A method and apparatus for extracting substation equipment from lidar point clouds

Abstract

This invention discloses a method and apparatus for extracting substation equipment from lidar point clouds. The method includes the following steps: acquiring raw lidar point clouds of a substation scene to form first point cloud data; enhancing the feature saliency of the first point cloud data based on a feature transformation function to obtain second point cloud data; segmenting the second point cloud data to obtain coarse-scale third point cloud data; extracting features from the coarse-scale third point cloud data using an offsetting attention mechanism to obtain fine-scale fourth point cloud data; and extracting substation equipment from the fine-scale fourth point cloud data to obtain the substation scene equipment distribution. This invention utilizes a method based on feature enhancement of raw point cloud data, point deformer neural networks, and edge convolution to achieve accurate and stable equipment extraction in complex substation scenes.

Description

Technical Field

[0001] This invention belongs to the technical field of smart grids, specifically relating to a method and apparatus for extracting substation equipment from lidar point clouds. Background Technology

[0002] As a crucial component of new power systems and smart grids, substation equipment requires periodic monitoring and maintenance. Intelligent monitoring can effectively ensure the accuracy of intelligent inspections of power equipment. Traditional power equipment monitoring methods mainly rely on manual field measurements and digital photogrammetry. Manual methods are significantly affected by factors such as high workload and safety risks. Digital photogrammetry, which uses drone images and surveillance camera images to acquire substation equipment information, is also limited by factors such as rain and snow, object obstruction, and image resolution, resulting in the accuracy of the extracted power equipment not meeting the requirements of online intelligent inspections.

[0003] Research has shown that LiDAR scanning can be used for mapping to quickly and accurately collect 3D information of large-scale substation scenes and extract power equipment of various shapes. For example, patent CN114782626A presents a substation scene mapping and localization optimization method based on laser and vision fusion. First, it acquires and preprocesses image data from single-line LiDAR, inertial measurement unit (IMU), odometer, and binoculars. Then, it performs local localization and mapping to construct a sub-map. Simultaneously, while performing laser-based localization and mapping, it processes the binocular images (i.e., the vision thread) to achieve feature point tracking and keyframe identification. Next, it performs target detection on the obtained keyframes. Finally, it generates semantic landmarks and projects them onto the sub-map. The sub-map is then optimized, and global optimization is performed. This method combines the advantages of graph-based optimization localization algorithms and deep learning, resulting in a complete and stable navigation map. It significantly shortens feature matching time, greatly increases the stability of localization in complex substation operation and maintenance scenarios, and makes navigation in substation scenes faster and more reliable.

[0004] For example, the semantic map construction method for substations based on laser SLAM and visual fusion, as given in patent CN114638909A, includes the following steps: S1-1, calibrating the intrinsic parameters of the depth camera and jointly calibrating the extrinsic parameters of the LiDAR and camera; S1-2, synchronously preprocessing the data acquired by the depth camera and LiDAR; S2-1, modeling the operation and maintenance environment map using point cloud data collected by the LiDAR and odometer information; S2-2, acquiring RGBD images from the depth camera, performing target recognition and scene information understanding through deep learning to obtain semantic information; S2-3, performing coordinate transformation to project the targets identified in step S2-2 onto the grid map, providing environmental awareness information for the substation; S3, repeating step S2 to complete the construction of the semantic map. Using the above technical solution, the algorithm has the advantage of high adaptability to different weather environments and lighting conditions during the mapping process. It can effectively remove laser motion distortion, improve mapping accuracy, and reduce cumulative errors.

[0005] However, as with the existing technologies mentioned above, current methods for extracting power equipment are still limited to morphological analysis and deep learning, which are subject to the following limitations: (1) there is a diversity of types of power equipment in substations, and the shape of substations is relatively complex; (2) there is interference and occlusion between power equipment; (3) the amount of high-quality power equipment point cloud labeled data is limited, which cannot meet the requirements for accurate feature extraction of power equipment. Based on the above limitations, existing methods cannot yet meet the requirements for extracting power equipment in complex scenarios of large-scale substations.

