Substation three-dimensional model construction method based on deep learning

Abstract

The invention belongs to the technical field of transformer substations, and particularly relates to a transformer substation three-dimensional model construction method based on deep learning, which comprises the following specific steps: acquiring point cloud data of a transformer substation and constructing a transformer substation equipment model library; preprocessing the point cloud data of the transformer substation; segmenting the preprocessed point cloud data of the transformer substation to obtain a single point cloud set; and identifying the single point cloud set, and then searching a point cloud model of corresponding equipment in a substation equipment model library. According to the method, the PCSCK method is designed, the requirement for computing resources can be greatly reduced, meanwhile, the segmentation efficiency is remarkably improved, and rapid and accurate segmentation of the substation equipment is achieved. According to the point cloud model registration method provided by the invention, deep learning networks SERN and KPSM are utilized, the efficiency and robustness of transformer substation equipment registration are improved, and the problem that a registration algorithm falls into local optimum is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of substations, and in particular to a method for constructing a three-dimensional model of a substation based on deep learning. Background Art

[0002] Substations play a key role in the power system, responsible for the transmission, regulation and protection of electric energy, ensuring the stable operation of the power grid and the quality of electric energy, which is crucial to the industry and daily life of modern society. However, the current level of information management of various substations is uneven, and most of them are still managed based on two-dimensional plane diagrams. This method cannot fully and truly reflect the complex environment, line layout, equipment intersection and three-dimensional spatial properties of the substation site, resulting in low management efficiency.

[0003] In the context of big data, building a 3D model of a substation is a key technical foundation for achieving efficient operation and maintenance of the substation. Through 3D modeling, the distribution of equipment and lines in the substation can be presented intuitively, and maintenance personnel can locate and handle problems more accurately, improving maintenance efficiency. In addition, different operation and failure scenarios can be simulated to provide a virtual training platform for operation and maintenance personnel, improving their emergency response capabilities and technical level.

[0004] At present, the general process of substation modeling based on 3D laser point cloud data is to preprocess the original point cloud data of the substation, segment it to obtain an independent point cloud model, match it with the standard model library, and finally place the model in the standard model library in the scene to obtain the scene model. This process involves background noise removal, equipment segmentation and recognition, feature extraction technology and model matching. However, substation equipment is numerous and complex in structure, and point cloud data has the characteristics of large quantity, uneven distribution and high noise.

[0005] The traditional processing method has the following problems:

[0006] 1. When processing large-scale point cloud data, the improved Euclidean clustering may perform repeated calculations on the point cloud data, resulting in low computational efficiency, which makes device segmentation difficult.

[0007] 2. The traditional ICP registration algorithm has a large computational overhead for finding the nearest corresponding point. It only considers the distance between points and lacks the use of point cloud structure information. It is sensitive to initial transformations and easily falls into local optimal solutions, resulting in low device registration efficiency.

[0008] 3. Traditional modeling methods are inefficient, time-consuming and costly in reconstructing three-dimensional substation models.

[0009] In order to solve the above problems and realize rapid modeling of substations, the present invention proposes a substation three-dimensional model construction method based on deep learning, which improves the accuracy of the substation three-dimensional model, reduces costs and reduces modeling time. Summary of the invention

[0010] In view of the above-mentioned and / or existing problems in a method for constructing a three-dimensional model of a substation based on deep learning, the present invention is proposed.

[0011] Therefore, the purpose of the present invention is to provide a method for constructing a three-dimensional model of a substation based on deep learning, which can solve the above-mentioned existing problems.

[0012] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0013] A method for constructing a three-dimensional model of a substation based on deep learning, which includes the following specific steps:

[0014] S1: Collect substation cloud data and build substation equipment model library;

[0015] S2: pre-processing the substation cloud data;

[0016] S3: segment the pre-processed substation cloud data to obtain a single point cloud set;

[0017] S4: Identify the single point cloud set, and then search for the point cloud model of the corresponding equipment in the substation equipment model library;

[0018] S5: registering the single point cloud set and the point cloud model of the corresponding device to obtain a three-dimensional model of the target device;

[0019] S6: Place the three-dimensional models of all target devices to obtain a three-dimensional model of the substation.

