A transformer component point cloud segmentation and three-dimensional reconstruction method based on deep learning

Abstract

The application provides a transformer component point cloud segmentation and three-dimensional reconstruction method based on deep learning, relates to three-dimensional point cloud processing, and comprises the following steps: collecting point cloud data of a transformer component under different viewing angles, performing registration, denoising and coordinate unification processing on the point cloud data, forming component point cloud data, and extracting geometric structure information in the point cloud; in combination with the structural features and engineering semantic information of the transformer component, the component point cloud is regionally divided, the point cloud regions corresponding to different functional components are distinguished, and the corresponding relationship between the component point cloud and the actual engineering parts is established; the positional relationship and connection relationship between the component point cloud regions are analyzed, the assembly correlation relationship between the components is determined, and the incomplete regions caused by shielding or collection loss are structurally supplemented, high-confidence repair of the missing regions is realized, the geometric integrity of the component point cloud is ensured, and the reconstruction failure or the decrease in reconstruction precision caused by the point cloud loss is effectively avoided.

Description

Technical Field

[0001] This application relates to 3D point cloud processing, and more specifically, to a deep learning-based method for point cloud segmentation and 3D reconstruction of transformer components. Background Technology

[0002] Three-dimensional digital modeling of transformer components is an important technical foundation for intelligent operation and maintenance and condition assessment of power equipment. With the rapid development of three-dimensional laser scanning technology, point cloud-based three-dimensional reconstruction methods have gradually become the main means of obtaining geometric information of transformer components. Deep learning technology has made significant progress in the field of point cloud processing. Point cloud neural networks can directly learn spatial features from disordered point clouds and have shown strong feature extraction capabilities in semantic segmentation and instance segmentation tasks in general three-dimensional scenes.

[0003] However, the following prominent problems exist in the point cloud segmentation and 3D reconstruction technology of transformer components based on deep learning:

[0004] (1) Transformer components include various functional parts such as bushings, heat sinks, windings, oil tanks, and insulators. Each component has a highly regular geometric shape, such as cylinders, thin-walled plates, and rings. However, existing general point cloud segmentation networks do not make full use of the geometric prior knowledge specific to this field in the feature extraction stage, resulting in insufficient segmentation accuracy for slender and thin-walled structures. Especially in areas with many component contact surfaces and blurred boundaries, the segmentation results are prone to missegmentation and omission, which directly restricts the integrity and accuracy of subsequent three-dimensional reconstruction.

[0005] (2) The actual point cloud data is inevitably affected by multiple factors such as acquisition error, manufacturing error and assembly error. Existing technologies lack effective separation and correction mechanisms for the above-mentioned multi-source errors, which makes it difficult to guarantee the geometric accuracy of the reconstructed model. In existing methods, point cloud segmentation and 3D reconstruction are usually treated as independent serial links. Segmentation error will be transmitted unidirectionally to the reconstruction link and cannot be corrected. There is a lack of closed-loop feedback optimization mechanism between segmentation and reconstruction.

[0006] Therefore, there is an urgent need for a deep learning-based method for point cloud segmentation and 3D reconstruction of transformer components. Summary of the Invention

[0007] To address the aforementioned technical problems, this application is proposed. The first aspect of this application provides a deep learning-based method for point cloud segmentation and 3D reconstruction of transformer components, comprising the following specific steps: collecting point cloud data of the transformer component from different perspectives; performing registration, denoising, and coordinate unification processing on the point cloud data to form component point cloud data; and extracting geometric structure information from the point cloud.

[0008] Combining the structural features and engineering semantic information of transformer components, the point cloud of the components is divided into regions to distinguish the point cloud regions corresponding to different functional components and to establish the correspondence between the component point cloud and the actual engineering parts.

[0009] Analyze the positional and connectivity relationships between the point cloud regions of each component, determine the assembly association relationships between components, and supplement the structure of incomplete regions caused by occlusion or missing data acquisition.

[0010] Based on the structural characteristics of the components, a parametric description method is established to separate and analyze point cloud acquisition errors, manufacturing errors, and assembly errors, and to correct the impact of these errors on the geometric description of the components.

[0011] The corrected component geometric description is matched and verified with the actual acquired point cloud. The error correction parameters and surface fitting process are adjusted according to the matching results. If the verification finds obvious errors in the component segmentation, the point cloud segmentation step is returned to be reprocessed to form the transformer component point cloud segmentation and 3D reconstruction results.

[0012] Preferably, the extraction of geometric structure information from the point cloud includes;

[0013] Neighborhood spatial analysis is performed on the preprocessed component point cloud. Based on the spatial distance relationship, distribution relationship and local topological relationship between sampling points, the spatial structure characteristics of the point cloud region are established.

[0014] Based on the changes in the surface morphology of the point cloud, the curvature features, boundary features, and spatial orientation features within the point cloud region of the component are calculated, and the surface structure of the component is described in combination with the local geometric change rules.

[0015] Preferably, establishing the correspondence between the component point cloud and the actual engineering component includes:

[0016] The preprocessed transformer component point cloud is divided into regions according to spatial connectivity. For multiple independent point cloud regions formed due to occlusion, scanning angle limitation or surface discontinuity, a region-level structure description is established.

[0017] Within each independent point cloud region, the spatial distribution of the point cloud is analyzed to obtain the main direction, surface change trend, spatial scale parameters, and edge contour features of the point cloud within the region.

[0018] For the component association relationships between different independent point cloud regions, a matching analysis is performed on the structural orientation consistency, size ratio relationship and spatial arrangement relationship of each region. When multiple independent point cloud regions meet the preset structural association conditions, the multiple independent point cloud regions are associated as the same component region.

[0019] For point cloud regions that are in contact with each other but belong to different engineering components, the surface orientation changes, structural continuity and connection forms at the boundary of the regions are analyzed. When the adjacent regions do not meet the structural continuity condition, the adjacent regions are separated.

[0020] Preferably, determining the assembly association between components includes:

[0021] After the point cloud region is divided, spatial position analysis is performed on the different component regions, and spatial relationships are established based on the distance, direction and relative position relationships between the component regions.

[0022] Based on the structural characteristics of transformer assembly installation, the connection boundaries, installation directions and spatial constraints between assembly areas are determined, and the assembly relationships are determined according to the direct connection relationships or spacing relationships between the assemblies.

[0023] Preferably, the structural supplementation of incomplete areas caused by occlusion or missing data collection includes:

[0024] For the associated component point cloud region, analyze the spatial connection status between the component and adjacent components to determine the installation position relationship, connection direction relationship and mutual constraint relationship between the components;

[0025] Based on the structural characteristics of transformer components, the engineering constraints corresponding to different types of components are determined. Specifically, for components with axial installation characteristics, the axial direction and end connection positions are determined; for components with array arrangement characteristics, the arrangement direction, spacing relationship, and quantity constraints are determined; and for components with symmetrical structural characteristics, the structural center position and symmetry relationship are determined.

[0026] For occluded or missing regions in point clouds where no valid sampling points exist, the structural change trends around the missing regions are correlated with the engineering constraints of the corresponding components to determine the spatial location and extent of the missing regions.

[0027] Based on the spatial location and range, the geometric change trend of the point cloud collected around the missing area is analyzed, and the surface orientation, boundary contour and spatial size information corresponding to the missing area are supplemented.

[0028] The connection positions, installation directions, and spatial spacing between the supplemented component structure and adjacent components are checked. When connection conflicts or inconsistent spatial constraints occur, the process of supplementing the missing areas is adjusted.