[0006] Therefore, how to design a method and apparatus for extracting substation equipment from lidar point clouds, so as to achieve stable, accurate and robust extraction of substation equipment of various shapes from high-density, disordered massive lidar point cloud data, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for extracting substation equipment from lidar point clouds. It utilizes a high-precision lidar point cloud to reconstruct a 3D real-world substation scene, ensuring that each device, component, and camera in the substation scene has accurate spatial coordinate information. An algorithm can calculate the correlation between inspection points and cameras. By employing methods based on feature enhancement of original point cloud data, point deformer neural networks, and edge convolution, the complex and densely located equipment in the substation, as well as the spatial occlusion relationships between devices, components, and points, are effectively resolved, achieving accurate and stable equipment extraction in complex substation scenarios.

[0008] In a first aspect, the present invention provides a method for extracting substation equipment from lidar point clouds, comprising the following steps:

[0009] Collect raw point cloud data from lidar in the substation scene to form the first point cloud data;

[0010] Based on the feature transformation function, the feature saliency of the first point cloud data is enhanced to obtain the second point cloud data;

[0011] The second point cloud data is segmented to obtain a coarse-scale third point cloud data.

[0012] By using an offsetting attention mechanism to extract features from the coarse-scale third point cloud data, fine-scale fourth point cloud data is obtained.

[0013] Substation equipment is extracted from the fine-scale fourth-point cloud data to obtain the scene equipment distribution of the substation.

[0014] Furthermore, based on the feature transformation function, the feature saliency of the first point cloud data is enhanced to obtain the second point cloud data, specifically including the following steps:

[0015] Define the coordinate system orientation of the original point cloud, where the coordinate system includes three axes: x, y, and z.

[0016] Based on the given feature transformation relationship, feature enhancement is performed on the first point cloud data.

[0017] Furthermore, the coordinate system includes three axes: x, y, and z. Specifically, the x-axis represents the front of the lidar device as it travels, the y-axis represents the right side of the lidar device, and the z-axis represents the top of the lidar device.

[0018] Based on the given feature transformation relationship, feature enhancement is performed on the first point cloud data, specifically including the following steps:

[0019] The elevation and intensity context features of each original point are analyzed, and the specific calculation relationships are as follows:

[0020]

[0021] HCF i For the elevation context feature of the i-th original point, ICF i Let z be the intensity context feature of the i-th original point, cot be the cotangent function, δ be the preset transformation degree, and z be the intensity context feature of the i-th original point. i Let r be the z-axis coordinate of the i-th original point in the substation scenario. i Let be the intensity of the i-th original point in the substation scenario;

[0022] The density context features of each original point are determined by the number of neighboring points within a predetermined neighborhood. The specific determination relationship is as follows:

[0023] DCF i=S(R,range(0,1))

[0024] Among them, DCF i Let S be the density context feature of the i-th original point, S be the normalization function, R be the spherical neighborhood search radius of the i-th original point, and range be the normalization range.

[0025] The contextual features of the normal vector for each point are given, and the specific relationships are as follows:

[0026]

[0027] Among them, NCF i For the context features of the normal vector of the i-th original point, N xi Let N be the x-axis coordinate of the normal vector of the i-th original point. yi Let N be the y-axis coordinate of the normal vector of the i-th original point. zi Let z be the z-axis coordinate of the normal vector of the i-th original point;

[0028] By merging elevation context features, intensity context features, density context features, and normal vector context features, feature enhancement of the first point cloud data is achieved.

[0029] Furthermore, the second point cloud data is segmented to obtain a coarse-scale third point cloud data, specifically including the following steps:

[0030] Feature embedding is performed on the second point cloud data using multiple cascaded deformers;

[0031] In each cascaded deformer, based on the multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data to obtain multi-dimensional semantic features;

[0032] Multidimensional semantic features are fed into a fully connected layer and subjected to summation linear functions and batch normalization to obtain coarse-scale third point cloud data.

[0033] Furthermore, based on a multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data, specifically including:

[0034] The semantic information extracted from the second point cloud data includes three feature matrices: Query, Key, and Value.

[0035] The specific formulas for calculating the above three feature matrix values ​​are as follows:

[0036] (Q, K, V) = F in ·(W q W k W v )

[0037] Where Q is the Query feature matrix value, K is the Key feature matrix value, V is the Value feature matrix value, and F is the Key feature matrix value. in For the input function, W q W represents the learnable shared weights in the Query feature matrix. k W represents the learnable shared weights in the Key feature matrix. v These are the learnable shared weights in the Value feature matrix;

[0038] Based on the Query feature matrix values ​​and Key feature matrix values, the attention weights of the three feature matrices are obtained through matrix dot product, and the calculation formula is as follows:

[0039] A = Q·K T

[0040] Where A represents the attention weights of the three feature matrices, and T is the transpose function;

[0041] Based on the attention weights of the three feature matrices, semantic features are obtained by weighted summation and normalization.

