Transformer substation scene-oriented point cloud denoising method and system
By using a deep learning-based point cloud iterative enhancement filtering network, the efficiency and accuracy issues of point cloud denoising in substation scenarios are solved, adapting to complex scenarios and supporting high-precision 3D modeling of smart grids.
Patent Information
- Application Number
- CN202511183554.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-22
- Publication Date
- 2025-11-25
AI Technical Summary
Existing point cloud denoising methods have failed to effectively cope with the high-density equipment layout and heterogeneous structural complexity in substation scenarios, resulting in insufficient noise suppression capabilities, insufficient preservation of local features, and insufficient adaptability to complex scenarios, which affects the accuracy and efficiency of 3D modeling.
We employ a deep learning-based iterative enhancement filtering network for point clouds. Through multi-source remote sensing data preprocessing and flow field-driven convolutional filtering, we achieve detail preservation and few-sample generalization modeling for point cloud data in substation scenarios, thus constructing a highly adaptive denoising method.
It improves the efficiency, accuracy, and adaptability of point cloud denoising in substation scenarios, achieving efficient point cloud denoising effects and supporting the high-precision 3D modeling requirements of smart grids.
Smart Images

Figure CN121010523A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional point cloud data processing technology, and in particular to a method and system for point cloud denoising in substation scenarios. Background Technology
[0002] With the rapid development of smart grid construction, the power grid system is undergoing a transformation and upgrade centered on digitalization and intelligence. As a key hub of the power transmission network, the scale of equipment in substations continues to expand, and the safety and operational efficiency of their equipment directly affect the stability of the power grid. Currently, substation operation and maintenance is transitioning from traditional manual methods to intelligent monitoring; however, significant challenges are faced in actual operation. The equipment layout within substations is complex, with narrow spacing between insulators, circuit breakers, current transformers, voltage transformers, and other equipment, which operate energized for extended periods. Manual inspections require close contact with energized areas, posing risks of electric shock and other safety hazards. Simultaneously, equipment status monitoring needs to meet the demands of high-frequency, 24 / 7 real-time monitoring, where traditional manual methods have significant limitations in terms of timeliness and continuity.
[0003] Achieving high-frequency, all-weather intelligent monitoring requires a high-precision 3D model as its digital foundation. Currently, the mainstream modeling technologies for substations mainly fall into two categories: one is active scanning based on ground-based or airborne LiDAR, which acquires 3D point clouds with centimeter-level accuracy by emitting laser pulses; the other is UAV oblique photogrammetry, which generates 3D models through multi-view image matching. Compared to LiDAR solutions, UAV oblique photogrammetry has become the preferred solution for substation modeling due to its advantages such as large coverage area per operation and low overall cost. However, in practical applications, this technology faces two challenges in substation scenarios: firstly, strong electromagnetic environments can easily interfere with UAV positioning signals, reducing flight stability; secondly, the complex background formed by dense equipment can easily cause mismatches of image feature points, compounded by factors such as reflections from equipment metal surfaces and flight path deviations caused by airflow disturbances, resulting in point cloud data filled with a large amount of redundant noise, outliers, and artifacts. This noise not only consumes storage resources but also introduces errors in subsequent mesh generation and feature extraction processes. Therefore, point cloud denoising has become an indispensable and crucial preprocessing step in the 3D modeling workflow.
[0004] Currently, existing denoising methods are mostly designed for general urban building or terrain mapping scenarios, and fail to fully consider the challenges brought about by the high-density layout and complex heterogeneous structure of substation equipment. Therefore, it is urgent to develop a dedicated point cloud denoising method for substation scenarios. By breaking through the core technical bottlenecks such as detail preservation under high-density layout and generalized modeling with few samples, it can better support the demand for high-precision and high-reliability 3D modeling in the digital transformation of smart grids. Summary of the Invention
[0005] This invention provides a point cloud denoising method and system for substation scenarios, solving the technical problem of how to improve the efficiency, accuracy, and adaptability of point cloud denoising in substation scenarios to complex scenarios.
[0006] The first aspect of this invention provides a point cloud denoising method for substation scenarios, comprising:
[0007] Acquire multi-source remote sensing data of substations;
[0008] The multi-source remote sensing data is processed into point cloud data to obtain target point cloud data;
[0009] The target point cloud data is subjected to flow field-driven convolutional filtering by multiple iterative filtering modules connected in series within the point cloud iterative enhancement filtering network, thereby obtaining the filtered point cloud result.
[0010] Optionally, the multi-source remote sensing data includes UAV oblique imagery and multi-station radar point cloud data. The step of performing point cloud normalization on the multi-source remote sensing data to obtain target point cloud data includes:
[0011] The UAV tilted imagery is subjected to aerial triangulation encryption to obtain optimized tilted imagery point cloud data;
[0012] The multi-station radar point cloud data is denoised and downsampled to obtain registered multi-station point cloud data.
[0013] The optimized oblique image point cloud data and the registered multi-site point cloud data are denoised and integrated to obtain the initial point cloud data;
[0014] The initial point cloud data is divided into grids to obtain the target point cloud data.
[0015] Optionally, the iterative filtering module includes an encoder and a decoder, and the flow field-driven convolutional filtering process specifically includes:
[0016] The encoder is used to analyze the flow field parameter features of the input target point cloud data to obtain coded point cloud data;
[0017] The decoder is used to decode the encoded point cloud data to obtain decoded point cloud data;
[0018] The decoded point cloud data is subjected to Gaussian weighted fusion of overlapping areas to obtain optimized target point cloud data.
[0019] Optionally, the encoder includes a first dynamic edge convolution module, a flow field motion parameter estimation module, and a second dynamic edge convolution module connected in sequence. The second dynamic edge convolution module includes multiple first dynamic edge convolution modules connected in sequence. The step of using the encoder to perform flow field parameter feature parsing on the input target point cloud data to obtain coded point cloud data includes:
[0020] The first dynamic edge convolution module is used to extract neighborhood graph convolution features from the input target point cloud data to obtain convolutional point cloud data;
[0021] The flow field motion parameter estimation module is used to calculate the flow field driving parameters of the convolutional point cloud data to obtain three-dimensional velocity field point cloud data.
[0022] The second dynamic edge convolution module is used to extract continuous neighborhood graph convolution features from the three-dimensional velocity field point cloud data to obtain coded point cloud data.
[0023] Optionally, the neighborhood graph convolutional feature extraction specifically includes:
[0024] A local neighborhood graph is constructed using the input point cloud data;
[0025] Nonlinear mapping is performed on the vertex features of the local neighborhood graph to obtain local semantic information;
[0026] Geometric residuals are calculated using the vertex features and associated neighborhood features to obtain residual coding information;
[0027] The extracted point cloud data is obtained by aggregating the local semantic information and the residual coding information.