[0029] Preferably, the separation and analysis of point cloud acquisition error, manufacturing error, and assembly error includes:

[0030] Error feature analysis was performed on the collected component point cloud data. Based on the point cloud coordinate offset law, point cloud distribution change characteristics and measurement environment parameters, a spatial change model corresponding to the point cloud acquisition error was established.

[0031] By combining the differences between component design parameters, processing dimension parameters, and actual point cloud geometric parameters, dimensional deviations and form and position deviations generated during component manufacturing are identified.

[0032] Based on the spatial relationships, connection postures, and assembly constraints between components, the positional offsets and posture errors generated during the assembly process are analyzed, and errors from different sources are classified and processed.

[0033] Preferably, the impact of the correction error on the component's geometric description includes:

[0034] Error analysis was performed on the coordinate offset, measurement noise, and point cloud density changes generated during point cloud acquisition. Based on the relationship between the error sources and spatial coordinate changes, the point cloud geometric parameters were corrected.

[0035] The component's dimensional parameters and spatial relationships are adjusted by taking into account dimensional and geometric deviations during component manufacturing, as well as positional offsets and attitude changes during component installation.

[0036] Preferably, the point cloud segmentation and 3D reconstruction results of the transformer component include:

[0037] Based on the component point cloud region division results, spatial coordinate transformation and structural relationship matching are performed on different component regions to establish the spatial positional relationship between components;

[0038] By combining the component's geometric structure information, assembly relationships, and error-corrected geometric parameters, the surface structure of the component is fitted with a surface, and a three-dimensional model of the transformer component is constructed based on the connection relationships between the components.

[0039] Preferably, multi-scale feature extraction is performed on the input transformer component point cloud to obtain local spatial features and overall structural features in the point cloud region, and the point cloud features at different scales are fused.

[0040] Based on the fused point cloud features, component classification features are constructed, and different transformer component regions are segmented according to the spatial feature relationships corresponding to each point cloud sampling point.

[0041] Preferably, local coordinate relationships are established based on the segmented component point cloud regions, key structural points in the component point cloud are identified, and the spatial structure of the component is determined based on the positional relationships between the key structural points.

[0042] By combining the structural features of the component with the preset geometric model parameters, surface fitting and structural matching are performed on the point cloud on the component surface to complete the construction of the component's three-dimensional model.

[0043] Compared with existing technologies, the deep learning-based transformer component point cloud segmentation and 3D reconstruction method provided in this application determines the positional relationship by calculating the centroid coordinates and bounding boxes of the point cloud regions of each component, determines the connection relationship based on the spatial proximity and normal vector consistency of the boundary point set, and determines the assembly association relationship by combining the engineering assembly specifications of the transformer components. This achieves complete modeling of the spatial topological relationship between components from both data-driven and knowledge-driven dimensions. Based on the geometric features of adjacent components and assembly constraints, it performs structural supplementation on incomplete areas caused by occlusion or missing data collection, achieving high-confidence repair of missing areas, ensuring the geometric integrity of the point clouds of each component, and effectively avoiding reconstruction failure or decreased reconstruction accuracy due to missing point clouds. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort. In the drawings:

[0045] Figure 1 This is a flowchart of a deep learning-based point cloud segmentation and 3D reconstruction method for transformer components according to an embodiment of this application. Detailed Implementation

[0046] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0047] As mentioned in the background section, transformer components include various functional parts such as bushings, heat sinks, windings, tanks, and insulators. Each component has a highly regular geometric shape, such as cylinders, thin-walled plates, and rings. However, existing general point cloud segmentation networks do not fully utilize the domain-specific geometric prior knowledge in the feature extraction stage, resulting in insufficient segmentation accuracy for slender and thin-walled structures. Especially in areas with many component contact surfaces and blurred boundaries, the segmentation results are prone to missegmentation and omissions, directly restricting the integrity and accuracy of subsequent 3D reconstruction. Moreover, the actual collected point cloud data is inevitably affected by multiple factors such as acquisition errors, manufacturing errors, and assembly errors. Existing technologies lack effective separation and correction mechanisms for these multi-source errors, making it difficult to guarantee the geometric accuracy of the reconstructed model. In existing methods, point cloud segmentation and 3D reconstruction are usually treated as independent serial steps, with segmentation errors being unidirectionally transmitted to the reconstruction step and unable to be corrected. There is a lack of a closed-loop feedback optimization mechanism between segmentation and reconstruction.

[0048] Therefore, there is an urgent need for a deep learning-based method for point cloud segmentation and 3D reconstruction of transformer components.

[0049] Figure 1 The flowchart below shows a method for point cloud segmentation and 3D reconstruction of a transformer component based on deep learning, according to an embodiment of this application. The method includes the following specific steps:

[0050] S1: Collect point cloud data of transformer components from different perspectives, perform registration, noise reduction and coordinate unification processing on the point cloud data to form component point cloud data, and extract geometric structure information from the point cloud.

[0051] In step S1, neighborhood space analysis is performed on the preprocessed component point cloud. Based on the spatial distance relationship, distribution relationship and local topological relationship between sampling points, the spatial structure features of the point cloud region are established.

[0052] Based on the changes in the surface morphology of the point cloud, the curvature features, boundary features, and spatial orientation features within the point cloud region of the component are calculated, and the surface structure of the component is described in combination with the local geometric change rules.

[0053] Furthermore, to improve the accuracy and stability of geometric structure information extraction from transformer component point cloud data, considering the influence of scanning distance, acquisition angle, component surface material and occlusion during the actual acquisition process, which leads to uneven density, local missing data and noise interference in point cloud data in different areas, deep geometric information mining is performed on the preprocessed component point cloud.

[0054] Considering that traditional point cloud feature extraction methods are usually based on fixed neighborhood ranges, when the neighborhood range is set small, although it can retain local details of the component surface, it is easily affected by measurement noise and outliers, causing large fluctuations in the calculated curvature, normal direction and other geometric parameters. When the neighborhood range is set large, although it can enhance the ability to describe the overall structure of the point cloud, it is easy to cause local structural features to be smoothed, making it impossible to accurately represent edges, connection positions and fine structural features. Based on the spatial distribution characteristics of different regions of transformer components, a multi-scale neighborhood analysis method is introduced. By comprehensively describing the distribution law of point cloud in different spatial ranges, the ability to extract geometric features of point cloud under complex structural conditions is improved.

[0055] Using the component point cloud data after registration, denoising, and coordinate unification as input, a local spatial neighborhood is established for each sampling point. Based on the spatial coordinate information of the current sampling point, multiple neighborhood regions of different scales are constructed with the center within a preset distance range. Among them, the small-scale neighborhood is mainly used to describe the local detail changes of the point cloud surface, the medium-scale neighborhood is used to analyze the local structural change trend of the component, and the large-scale neighborhood is used to obtain the overall spatial morphology information of the component.

[0056] Within the neighborhood at each scale, the spatial distance relationship, distribution density relationship, and spatial direction relationship between neighboring points are calculated respectively. The spatial aggregation degree of the point cloud is calculated based on the Euclidean distance between neighboring points, and the local structure direction is determined by statistically analyzing the number of point cloud distributions in different directions. At the same time, a local spatial covariance matrix is ​​established based on the changes in the coordinates of neighboring points. By performing eigenvalue decomposition on the covariance matrix, the main directional features of the point cloud in the current region are obtained.

[0057] Furthermore, the geometric variation of the point cloud region is calculated based on the proportional relationship between different eigenvalues. When the maximum eigenvalue is significantly higher than other eigenvalues, it indicates that the point cloud in the region is distributed along a certain direction, usually corresponding to the edge of the component, axial structure, or rod-shaped structure. When two eigenvalues ​​are close to and significantly higher than the third eigenvalue, it indicates that the region has obvious planar features, corresponding to the structure of the component shell, mounting surface, etc. When three eigenvalues ​​are close, it indicates that the spatial variation of the region is relatively complex, possibly corresponding to the curved transition region or connection region.