[0042] Furthermore, feature extraction is performed on the coarse-scale third point cloud data using an offsetting attention mechanism to obtain the fine-scale fourth point cloud data. This process includes the following steps:

[0043] Based on the offsetting attention mechanism, the differentiation features of coarse-scale third point cloud data are calculated by element subtraction to form a differentiation feature set.

[0044] The distinct feature sets are input into the ensemble network model to eliminate noise, resulting in fine-scale fourth point cloud data.

[0045] Furthermore, the differentiation characteristics of the coarse-scale third point cloud data are calculated using element-wise subtraction. The specific calculation formula is as follows:

[0046] F out =OA(F in ) = F in +LBR(F in -F mid )

[0047] Among them, F out For the third point cloud data, OA is the differential feature, and F is the offsetting attention function used to calculate the differential values ​​of the input features. in For the third point cloud data, F mid For the intermediate output features of the third point cloud data, and F in -F mid The LBR is a discrete Laplace operator, a set network model, and includes functions for linear, batch normalization, and ReLU activation of the data.

[0048] Furthermore, substation equipment is extracted from the fine-scale fourth-point cloud data to obtain the scene equipment distribution of the substation, specifically including the following steps:

[0049] Based on the K-nearest neighbor algorithm, local data points of any center point are determined in the fourth point cloud data at a fine scale;

[0050] Using a channel-widening symmetric aggregation function, the edge features of the local data point are calculated within a given range;

[0051] The edge features and the differentiation features of the third point cloud data are merged and input into a linear function, a batch normalization function, a ReLU function, and a Dropout function, respectively, to obtain the scene equipment distribution of the substation.

[0052] Furthermore, the edge features of this local data point are calculated within a given range, using the following formula:

[0053]

[0054]

[0055] in, h represents the edge features of local data point i. ψ It is a nonlinear transformation function. The vector context features of point i, Let i be the vector context feature of another point j within a given range of i, max be the maximum value function used as the symmetric aggregation function, and e be the edge connecting the points.

[0056] In a second aspect, the present invention also provides an apparatus for extracting substation equipment from lidar point clouds, employing the method described above for extracting substation equipment from lidar point clouds, comprising:

[0057] The acquisition module collects raw point clouds from the LiDAR in the substation scene to form the first point cloud data;

[0058] The analysis and processing module enhances the feature saliency of the first point cloud data based on the feature transformation function to obtain the second point cloud data. The second point cloud data is then segmented to obtain the third point cloud data at a coarse scale.

[0059] The feature extraction module extracts features from the coarse-scale third point cloud data through an offsetting attention mechanism to obtain the fine-scale fourth point cloud data. The fine-scale fourth point cloud data is then used to extract substation equipment to obtain the scene equipment distribution of the substation.

[0060] The present invention provides a method and apparatus for extracting substation equipment from lidar point clouds, which has at least the following beneficial effects:

[0061] (1) The three-dimensional real scene of the substation is reconstructed by high-precision laser point cloud, so that each device, component and camera in the substation scene has accurate spatial coordinate information. The relationship between the inspection point and the camera can be calculated by the algorithm. By using the method of feature enhancement based on the original point cloud data, point deformer neural network and edge convolution, the complex equipment and dense location in the substation, and the spatial occlusion relationship between equipment, components and points can be effectively solved.

[0062] (2) By using the method of feature enhancement based on original point cloud data, point deformer neural network and edge convolution, accurate and stable equipment extraction in complex substation scenarios is achieved.

[0063] (3) By employing feature enhancement, self-attention and edge convolution, the limitations of uneven distribution of density and reflectance values ​​are overcome, and the inherent features of power equipment point cloud data are effectively extracted, making the power equipment extraction results more robust and efficient.

[0064] (4) The use of feature enhancement, self-attention and edge convolution methods also improves the efficiency of large-scale point cloud data processing, improves the effectiveness of intelligent operation and maintenance and the security of urban power grid. Attached Figure Description

[0065] Figure 1 This invention provides a schematic flowchart of a method for extracting substation equipment from lidar point clouds.