[0020] As a preferred solution of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the specific steps of S1 are as follows:

[0021] S11, collect substation cloud data: according to the substation area, understand the substation layout and equipment distribution in detail, select the location to install the lidar equipment to ensure that the substation target area can be fully covered, and set the corresponding lidar layout plan based on the field of view occlusion, and arrange the lidar. At the same time, set the parameters of the lidar scanning range, scanning rate, and angular resolution to collect the substation cloud data N1. Each point cloud data in N1 is a triple<x,y,z> , where x is the x-axis coordinate in the unified coordinate system, y is the y-axis coordinate in the unified coordinate system, and z is the z-axis coordinate in the unified coordinate system;

[0022] S12, build a substation equipment model library: collect relevant information of various equipment in the substation, including drawings, photos and measurement data, use professional modeling software for manual modeling, and build three-dimensional models of various equipment in the substation; at the same time, use the constructed three-dimensional model to perform triangular meshing, and obtain a point cloud model after uniform density sampling; then save the models of various equipment into the substation equipment model library, where each type of model contains a three-dimensional model and a point cloud model, and the size of each type of model is consistent with the size of the actual equipment. Finally, a model library containing various substation equipment models is obtained.

[0023] As a preferred solution of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the specific steps of S2 are as follows:

[0024] S21, use DBSCAN algorithm to denoise the point cloud data: use DBSCAN algorithm to denoise N1 and obtain the denoised substation cloud data N2;

[0025] S22, use the RANSAC algorithm to remove the ground point cloud: use the RANSAC algorithm to remove the ground point cloud from N2, and obtain the substation cloud data N3 after the ground point cloud is removed and denoised.

[0026] As a preferred solution of the method for constructing a three-dimensional substation model based on deep learning described in the present invention, the specific steps of S3 are as follows:

[0027] S31: Establish an empty unvisited point set Q, the elements of which include all unvisited points, add each point in N3 to Q, and establish an empty visited point set P, the elements of which include all visited points, where the points with category information are visited points;

[0028] S32: For any point q in Q, a radius search method is used to search a sphere with q as the center and r as the radius, and all points within the sphere are added to a set R, which contains all points within the sphere with q as the center and r as the radius;

[0029] S33: Use the Kd-Tree algorithm to search all points in R, return the distance from each point to q, and sort them in ascending order, then select the k points farthest from q and form a set K;

[0030] S34: Determine whether the points in the above set K have been visited:

[0031] If there is no visited point in K, add a category information C to each point in R i, where i starts at 1 and is automatically increased by 1 each time a category is added; add each point in R to P, and then delete all visited points in Q;

[0032] If there are visited points in K, add all visited points to set B, which only contains the visited points queried in the current sphere, and find the earliest added category information C from B min , that is, the subscript min is the smallest of all category information, and the category information C of each point is queried in P p , if C p The category information C of any point in B b Same, that is, subscript p = b, then C p Set to C min , add all unvisited points in R to C min , then add to P, and finally delete all visited points in Q;

[0033] S35: Repeat S32 to S34 until Q is empty;

[0034] S36: For P, points with the same category information are clustered into a single point cloud set, and finally n single point cloud sets are obtained.

[0035] As a preferred solution of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the specific steps of S4 are as follows:

[0036] S41: input a single point cloud set;

[0037] S42, use the PointNet network to identify the monomer point cloud set: use the multi-layer perceptron MLP to extract features from each point in the monomer point cloud set, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128), Conv1d(128,512) and Conv1d(512,024) to extract features from each point in turn, expand its feature representation from 3 dimensions to 1024 dimensions, and then use the maximum pooling layer to obtain the global 1024-dimensional features of the monomer point cloud set. The global 1024-dimensional features of the monomer point cloud set are used as data input into several fully connected layers FC, and the FC converts the global features into an e-dimensional feature through linear transformation and nonlinear activation function. Finally, the e-dimensional features are normalized using the Softmax activation function to obtain the probability distribution of each category and determine what type of device it is;

[0038] S43: According to the identified equipment type, search for the corresponding type of substation equipment in the substation equipment model library, and finally obtain the point cloud model of the corresponding equipment.