[0058] After obtaining the local spatial structural features, the curvature features, boundary features, and spatial orientation features of the component region are calculated by combining the surface change law of the point cloud. The local curvature change trend is determined by analyzing the change of the normal vector of adjacent points. The continuous region of the component surface is judged based on the degree of continuous change of the normal vector. For the location where the normal vector changes significantly, it is considered that the region may have structural boundaries, connecting end faces, or component turning points, and the boundary contour information is extracted.

[0059] To address the coexistence of large-area regular surfaces and local complex structures in transformer components, geometric parameters obtained from neighborhoods at different scales are combined to form comprehensive descriptive parameters that include information on point cloud spatial distribution, surface variation, and structural orientation.

[0060] For example, for component areas with obvious directional characteristics, such as transformer bushings, heat sinks, and enclosure connectors, the component's axial direction, installation direction, and boundary position are determined by analyzing the main direction of the point cloud and the surface change trend. For large planar areas such as enclosure surfaces, the spatial orientation is determined by planar feature parameters. For complex areas such as connecting flanges and bolt installation areas, local curvature changes and boundary features are used for detailed description, transforming point cloud data from traditional discrete spatial coordinate information into geometric feature information with structural meaning. This provides a reliable data foundation for subsequent component area division, engineering semantic matching, and 3D model reconstruction.

[0061] To address the issue of inaccurate geometric feature extraction due to uneven sampling, noise interference, and complex structures, a multi-scale spatial analysis method is used to enhance the point cloud structure description capability. This enables the extraction results to more realistically reflect the actual surface morphology and spatial structure relationship of transformer components, thereby improving the accuracy of subsequent point cloud segmentation and 3D reconstruction processes.

[0062] S2: Combining the structural features and engineering semantic information of transformer components, the point cloud of the components is divided into regions to distinguish the point cloud regions corresponding to different functional components, and the correspondence between the component point cloud and the actual engineering parts is established.

[0063] In step S2, the preprocessed transformer component point cloud is divided into regions according to spatial connectivity. For multiple independent point cloud regions formed due to occlusion, scanning angle limitations, or surface discontinuities, a region-level structural description is established for each region.

[0064] Within each independent point cloud region, the spatial distribution of the point cloud is analyzed to obtain the main direction, surface change trend, spatial scale parameters, and edge contour features of the point cloud within the region.

[0065] To address the component association relationships between different independent point cloud regions, a matching analysis is performed on the structural orientation consistency, size ratio, and spatial arrangement relationships of each region. When multiple independent point cloud regions meet the preset structural association conditions, the multiple independent point cloud regions are associated as the same component region.

[0066] For point cloud regions that are in contact with each other but belong to different engineering components, the surface orientation changes, structural continuity and connection forms at the boundary of the regions are analyzed. When the adjacent regions do not meet the structural continuity condition, the boundary separation process is performed on the adjacent regions.

[0067] Furthermore, to improve the accuracy of distinguishing between different functional components during the point cloud region segmentation of transformer components, considering that actual transformer equipment has the characteristics of multiple components, multiple connections, and multiple scales, when region segmentation is based solely on point cloud spatial distance or connectivity, it is easy to cause adjacent components to be incorrectly merged and the same component to be incorrectly segmented due to occlusion or missing scans. Therefore, by combining the geometric structural features of the components and engineering semantic information, an optimized point cloud region segmentation for engineering component recognition is established.

[0068] Considering that transformer components have clear structural patterns during actual installation, such as bushings, radiators, housings, connecting pipes, and fasteners, which all have relatively fixed morphological characteristics, spatial orientation characteristics, and connection relationships, point cloud region division should not only rely on spatial positional relationships, but should also combine the structural characteristics of the components themselves and the engineering relationships between parts, so that the point cloud region division results can correspond to the functional components in the actual equipment.

[0069] By utilizing the spatial structure features of point clouds, initial region division is performed on the preprocessed component point clouds. Based on the spatial distance between point clouds, local normal consistency, and spatial connectivity, point clouds with continuous geometric features are divided into multiple independent regions. For discontinuous point cloud regions caused by scanning angle limitations, occlusion, or surface discontinuities, they are not directly identified as different components, but rather region-level structural description information is established for each.

[0070] Within each independent point cloud region, the distribution of point clouds within the region is analyzed, and regional spatial scale parameters, principal direction parameters, surface change trends, and edge contour features are extracted. Based on the spatial distribution relationship of the point cloud sampling points within the region, the principal direction features of the regional point cloud are calculated, and the spatial center of all sampling points within the region is calculated.

[0071] ;

[0072] in, Indicates the first The spatial center location of a point cloud region Indicates the number of sampling points in the area. Indicates the first The three-dimensional coordinates of each sampling point.

[0073] Furthermore, based on the positional changes of each sampling point within the region relative to the spatial center, the spatial distribution matrix of the point cloud is calculated:

[0074] ;

[0075] in, The feature matrix representing the spatial distribution of a region's point cloud is obtained by performing eigenvalue decomposition on the matrix:

[0076] ;

[0077] To obtain the degree of spatial variation in different directions, the eigenvector corresponding to the largest eigenvalue is:

[0078] ;

[0079] in, Represents the spatial distribution matrix The 1 eigenvalue, Indicates and The corresponding eigenvectors have their eigenvalues ​​sorted from largest to smallest. As the main direction of the current point cloud region, the spatial extension direction of different point cloud regions is obtained, providing directional constraints for subsequent determination of whether different regions belong to the same engineering component.

[0080] Among them, the spatial scale parameter is used to describe the overall size information of the region, such as length, width, and height; the main direction parameter is used to describe the component installation direction or the structural extension direction; the surface change trend is used to determine whether the region belongs to a planar structure, a curved structure, or a complex connection structure; and the edge contour feature is used to describe the boundary shape of the component.

[0081] Furthermore, to address the issue of multiple independent point cloud regions formed by the same component due to occlusion, an association judgment method based on regional structural consistency is used to match and analyze the structural orientation, size ratio, and spatial position relationship between different independent regions. When multiple regions have similar main orientations, similar size ratios, and spatial position relationships that conform to engineering installation relationships, it is considered that multiple regions may belong to the same component.

[0082] For two independent point cloud regions and Calculate the degree of directional consistency based on the obtained principal direction vector:

[0083] ;

[0084] in, This represents the angle between the principal directions of two point cloud regions. , The first and the The principal direction vector of a point cloud region, when When two regions share the same structural extension direction, structural continuity is determined based on the projection distance between their centers along the main direction: the average direction of the two regions' main directions is taken as the reference direction, and the projection distance between the centers of the two regions along this direction is calculated and recorded as the distance along the main direction. If the distance is less than a preset threshold, the two regions are considered continuous in the extension direction. Simultaneously, the offset distance between the centers of the two regions on a plane perpendicular to the reference direction is calculated. If the offset distance is less than another preset threshold, the two regions are considered to belong to the same structural line, avoiding erroneous merging of parallel components. When the included angle, the projection distance along the main direction, the vertical offset, and the size difference are all less than a threshold, the two independent point cloud regions are deemed to meet the component association conditions.

[0085] Combined with regional size characteristics:

[0086] ;

[0087] in, , , They represent the first The length, width, and height dimensions of each region, when When, it indicates that the two regions have similar structural scales, therefore, when and and When two independent point cloud regions meet the component association conditions, multiple regions are associated as the same component region. This avoids the problem of traditional methods that rely solely on spatial distance, which leads to the incorrect merging of adjacent but different functional components. At the same time, it can identify multiple separate point cloud regions of the same component caused by occlusion.