[0066] Figure 2 A schematic diagram of the offsetting attention mechanism for extracting coarse-scale third point cloud data in one embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram illustrating the calculation of edge features of local data points within a given range in one embodiment of the present invention.

[0068] Figure 4 This invention provides a schematic diagram of a device for extracting substation equipment from lidar point clouds. Detailed Implementation

[0069] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0070] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0071] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0072] The substation is scanned using 3D laser scanning technology (LiDAR) to obtain the spatial geometric information of the relevant power equipment in the substation. Then, the digitized power equipment is imported into 3D design software, and a model is built based on the digitized information.

[0073] Regarding the extraction of spatial geometric information of substation equipment, reverse modeling based on point cloud data, such as... Figure 1 As shown, the present invention provides a method for extracting substation equipment from lidar point clouds, comprising the following steps:

[0074] Collect raw point cloud data from lidar in the substation scene to form the first point cloud data;

[0075] Based on the feature transformation function, the feature saliency of the first point cloud data is enhanced to obtain the second point cloud data;

[0076] The second point cloud data is segmented to obtain a coarse-scale third point cloud data.

[0077] By using an offsetting attention mechanism to extract features from the coarse-scale third point cloud data, fine-scale fourth point cloud data is obtained.

[0078] Substation equipment is extracted from the fine-scale fourth-point cloud data to obtain the scene equipment distribution of the substation.

[0079] The 3D real-world view of the substation reconstructed using high-precision laser point cloud ensures that each device, component, and camera in the substation scene has accurate spatial coordinate information. The algorithm can calculate the relationship between the inspection point and the camera. By using the method of feature enhancement based on the original point cloud data, point deformer neural network, and edge convolution, the spatial occlusion relationship between the complex and densely located equipment, components, and points in the substation can be effectively solved.

[0080] Based on the feature transformation function, the feature saliency of the first point cloud data is enhanced to obtain the second point cloud data. The specific steps include the following:

[0081] Define the coordinate system orientation of the original point cloud, where the coordinate system includes three axes: x, y, and z.

[0082] Based on the given feature transformation relationship, feature enhancement is performed on the first point cloud data.

[0083] Furthermore, the coordinate system includes three axes: x, y, and z. Specifically, the x-axis represents the front of the lidar device as it travels, the y-axis represents the right side of the lidar device, and the z-axis represents the top of the lidar device.

[0084] Based on the given feature transformation relationship, feature enhancement is performed on the first point cloud data, specifically including the following steps:

[0085] The elevation and intensity context features of each original point are analyzed, and the specific calculation relationships are as follows:

[0086]

[0087] HCF i For the elevation context feature of the i-th original point, ICF i Let z be the intensity context feature of the i-th original point, cot be the cotangent function, δ be the preset transformation degree, and z be the intensity context feature of the i-th original point. i Let r be the z-axis coordinate of the i-th original point in the substation scenario. i Let be the intensity of the i-th original point in the substation scenario;

[0088] The density context features of each original point are determined by the number of neighboring points within a predetermined neighborhood. The specific determination relationship is as follows:

[0089] DCF i =S(R,range(0,1))

[0090] Among them, DCF i Let S be the density context feature of the i-th original point, S be the normalization function, R be the spherical neighborhood search radius of the i-th original point, and range be the normalization range.

[0091] The contextual features of the normal vector for each point are given, and the specific relationships are as follows:

[0092]

[0093] Among them, NCF i For the context features of the normal vector of the i-th original point, N xi Let N be the x-axis coordinate of the normal vector of the i-th original point.yi Let N be the y-axis coordinate of the normal vector of the i-th original point. zi Let z be the z-axis coordinate of the normal vector of the i-th original point;

[0094] By merging elevation context features, intensity context features, density context features, and normal vector context features, feature enhancement of the first point cloud data is achieved.

[0095] The value of the preset conversion degree δ can be set according to the needs of feature enhancement. In one embodiment, δ is set to 1. Of course, other values ​​can also be set according to the effects of elevation context features and intensity context features. No specific limitation is made here.

[0096] The value of the spherical neighborhood search radius R can be preset. In one embodiment, it can be set to 0.5m. Of course, other values ​​can also be set according to the differences in different substation scenarios. No specific limitation is made here.

[0097] Furthermore, the second point cloud data is segmented to obtain a coarse-scale third point cloud data, specifically including the following steps:

[0098] Feature embedding is performed on the second point cloud data using multiple cascaded deformers;

[0099] In each cascaded deformer, based on the multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data to obtain multi-dimensional semantic features;

[0100] Multidimensional semantic features are fed into a fully connected layer and subjected to summation linear functions and batch normalization to obtain coarse-scale third point cloud data.