[0039] As a preferred solution of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the specific steps of S5 are as follows:

[0040] S51: The point cloud model of the corresponding equipment in the substation equipment model library is used as the source point cloud data S, and the monomer point cloud set is used as the sample point cloud data T, which are input into the deep learning registration network SERN;

[0041] S52: SERN extracts, connects and aligns S and T to obtain a seven-tuple<x,y,z,q0,q1,q2,q3> ; In the seven-tuple, the first element x represents the distance moved along the x-axis; the second element y represents the distance moved along the y-axis; the third element z represents the distance moved along the z-axis; the last four elements are q0, q1, q2, and q3, respectively, where q0 is the cosine component of the rotation angle θ, cos(θ / 2), i.e., q0 = cos(θ / 2), θ refers to the angle of rotation of a point around the rotation axis according to the right-hand rule, and the rotation axis is a direction in three-dimensional space, represented as a unit vector Q1 is The component u in the x-axis direction x The product of and sin(θ / 2); q2 is The component u in the y-axis direction y The product of and sin(θ / 2); q3 is The component u in the z-axis direction z The product of q1 and sin(θ / 2), that is, q1 = u x sin(θ / 2),q2=u y sin(θ / 2),q3=u z sin(θ / 2); Direction by u x 、u y 、u z Jointly determined, the calculation formula is: That is, the rotation angle is obtained from q0, the direction of the rotation axis is obtained from q1, q2, and q3, and q0 2 +q1 2 +q2 2 +q3 2 =1;

[0042] S53: Using the seven-tuple, calculate the rotation and translation matrix; through the seven-tuple<x,y,z,q0,q1,q2,q3> The first element x, the second element y, and the third element z in the matrix are used to construct a translation matrix, and then the rotation matrix is ​​constructed according to the fourth to seventh elements q0, q1, q2, and q3. Finally, the two matrices are multiplied to calculate the rotation and translation matrix.

[0043] S531: Use x, y, and z to construct a translation matrix, which is:

[0044]

[0045] S532: Use q0, q1, q2, and q3 to construct a rotation matrix, which is:

[0046]

[0047] S533: Multiply the translation matrix and the rotation matrix to obtain a rotation-translation matrix, whose rotation-translation matrix A j for:

[0048]

[0049] S54: Based on A j , get the new source point cloud data S'; for each point in the point cloud data<a,b,c> Perform rotation and translation calculations to obtain the coordinates of the corresponding points<a',b',c'> , and its calculation formula is:

[0050]

[0051] After all points in S are rotated and translated, a new source point cloud data S' with all corresponding points is obtained, that is, the point cloud model of the corresponding equipment in the substation equipment model library is rotated and translated;

[0052] S55: Calculate the average Euclidean distance of the corresponding key points;

[0053] S551: Use the key point extraction module KPSM to extract the key points of S' and T; input the point cloud data T or S' into the key point extraction module in turn, use MLP to extract features for each point v, that is, use Conv1d(3,64) to extract features for v, and expand the feature representation of v from 3 dimensions to 64 dimensions; then input v into FC, the number of FC layers is (64,32,1), where the activation function of each layer is LeakyRelu, and each layer is normalized, and finally the importance score of v is obtained; the higher the importance score, the stronger the representativeness of the point, and vice versa, the weaker the representativeness, select the h points with higher rankings, and sort them, and then find the corresponding h points in turn, and finally get the three-dimensional coordinates of the h points.<x,y,z> ;

[0054] S552: Calculate the average Euclidean distance of the corresponding key points; for T and S', the key points in T are expressed as follows: <x 1i ,y 1i ,z 1i >, the key points of S' are expressed as <x 2i ,y 2i ,z 2i>, i = 1, 2, 3...h; each key point of T has a unique key point of S' corresponding to it, that is, key points with the same i correspond to each other;

[0055] Calculate the Euclidean distance d of the key points corresponding to S' and T in turn i , the formula is as follows:

[0056]

[0057] Calculate the average Euclidean distance d of the corresponding key points, the formula is as follows:

[0058]

[0059] S56: Set the threshold σ. If d>σ, T and S' are used as inputs and S52-S55 are repeated. Otherwise, the iteration ends.

[0060] S57: After the iteration is completed, all the rotation and translation matrices A generated during the iteration are j Multiply them together to get a final rotation and translation matrix A, that is j is the number of iterations, and then the three-dimensional model in the substation equipment model library is rotated and translated according to A, and finally the three-dimensional model of the target equipment is obtained. The target equipment is a substation equipment with an absolute position in a unified coordinate system after a specific type of equipment in the substation equipment model library is rotated and translated.