[0028] Optionally, the flow field motion parameter estimation module includes a velocity module, a distance module, and a multiplicative modulation module, and the flow field driving parameter solution is specifically as follows:
[0029] The velocity module is used to perform hierarchical convolutional feature upscaling encoding on the convolutional point cloud data to obtain the flow velocity vector;
[0030] The distance module is used to perform multi-level aggregation regression encoding on the convolutional point cloud data to obtain a distance scalar;
[0031] The flow velocity vector and the distance scalar are input into the multiplicative modulation module to determine the three-dimensional velocity field point cloud data.
[0032] The second aspect of this invention provides a point cloud denoising system for substation scenarios, comprising:
[0033] The acquisition module is used to acquire multi-source remote sensing data of the substation;
[0034] The processing module is used to perform point cloud normalization on the multi-source remote sensing data to obtain target point cloud data;
[0035] The output module is used to sequentially perform flow field-driven convolutional filtering on the target point cloud data through multiple iterative filtering modules connected in series within the point cloud iterative enhancement filtering network, so as to obtain the filtered point cloud result.
[0036] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the point cloud denoising method for substation scenarios as described in any of the preceding claims.
[0037] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the point cloud denoising method for substation scenarios as described in any of the preceding claims.
[0038] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the point cloud denoising method for substation scenarios as described in any of the preceding claims.
[0039] As can be seen from the above technical solutions, the present invention has the following advantages:
[0040] This invention acquires multi-source remote sensing data from substations, generates target point clouds through point cloud normalization, and performs flow field-driven convolutional filtering based on iterative filtering modules connected in series within the point cloud iterative enhancement filtering network. It guides the migration of noise straight-line paths through flow field modeling and performs local-global correction through multi-module iterative optimization, adapting to the high-density layout characteristics of substation equipment and improving the efficiency, accuracy, and adaptability of point cloud denoising to complex scenarios. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 A flowchart illustrating the steps of a point cloud denoising method for substation scenarios provided in this embodiment of the invention;
[0043] Figure 2 This is a schematic diagram of a point cloud iterative enhancement filter network structure that adopts a four-level serial iterative module architecture.
[0044] Figure 3 This invention provides a structural block diagram of a point cloud denoising system for substation scenarios.
[0045] Figure 4This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0046] This invention provides a point cloud denoising method and system for substation scenarios, which addresses the technical problem of how to improve the efficiency, accuracy, and adaptability of point cloud denoising in substation scenarios to complex scenarios.
[0047] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0048] In substation scenarios, while filtering algorithms have effectively improved the reliability of substation equipment condition monitoring and transmission line fault diagnosis, their practical engineering deployment still faces multiple technical bottlenecks. First, existing algorithms have significant shortcomings in real-time performance, edge adaptability, and robustness. High-order filtering algorithms require substantial computational resources in complex substation scenarios, making it difficult to meet the real-time monitoring needs of power systems. Second, some filtering algorithms require initial state estimation; inaccurate initial values can lead to prolonged convergence or error accumulation, thus affecting the overall filtering accuracy. Third, traditional manual filtering methods in substation scenarios generally suffer from low efficiency and high labor costs.
[0049] In recent years, the application of deep learning-based point cloud denoising algorithms in substation scenarios has faced significant technical bottlenecks. Existing methods are mostly designed for general urban building or topographic mapping scenarios, failing to fully consider the challenges posed by the high-density layout and complex heterogeneous structures of substation equipment: traditional voxel grid filtering methods rely on preset voxel sizes, making it difficult to adaptively preserve details of tiny gaps between devices; classification denoising models based on random forests or support vector machines require extensive labeled data training, resulting in insufficient ability to represent the diversity of heterogeneous equipment in substations.
[0050] In addition, point clouds generated by oblique photogrammetry usually contain jagged edges and texture noise from multi-view image stitching. While existing algorithms filter out such noise, they are prone to losing key structural information of the device surface, resulting in increased geometric deviation between the modeling results and the real physical entity.
[0051] Therefore, there is an urgent need to develop a dedicated point cloud denoising method for substation scenarios. This method should overcome key technical bottlenecks such as detail preservation under high-density arrangement and generalized modeling with few samples to better support the demand for high-precision and high-reliability 3D modeling in the digital transformation of smart grids.
[0052] To address the aforementioned technical challenges, this invention proposes a deep learning-based point cloud denoising method specifically for substations. It constructs an iterative enhancement filtering network structure for point clouds, effectively overcoming the shortcomings of traditional methods in noise suppression, local feature preservation, and adaptability to complex scenes, particularly in densely populated areas and complex structural connections. This method achieves breakthroughs in efficiency, accuracy, and generalization, resulting in more efficient point cloud denoising.
[0053] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a point cloud denoising method for substation scenarios provided in this embodiment of the invention.
[0054] This invention provides a point cloud denoising method for substation scenarios, comprising:
[0055] Step 101: Obtain multi-source remote sensing data of the substation.
[0056] Multi-source remote sensing data refers to a data set that combines UAV oblique RGB imagery with ground-based lidar point clouds for substation scenarios.
[0057] In this embodiment of the invention, the multi-source remote sensing data includes UAV oblique images and multi-station radar point cloud data. The UAV equipped with an RGB camera performs multi-angle surround shooting of all equipment in the substation along a preset route, recording multi-view optical information of the equipment's appearance texture and geometric contour as UAV oblique images. The substation is scanned and stitched by ground-based lidar to obtain three-dimensional spatial coordinate data of the equipment surface with millimeter-level precision as multi-station radar point cloud data.
[0058] Step 102: Perform point cloud normalization on the multi-source remote sensing data to obtain the target point cloud data.
[0059] Point cloud normalization refers to the preprocessing operations performed on the raw point clouds generated from multi-source remote sensing data, including image point cloud optimization, radar point cloud preprocessing, cross-source fusion normalization, and grid structuring, to standardize the spatial distribution, data format, and noise level of the point clouds.
[0060] In this embodiment of the invention, the spatial positioning deviation of the UAV tilted image point cloud is specifically corrected by aerial triangulation. After denoising, downsampling and registration processing, the coordinate reference and data density of the multi-station radar point cloud are unified. Then, the redundant noise and format differences between the two types of point clouds are eliminated by cross-source fusion. Finally, a regular spatial neighborhood structure is constructed by grid division to form target point cloud data that meets the requirements of geometric accuracy, noise control and data consistency.
[0061] Furthermore, the multi-source remote sensing data includes UAV oblique imagery and multi-station radar point cloud data, and step 102 may include the following sub-steps:
[0062] S11. Perform aerial triangulation on the UAV tilted imagery to obtain optimized tilted imagery point cloud data.