[0101] In one embodiment, four cascade deformers can be set. Based on the characteristics of the second point cloud data of the substation, the data feature output dimension of each cascade deformer is 256-dimensional. By merging the output results of all the above cascade deformers, a total of 1024-dimensional semantic features are obtained.

[0102] A method for capturing and rapidly identifying targets in power equipment based on a multi-cascade classifier utilizes a funnel-shaped cascade classifier as the target detection algorithm and a trained deep neural network for target recognition. These two algorithms are then integrated into a single system. This algorithm is based on embedded deep learning technology and exhibits a certain degree of resistance to interference caused by changes in the orientation and lighting of substation equipment targets.

[0103] Furthermore, based on a multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data, specifically including:

[0104] The semantic information extracted from the second point cloud data includes three feature matrices: Query, Key, and Value.

[0105] The specific formulas for calculating the above three feature matrix values ​​are as follows:

[0106] (Q, K, V) = F in ·(W q W k W v )

[0107] Where Q is the Query feature matrix value, K is the Key feature matrix value, V is the Value feature matrix value, and F is the Key feature matrix value. in For the input function, W q W represents the learnable shared weights in the Query feature matrix. k W represents the learnable shared weights in the Key feature matrix. v These are the learnable shared weights in the Value feature matrix;

[0108] Based on the Query feature matrix values ​​and Key feature matrix values, the attention weights of the three feature matrices are obtained through matrix dot product, and the calculation formula is as follows:

[0109] A = Q·K T

[0110] Where A represents the attention weights of the three feature matrices, and T is the transpose function;

[0111] Based on the attention weights of the three feature matrices, semantic features are obtained by weighted summation and normalization.

[0112] like Figure 2 As shown, feature extraction is performed on the coarse-scale third point cloud data using an offsetting attention mechanism to obtain the fine-scale fourth point cloud data. The specific steps include:

[0113] Based on the offsetting attention mechanism, the differentiation features of coarse-scale third point cloud data are calculated by element subtraction to form a differentiation feature set.

[0114] The distinct feature sets are input into the ensemble network model to eliminate noise, resulting in fine-scale fourth point cloud data.

[0115] Furthermore, the differentiation characteristics of the coarse-scale third point cloud data are calculated using element-wise subtraction. The specific calculation formula is as follows:

[0116] F out =OA(F in ) = F in +LBR(F in -F mid )

[0117] Among them, F out For the third point cloud data, OA is the differential feature, and F is the offsetting attention function used to calculate the differential values ​​of the input features. in For the third point cloud data, F mid For the intermediate output features of the third point cloud data, and F in -F mid The LBR is a discrete Laplace operator, a set network model, and includes functions for linear, batch normalization, and ReLU activation of the data.

[0118] Feature extraction also includes extracting the size information of features, such as the width and depth of grooves, and the radius and depth of holes.

[0119] like Figure 3 As shown, the equipment distribution of the substation is obtained by extracting substation equipment from the fine-scale fourth point cloud data, specifically including the following steps:

[0120] Based on the K-nearest neighbor algorithm, local data points of any center point are determined in the fourth point cloud data at a fine scale;

[0121] Using a channel-widening symmetric aggregation function, the edge features of the local data point are calculated within a given range;

[0122] The edge features and the differentiation features of the third point cloud data are merged and input into a linear function, a batch normalization function, a ReLU function, and a Dropout function, respectively, to obtain the scene equipment distribution of the substation.

[0123] Furthermore, the edge features of this local data point are calculated within a given range, using the following formula:

[0124]

[0125]

[0126] in, h represents the edge features of local data point i. ψ It is a nonlinear transformation function. The vector context features of point i, Let i be the vector context feature of another point j within a given range of i, max be the maximum value function used as the symmetric aggregation function, and e be the edge connecting the points.

[0127] The constraint relationship between adjacent data points within a preset range is determined. Using the above relationship discrimination method, all local data points are traversed to determine the adjacency relationship between the edges of substation equipment, and finally the scene equipment distribution of the substation is obtained.

[0128] This invention provides a device for extracting substation equipment from lidar point clouds. By utilizing a method based on feature enhancement of original point cloud data, point deformer neural network, and edge convolution, it achieves accurate and stable equipment extraction in complex substation scenarios.