[0061] As a preferred solution of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the operation process of the deep learning registration network SERN is as follows:

[0062] Process 1: Use MLP to extract features from each point in S and T, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128) and Conv1d(128,512) to extract features from each point in turn, and expand its feature representation from 3 dimensions to 512 dimensions;

[0063] Process 2: Aggregate the features into a 512-dimensional global feature through the hybrid pooling layer;

[0064] Process 3: Use the Concatenate layer to aggregate the 512-dimensional global features of T and S into a 1024-dimensional feature, and convert the 1024-dimensional feature into a 7-dimensional feature through FC. The specific hierarchical structure is: FC(1024,512), FC(512,512), FC(512,256), FC(256,7);

[0065] Process 4: Convert 7-dimensional features into 7-tuples<x,y,z,q0,q1,q2,q3> .

[0066] As a preferred scheme of the method for constructing a three-dimensional model of a substation based on deep learning described in the present invention, the specific steps of S6 are as follows: the three-dimensional models of all target equipment are placed to obtain a three-dimensional model of the substation; in a unified coordinate system, the three-dimensional models of all target equipment are placed in sequence according to the absolute position information of the three-dimensional model of the target equipment, and finally the three-dimensional model of the substation is obtained, that is, in a professional modeling software, a three-dimensional space is set up, and the origin position is determined, and the three-dimensional models of all target equipment are placed according to the absolute coordinates of the three-dimensional model of the target equipment, and finally the three-dimensional model of the substation is obtained.

[0067] Compared with existing technologies:

[0068] 1. The present invention designs a PCSCK method, which can greatly reduce the demand for computing resources, while significantly improving the segmentation efficiency, and realizing fast and accurate segmentation of substation equipment.

[0069] 2. The present invention trains the deep learning network PointNet, which improves the recognition accuracy of substation equipment and has high reusability, effectively reducing costs.

[0070] 3. The present invention proposes a point cloud model registration method, which uses deep learning networks SERN and KPSM to improve the efficiency and robustness of substation equipment registration and avoid the problem of the registration algorithm falling into local optimality. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic diagram of the process of the present invention;

[0072] Figure 2 It is a schematic diagram of the PCSCK algorithm flow of the present invention;

[0073] Figure 3 This is a schematic diagram of the process of identifying a single point cloud set in the present invention;

[0074] Figure 4 This is a schematic diagram of the substation equipment registration process of the present invention;

[0075] Figure 5 A schematic diagram of the SERN registration network of the present invention;

[0076] Figure 6 It is a schematic diagram of the KPSM process of the present invention. DETAILED DESCRIPTION

[0077] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0078] The present invention provides a method for constructing a three-dimensional model of a substation based on deep learning. Figure 1-Figure 6 , including the following specific steps:

[0079] S1: Collect substation cloud data and build substation equipment model library;

[0080] The specific steps of S1 are as follows:

[0081] S11, collect substation cloud data: according to the substation area, understand the substation layout and equipment distribution in detail, select the location to install the lidar equipment to ensure that the substation target area can be fully covered, and set the corresponding lidar layout plan based on the field of view occlusion, and arrange the lidar. At the same time, set the parameters of the lidar scanning range, scanning rate, and angular resolution to collect the substation cloud data N1. Each point cloud data in N1 is a triple<x,y,z> , where x is the x-axis coordinate in the unified coordinate system, y is the y-axis coordinate in the unified coordinate system, and z is the z-axis coordinate in the unified coordinate system;

[0082] The details are as follows:

[0083] (1) Collect the area, layout, and equipment distribution information of the substation to be modeled;

[0084] (2) Determine the model of laser radar equipment to be used (Faro laser scanner) and the location of its placement;

[0085] (3) Set the parameters of the laser radar scanning range, scanning rate, and angular resolution;

[0086] (4) Collecting initial point cloud data of the target;

[0087] (5) Re-collect the occlusion and missing information;

[0088] (6) Obtain the complete substation cloud dataset N1;

[0089] S12, building a substation equipment model library: collect relevant information of various equipment in the substation, including drawings, photos and measurement data, use professional modeling software to manually model, and build three-dimensional models of various substation equipment; at the same time, use the constructed three-dimensional model to perform triangular meshing, and obtain a point cloud model after uniform density sampling; then save the models of various equipment to the substation equipment model library, where each type of model contains a three-dimensional model and a point cloud model, and the size of each type of model is consistent with the size of the actual equipment, and finally obtain a model library containing various substation equipment models;

[0090] The details are as follows:

[0091] Use AutoCAD to model various equipment in substations, including but not limited to: current transformers, voltage transformers, transformers, insulators, circuit breakers, disconnectors, lightning rods, capacitors, reactors, etc.;

[0092] To model the insulators in the substation equipment, the process is as follows:

[0093] (1) Obtain drawings, specifications, photos and dimensions of insulators;

[0094] (2) Create a new project file in AutoCAD and configure the basic parameters of the project;

[0095] (3) If there are existing 2D drawings, they can be imported into AutoCAD; draw the basic plan of the insulator based on the 2D drawings and determine the main structural dimensions;

[0096] (4) Use 3D modeling tools (such as "extend" and "rotate") to start building the basic 3D shape of the insulator, add other components on the insulator, such as support frames, terminals, etc., and adjust the size and proportion of the 3D model according to actual specifications;

[0097] (5) Enhance the details of the model as needed to more accurately reflect the actual insulator;

[0098] (6) Obtaining a three-dimensional model of the insulator, meshing the three-dimensional model into triangular surfaces, and then uniformly sampling;

[0099] (7) Obtain the insulator point cloud model;

[0100] (8) Output and save the three-dimensional model and point cloud model of the insulator;

[0101] Among them: Triangular meshing: the process of converting a 3D model into a mesh composed of triangles; triangular meshing is mainly divided into the following steps:

[0102] (1) Sampling, i.e. uniformly sampling a set of points on the 3D model; these sampling points will serve as the vertices of the mesh;

[0103] (2) triangulation, connecting the sampling points into triangles to generate a mesh;

[0104] (3) Fitting: aligning the triangles connected by the sampling points with the surface of the 3D model through interpolation and optimization algorithms;

[0105] (4) Optimization: Optimize the generated mesh to make it have better quality;

[0106] Triangular meshing is mature and can better retain the characteristics of the device. After the device is meshed, density sampling is performed, that is, points are uniformly and randomly generated on each grid to form a point cloud model.

[0107] 3D model: A three-dimensional model built by 3D software, which can be rotated and translated;

[0108] Point cloud model: A data structure used to represent objects or scenes in three-dimensional space, containing multiple three-dimensional coordinate points<x,y,z> A collection of;

[0109] The professional modeling software used in this step include: AutoCAD, Blender, Unity, etc.;

[0110] S2: pre-processing the substation cloud data;

[0111] The specific steps of S2 are as follows:

[0112] S21, use DBSCAN algorithm to denoise the point cloud data: use DBSCAN algorithm to denoise N1 and obtain the denoised substation cloud data N2;

[0113] The details are as follows:

[0114] (1) Set the domain radius r and the minimum number of points MinPts within the domain radius to become a core object;

[0115] (2) Start by randomly selecting a point p in N1 and marking it as “visited”;

[0116] (3) Check whether it is a core point (i.e., the r neighborhood of p has at least MinPts points). If it is not a core point, mark it as a noise point and go to (5); otherwise, create a new cluster C for p and put all the points in the r neighborhood of p into the candidate set R;

[0117] (4) Iteratively add points in R that do not belong to other clusters to C. In this process, for a point q marked as "unvisited" in R, mark it as "visited" and check its r neighborhood. If q is also a core point, all points in q's r neighborhood are added to R. Continue to add point q to C until C cannot be expanded, that is, until R is empty. At this point, C is completely generated.

[0118] (5) Randomly select the next unvisited point from N1 and repeat the process (3) until all points have been visited;

[0119] (6) Delete the points marked as noise points to obtain the denoised substation cloud data N2;

[0120] DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a publicly available density-based clustering algorithm that uses density features as the point cluster classification standard, determines the density around the point to distinguish the regional density, and records the points in the low-density area as outliers, i.e., noise; the DBSCAN algorithm has two main parameters: the domain radius: r; and the minimum number of points within the domain radius to become a core object: MinPts;

[0121] S22, using the RANSAC algorithm to remove the ground point cloud: using the RANSAC algorithm to remove the ground point cloud from N2, and obtaining the substation cloud data N3 after the ground point cloud is removed and denoised;

[0122] The details are as follows:

[0123] (1) In N2, randomly select three non-collinear points p1 =<x1,y1,z1> , p2=<x2,y2,z2> , p3=<x3,y3,z3> As the minimum sample subset, and vectorized representation, calculate the normal vector of the plane where the three points are located The calculation formula is as follows:

[0124]

[0125] (2) After the plane is determined, calculate the distance d from each point in N2 except p1, p2, and p3 to the plane. i , the calculation formula is as follows:

[0126]

[0127] in, for The transpose of i is any point in N2 (except p1, p2, p3), is the vector modulus, representing the length of the vector, i = 1, 2, 3, ..., m-3, m is the number of substation clouds;