[0063] In this embodiment of the invention, distortion correction is first performed on the UAV tilted image to eliminate camera intrinsic parameter errors. Then, the same-name feature points of the multi-view image data of the UAV tilted image are extracted and matched using the Scale-Invariant Feature Transform (SIFT) algorithm to obtain a set of same-name point matching. Based on the same-name point matching set, geometric constraint relationships between images are constructed. The camera exterior orientation elements (position, attitude) and the three-dimensional coordinates of the densified points are used as optimization variables to obtain a geometric constraint model. Then, the geometric constraint model is solved iteratively by bundle adjustment to optimize the camera exterior orientation elements (position, attitude) and the three-dimensional coordinates of the densified points, and the optimized parameter set (including camera parameters and densified point coordinates) is output to obtain the Digital Surface Model (DSM) and the Digital Elevation Model (DEM). The Model (DEM) should be explained as follows: DSM / DEM is a model representation of the scene surface based on an optimized parameter set. Then, using the optimized parameter set, a semi-global matching (SGM) algorithm is used to calculate pixel disparity, converting it into a depth map and solving for 3D point coordinates. Image RGB information is then fused to obtain initial color point cloud data. This initial color point cloud data is then imported into 3D point cloud and mesh processing software. Local curvature analysis is used to extract the edge contour features of substation equipment, resulting in point cloud data with edge features. Finally, based on Ground Control Points (GCPs), spatial position errors are calculated on the point cloud data with edge features, combined with Iterative Closest Points (ICPs). The consistency verification of the registration between the Point (ICP) algorithm and the LiDAR point cloud is performed, and the point-by-point error index of the point cloud data is extracted. Combined with the geometric accuracy requirements of substation equipment (such as setting the spatial error threshold of equipment area to ≤5cm and the threshold of terrain area to ≤10cm), the point cloud with edge features is traversed, and the distance error (relative to GCP) and registration residual (relative to LiDAR point cloud) of each point are calculated. Point cloud data with both types of errors exceeding the limit are marked. After removing the marked points, the remaining point cloud is uniformly resampled to optimize the spatial density, and finally the optimized oblique image point cloud data is obtained.
[0064] It is worth mentioning that distortion correction eliminates camera intrinsic parameter errors, ensuring geometric consistency in matching corresponding feature points across multi-view images. Combined with bundle adjustment iterative optimization of camera exterior orientation elements and densified point coordinates, this not only supports the accurate characterization of substation scene surface morphology by digital surface models (DSM) and digital elevation models (DEM), but also provides reliable parameters for the semi-global matching (SGM) algorithm to solve depth point clouds, guaranteeing the spatial coordinate accuracy of the initial color point cloud. Simultaneously, local curvature analysis preserves equipment edge contour features, accurately extracting substation equipment edge contour features, providing a basis for subsequent iterative enhancement filtering network identification of equipment structures. The construction of "surface continuity" constraints (to avoid denoising that destroys the true contours) provides a key basis for this process. Then, through ground control point (GCP) positional error calculation and iterative nearest point (ICP) cross-source registration verification, combined with dual error threshold screening for equipment areas (≤5cm) and terrain areas (≤10cm), the absolute accuracy and cross-source geometric consistency of the point cloud are strictly controlled. This ensures that the filtering network has a reliable coordinate reference when locating spatial offsets of noise points. Finally, through uniform resampling, the oblique image point cloud meets the geometric accuracy requirements for fine modeling of substation equipment while also possessing complete equipment features and uniform density suitable for multi-source fusion. This step comprehensively improves the quality of the oblique image point cloud from three dimensions: coordinate accuracy, structural features, and density distribution.
[0065] S12. Denoise and downsample the multi-station radar point cloud data to obtain registered multi-station point cloud data.
[0066] In this embodiment of the invention, the original multi-station radar point cloud data is analyzed using a statistical outlier filtering algorithm to identify and remove outlier noise points caused by electromagnetic interference, resulting in a de-noiseed point cloud. For the de-noiseed point cloud, a voxel mesh adapted to the scale of substation equipment is constructed. Downsampling is achieved through voxel point cloud compression to reduce data redundancy and retain key structures, resulting in a downsampled point cloud. For the downsampled point cloud, at least four pairs of substation feature points, such as transformer bushing vertices and equipment base corners, are selected in the 3D point cloud and mesh processing software. The initial rigid body transformation matrix is solved based on Singular Value Decomposition (SVD) to complete the coarse-scale coordinate alignment of the multi-station cloud, resulting in registered multi-station cloud data. The multi-station spatial reference is unified through coarse registration.
[0067] It is worth mentioning that statistical outlier filtering specifically identifies and removes outlier noise points caused by electromagnetic interference, avoiding false point clouds generated by electromagnetic interference in radar signals, and preventing noise from being incorrectly retained or amplified during downsampling and registration. Voxel mesh downsampling is adapted to the scale of substation equipment. For example, the voxel resolution is designed according to the size of equipment such as transformers and switchgear. While compressing the amount of data, the voxel point cloud retains the key structural features of the equipment, such as the geometric features of bushing vertices and base corners, providing clear feature support for the selection of corresponding point pairs and coarse registration calculation. Based on the coarse registration of equipment feature points, and with the help of substation-specific geometric features such as transformer bushing vertices and equipment base corners, the rigid body transformation matrix is solved by SVD to quickly unify the spatial reference of multi-station point clouds. This ensures the geometric consistency of the radar point clouds themselves. This step comprehensively improves the quality of radar point clouds from three dimensions: noise purification, data lightweighting, and spatial reference unification.
[0068] S13. Denoise and integrate the optimized oblique image point cloud data and the registered multi-site point cloud data to obtain the initial point cloud data.
[0069] In this embodiment of the invention, based on the feature points of substation equipment, including but not limited to transformer bushing vertices and equipment base corners, cross-source registration is performed on optimized tilted image point cloud data and registered multi-site point cloud data. The geometric coordinates and texture information of the two types of point clouds are fused to obtain a cross-source fused point cloud. For the cross-source fused point cloud, outliers remaining from the registration are manually removed using 3D point cloud and mesh processing software, such as stray points introduced by registration errors. Details such as equipment edges and terrain transition areas are manually refined to obtain a manually refined point cloud. Using the manually refined point cloud as the ground truth and the original cross-source fused point cloud as the data to be processed, a point cloud filtering training dataset is constructed (the point cloud filtering training dataset is used to support the training of subsequent iterative enhancement filtering networks). At the same time, the spatial structure and texture features of the two types of point clouds are integrated to obtain the initial point cloud data.
[0070] It is worth mentioning that cross-source registration relies on the feature points of substation equipment to accurately fuse the texture details (RGB) of the tilted image with the geometric accuracy of the radar point cloud, providing a geometric-texture collaborative constraint basis for denoising. This avoids the loss of information from a single data source. Manual fine-tuning is used to specifically remove residual noise from registration and repair fine structures such as equipment edges. This provides high-precision ground truth annotation for the training dataset and ensures that the true shape of the point cloud is not destroyed by subsequent filtering. The training set and the initial point cloud are constructed simultaneously, which allows the iterative enhancement filtering network to learn the mapping rules of "noise-ground truth" to improve denoising accuracy. Furthermore, by integrating dual-source point clouds, it provides multi-source complementary input data for the filtering process. This step strengthens the core support for point cloud denoising in substation scenarios from three aspects: data quality, algorithm training, and process connection.
[0071] S14. Divide the initial point cloud data into a grid to obtain the target point cloud data.