[0088] For example, in a transformer heat sink assembly, since some areas may be blocked by other components, multiple separate point cloud regions are formed after scanning. At this time, by analyzing the arrangement direction, spacing relationship and overall size ratio of each region, it can be determined that multiple regions belong to the same heat dissipation structure and will be re-associated as a complete component region.

[0089] Furthermore, for point cloud regions that are spatially adjacent but actually belong to different engineering components, region separation is determined by analyzing the structural continuity at the region boundaries. Specifically, the degree of normal variation, surface continuity, and connection boundary characteristics between adjacent regions are calculated. When the surface orientation changes continuously between two regions and satisfy the actual structural connection relationship, they are considered to belong to the same component structure; when there are obvious abrupt changes in orientation, discontinuous boundaries, or significant differences in dimensional characteristics between two regions, they are considered to belong to different engineering components, and the independent region division results are retained.

[0090] To distinguish between spatially connected but functionally different engineering components, calculate the normal variation relationship at the boundary of adjacent regions:

[0091] ;

[0092] in, and These represent the surface normal directions at the boundary of two adjacent regions. This represents the angle between two normal vectors.

[0093] when When the surface changes of the two regions are continuous, it can be considered that there is a structural continuity relationship; when If a significant structural change occurs between two regions, they are identified as different engineering components, and boundary separation processing is performed to improve the accuracy of segmentation between contact components.

[0094] To improve the correspondence between the region division results and the engineering objects, engineering semantic constraint information is introduced. Based on the structural characteristics of transformer components, the component types that may correspond to different regions are constrained. For example, regions with cylindrical structures that extend along a fixed direction can be candidate regions for bushing-type components; regions with large-area continuous planar structures can be candidate regions for box structures; and regions with multiple repeating units can be candidate regions for heat dissipation structures.

[0095] By establishing a matching relationship between regional feature description information and engineering component structural features, the initial regional division results are optimized and adjusted so that the final point cloud region not only represents a continuous spatial region, but also corresponds to the functional components in the actual project. This enables the system to identify incomplete component regions caused by occlusion and to distinguish between spatially connected but functionally different engineering components. The mapping relationship between point cloud data and actual transformer components is established, improving the accuracy and engineering applicability of point cloud region division in complex equipment environments.

[0096] The region segmentation of the component point cloud is achieved using a deep learning-based point cloud semantic segmentation model. Local geometric features of the point cloud are extracted and restored to their original resolution through multi-scale grouping and feature propagation modules. A Transformer self-attention layer is introduced after the backbone network to enhance the global modeling capability of spatial relationships between components. The segmentation head consists of a multi-layer perceptron, outputting the probability of each point belonging to each component category. Training data is obtained by semantically annotating each point in actual collected and synthesized point clouds. Annotated categories include iron core, high-voltage winding, low-voltage winding, bushing, radiator, oil tank, and other accessories. Data augmentation strategies such as random rotation, translation, scaling, downsampling, adding noise, and simulated occlusion are employed during training. The model training uses a combined loss function, including weighted cross-entropy loss, Dice loss, and boundary-aware loss, to alleviate class imbalance and improve boundary segmentation accuracy.

[0097] S3: Analyze the positional and connectivity relationships between the point cloud regions of each component, determine the assembly association relationships between components, and supplement the structure of incomplete regions caused by occlusion or missing data acquisition.

[0098] In step S3, spatial position analysis is performed on the different component regions after the point cloud region is divided, and spatial association is established based on the distance relationship, direction relationship and relative position relationship between the component regions.

[0099] Based on the structural characteristics of transformer assembly installation, the connection boundaries, installation directions and spatial constraints between assembly areas are determined, and the assembly relationships are determined according to the direct connection relationships or spacing relationships between the assemblies.

[0100] For the associated component point cloud region, analyze the spatial connection status between the component and adjacent components to determine the installation position relationship, connection direction relationship and mutual constraint relationship between the components.

[0101] Based on the structural characteristics of transformer components, the engineering constraints corresponding to different types of components are determined. Specifically, for components with axial mounting characteristics, the axial direction and end connection positions are determined. For components with array arrangement characteristics, the arrangement direction, spacing relationships, and quantity constraints are determined. For components with symmetrical structural characteristics, the structural center position and symmetry relationships are determined.

[0102] For occluded or missing regions in point clouds where no valid sampling points exist, the structural change trends around the missing regions are correlated with the engineering constraints of the corresponding components to determine the spatial location and extent of the missing regions.

[0103] Based on the spatial location and extent, the geometric change trend of the point cloud collected around the missing area is analyzed, and the surface orientation, boundary contour and spatial size information of the missing area are supplemented.

[0104] The connection positions, installation directions, and spatial spacing between the supplemented component structure and adjacent components are checked. When connection conflicts or inconsistent spatial constraints occur, the process of supplementing the missing areas is adjusted.

[0105] Furthermore, to improve the accuracy of component assembly relationship identification during the reconstruction of the three-dimensional structure of point cloud of complex transformer equipment, considering the influence of factors such as limited scanning angle, equipment occlusion, mutual coverage between components, and the invisibility of some areas in the actual acquisition environment, the acquired point cloud data usually has local missing parts, structural discontinuities, and problems with component spatial relationships. Therefore, it is necessary to analyze the spatial relationship between different component regions after completing the division of component point cloud regions, and use existing structural information to reasonably supplement the missing areas.

[0106] Traditional point cloud reconstruction methods typically perform surface fitting directly on the point cloud surface. When there are missing areas in the point cloud, they can only perform simple interpolation based on local surface changes, which can easily produce compensation results that do not conform to the actual engineering structure. For example, in transformer equipment, components such as bushings, radiators, connecting pipes, and enclosures are not randomly distributed, but are constrained by installation location, connection direction, size, and structural design rules. Relying solely on the spatial proximity of the point cloud cannot accurately determine the actual assembly relationship between components, nor can it guarantee the structural rationality after the missing areas are restored.

[0107] After completing the division of the component point cloud region, different component regions are treated as independent structural units. By analyzing the positional, directional, and connection relationships between components, the point cloud data is transformed from a simple geometric representation into a structural representation with engineering assembly implications.

[0108] For each component point cloud region after region division, calculate the spatial center position, principal direction vector, boundary range, and structural feature parameters for each component region. Also calculate the component spatial center position for each component region.

[0109] ;

[0110] And in conjunction with the main direction:

[0111] ;

[0112] Together they describe the component space state, where, Used to indicate the spatial installation location of components. Used to indicate the direction of extension of the component structure.

[0113] Spatial location and orientation information provide a basis for determining the connection relationship of components. The spatial center position is used to describe the overall spatial distribution of components, the main direction vector is used to represent the installation direction of components or the extension direction of structures, and the boundary range is used to determine the location area where there may be connections or constraints between components.

[0114] Furthermore, based on the spatial relationship between different component areas, the spatial distance relationship between any two component areas is calculated. By analyzing the distance between the center points of the two areas, the minimum distance between the boundaries, and the degree of spatial overlap, it is determined whether there is a direct connection or installation relationship between the two components.

[0115] When the distance between two component regions is small and the boundary region shows a continuous changing trend, the directional consistency between the two is analyzed. By comparing the angle relationship between the main direction vectors of the two components, it is determined whether the components meet the engineering installation direction constraints.

[0116] For two component regions that have a potential connection relationship and Consistency in calculation direction:

[0117] ;

[0118] in, This indicates the angle between the main directions of the two components. , The first and the The main direction vector of each component region, when When the two components are aligned, it means that they meet the condition of having the same direction.