[0128] (3) Set an appropriate threshold σ (need to be set according to the actual scenario of the substation, determine the σ value, and the initial σ value can be set to 0.1 m, and adjusted according to the effect of filtering out the ground point cloud data). i Points <σ are classified as the plane, and the number of points in the plane is counted;

[0129] (4) Repeat (1)-(3) g times (need to be set according to the actual scenario of the substation, g can be initially set to 1000, and adjusted according to the effect of filtering out the ground point cloud data. For example, if there are too many ground points extracted or incomplete ground points, the effect is poor, and g can be gradually increased);

[0130] (5) After the iteration is completed, the plane with the most points is selected as the ground point cloud. After removing the ground point cloud from N2, the substation cloud data N3 after removing the ground point cloud and denoising is obtained;

[0131] Among them: RANSAC (Random sample consensus) is a publicly available algorithm based on statistical probability. After obtaining the shape of the object in advance, the parameters representing the mathematical model are solved iteratively to finally obtain a point cloud set representing the object. The key parameters are the threshold σ and the number of iterations k;

[0132] S3: Segment the pre-processed substation cloud data to obtain a single point cloud set; the present invention uses the self-designed PCSCK algorithm (Point Cloud Spatial Clustering With Kd-Tree) to segment N3, thereby obtaining several single point cloud sets, as shown in the following figure: Figure 2 As shown;

[0133] The specific steps of S3 are as follows:

[0134] S31: Establish an empty unvisited point set Q, the elements of which include all unvisited points, add each point in N3 to Q, and establish an empty visited point set P, the elements of which include all visited points, where the points with category information are visited points;

[0135] S32: For any point q in Q, a radius search method is used to search a sphere with q as the center and r as the radius, and all points within the sphere are added to a set R, which contains all points within the sphere with q as the center and r as the radius;

[0136] S33: Use the Kd-Tree algorithm to search all points in R, return the distance from each point to q, and sort them in ascending order, then select the k points farthest from q and form a set K;

[0137] S34: Determine whether the points in the above set K have been visited:

[0138] If there is no visited point in K, add a category information C to each point in R i, where i starts at 1 and is automatically increased by 1 each time a category is added; add each point in R to P, and then delete all visited points in Q;

[0139] If there are visited points in K, add all visited points to set B, which only contains the visited points queried in the current sphere, and find the earliest added category information C from B min , that is, the subscript min is the smallest of all category information, and the category information C of each point is queried in P p , if C p The category information C of any point in B b Same, that is, subscript p = b, then C p Set to C min , add all unvisited points in R to C min , then add to P, and finally delete all visited points in Q;

[0140] S35: Repeat S32 to S34 until Q is empty;

[0141] S36: For P, points with the same category information are clustered into a single point cloud set, and finally n single point cloud sets are obtained;

[0142] Among them: each point of the substation data is only visited once, and Kd-Tree is used to improve the speed of obtaining the single point cloud set;

[0143] Radius search method: take the query point as the center of the sphere and the query distance as the radius, and find all data from the data set whose distance from the center of the sphere is less than the threshold r (the threshold is also the radius r);

[0144] Kd-Tree algorithm: Kd-Tree (also known as kd tree) is a high-dimensional spatial index structure used to partition k-dimensional data space. Its essence is a constrained binary search tree. It is mainly used to quickly find the nearest neighbor or range search of data points in k-dimensional space. The core idea of ​​Kd-Tree is to recursively divide the space into smaller subspaces to speed up the search.

[0145] The algorithm flow of using Kd-Tree to query the farthest k points is as follows:

[0146] 1. Construction of Kd-Tree:

[0147] 1. Select the partitioning dimension: When constructing a Kd-Tree, first select a dimension as the partitioning axis; usually select dimensions in turn (in three-dimensional space, the root node is the partition plane perpendicular to the x-axis, its child nodes are all partition planes perpendicular to the y-axis, its grandchild nodes are all partition planes perpendicular to the z-axis, its great-grandchild nodes are all partition planes perpendicular to the x-axis, and so on);

[0148] 2. Select the split point: On the selected split dimension, split the data set into two parts according to the median of the dimension (or other split point); the selection of the median makes the amount of data after splitting more uniform and the depth of the tree smaller;

[0149] 3. Recursively construct subtrees: After dividing the space into two parts, recursively construct subtrees for the data in these two parts respectively; the left subtree corresponds to the subspace to the left of the split point, and the right subtree corresponds to the subspace to the right of the split point;