[0129] like Figure 4 As shown, the present invention also provides an apparatus for extracting substation equipment from lidar point clouds, employing the method described above for extracting substation equipment from lidar point clouds, comprising:

[0130] The acquisition module collects raw point clouds from the LiDAR in the substation scene to form the first point cloud data;

[0131] The analysis and processing module enhances the feature saliency of the first point cloud data based on the feature transformation function to obtain the second point cloud data. The second point cloud data is then segmented to obtain the third point cloud data at a coarse scale.

[0132] The feature extraction module extracts features from the coarse-scale third point cloud data through an offsetting attention mechanism to obtain the fine-scale fourth point cloud data. The fine-scale fourth point cloud data is then used to extract substation equipment to obtain the scene equipment distribution of the substation.

[0133] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for extracting substation equipment from lidar point clouds, characterized in that, Includes the following steps: Collect raw point cloud data from lidar in the substation scene to form the first point cloud data; Based on the feature transformation function, the feature saliency of the first point cloud data is enhanced to obtain the second point cloud data; The second point cloud data is segmented to obtain a coarse-scale third point cloud data. By using an offsetting attention mechanism to extract features from the coarse-scale third point cloud data, fine-scale fourth point cloud data is obtained. Substation equipment is extracted from the fine-scale fourth-point cloud data to obtain the scene equipment distribution of the substation; The process of enhancing the feature saliency of the first point cloud data based on a feature transformation function to obtain the second point cloud data includes the following steps: Define the coordinate system orientation of the original point cloud, where the coordinate system includes three axes: x, y, and z. The x-axis represents the front of the lidar device, the y-axis represents the right side of the lidar device, and the z-axis represents the top of the lidar device. The elevation and intensity context features of each original point are analyzed, and the specific calculation relationships are as follows: ； HCF i For the elevation context feature of the i-th original point, ICF i Let z be the intensity context feature of the i-th original point, cot be the cotangent function, δ be the preset transformation degree, and z be the intensity context feature of the i-th original point. i Let r be the z-axis coordinate of the i-th original point in the substation scenario. i Let be the intensity of the i-th original point in the substation scenario; The density context features of each original point are determined by the number of neighboring points within a predetermined neighborhood. The specific determination relationship is as follows: ； Among them, DCF i Let S be the density context feature of the i-th original point, S be the normalization function, R be the spherical neighborhood search radius of the i-th original point, and range be the normalization range. The contextual features of the normal vector for each point are given, and the specific relationships are as follows: ； Among them, NCF i For the context features of the normal vector of the i-th original point, N xi Let N be the x-axis coordinate of the normal vector of the i-th original point. yi Let N be the y-axis coordinate of the normal vector of the i-th original point. zi Let z be the z-axis coordinate of the normal vector of the i-th original point; By merging elevation context features, intensity context features, density context features, and normal vector context features, feature enhancement of the first point cloud data is achieved.

2. The method for extracting substation equipment from lidar point clouds as described in claim 1, characterized in that, The second point cloud data is segmented to obtain a coarse-scale third point cloud data, specifically including the following steps: Feature embedding is performed on the second point cloud data using multiple cascaded deformers; In each cascaded deformer, based on the multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data to obtain multi-dimensional semantic features; Multidimensional semantic features are fed into a fully connected layer and subjected to summation linear functions and batch normalization to obtain coarse-scale third point cloud data.

3. The method for extracting substation equipment from lidar point clouds as described in claim 2, characterized in that, Based on a multi-head self-attention mechanism, features are extracted from different semantic information of the second point cloud data, specifically including: The semantic information extracted from the second point cloud data includes three feature matrices: Query, Key, and Value. The specific formulas for calculating the above three feature matrix values ​​are as follows: ； Where Q is the Query feature matrix value, K is the Key feature matrix value, V is the Value feature matrix value, and F is the Key feature matrix value. in For the input function, W q W represents the learnable shared weights in the Query feature matrix. k W represents the learnable shared weights in the Key feature matrix. v These are the learnable shared weights in the Value feature matrix; Based on the Query feature matrix values ​​and Key feature matrix values, the attention weights of the three feature matrices are obtained through matrix dot product, and the calculation formula is as follows: ； Where A represents the attention weights of the three feature matrices, and T is the transpose function; Based on the attention weights of the three feature matrices, semantic features are obtained by weighted summation and normalization.