[0072] In this embodiment of the invention, the geometric center of the substation is selected as the reference point O, and an XY orthogonal grid coordinate system with O as the origin is constructed. The X-axis is defined to point out of the substation, and the Y-axis is perpendicular to the main transformer array axis, thus clarifying the reference framework for point cloud spatial division. The grid is synchronously expanded along the negative X-axis and Y-axis, and the grid interval is set to 40m to generate standard grid units covering the entire substation area, forming a grid division rule set. According to the grid division rule set, the initial point cloud data is allocated to the corresponding grid units according to spatial affiliation, realizing the block organization of the point cloud, reducing the amount of point cloud data input to the filter network in a single operation, and obtaining the target point cloud data. By encapsulating the data in grid units, it is adapted to the block parallel processing of the subsequent iterative enhancement filter network.
[0073] It is worth mentioning that by taking the substation outgoing line direction as the X-axis and the perpendicular line of the main transformer array axis as the Y-axis, the grid naturally fits the equipment layout, avoiding grid cutting of the equipment body. This provides a spatial basis for the iterative enhancement filtering network to preserve the complete geometric topology of the equipment. The 40m interval is used to accurately match the spatial characteristics of the substation: "equipment cluster (5~20m scale) + open space". This not only reduces the amount of data by aggregating redundant point clouds in open areas through grids, thereby improving filtering efficiency, but also ensures that key features such as bushings and bases are not destroyed by dimensionality reduction by accommodating complete equipment at the unit scale. The point cloud is decomposed into grid units, which on the one hand adapts to the block parallel computing architecture of the iterative enhancement filtering network, thereby breaking through the memory bottleneck of global processing and accelerating iterative convergence. On the other hand, grid spatial indexing builds a connecting bridge for subsequent cross-block feature fusion.
[0074] Step 103: The target point cloud data is processed by flow field driven convolutional filtering through multiple iterative filtering modules connected in series in the point cloud iterative enhancement filtering network to obtain the filtered point cloud result.
[0075] Point cloud iterative enhancement filtering network refers to a deep learning network composed of multiple iterative filtering modules connected in series. Through multiple rounds of "encoding-decoding-fusion" flow field driven filtering, the filtering result of the previous round is used as the input of the next round to iteratively optimize the point cloud denoising accuracy (such as gradually converging noisy points to a clean surface), thereby achieving an "enhanced" improvement in point cloud quality.
[0076] The iterative filtering module refers to the core functional unit of the point cloud iterative enhancement filtering network, the smallest execution module for a single round of filtering. It has a built-in encoder (flow field parameter parsing), decoder (point cloud reconstruction), and integrates Gaussian weighted fusion of overlapping areas to complete the complete logic of a single round of "flow field driven - point cloud filtering". Multiple modules are connected in series to form an iterative mechanism.
[0077] Flow field driven convolutional filtering refers to the process of sequentially performing flow field parameter analysis, point cloud reconstruction, and fusion optimization.
[0078] It should be noted that the multiple iterative filtering modules connected in series sequentially perform flow-driven convolutional filtering on the target point cloud data. Specifically, the target point cloud data is first input into the first iterative filtering module for flow-driven convolutional filtering, and then the output of the first iterative filtering module is input into the second iterative filtering module for flow-driven convolutional filtering, and so on. This will not be elaborated further here.
[0079] It should be noted that the point cloud iterative enhancement filter network adopts a multi-level cascaded iterative module architecture. For easier understanding, please refer to [link to relevant documentation]. Figure 2 , Figure 2 This diagram illustrates a point cloud iterative enhancement filtering network structure employing a four-level cascaded iterative module architecture. The original noisy point cloud in the diagram refers to the target point cloud data. The network consists of four iterative filtering modules, which progressively optimize the point cloud positions through step-by-step filtering. After four iterations, the final denoised 3D point coordinates are output. This design achieves a balance between computational efficiency and applicability to real-world scenarios, effectively improving point cloud denoising through a progressive optimization strategy.
[0080] Furthermore, the iterative filtering module includes an encoder and a decoder, and step 103 may include the following sub-steps:
[0081] It's worth noting that before S21, the process also includes: dividing the input target point cloud data into point cloud segments to obtain multiple local target point cloud data sets. In the specific implementation, for the input point cloud... Divide into multiple local blocks, select Reference points This ensures coverage of the entire point cloud surface. For each reference point... Extract its k nearest neighbors (k=1000) to form a local block. ,Right now: Then normalize it. Translate to The purpose of using it as the center is to eliminate absolute positional deviation.
[0082] It should be noted that the point cloud partitioning here and the grid partitioning in step 102 are two different levels of partitioning. The grid partitioning in step 102 is a macroscopic level data block division, the purpose of which is to break down the large substation site cloud into small, processable blocks. Then, the local block partitioning performed within a single grid block in this step is a microscopic level feature extraction unit, the purpose of which is to capture the local geometric features of the equipment.
[0083] S21. The encoder is used to analyze the flow field parameter features of the input target point cloud data to obtain the encoded point cloud data.
[0084] Flow field parameter feature analysis refers to the process by which an encoder extracts parameters that quantify the correlation between the local and global structure of a point cloud by analyzing the spatial geometric distribution of the target point cloud data. These parameters can be compared to the spatial trends of "fluid flow," accurately characterizing the structural features and noise properties of the point cloud.
[0085] Furthermore, the encoder includes a first dynamic edge convolution module, a flow field motion parameter estimation module, and a second dynamic edge convolution module connected in sequence. The second dynamic edge convolution module includes multiple first dynamic edge convolution modules connected in sequence. S21 may include the following sub-steps:
[0086] It should be noted that the first dynamic edge convolution module is a feature extraction layer based on graph convolution. Its core structure consists of two layers of perceptron (MLP) and a dynamic neighbor relationship aggregation mechanism.
[0087] S211. The first dynamic edge convolution module is used to extract neighborhood graph convolution features from the input target point cloud data to obtain convolution point cloud data.
[0088] Neighborhood graph convolutional feature extraction refers to the process of extracting local structured features and global relational features from unstructured point cloud data by first constructing a graph structure with "center-neighborhood" relationships, then calculating dynamic edge features between points to quantify the geometric correlation between points and their neighborhoods, and finally aggregating neighborhood features through graph convolution operations. Its core is to utilize the graph structure to adapt to the unordered nature of point clouds, capture local geometric differences through dynamic edge features, and enhance the structural expression of features through convolutional aggregation.
[0089] The specific steps for neighborhood graph convolution feature extraction are as follows:
[0090] A local neighborhood graph is constructed using the input point cloud data;
[0091] A local neighborhood graph refers to a graph structure that dynamically constructs local neighborhood relationships for each vertex. It should be noted that a directed graph is a specific implementation of a local neighborhood graph, describing the topological relationships between the central vertex and its neighboring vertices through directed edges.
[0092] It should be noted that when this is the first round of filtering, the input point cloud data here refers to the target point cloud data. When this is the second round of filtering, the input point cloud data here refers to the optimized target point cloud data output from the first round of filtering, and so on. This will not be elaborated further here. Similarly, after each iterative filtering module is input, point cloud partitioning is required, and then flow field-driven convolutional filtering is performed.