[0119] Calculate the distance between component boundaries :

[0120] ;

[0121] in, and These represent the boundary points of the two components, , These are the boundary point sets of the point cloud regions of the two components, when and If so, then the two components are considered to satisfy the installation connection relationship.

[0122] For example, the connection between a sleeve and a mounting flange typically has the same axial direction; the connection between a radiator and a housing typically has a connection surface with a fixed direction. Therefore, directional consistency analysis can determine whether the components meet the actual assembly rules.

[0123] Furthermore, for components with axial installation characteristics, such as sleeves and connecting pipes, the central axis of the component's point cloud region is first extracted, and the component's axial direction is determined by analyzing the distribution pattern of the point cloud along the axial direction. Then, based on the axial extension direction and the end boundary position, the component's mounting end face and connection position are determined.

[0124] By detecting the point cloud at the edge of the component area, the end contour information is obtained. Combined with the distance relationship between the boundaries of adjacent components, it is determined whether there is an axial connection relationship between the two components. When the axial directions of the two components are consistent, the end distance meets the preset range, and the connection boundary has a structural matching relationship, it is determined that there is an installation connection relationship between the two.

[0125] For components with array arrangement characteristics, such as transformer heat sinks and arrayed connectors, the spatial distance between the center points of each region is calculated based on the positional distribution pattern of multiple component regions. Then, the spatial arrangement axis is fitted according to the arrangement direction of the center points to obtain the component arrangement direction, arrangement spacing and quantity relationship.

[0126] Even if some components are missing from the point cloud due to occlusion, the location range of the missing components can be inferred by the arrangement pattern among other complete components, providing constraints for subsequent structural compensation.

[0127] For components with symmetrical structural features, such as the structures on both sides of a box or symmetrically installed parts, the center position of the overall structure is determined by analyzing the positional relationship between multiple component regions based on the spatial distribution of the point cloud region, and the symmetry constraints are used to determine the possible location and shape of the missing region.

[0128] Furthermore, for point cloud missing areas caused by occlusion or scanning angle limitations, instead of using a simple data filling method, structural restoration is performed by combining component engineering structural constraints. The missing positions in the point cloud area are detected, and the boundary range of the missing area is determined by analyzing the spatial distribution trend of the existing point clouds around the missing area. The boundary extension direction is determined based on the changes in the normal direction, curvature, and boundary extension direction of the surrounding point clouds.

[0129] For example, when the missing region is located in a planar structure, the planar extension of the missing region is performed based on the planar parameters obtained by fitting the surrounding point cloud; when the missing region is located in a regular curved surface structure, a curved surface fitting is established based on the curvature change trend of the surrounding point cloud; when the missing region belongs to a connected transition region, the transition structure is determined by combining the connection constraint relationship between adjacent components.

[0130] Furthermore, after completing the initial compensation for the missing regions, the structural consistency of the compensation results is checked. By comparing the geometric continuity between the compensated region and the surrounding valid point cloud, the compensation results are judged, and the curvature of the local region is calculated. :

[0131] ;

[0132] in, , , The eigenvalues ​​represent the spatial distribution matrix of the local point cloud.

[0133] Calculate the curvature difference between the compensation region and the adjacent region :

[0134] ;

[0135] in, To compensate for the curvature of the region, For the curvature of the neighborhood of the original point cloud, when When this is the case, it indicates that the compensation area meets the surface continuity requirement.

[0136] Simultaneously calculate the spatial position deviation:

[0137] ;

[0138] in, This represents the Euclidean distance between the center of the compensation region and the center of the prediction region. To compensate the regional center, As the original structure prediction center, when When the compensation result satisfies the spatial constraints, the restoration result of the missing region is verified to avoid unreasonable structures caused by simple interpolation.

[0139] The verification process includes: checking the surface continuity between the compensation area and the surrounding known point cloud to determine if there are any obvious curvature abrupt changes; checking whether the spatial distance between the compensated component and adjacent components meets the installation constraints; checking whether the orientation of the compensated component conforms to the actual equipment structure relationship; and when connection conflicts, unreasonable spatial positions, or insufficient structural continuity are detected in the compensation results, the range of the missing area is readjusted or the compensation parameters are optimized until the preset engineering constraints are met.

[0140] This enables the transformation from point cloud spatial relationship analysis to engineering assembly relationship reasoning. By introducing engineering constraints to assist in the recovery of missing areas, the reconstruction results no longer rely on simple data fitting, but conform to the design rules of real transformer structures.

[0141] S4: Based on the structural characteristics of the components, a parametric description method is established to separate and analyze point cloud acquisition errors, manufacturing errors, and assembly errors, and to correct the impact of these errors on the geometric description of the components.

[0142] In step S4, error feature analysis is performed on the collected component point cloud data. Based on the point cloud coordinate offset pattern, point cloud distribution change characteristics, and measurement environment parameters, a spatial change model corresponding to the point cloud acquisition error is established.

[0143] By combining the differences between component design parameters, processing dimension parameters, and actual point cloud geometric parameters, dimensional deviations and geometric deviations generated during component manufacturing are identified.

[0144] Based on the spatial relationships, connection postures, and assembly constraints between components, the positional offsets and posture errors generated during the assembly process are analyzed, and errors from different sources are classified and processed.

[0145] Error analysis was performed on the coordinate offset, measurement noise, and point cloud density changes generated during point cloud acquisition. Based on the relationship between the error sources and spatial coordinate changes, the point cloud geometric parameters were corrected.

[0146] The component's dimensional parameters and spatial relationships are adjusted by taking into account dimensional and geometric deviations during component manufacturing, as well as positional offsets and attitude changes during component installation.

[0147] Furthermore, in order to improve the realism and accuracy of the three-dimensional geometric description results of transformer components, considering that the data obtained in the actual point cloud acquisition process does not fully reflect the actual structure of the components and contains a variety of error factors introduced by the measurement equipment, acquisition environment, component manufacturing process and actual assembly process, it is necessary to identify, separate and correct errors from different sources to avoid the superposition of different error types causing the geometric parameter description of the components to deviate from the actual situation.

[0148] Traditional point cloud processing methods typically treat all point cloud deviations as random noise and filter them out. However, in actual transformer component testing, some errors are not invalid information. For example, dimensional deviations, processing errors, and form and position errors generated during component manufacturing are part of the actual state of the component. Simply removing them would fail to reflect the true operating state of the equipment. Measurement errors caused by the accuracy of the scanning equipment, changes in the acquisition distance, and environmental factors are non-structural errors that need to be corrected. Based on the different sources of error, point cloud errors are divided into acquisition errors, manufacturing errors, and assembly errors to achieve decoupling processing of different error factors.

[0149] This paper analyzes the measurement errors generated during point cloud acquisition. Since the same component area may be scanned from multiple perspectives, the point cloud data acquired from different perspectives usually have a certain positional offset in the spatial overlapping area. The overlapping area between point clouds from multiple perspectives is used as the basis for error analysis. By comparing the spatial coordinate differences between corresponding points under different perspectives, the offset direction and offset magnitude of the point cloud are analyzed.

[0150] Furthermore, based on the variation law of point cloud offset in the spatial coordinate system, the point cloud coordinate offset is decomposed into error components in three directions: X, Y and Z. The error variation trend at different spatial locations is analyzed. When the error shows a consistent offset throughout the entire component area, it is judged as coordinate transformation error or scanning reference error. When the error exists only in a local area and fluctuates with the change of spatial location, it is judged as error caused by local measurement noise or point cloud density change.