[0150] 4. Termination condition: The recursion continues until there are no data points in the node or only one data point remains;

[0151] 2. Initialize the small top heap:

[0152] Use a small top heap to store the k points farthest from the center of the sphere found so far; the size of the heap is k, and the top of the heap stores the point closest to the center of the sphere among the k points;

[0153] 3. Recursively traverse Kd-Tree:

[0154] 1. Starting from the root node, recursively traverse the entire Kd-Tree;

[0155] 2. Calculate the distance: For each point on the node, calculate the Euclidean distance from the point to the center of the sphere;

[0156] 3. Update the small top heap:

[0157] (1) If the size of the heap is less than k, directly add the current point to the heap;

[0158] (2) If the size of the heap is already k, and the distance of the current point is greater than the distance of the heap vertex (that is, the smallest distance in the heap), remove the heap vertex and add the current point to the heap;

[0159] 4. Select a subtree to continue traversing:

[0160] (1) If the distance from the center of the sphere to the splitting plane of the current node is greater than the distance to the heap vertex, there may be a farther point, so the corresponding subtree needs to be recursively checked;

[0161] (2) If the distance from the center of the sphere to the splitting plane is less than or equal to the distance from the heap vertex, the subtree can be skipped to avoid unnecessary calculations;

[0162] 5. Repeat the above steps until all nodes have been visited;

[0163] 4. Termination conditions:

[0164] When the recursive traversal ends, the k points in the small top heap are the k points farthest from the target point;

[0165] S4: Identify the single point cloud set, and then search for the point cloud model of the corresponding equipment in the substation equipment model library;

[0166] The specific steps of S4 are as follows:

[0167] S41: input a single point cloud set;

[0168] S42, use the PointNet network to identify the monomer point cloud set: use the multi-layer perceptron MLP to extract features from each point in the monomer point cloud set, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128), Conv1d(128,512) and Conv1d(512,024) to extract features from each point in turn, expand its feature representation from 3 dimensions to 1024 dimensions, and then use the maximum pooling layer (MaxPool operation means selecting the maximum value in each dimension, thereby retaining the most significant features of the input point cloud) to obtain the global 1024-dimensional features of the monomer point cloud set. The global 1024-dimensional features of the monomer point cloud set are input as data into several fully connected layers FC, and the FC converts the global features into an e-dimensional feature through linear transformation and nonlinear activation function (ReLU). Finally, the e-dimensional features (where e is the number of device categories, which is a known parameter) are normalized using the Softmax activation function to obtain the probability distribution of each category and determine what type of device it is;

[0169] S43: searching for a substation device of a corresponding type in a substation device model library according to the identified device type, and finally obtaining a point cloud model of the corresponding device;

[0170] in:

[0171] Multilayer Perceptron (MLP) is a feed-forward artificial neural network model that maps multiple input data sets to a single output data set.

[0172] Conv1d: represents a one-dimensional convolution layer, which is usually used to process sequence data or one-dimensional signals. In Conv1d(3,64), 3 is the number of input channels. For point cloud data, 3 usually represents the 3D coordinates (x, y, z) of each point, so the number of input channels is 3. 64 is the number of output channels. The function of the convolution layer is to perform convolution operations on the input data through several convolution kernels, mapping the input features to new features. The number of output channels 64 means that the convolution layer uses 64 different convolution kernels, each of which generates an output channel, and finally maps the input 3D features to 64-dimensional features.

[0173] Training PointNet network

[0174] (1) Constructing a training sample set

[0175] Use the point cloud models of the device constructed in Step 1, a total of E, randomly add Gaussian noise points to each device point cloud model, a total of E; randomly scale each device point cloud model, a total of E; randomly rotate each device point cloud model to simulate different perspectives, a total of E; then use them together as a training set, a total of 4E device point cloud models in the training set, and assign a category label to each point cloud model in the sample set, for example: the insulator point cloud model is labeled as insulator;

[0176] (2) Train the PointNet network to obtain a network model that can automatically identify substation equipment

[0177] First, randomly initialize the network parameters and set hyperparameters, such as learning rate and optimizer; here, set the learning rate to 0.001 and use the Adam optimizer;

[0178] Then, each point cloud model in the sample set is input into the PointNet network; the input of the PointNet network is an N x 3 matrix, where N is the number of points and 3 is the 3D coordinates (x, y, z) of each point; the network first extracts features from each point through a multi-layer perceptron (MLP); specifically, the MLP contains several 1D convolutional layers and activation functions (such as ReLU) to gradually increase the feature dimension of each point to 1024 dimensions;