[0093] In this embodiment of the invention, a local neighborhood graph is dynamically constructed for each vertex based on the geometric distribution of the input point cloud data. Specifically, for each vertex of the input target point cloud... A directed graph is constructed around the center using the k-nearest neighbor algorithm (k=32). Vertex set , The input is a point cloud, represented by spatial discrete vertices and edge sets. Define the relationship between vertices and their neighborhoods, so that the neighborhood topology adapts adaptively to the spatial distribution of the point cloud for each vertex. Connect its 32 nearest neighbors to form a set of directed edges.
[0094] Nonlinear mapping is performed on the vertex features of the local neighborhood graph to obtain local semantic information;
[0095] In this embodiment of the invention, a multilayer perceptron (MLP) is invoked. ) for the current layer vertex features ( Perform nonlinear transformations (such as ReLU activation) on the network layers to extract local semantic information. .
[0096] Geometric residuals are calculated using vertex features and associated neighborhood features to obtain residual coding information;
[0097] In this embodiment of the invention, the vertex is calculated. Its neighboring vertices Feature difference The residual and vertex features Concatenation, the symbol || indicates concatenation of feature dimensions, input to another multilayer perceptron (MLP) Encoding captures local geometric differences to obtain residual coding information. .
[0098] The extracted point cloud data is obtained by aggregating local semantic information and residual coding information.
[0099] In this embodiment of the invention, for vertices All neighborhood edges The residual encoding information is aggregated and fused with local semantic information. The formula for updating vertex features of edge convolution is as follows:
[0100]
[0101] In the formula, Indicates the first The point features of the layer, i.e., the first dynamic edge convolution module, Point In the The feature representation of a layer serves as the input to the next layer or the next iterative filtering module, where || represents the feature dimension concatenation operation.
[0102] Output the updated vertex feature set This refers to convolutional point cloud data that contains local semantic-geometric structural features.
[0103] It should be noted that in the substation scenario, firstly, substation equipment (such as circuit breakers and insulators) contains a large number of high-curvature surfaces, small parts, and nested structures, resulting in significant local density differences in the point cloud, which can easily lead to the loss of geometric details by traditional filtering methods. Secondly, the dense arrangement of substation equipment creates scanning blind spots, and the point cloud contains holes and incomplete surfaces. Thirdly, substation equipment needs to be monitored for deformation and displacement in real time, requiring high spatiotemporal consistency of the point cloud after denoising, while traditional iterative methods have insufficient convergence speed. To address these issues, this invention introduces two modules: a velocity module and a distance module. The concept of velocity is introduced to simulate physical motion trajectories, making the movement of the point cloud more consistent with the deformation or displacement of actual equipment. Substation equipment monitoring requires tracking deformation, and the concept of velocity can better model continuous changes rather than discrete adjustments at each step. In point cloud processing, especially in denoising and reconstruction, the concept of velocity may be closer to the motion model of the physical world, helping the model better understand how the point cloud should move to achieve the best denoising effect. Breakthroughs are achieved in three dimensions: noise robustness, geometric fidelity, and convergence speed, providing a reliable technical foundation for the digital inspection of power equipment.
[0104] S212. The flow field motion parameter estimation module is used to calculate the flow field driving parameters of the convolutional point cloud data to obtain the three-dimensional velocity field point cloud data.
[0105] Flow field driving parameter calculation refers to the process of modeling the "noise offset law" and "real structural stability trend" of point clouds based on the structured features (local geometric correlation, global structural constraints) of convolutional point clouds through network modules, and then resolving the three-dimensional velocity field parameters (displacement correction vector of each point) similar to fluid motion.
[0106] Furthermore, the flow field motion parameter estimation module includes a velocity module, a distance module, and a multiplicative modulation module. The specific solution of the flow field driving parameters is as follows:
[0107] A velocity module is used to perform hierarchical convolutional feature up-dimensional encoding on the convolutional point cloud data to obtain the flow velocity vector;
[0108] Layered convolutional feature upscaling encoding refers to the encoding process that relies on Dynamic Graph Convolutional Network (DGCNN) to gradually improve the dimensionality and spatial relational representation of point cloud features through stacking three layers of convolutions: "local-medium-global".
[0109] In this embodiment of the invention, the velocity module is constructed based on a Dynamic Graph Convolutional Neural Network (DGCNN). Convolutional point cloud data is input into the velocity module, which employs three stacked convolutional layers to progressively increase the feature level. The first convolutional layer takes 3D coordinates (x, y, z) as input and traverses the neighborhood associations of each point through a 3×3 convolutional kernel to capture basic geometric attributes such as distance between points and density gradient, outputting 64-dimensional features. The second convolutional layer takes 64-dimensional features as input and expands the receptive field of the convolution to associate more distant neighbor point pairs, fusing multi-scale local features to output 128-dimensional features. The third convolutional layer takes 128-dimensional features as input for further dimensionality increase, aggregating the topological relationships of the global point cloud to output 256-dimensional global features. Finally, a fully connected layer maps the 256-dimensional global features to perform global max pooling, compressing them into patch-level feature vectors. The fully connected layer then progressively reduces the dimensionality, first mapping to 128-dimensional intermediate features, and then reducing the dimensionality to 3-dimensional flow velocity vectors (corresponding to the flow velocity components in the x, y, and z directions, respectively), and calculates the predicted flow velocity vector for each point.
[0110] It is worth mentioning that, in order to make the predicted velocity as close as possible to the actual displacement, the velocity module is optimized during the training phase using a loss function, which is as follows:
[0111]
[0112] In the formula, The loss function of the velocity module. Indicates uniform distribution in the interval The variable t within takes the expected value. This indicates the predicted point cloud in state. The flow velocity field at that time, This represents the true displacement vector field, specifically the coordinate difference between the manually refined point cloud and the original point cloud, representing the true target of noise correction.
[0113] By minimizing the error between the predicted flow velocity and the actual displacement, the velocity module learns the "noise point offset trend and structural point stability characteristics," ensuring that the flow velocity vector output after pre-training can accurately guide the displacement correction in the denoising process.
[0114] A distance module is used to perform multi-level aggregation regression encoding on the convolutional point cloud data to obtain a distance scalar.
[0115] Multi-level aggregation regression encoding is a process for processing convolutional point cloud data. It encodes the geometric correlation features of the point cloud into quantized scalars (distance scalars) through a multi-stage processing flow of "feature reuse - multi-scale mining - hierarchical aggregation - regression mapping".