[0151] The spatial variation model takes local features of the point cloud as input and the 3D coordinate offset of each sampling point as output. It establishes a mapping relationship between input features and output offset through a multilayer perceptron network structure. The model training adopts a supervised learning approach, using training samples with known coordinate offsets for optimization. During the training process, error regression loss, local smoothing loss, and anomaly robustness loss are used in combination to ensure the accuracy of the predicted offset, the continuous change of the prediction results of adjacent points, and to suppress the impact of large-scale errors on the model, respectively. After training, the coordinate error of each point can be predicted based on the local features of the input point cloud, which is used for subsequent point cloud coordinate correction and error separation.

[0152] After completing the acquisition error analysis, the dimensional and geometric errors generated during the component manufacturing process are identified, and the point cloud geometric parameters after acquisition error correction are compared with the component's theoretical design parameters.

[0153] For components with regular geometric features, such as cylindrical, planar, and symmetrical structures, actual dimensional parameters are obtained by fitting the actual point cloud surface and compared with theoretical dimensional parameters. Manufacturing error indicators such as component diameter deviation, length deviation, flatness deviation, and roundness deviation are calculated. For complex curved surface structures, the distribution of local morphological deviations is obtained by analyzing the distance change between the fitted surface of the point cloud and the theoretical surface. For cylindrical components, the least squares method can be used to fit the cylinder axis to calculate the diameter deviation and roundness deviation. For planar components, the flatness deviation can be calculated by fitting the plane. For complex curved surface structures, the distribution of local morphological deviations can be statistically analyzed by calculating the normal distance from the point cloud to the theoretical surface.

[0154] Furthermore, the assembly errors generated during the transformer assembly installation process are analyzed. Since each component may experience certain positional offsets and attitude changes during actual equipment installation, the installation error cannot be determined solely by the geometric characteristics of a single component; it is necessary to analyze the spatial constraints between components.

[0155] The theoretical connection relationships and actual spatial position relationships between components are obtained. By calculating the transformation relationship between the two component coordinate systems, the spatial offset of the components is decomposed into translation error and rotation error. Translation error describes the spatial offset between the actual installation position and the theoretical installation position of the component. Rotation error describes the angular deviation between the installation direction and the theoretical direction of the component. For example, for a sleeve assembly with an axial installation relationship, the installation posture error can be obtained by comparing the angle between the actual axial direction and the theoretical installation direction. For an array of heat dissipation components, the arrangement spacing error and the overall installation offset are obtained by analyzing the deviation of the center position of each component. Translation error can be described by displacement components in three-dimensional space, and rotation error can be described by rotation angle parameters. The error parameters can be obtained by solving the rigid transformation between the two component coordinate systems.

[0156] Furthermore, based on the spatial characteristics and generation mechanisms of different error types, errors are classified and corrected. For acquisition errors, correction is achieved through coordinate transformation, spatial registration optimization, and point cloud relocation to unify point clouds from different perspectives into an accurate spatial coordinate system. For manufacturing errors, the true structural information is retained instead of being simply eliminated; instead, it is recorded during the parameter description process to reflect the actual manufacturing state of the components. For assembly errors, optimization and adjustment are performed based on the engineering constraints between components to ensure that the spatial relationships of the components meet the actual installation requirements. In the absence of external high-precision ground truth or standard CAD models, classification and correction are achieved through joint optimization: the distance error between corresponding points in the overlapping areas of multi-view point clouds, the statistical deviation of component repeating structural parameters and the violation of soft constraints of design parameters, and the constraint violation between component assembly features are comprehensively weighted to construct a joint optimization objective. By alternately optimizing and solving this objective, various error components are gradually separated, realizing the transformation of point cloud errors from a mixed state to a classified state.

[0157] During the error correction process, the corrected component geometric parameters are re-verified. When there is still a large deviation between the correction and the original point cloud, the source of error is re-determined based on the deviation distribution, and the corresponding error correction parameters are adjusted to realize the transformation of point cloud error from "mixed state" to "classified state". This allows different error sources to be processed separately, avoiding the problem of loss of real structural information due to over-filtering in traditional methods.

[0158] The spatial variation model takes local point cloud features as input and the 3D coordinate offsets of each sampling point as output. A multilayer perceptron network structure establishes a mapping relationship between input features and output offsets. Supervised learning is employed for model training, utilizing training samples with known coordinate offsets for optimization. During training, error regression loss, local smoothing loss, and anomaly robustness loss are used comprehensively to ensure accurate predicted offsets, continuous variation in predicted results for adjacent points, and to suppress the impact of large-scale errors on the model, respectively. After training, the coordinate errors of each point can be predicted based on the local features of the input point cloud for subsequent point cloud coordinate correction and error separation.

[0159] S5: Match and verify the corrected component geometric description with the actual collected point cloud, and adjust the error correction parameters and surface fitting process according to the matching results; if the verification finds obvious errors in the component segmentation, return to the point cloud segmentation step for reprocessing to form the transformer component point cloud segmentation and 3D reconstruction results.

[0160] In step S5, based on the component point cloud region division results, spatial coordinate transformation and structural relationship matching are performed on different component regions to establish the spatial positional relationship between components.

[0161] By combining component geometric information, assembly relationships, and error-corrected geometric parameters, surface fitting is performed on the component surface structure, and the process is completed based on the connection relationships between components.

[0162] Furthermore, in order to accurately reflect the actual equipment structure status, considering that the structural supplementation and error correction process may generate certain cumulative deviations, it is necessary to establish a matching and verification mechanism based on the actual collected point cloud feedback to comprehensively verify and dynamically optimize the error correction and surface fitting process. The component segmentation result, as the underlying data-driven result, will not be adjusted through reconstruction feedback once determined. If the verification finds that the segmentation is incorrect, it is necessary to return to the point cloud segmentation and reprocess it, thus forming a closed-loop point cloud segmentation and 3D reconstruction process.

[0163] Considering that traditional 3D reconstruction methods usually adopt a one-time processing flow, that is, directly outputting a 3D model after completing point cloud segmentation and fitting, without feedback verification of intermediate processing results, structural deviations may occur when there are errors in the initial region segmentation, unreasonable compensation for missing regions, or inaccurate error correction parameters. However, these deviations cannot be corrected through subsequent steps. Therefore, a point cloud-model bidirectional matching and verification mechanism is introduced to perform reverse verification by actually collecting point clouds and adjust the preprocessing process based on the verification results.

[0164] The geometric descriptions of the components, after being divided into regions, supplemented with structural details, and corrected for errors, are transformed into a unified spatial coordinate system. The overall spatial assembly is then completed based on the assembly relationships between the components. Surface fitting is performed on each component region, with the appropriate geometric fitting method selected according to the different structural characteristics of the components: for regular planar regions such as housings and support structures, planar fitting is used; for rotating structures such as bushings and pipes, axis extraction and surface fitting are used; for complex connection regions, a continuous transition structure is established through local surface fitting and boundary constraints. Based on the component relationships, different components are spatially assembled. By adjusting the positional relationships, connection directions, and spatial spacing between components, the actual transformer structural layout is achieved.

[0165] Furthermore, the generated 3D model is remapped to the point cloud acquisition coordinate system, and spatial matching analysis is performed with the original acquired point cloud. The distance error between the surface and the actual point cloud is calculated, and matching evaluation indicators are established, including: the average distance between surface points and the actual point cloud; the maximum deviation in local areas; the degree of matching of component boundary positions; and the spatial consistency of component connection positions. The indicators are used to determine whether the 3D reconstruction results meet the preset accuracy requirements.

