perform hole residual mapping on the occlusion region connectivity data to obtain hole residual data.

[0155] The category-specific graph inference module is used to perform graph neural network inference on the punctuous residual data to obtain category inference data.

[0156] The physical rule-driven completion modeling module is used to fuse physical rules based on category reasoning data to obtain point cloud completion data; and to construct a point cloud model based on the point cloud completion data to obtain the point cloud model.

[0157] 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.

[0158] 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 point cloud data based intelligent modeling method, characterized in that, The method comprises: S1, high-density multi-angle scanning is performed on the inside of the transformer substation to obtain visual field overlap map data; point cloud density distribution field data is obtained by constructing a point cloud density distribution field based on the visual field overlap map data; S2, occlusion zone data is obtained by extracting the occlusion zone from the point cloud density distribution field data; occlusion zone connectivity data is obtained by calculating the structural connectivity of the point cloud density distribution field data based on the occlusion zone data; cavity residual data is obtained by mapping the cavity residues of the occlusion zone connectivity data; S3, category inference data is obtained by performing graph neural network inference on the cavity residual data; S4, point cloud completion data is obtained by fusing physical rules based on the category inference data; a point cloud model is constructed based on the point cloud completion data to obtain the point cloud model.

2. The method of claim 1, wherein, The high-density multi-angle scanning comprises: The scanner is controlled to perform multi-angle acquisition through a preset scanning posture and a scanning movement path to obtain scanning data; The scanning data is subjected to pose calibration to obtain calibration data; The calibration data is subjected to visual field overlap relationship calculation to obtain visual field overlap map data.

3. The method of claim 1, wherein, The point cloud density distribution field construction comprises: The visual field overlap map data is subjected to spatial grid division to obtain spatial grid data; The spatial grid data is subjected to point cloud mapping based on the visual field overlap map data to obtain point cloud mapping data; The point cloud mapping data is subjected to visual field overlap weighted correction to obtain point cloud mapping data; The point cloud density distribution field data is obtained by calculating the density gradient field based on the point cloud mapping data.

4. The method of claim 1, wherein, The occlusion zone extraction comprises: The point cloud density distribution field data is subjected to scanning visual field simulation calculation to obtain scanning visual field data; The visual field comparison residual data is obtained by comparing the visual fields based on the point cloud density distribution field data and the scanning visual field data; The local occlusion chain atlas data is obtained by topological occlusion chain analysis of the visual field comparison residual data; The occlusion potential data is obtained by calculating the occlusion potential based on the local occlusion chain atlas data; The occlusion zone data is obtained by cross-time domain occlusion processing based on the occlusion potential data.

5. The method of claim 4, wherein, The occlusion potential calculation comprises: The occlusion chain tension distribution data is obtained by calculating the occlusion chain tension distribution based on the local occlusion chain atlas data; The occlusion chain return coefficient data is obtained by calculating the occlusion chain return coefficient based on the occlusion chain tension distribution data; The time sequence fluctuation residual data is obtained by calculating the scanning time sequence fluctuation residual based on the local occlusion chain atlas data; The visual field penetration degree coefficient data is obtained by calculating the visual field penetration degree coefficient based on the time sequence fluctuation residual data; The occlusion potential data is obtained by potential segmental coherence mapping of the local occlusion chain atlas data based on the occlusion chain return coefficient data and the visual field penetration degree coefficient data.

6. The method of claim 1, wherein, The structural connectivity calculation comprises: The heteroscedastic density feature data is obtained by extracting the heteroscedastic density features from the point cloud density distribution field data based on the occlusion zone data; The limited adjacency graph data is obtained by constructing a limited adjacency graph based on the heteroscedastic density feature data; The occlusion zone connectivity data is obtained by guided connectivity propagation of the limited adjacency graph data.

7. The method of claim 1, wherein, The cavity residual mapping comprises: The occlusion zone connectivity potential field data is obtained by generating an occlusion zone connectivity potential field based on the occlusion zone connectivity data; The local density missing kernel mapping is performed on the occlusion area connected potential field data to obtain local density missing kernel data; The hollow residual path backtracking is performed according to the local density missing kernel data to obtain hollow backtracking data; The global hollow residual field construction is performed according to the hollow backtracking data to obtain hollow residual data.

8. The method of claim 1, wherein, S3 comprises: The residual graph structure construction is performed on the hollow residual data to obtain residual graph structure data; The category prior label injection is performed on the residual graph structure data to obtain graph label data; The category-specific graph convolution reasoning is performed according to the graph label data to obtain category reasoning data.

9. The method of claim 1, wherein, S4 comprises: The physical constraint rule set dynamic loading is performed according to the category reasoning data to obtain physical constraint data; The rule-driven completion candidate generation is performed according to the physical constraint data to obtain completion candidate data; The multi-rule collaborative optimization completion is performed according to the completion candidate data to obtain point cloud completion data.

10. An intelligent modeling system based on point cloud data, characterized by, The intelligent modeling system based on point cloud data is used to perform the intelligent modeling method based on point cloud data, and the intelligent modeling system based on point cloud data comprises: A multi-view high-density point cloud field construction module is configured to perform high-density multi-angle scanning on the inside of a substation to obtain view overlap graph data, and to perform point cloud density distribution field construction on the view overlap graph data to obtain point cloud density distribution field data. An occlusion area connected residual mapping module is configured to extract an occlusion area from the point cloud density distribution field data to obtain occlusion area data, to perform structure connectivity calculation on the point cloud density distribution field data according to the occlusion area data to obtain occlusion area connected data, and to perform hollow residual mapping on the occlusion area connected data to obtain hollow residual data. A category-specific graph reasoning module is configured to perform graph neural network reasoning on the hollow residual data to obtain category reasoning data. A physical rule-driven completion modeling module is configured to perform physical rule fusion according to the category reasoning data to obtain point cloud completion data, and to perform point cloud model construction according to the point cloud completion data to obtain a point cloud model.

Images