[0116] In this embodiment of the invention, convolutional point cloud data is input into the distance module to ensure feature homology with the velocity module and avoid cross-module feature bias. Through three layers of convolutional stacking, multi-scale geometric correlation features are extracted layer by layer, outputting 128-dimensional depth features. These features are then subjected to dimensionality reduction by a fully connected layer, first mapped to 128-dimensional intermediate features, then compressed to 64 dimensions, finally outputting a 1-dimensional initial scalar. Global max pooling is then performed on the point-level features output by the fully connected layer to aggregate them into patch-level global features. A Sigmoid activation function is introduced to constrain the scalar to the [0, 1] interval, generating a distance scalar. The specific formula is as follows:
[0117]
[0118] In the formula, Represents a distance scalar. This represents convolutional point cloud data, that is, intermediate state point cloud blocks after initial denoising. This represents a feature extraction network. This represents the normalized activation function.
[0119] It is worth mentioning that, in order to optimize the prediction accuracy of the distance scalar, the distance module is optimized using a loss function during the training phase. The specific loss function is as follows:
[0120]
[0121] In the formula, It is the value of the loss function. The desired operation is uniform sampling time step. , yes Point clouds with noise added at all times. It is a learnable distance metric function. (For network parameters) It is a point cloud To the target truth point cloud The displacement vector field, It is a point cloud Point cloud reconstructed by denoising network The displacement vector field, Is it the target truth point cloud? Point cloud reconstructed by denoising network The displacement vector field, It is a point cloud reconstructed by a denoising network. It is the target truth point cloud. These are regularization weight coefficients. It is a scalar of forced prediction Approximating the theoretical relative distance ensures the accuracy of adaptive step size adjustment. Avoid residual noise or geometric distortion.
[0122] The flow velocity vector and distance scalar are input into the multiplicative modulation module to determine the three-dimensional velocity field point cloud data.
[0123] It should be noted that, It is a quantification of "static geometric features". It is a fusion of "static basis + dynamic iteration strategy". In short, these two letters are different expressions of the distance scalar in the "original prediction - dynamic modulation" stage.
[0124] In this embodiment of the invention, the flow velocity vector and the distance scalar are input to a multiplicative modulation module for correction to determine the three-dimensional velocity field point cloud data, specifically as follows:
[0125]
[0126] In the formula, This represents the corrected flow velocity vector at a single point. This indicates that the k-th velocity module represents the predicted flow velocity vector for the current single point. The distance module represents the distance scalar output for the current single point. A single point refers to the smallest unit that constitutes the convolutional point cloud data, a discrete sampling point with three-dimensional spatial coordinates (x, y, z), specifically a surface sampling point of substation equipment. T represents the time step normalization factor, specifically representing the total number of time steps. Indicates a time step.
[0127] It should be noted that a time step normalization factor is introduced for time step normalization to ensure stable convergence and high accuracy of substation cloud denoising.
[0128] Once the corrected flow velocity vectors at all points have been adjusted point by point, the corrected flow velocity vectors at all points are summed up to form three-dimensional velocity field point cloud data.
[0129] It is worth mentioning that, based on the 3D velocity field point cloud data generated by the multiplicative modulation module, the Euler method is introduced to iteratively update the position point by point along the straight trajectory. This can reduce the discretization error of the point cloud iterative enhancement filter network and accelerate convergence, thus facilitating subsequent decoding. The specific Euler iteration formula is as follows:
[0130]
[0131] In the formula, This represents the target of the three-dimensional velocity field output by the multiplicative modulation module, i.e., the current state coordinates of each point, including device points and noise points. Indicates a single point at the th Coordinates of the next local iteration This represents the distance scalar output by the distance module. This represents the flow velocity vector output by the velocity module. Global time step of the representation modulation module, This indicates the number of iterations for a single point.
[0132] It should be noted that the coordinates of the local iterations are not the final output, but rather key intermediate states in the network optimization process. These coordinates serve as input to the outputs of the velocity and distance modules to calculate the new point coordinates.
[0133] The input is then fed into the subsequent dynamic edge convolution module to extract higher-level features. After being processed by the decoder, the output displacement vector is used as the input to the next iterative filtering module. After four iterations, the final denoised point cloud coordinates are output.
[0134] S213. The second dynamic edge convolution module is used to extract continuous neighborhood graph convolution features from the three-dimensional velocity field point cloud data to obtain coded point cloud data.
[0135] In this embodiment of the invention, the second dynamic edge convolution module includes a plurality of first dynamic edge convolution modules connected in sequence. (See also: [link to relevant documentation]). Figure 2 It can be seen that the second dynamic edge convolution module includes two first dynamic edge convolution modules connected in sequence. The three-dimensional velocity field point cloud data output by the multiplicative modulation module is processed by the first dynamic edge convolution module of the first layer for neighborhood graph convolution feature extraction. The extracted result is then input into the second dynamic edge convolution module of the second layer for neighborhood graph convolution feature extraction. The neighborhood graph convolution feature extraction is the same as the neighborhood graph convolution feature extraction in S211 above, and will not be repeated here. After two consecutive neighborhood graph convolutions, the output encoded point cloud data achieves feature dimensionality enhancement of "local velocity details - global velocity pattern - structural constraint enhancement", providing the decoder with high-dimensional features of "velocity field - geometric structure" collaboration.
[0136] S22. Use a decoder to decode the encoded point cloud data to obtain decoded point cloud data.
[0137] In this embodiment of the invention, the decoder takes the 256-dimensional encoded point cloud data (including global velocity field patterns, local geometric correlations, and equipment structural constraints) output by the second dynamic edge convolution module as input, and completes the decoding through a coherent process of multi-scale feature dimensionality reduction, skip connection fusion, and coordinate regression. First, the 256-dimensional features are reduced to 32 dimensions in layers, and the semantic prior of the substation is embedded to strengthen the weight of the equipment features. Then, the encoder local features are fused through skip connections to repair details such as equipment edges and breaks. Finally, the three-dimensional coordinate offset is mapped and combined with the accuracy of the manually refined point cloud constraints (equipment area error ≤ 5cm) to generate decoded point cloud data that combines noise correction and structural integrity.
[0138] S23. Perform Gaussian weighted fusion on the overlapping area of the decoded point cloud data to obtain optimized target point cloud data.
[0139] Overlapping region Gaussian weighted fusion refers to the operation of fusing the filtering results of overlapping regions to ensure the geometric continuity of the global point cloud, addressing the inconsistency of inter-block boundaries caused by independent optimization of local point cloud blocks processed by the iterative filtering module. This is achieved through a post-processing method of "distance Gaussian weight mapping - multi-block displacement competition - boundary continuity repair". The weights are calculated based on the spatial distance between the point and the patch reference point, and the filtering results of the overlapping region are fused.
[0140] In this embodiment of the invention, based on the 40m grid division rule in step 102, the center coordinates of each grid block are used as patch reference points. The overlapping area of adjacent grid blocks is the block edge. m, for each point (coordinates) within the overlapping region (Originating from the iterative update position of the decoded point cloud), traversing all its grid blocks, and calculating the Gaussian weight of the point in each block according to the formula:
[0141]
[0142] In the formula, The Gaussian weights represent the coordinates of the current point. Indicates the coordinates of the current point. This indicates the reference point for the patch, with the radius parameter set to 1 / 3 of the patch radius. , This indicates the grid side length.