[0166] When a significant deviation is detected between the data and the point cloud, the cause is analyzed based on the error distribution. If the deviation is mainly concentrated in the component boundary area, it indicates that there may be an error in the component boundary identification during the region division process, and the region division result needs to be re-optimized. If the deviation is mainly concentrated in the missing area, it indicates that the structural compensation process does not conform to the actual geometric changes, and it needs to be readjusted. If the deviation is manifested as an overall spatial offset, it indicates that there is a parameter deviation in the error correction process, and it needs to be re-optimized. This transforms the entire 3D reconstruction process from a unidirectional processing to a cyclic optimization process, thereby improving accuracy.

[0167] Furthermore, after meeting the matching accuracy requirements, structural information is integrated, and the component geometric parameters, spatial position relationships, assembly relationships, and error correction information are stored in a unified manner. This not only represents the three-dimensional shape of the equipment surface, but also reflects the connection relationships, installation relationships, and actual error states between components. This provides basic data for subsequent equipment inspection, fault analysis, and intelligent maintenance, avoiding the distortion problem caused by the accumulation of local errors in traditional methods. This improves the completeness, accuracy, and engineering applicability of the three-dimensional reconstruction results of complex transformer equipment.

[0168] A 3D model refers to a 3D surface representation obtained by surface fitting and structural reconstruction of component point clouds. It adopts a neural network structure based on point cloud encoding and decoding. The encoder is used to extract global geometric features from the point cloud, and the decoder is used to generate distance field values ​​or occupancy probabilities of each point in space based on the features, thereby extracting the 3D mesh surface. The training adopts a supervised learning method and optimizes using a dataset with known 3D surface shapes. The loss function includes point cloud distance loss, normal consistency loss, and edge smoothing loss, which are used to constrain the reconstructed surface to be consistent with the actual point cloud in terms of geometric shape, surface orientation, and boundary continuity. After training, the component surface model can be automatically generated based on the input point cloud, realizing the conversion from discrete point cloud to 3D model.

[0169] Example 2

[0170] Figure 1 The flowchart below shows a deep learning-based method for point cloud segmentation and 3D reconstruction of transformer components according to an embodiment of this application. The method includes: extracting multi-scale features from the input point cloud of the transformer component, obtaining local spatial features and overall structural features in the point cloud region respectively, and fusing the point cloud features at different scales.

[0171] Based on the fused point cloud features, component classification features are constructed, and different transformer component regions are segmented according to the spatial feature relationships corresponding to each point cloud sampling point.

[0172] Establish local coordinate relationships based on the segmented component point cloud regions, identify key structural points in the component point cloud, and determine the component spatial structure based on the positional relationships between key structural points.

[0173] By combining the structural features of the component with the preset geometric model parameters, surface fitting and structural matching are performed on the point cloud on the component surface to complete the construction of the component's three-dimensional model.

[0174] Furthermore, to improve the accuracy of feature extraction and component region identification in complex transformer component point cloud data, considering that the actual transformer component point clouds usually have irregular distribution characteristics and significant differences in point cloud density in different regions, and are also affected by equipment occlusion, scanning angle changes, and the complexity of component surface structure, traditional fixed-scale feature description methods are difficult to simultaneously take into account local structural details and overall component morphology information. Therefore, a multi-scale feature learning approach based on deep learning is adopted. By automatically extracting and fusing point cloud features at different spatial scales, the network can fully learn the local geometric change features and overall structural features contained in the transformer component point cloud.

[0175] The preprocessed transformer component point cloud is used as input data. The neighborhood relationship of the point cloud is established based on the spatial coordinate information of each sampling point. Considering that the point cloud data does not have a fixed arrangement order, the spatial position relationship of the sampling points is used as the basis for feature learning during the feature extraction process. By analyzing the changes in spatial distance, direction and local structure between each sampling point and its surrounding neighboring points, basic feature information that can reflect the surface morphology of the component is extracted.

[0176] Furthermore, to avoid the problem of missing local information or insufficient expression of the overall structure during single-scale feature extraction, neighborhood information of different ranges is collected for the point cloud region according to the different structural scale characteristics of the transformer components. For neighborhood points in a smaller spatial range, local change information of the component surface is obtained through feature learning to describe detailed structures such as connection edges, mounting holes, and curved transition areas. For neighborhood points in a larger spatial range, the overall spatial distribution relationship of the component is learned, such as the component length direction, installation direction, and overall contour features.

[0177] During feature extraction, point cloud feature information obtained at different scales is comprehensively analyzed, and local spatial features are associated with overall structural features, enabling the network to understand both the detailed structure and the overall component morphology in the point cloud. For the same sampling point, feature information from different neighborhood ranges is fused to form a comprehensive feature description that includes spatial positional relationships, geometric change trends, and structural context information.

[0178] To avoid the missegmentation problem caused by traditional point cloud segmentation methods that rely solely on local geometric information, for example, at the location where the transformer housing and connecting accessories are in contact, since the spatial distance between the two areas is small, it is easy to classify them as the same area based solely on the spatial distance. However, the overall structural information obtained through deep learning can be combined with the complete shape of the component to determine that the two belong to different functional areas, thereby improving the component boundary recognition capability.

[0179] Furthermore, after obtaining the fused point cloud features, the transformer component region is classified and segmented according to the spatial feature relationships corresponding to different sampling points. Each point cloud sampling point is taken as the object to be classified, and the component category to which the sampling point belongs is determined according to the corresponding spatial features, local geometric features and overall structural features.

[0180] For regions with obvious regular structural features, such as the outer shell, heat dissipation structure, and sleeve structure, the system learns the spatial distribution patterns and geometric differences between different components to automatically distinguish between different component regions. For regions with occlusion or partial missing parts, the system combines the feature information of the surrounding complete region to infer the component category to which the missing region belongs.

[0181] Furthermore, to improve the correspondence between point cloud segmentation results and actual engineering components, after completing the initial point cloud region division, the internal structure of each component region is analyzed. Based on the segmented component point cloud regions, local spatial relationships are established, and the positional distribution, directional changes, and structural continuity of the point clouds within the regions are analyzed to determine the key structural points that can reflect the overall shape of the components.

[0182] Key structural points can include component boundary points, structural turning points, connection points, and points with significant geometric changes. By identifying key structural points, the impact of a large number of redundant sampling points in point cloud data on structural description can be reduced, while improving the computational efficiency in the subsequent 3D model construction process.

[0183] Furthermore, the spatial structure of the components is determined based on the positional relationships between key structural points. In practice, the distance, direction, and spatial arrangement relationships between different key structural points are analyzed to obtain structural information such as the component's length direction, installation direction, boundary range, and connection positions. For example, for a transformer bushing assembly with axial structural features, the direction of the component's central axis can be determined by analyzing the positional relationships between the key structural points at both ends; for a heat dissipation assembly with repeating arrangement features, the arrangement pattern of the components can be determined by analyzing the spatial spacing between multiple key structural points; and for components with connection relationships, the installation relationship between components is determined based on the positional relationships between key points at the connection positions.

[0184] Furthermore, after obtaining the component structure information, the component structure features and preset geometric parameters are combined to perform surface fitting and structural matching on the component surface point cloud. The corresponding geometric description method is selected according to different component types. For component regions with regular geometric shapes, basic geometric parameters are determined based on key structural points, and the point cloud surface information is used for fitting to make the generated surface more closely match the actual collected data. For complex curved surface structures, continuous surface fitting is performed on local areas according to the point cloud distribution pattern, and structural constraints are used to ensure the consistency of the connection relationship between different regions.

[0185] During the structural matching process, the fitted component geometry is compared with the actual point cloud data. The parameters are adjusted based on the positional deviation, surface continuity, and boundary consistency between the two. When a large deviation is found in a local area, the process of point cloud feature extraction, component segmentation, or key structure identification is returned to optimize the process based on the deviation location, so that the final 3D model can more accurately reflect the actual component structure.