4. The method for extracting substation equipment from lidar point clouds as described in claim 1, characterized in that, Feature extraction of coarse-scale third point cloud data is performed using an offsetting attention mechanism to obtain fine-scale fourth point cloud data. The specific steps include the following: Based on the offsetting attention mechanism, the differentiation features of coarse-scale third point cloud data are calculated by element subtraction to form a differentiation feature set. The distinct feature sets are input into the ensemble network model to eliminate noise, resulting in fine-scale fourth point cloud data.

5. The method for extracting substation equipment from lidar point clouds as described in claim 4, characterized in that, The differentiation characteristics of coarse-scale third-point cloud data are calculated using element-wise subtraction. The specific calculation formula is as follows: ； Among them, F out For the third point cloud data, OA is the differential feature, and F is the offsetting attention function used to calculate the differential values ​​of the input features. in For the third point cloud data, F mid For the intermediate output features of the third point cloud data, and F in -F mid The LBR is a discrete Laplace operator, a set network model, and includes functions for linear, batch normalization, and ReLU activation of the data.

6. The method for extracting substation equipment from lidar point clouds as described in claim 4, characterized in that, The substation equipment is extracted from the fine-scale fourth-point cloud data to obtain the scene equipment distribution of the substation. The specific steps include the following: Based on the K-nearest neighbor algorithm, local data points of any center point are determined in the fourth point cloud data at a fine scale; Using a channel-widening symmetric aggregation function, the edge features of the local data point are calculated within a given range; The edge features and the differentiation features of the third point cloud data are merged and input into a linear function, a batch normalization function, a ReLU function, and a Dropout function, respectively, to obtain the scene equipment distribution of the substation.

7. The method for extracting substation equipment from lidar point clouds as described in claim 6, characterized in that, Calculate the edge features of the local data point within a given range using the following formula: ； in, For the edge features of local data point i, It is a nonlinear transformation function. The vector context features of point i, Let i be the vector context feature of another point j within a given range of i, max be the maximum value function used as the symmetric aggregation function, and e be the edge connecting the points.

8. An apparatus for extracting substation equipment from lidar point clouds, characterized in that, The method for extracting substation equipment from lidar point clouds as described in any one of claims 1-7 includes: The acquisition module collects raw point clouds from the LiDAR in the substation scene to form the first point cloud data; The analysis and processing module, based on a feature transformation function, enhances the feature saliency of the first point cloud data to obtain the second point cloud data. The second point cloud data is then segmented to obtain a coarse-scale third point cloud data. Specifically, the process of enhancing the feature saliency of the first point cloud data to obtain the second point cloud data, based on the feature transformation function, includes the following steps: Define the coordinate system orientation of the original point cloud, where the coordinate system includes three axes: x, y, and z. The x-axis represents the front of the lidar device, the y-axis represents the right side of the lidar device, and the z-axis represents the top of the lidar device. The elevation and intensity context features of each original point are analyzed, and the specific calculation relationships are as follows: ； HCF i For the elevation context feature of the i-th original point, ICF i Let z be the intensity context feature of the i-th original point, cot be the cotangent function, δ be the preset transformation degree, and z be the intensity context feature of the i-th original point. i Let r be the z-axis coordinate of the i-th original point in the substation scenario. i Let be the intensity of the i-th original point in the substation scenario; The density context features of each original point are determined by the number of neighboring points within a predetermined neighborhood. The specific determination relationship is as follows: ； Among them, DCF i Let S be the density context feature of the i-th original point, S be the normalization function, R be the spherical neighborhood search radius of the i-th original point, and range be the normalization range. The contextual features of the normal vector for each point are given, and the specific relationships are as follows: ； Among them, NCF i For the context features of the normal vector of the i-th original point, N xi Let N be the x-axis coordinate of the normal vector of the i-th original point. yi Let N be the y-axis coordinate of the normal vector of the i-th original point. zi Let z be the z-axis coordinate of the normal vector of the i-th original point; By merging elevation context features, intensity context features, density context features, and normal vector context features, feature enhancement of the first point cloud data is achieved. The feature extraction module extracts features from the coarse-scale third point cloud data through an offsetting attention mechanism to obtain the fine-scale fourth point cloud data. The fine-scale fourth point cloud data is then used to extract substation equipment to obtain the scene equipment distribution of the substation.

Images