[0143] It should be noted that the weight is negatively correlated with the distance from the point to the patch reference point. The closer the distance, the higher the weight, which is suitable for the continuous structural characteristics of substation equipment across blocks.
[0144] For the same spatial point (such as the vertex of the main variable span block), compare its Gaussian weights in different grid blocks, select the filter displacement (the coordinate update amount of the decoded point cloud) corresponding to the grid block with the highest weight as the final displacement, and correct the coordinates of the point.
[0145] After traversing all overlapping points and completing displacement competition fusion, coordinate abrupt changes at the boundaries between blocks are eliminated, such as the breakage of equipment edges across blocks and the stepped error in the terrain transition zone. Optimized target point cloud data is output, which retains the fine denoising results of local blocks while ensuring global geometric continuity.
[0146] It is worth mentioning that, after one iteration of the filtering module, taking the local block of the point cloud as the object, the loss of a single point is defined as the scalar constraint between the predicted displacement and distance. error:
[0147]
[0148] In the formula, It represents the displacement of the current point to the nearest neighbor of the target block (corresponding to the correction amount predicted by the velocity module). Indicates the first The noisy local point cloud patch of the next iteration Indicates the iteration module round, Indicates the first In this iteration, the velocity module represents the displacement correction for a single point. Indicates the first In each iteration, the current coordinates of a single point. Indicates a single point from To target block The actual displacement of the nearest neighbor point.
[0149] After four iterations, the losses from each round are summed using a weighted average:
[0150]
[0151] in, Let T = 4, representing the weighted sum of losses over four iterations.
[0152] On the one hand, it serves as a training objective to guide the network to learn the mapping law between "noise shift trend and structural stability constraint". On the other hand, it verifies the iteration accuracy during the inference stage to ensure that the denoising error of the optimized point cloud meets the requirements of substation scenarios where the equipment area is ≤5cm and the terrain area is ≤10cm. Ultimately, it supports the network output of filtered point cloud results that have both geometric continuity and accuracy constraints.
[0153] In this embodiment of the invention, the iterative filtering modules of the point cloud iterative enhancement filtering network are connected in series. First, the encoder parses the flow field parameters of the target point cloud, then the decoder reconstructs and optimizes the point cloud. The results of each round are passed to the next round module for continuous iteration. Finally, the denoised point cloud coordinates output by the point cloud iterative enhancement filtering network are used as the final filtered result. Each iterative filtering module gradually corrects the position of noise points by predicting the displacement vector, so that the point cloud converges to the target clean surface after multiple internal iterations.
[0154] This invention has the following advantages:
[0155] 1) This invention uses a 40m grid to fit the layout design of the main transformer array and the outgoing line direction (avoiding cutting the main body of the equipment), combined with local block extraction of 1000 nearest neighbor points to capture the nesting details of the equipment, combined with encoder dynamic graph convolution to preserve the features of high curvature edges (such as bushing vertices), decoder jump connection to repair the gaps between dense equipment blocks, and Gaussian fusion in the overlapping area to eliminate boundary abrupt changes. It adapts to the high-density layout characteristics of substation equipment, preserves key surface structures, reduces geometric deviations in modeling, and fully considers the high-density layout characteristics of substation equipment. It can preserve key structural information of the equipment surface and reduce the geometric deviation between the modeling results and the real physical entity.
[0156] 2) This invention uses constant flow field modeling and Euler line iteration to drive noise point migration with a straight path (replacing random trajectories), combined with the constant flow velocity direction predicted by the velocity module, to reduce discretization error and accelerate noise convergence to the clean surface of the equipment.
[0157] 3) Design a lightweight DGCNN speed module, which directly infers the constant flow field from noise points to the clean surface through three layers of convolution, and couples a four-level iterative module for series optimization to enhance path linearity and improve denoising accuracy.
[0158] 4) Introduce a distance module to estimate the geometric deviation scalar and dynamically modulate the flow velocity amplitude (specifically, a multiplicative control formula) to avoid device overshoot and ensure convergence stability.
[0159] 5) A stacked iterative module architecture is proposed, in which each module integrates dynamic graph convolution and velocity / distance modules (to solve flow field parameters), and updates point cloud displacement round by round to adapt to local fine correction under high-density layout of substation equipment.
[0160] 6) Achieve end-to-end optimal trajectory learning based on deep learning, utilize graph convolutional networks to capture the geometric features of point clouds, and drive noise points to converge towards the target distribution along a straight path, providing a theoretically rigorous and highly efficient solution for point cloud denoising in complex substation scenarios.
[0161] Please see Figure 3 , Figure 3 This is a structural block diagram of a point cloud denoising system for substation scenarios provided in an embodiment of the present invention.
[0162] This invention provides a point cloud denoising system for substation scenarios, comprising:
[0163] The acquisition module 301 is used to acquire multi-source remote sensing data of the substation.
[0164] Processing module 302 is used to perform point cloud normalization on multi-source remote sensing data to obtain target point cloud data;
[0165] The output module 303 is used to sequentially perform flow field-driven convolutional filtering on the target point cloud data through multiple iterative filtering modules connected in series within the point cloud iterative enhancement filtering network, so as to obtain the filtered point cloud result.
[0166] Furthermore, the multi-source remote sensing data includes UAV oblique imagery and multi-station radar point cloud data, and the processing module 302 includes:
[0167] The sub-module for optimizing oblique image point cloud data is used to perform aerial triangulation encryption on UAV oblique images to obtain optimized oblique image point cloud data.
[0168] The multi-site cloud data registration submodule is used to denoise and downsample multi-site radar point cloud data to obtain registered multi-site cloud data.
[0169] The initial point cloud data submodule is used to denoise and integrate the optimized oblique image point cloud data and the registered multi-site point cloud data to obtain the initial point cloud data.
[0170] The target point cloud data submodule is used to divide the initial point cloud data into grids to obtain the target point cloud data.
[0171] Furthermore, the iterative filtering module includes an encoder and a decoder, and the flow-driven convolutional filtering process is specifically as follows:
[0172] An encoder is used to analyze the flow field parameter features of the input target point cloud data to obtain coded point cloud data;
[0173] The encoded point cloud data is decoded using a decoder to obtain the decoded point cloud data;
[0174] The decoded point cloud data is subjected to Gaussian weighted fusion of overlapping areas to obtain optimized target point cloud data.
[0175] Furthermore, the encoder includes a first dynamic edge convolution module, a flow field motion parameter estimation module, and a second dynamic edge convolution module connected in sequence. The second dynamic edge convolution module includes multiple first dynamic edge convolution modules connected in sequence. The encoder performs flow field parameter feature parsing on the input target point cloud data to obtain coded point cloud data, including:
[0176] The first dynamic edge convolution module is used to extract neighborhood graph convolution features from the input target point cloud data to obtain convolutional point cloud data;
[0177] The flow field motion parameter estimation module is used to solve the flow field driving parameters of the convolutional point cloud data to obtain three-dimensional velocity field point cloud data;
[0178] The second dynamic edge convolution module is used to extract continuous neighborhood graph convolution features from the three-dimensional velocity field point cloud data to obtain coded point cloud data.