[0186] By using deep learning methods to automatically learn multi-scale spatial information from point clouds of transformer components, and combining key structure identification and geometric structure matching processes, the automatic conversion from disordered point cloud data to 3D models of engineering components is achieved.

[0187] To address the shortcomings of traditional point cloud processing methods, such as reliance on manually designed features, difficulty in adapting to complex equipment structural changes, and insufficient segmentation accuracy under occlusion conditions, deep learning is used to automatically extract spatial features from point clouds, improving the ability to identify component regions. At the same time, structural information is combined to complete the 3D model reconstruction, resulting in a transformer component model with higher integrity, accuracy, and engineering application value.

[0188] A geometric model is a parameterized representation of the basic shape and size of a transformer component. It reflects the structural characteristics of the component through a set of geometric parameters. A lightweight encoder-regressor network structure is adopted. The encoder is used to extract structural features from the point cloud, and the regressor is used to predict the geometric parameters of the corresponding component. Different types of components output different parameter combinations. The training adopts a supervised learning method, using point cloud data with manually labeled geometric parameters as the training set. During the training process, parametric regression loss, surface consistency loss, and engineering constraint loss are used in combination to ensure the accuracy of the predicted parameters, the closeness of the geometric surface generated based on the predicted parameters to the actual point cloud, and the compliance of the parameter range with engineering specifications, respectively. After training, the model can directly predict the geometric parameters of the component based on the input point cloud and match them with the preset geometric model to complete surface fitting and structural matching.

Claims

1. A method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning, characterized in that, The specific steps include the following: Point cloud data of transformer components from different perspectives are collected, and the point cloud data is registered, denoised and coordinate unified to form component point cloud data, and geometric structure information in the point cloud is extracted. Combining the structural features and engineering semantic information of transformer components, the point cloud of the components is divided into regions to distinguish the point cloud regions corresponding to different functional components and to establish the correspondence between the component point cloud and the actual engineering parts. Analyze the positional and connectivity relationships between the point cloud regions of each component, determine the assembly association relationships between components, and supplement the structure of incomplete regions caused by occlusion or missing data acquisition. Based on the structural characteristics of the components, a parametric description method is established to separate and analyze point cloud acquisition errors, manufacturing errors, and assembly errors, and to correct the impact of these errors on the geometric description of the components. The corrected component geometric description is matched and verified with the actual collected point cloud. Based on the matching results, the error correction parameters and surface fitting process are adjusted. If the verification reveals obvious errors in the component segmentation, the process returns to the point cloud segmentation step for reprocessing, resulting in the point cloud segmentation and 3D reconstruction of the transformer component.

2. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The extracted geometric structure information from the point cloud includes: Neighborhood spatial analysis is performed on the preprocessed component point cloud. Based on the spatial distance relationship, distribution relationship and local topological relationship between sampling points, the spatial structure characteristics of the point cloud region are established. Based on the changes in the surface morphology of the point cloud, the curvature features, boundary features, and spatial orientation features within the point cloud region of the component are calculated, and the surface structure of the component is described in combination with the local geometric change rules.

3. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The establishment of the correspondence between component point clouds and actual engineering components includes: The preprocessed transformer component point cloud is divided into regions according to spatial connectivity. For multiple independent point cloud regions formed due to occlusion, scanning angle limitation or surface discontinuity, a region-level structure description is established. Within each independent point cloud region, the spatial distribution of the point cloud is analyzed to obtain the main direction, surface change trend, spatial scale parameters, and edge contour features of the point cloud within the region. For the component association relationships between different independent point cloud regions, a matching analysis is performed on the structural orientation consistency, size ratio relationship and spatial arrangement relationship of each region. When multiple independent point cloud regions meet the preset structural association conditions, the multiple independent point cloud regions are associated as the same component region. For point cloud regions that are in contact with each other but belong to different engineering components, the surface orientation changes, structural continuity and connection forms at the boundary of the regions are analyzed. When the adjacent regions do not meet the structural continuity condition, the adjacent regions are separated.

4. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The determination of the assembly association relationships between components includes: After the point cloud region is divided, spatial position analysis is performed on the different component regions, and spatial relationships are established based on the distance, direction and relative position relationships between the component regions. Based on the structural characteristics of transformer assembly installation, the connection boundaries, installation directions and spatial constraints between assembly areas are determined, and the assembly relationships are determined according to the direct connection relationships or spacing relationships between the assemblies.

5. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The structural supplementation of incomplete areas caused by occlusion or missing data includes: For the associated component point cloud region, analyze the spatial connection status between the component and adjacent components to determine the installation position relationship, connection direction relationship and mutual constraint relationship between the components; Based on the structural characteristics of transformer components, the engineering constraints corresponding to different types of components are determined. Specifically, for components with axial installation characteristics, the axial direction and end connection positions are determined; for components with array arrangement characteristics, the arrangement direction, spacing relationship, and quantity constraints are determined; and for components with symmetrical structural characteristics, the structural center position and symmetry relationship are determined. For occluded or missing regions in point clouds where no valid sampling points exist, the structural change trends around the missing regions are correlated with the engineering constraints of the corresponding components to determine the spatial location and extent of the missing regions. Based on the spatial location and range, the geometric change trend of the point cloud collected around the missing area is analyzed, and the surface orientation, boundary contour and spatial size information corresponding to the missing area are supplemented. The connection positions, installation directions, and spatial spacing between the supplemented component structure and adjacent components are checked. When connection conflicts or inconsistent spatial constraints occur, the process of supplementing the missing areas is adjusted.

6. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The separation and analysis of point cloud acquisition error, manufacturing error, and assembly error includes: Error feature analysis was performed on the collected component point cloud data. Based on the point cloud coordinate offset law, point cloud distribution change characteristics and measurement environment parameters, a spatial change model corresponding to the point cloud acquisition error was established. By combining the differences between component design parameters, processing dimension parameters, and actual point cloud geometric parameters, dimensional deviations and form and position deviations generated during component manufacturing are identified. Based on the spatial relationships, connection postures, and assembly constraints between components, the positional offsets and posture errors generated during the assembly process are analyzed, and errors from different sources are classified and processed.

7. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The impact of the correction error on the component's geometric description includes: Error analysis was performed on the coordinate offset, measurement noise, and point cloud density changes generated during point cloud acquisition. Based on the relationship between the error sources and spatial coordinate changes, the point cloud geometric parameters were corrected. The component's dimensional parameters and spatial relationships are adjusted by taking into account dimensional and geometric deviations during component manufacturing, as well as positional offsets and attitude changes during component installation.

8. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, The point cloud segmentation and 3D reconstruction results of the transformer components include: Based on the component point cloud region division results, spatial coordinate transformation and structural relationship matching are performed on different component regions to establish the spatial positional relationship between components; By combining the component's geometric structure information, assembly relationships, and error-corrected geometric parameters, the surface structure of the component is fitted with a surface, and a three-dimensional model of the transformer component is constructed based on the connection relationships between the components.

9. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, Multi-scale feature extraction is performed on the input transformer component point cloud to obtain local spatial features and overall structural features in the point cloud region, and the point cloud features at different scales are fused. Based on the fused point cloud features, component classification features are constructed, and different transformer component regions are segmented according to the spatial feature relationships corresponding to each point cloud sampling point.

10. The method for point cloud segmentation and 3D reconstruction of transformer components based on deep learning according to claim 1, characterized in that, Local coordinate relationships are established based on the segmented component point cloud regions, key structural points in the component point cloud are identified, and the spatial structure of the component is determined based on the positional relationships between the key structural points. By combining the structural features of the component with the preset geometric model parameters, surface fitting and structural matching are performed on the point cloud on the component surface to complete the construction of the component's three-dimensional model.

Images