[0179] Furthermore, the neighborhood graph convolution feature extraction specifically involves:
[0180] A local neighborhood graph is constructed using the input point cloud data;
[0181] Nonlinear mapping is performed on the vertex features of the local neighborhood graph to obtain local semantic information;
[0182] Geometric residuals are calculated using vertex features and associated neighborhood features to obtain residual coding information;
[0183] The extracted point cloud data is obtained by aggregating local semantic information and residual coding information.
[0184] Furthermore, the flow field motion parameter estimation module includes a velocity module, a distance module, and a multiplicative modulation module. The specific solution of the flow field driving parameters is as follows:
[0185] A velocity module is used to perform hierarchical convolutional feature up-dimensional encoding on the convolutional point cloud data to obtain the flow velocity vector;
[0186] A distance module is used to perform multi-level aggregation regression encoding on the convolutional point cloud data to obtain a distance scalar.
[0187] The flow velocity vector and distance scalar are input into the multiplicative modulation module to determine the three-dimensional velocity field point cloud data.
[0188] Given the complex equipment layout, narrow spacing, and long-term energized operation within substations, manual inspections require close contact with live areas, posing risks of electric shock and other safety hazards. This invention designs a deep learning-based point cloud filtering technique that automatically learns noise features through end-to-end training, reducing reliance on manual rules and significantly improving the algorithm's robustness in densely distributed and occluded scenarios.
[0189] By simulating real iterations during the training phase through stacked iterative modules, with each module corresponding to a gradual noise reduction process, the model can learn the dynamic relationships between iterations, significantly accelerating the convergence speed. Dynamically aggregating neighborhood information through graph convolution, combined with iterative optimization strategies, demonstrates better preservation of sharp features and complex structures in both synthetic and real scanned data. It solves the problems of traditional methods such as normal dependence, loss of geometric details, and slow convergence speed, significantly outperforming traditional filtering methods in accuracy, efficiency, and robustness.
[0190] Please see Figure 4 , Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of the present invention.
[0191] An electronic device according to an embodiment of the present invention includes: a memory 401 and a processor 402. The memory 401 stores a computer program. When the computer program is executed by the processor 402, the processor 402 performs a point cloud denoising method for substation scenarios as described in any of the above embodiments.
[0192] Memory 401 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 401 has storage space 403 for program code 413 for performing any of the method steps described above. For example, storage space 403 for program code may include individual program codes 413 for implementing the various steps in the methods described above. This program code can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. This program code can be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When this code is run by a computing device, it causes the device to perform the various steps in the point cloud denoising method for substation scenarios described above.
[0193] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the point cloud denoising method for substation scenarios as described in any of the above embodiments.
[0194] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, wherein when the program instructions are executed by a computer, the computer performs the point cloud denoising method for substation scenarios as described in any of the above embodiments.
[0195] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0196] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.
[0197] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0198] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0199] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0200] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A point cloud denoising method for substation scenarios, characterized in that, include: Acquire multi-source remote sensing data of substations; The multi-source remote sensing data is processed into point cloud data to obtain target point cloud data; The target point cloud data is subjected to flow field-driven convolutional filtering by multiple iterative filtering modules connected in series within the point cloud iterative enhancement filtering network, thereby obtaining the filtered point cloud result.
2. The point cloud denoising method for substation scenarios according to claim 1, characterized in that, The multi-source remote sensing data includes UAV oblique imagery and multi-station radar point cloud data. The step of performing point cloud normalization on the multi-source remote sensing data to obtain target point cloud data includes: The UAV tilted imagery is subjected to aerial triangulation encryption to obtain optimized tilted imagery point cloud data; The multi-station radar point cloud data is denoised and downsampled to obtain registered multi-station point cloud data. The optimized oblique image point cloud data and the registered multi-site point cloud data are denoised and integrated to obtain the initial point cloud data; The initial point cloud data is divided into grids to obtain the target point cloud data.
3. The point cloud denoising method for substation scenarios according to claim 1 or 2, characterized in that, The iterative filtering module includes an encoder and a decoder, and the flow-driven convolutional filtering process specifically involves: The encoder is used to analyze the flow field parameter features of the input target point cloud data to obtain coded point cloud data; The decoder is used to decode the encoded point cloud data to obtain decoded point cloud data; The decoded point cloud data is subjected to Gaussian weighted fusion of overlapping areas to obtain optimized target point cloud data.
4. The point cloud denoising method for substation scenarios according to claim 3, characterized in that, The encoder includes a first dynamic edge convolution module, a flow field motion parameter estimation module, and a second dynamic edge convolution module connected in sequence. The second dynamic edge convolution module includes multiple first dynamic edge convolution modules connected in sequence. The step of using the encoder to perform flow field parameter feature parsing on the input target point cloud data to obtain coded point cloud data includes: The first dynamic edge convolution module is used to extract neighborhood graph convolution features from the input target point cloud data to obtain convolutional point cloud data; The flow field motion parameter estimation module is used to calculate the flow field driving parameters of the convolutional point cloud data to obtain three-dimensional velocity field point cloud data. The second dynamic edge convolution module is used to extract continuous neighborhood graph convolution features from the three-dimensional velocity field point cloud data to obtain coded point cloud data.
5. The point cloud denoising method for substation scenarios according to claim 4, characterized in that, The neighborhood graph convolution feature extraction specifically involves: A local neighborhood graph is constructed using the input point cloud data; Nonlinear mapping is performed on the vertex features of the local neighborhood graph to obtain local semantic information; Geometric residuals are calculated using the vertex features and associated neighborhood features to obtain residual coding information; The extracted point cloud data is obtained by aggregating the local semantic information and the residual coding information.
6. The point cloud denoising method for substation scenarios according to claim 4, characterized in that, The flow field motion parameter estimation module includes a velocity module, a distance module, and a multiplicative modulation module. The flow field driving parameter solution is specifically as follows: The velocity module is used to perform hierarchical convolutional feature upscaling encoding on the convolutional point cloud data to obtain the flow velocity vector; The distance module is used to perform multi-level aggregation regression encoding on the convolutional point cloud data to obtain a distance scalar; The flow velocity vector and the distance scalar are input into the multiplicative modulation module to determine the three-dimensional velocity field point cloud data.
7. A point cloud denoising system for substation scenarios, characterized in that, include: The acquisition module is used to acquire multi-source remote sensing data of the substation; The processing module is used to perform point cloud normalization on the multi-source remote sensing data to obtain target point cloud data; The output module is used to sequentially perform flow field-driven convolutional filtering on the target point cloud data through multiple iterative filtering modules connected in series within the point cloud iterative enhancement filtering network, so as to obtain the filtered point cloud result.
8. An electronic device, characterized in that, The method includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the point cloud denoising method for substation scenarios as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the point cloud denoising method for substation scenarios as described in any one of claims 1-6.
10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the point cloud denoising method for substation scenarios as described in any one of claims 1-6.