[0179] Next, the MaxPool operation is used to aggregate the 1024-dimensional features of each point into a 1x1024 global feature vector; the MaxPool operation selects the maximum value in each dimension, thereby retaining the most significant features of the input point cloud; the key to this step is to capture the overall shape information of the point cloud through global feature aggregation; the resulting 1x1024 global feature vector is further processed through one or more fully connected layers (FC); these fully connected layers convert the global features into an e-dimensional feature vector through linear transformations and nonlinear activation functions (such as ReLU), where e is the number of categories, which is a known parameter;

[0180] Next, the e-dimensional feature vector is normalized using the Softmax activation function to obtain the probability distribution of each category; the Softmax function converts the output of the network into a probability vector, each element of which represents the probability that the input point cloud belongs to the corresponding category, and determines the label; the label predicted by the network is compared with the true label to calculate the loss value; it measures the difference between the predicted probability distribution and the true distribution; the gradient is calculated based on the loss value, and the gradient is back-propagated to each layer in the network to update the network parameters;

[0181] By repeating the above steps, the network parameters are gradually optimized, so that the performance of the network on the training set is continuously improved and the loss value is continuously reduced; after multiple rounds of training, a trained PointNet model is finally obtained, which can accurately recognize the input point cloud data;

[0182] (3) The PointNet network is a network used to process point cloud data. The network input is usually multiple point cloud 3D coordinates;

[0183] S5: registering the single point cloud set and the point cloud model of the corresponding device to obtain a three-dimensional model of the target device;

[0184] The specific steps of S5 are as follows:

[0185] S51: The point cloud model of the corresponding equipment in the substation equipment model library is used as the source point cloud data S, and the monomer point cloud set is used as the sample point cloud data T, which are input into the deep learning registration network SERN;

[0186] S52: SERN extracts, connects and aligns S and T to obtain a seven-tuple<x,y,z,q0,q1,q2,q3> ; In the seven-tuple, the first element x represents the distance moved along the x-axis; the second element y represents the distance moved along the y-axis; the third element z represents the distance moved along the z-axis; the last four elements are q0, q1, q2, and q3, respectively, where q0 is the cosine component of the rotation angle θ, cos(θ / 2), i.e., q0 = cos(θ / 2), θ refers to the angle of rotation of a point around the rotation axis according to the right-hand rule, and the rotation axis is a direction in three-dimensional space, represented as a unit vector Q1 is The component u in the x-axis direction x The product of and sin(θ / 2); q2 is The component u in the y-axis direction y The product of and sin(θ / 2); q3 is The component u in the z-axis direction z The product of q1 and sin(θ / 2), that is, q1 = u x sin(θ / 2),q2=u y sin(θ / 2),q3=u z sin(θ / 2); Direction by u x 、u y 、u z Jointly determined, the calculation formula is: That is, the rotation angle is obtained from q0, the direction of the rotation axis is obtained from q1, q2, and q3, and q0 2 +q1 2 +q2 2 +q3 2 =1;

[0187] Wherein: the operation process of the deep learning registration network SERN is as follows:

[0188] Process 1: Use MLP to extract features from each point in S and T, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128) and Conv1d(128,512) to extract features from each point in turn, and expand its feature representation from 3 dimensions to 512 dimensions;

[0189] Process 2: Aggregate the features into a 512-dimensional global feature through a mixed pooling layer (MixedPool mixed pooling means selecting the maximum value and the average value in each dimension and adding them to obtain a mixed value, retaining the most significant features in the input point cloud);

[0190] Process 3: Use the Concatenate layer to aggregate the 512-dimensional global features of T and S into a 1024-dimensional feature, and convert the 1024-dimensional feature into a 7-dimensional feature through FC. The specific hierarchical structure is: FC(1024,512), FC(512,512), FC(512,256), FC(256,7);

[0191] Process 4: Convert 7-dimensional features into 7-tuples<x,y,z,q0,q1,q2,q3> ;

[0192] S53: Using the seven-tuple, calculate the rotation and translation matrix; through the seven-tuple<x,y,z,q0,q1,q2,q3> The first element x, the second element y, and the third element z in the matrix are used to construct a translation matrix, and then the rotation matrix is ​​constructed according to the fourth to seventh elements q0, q1, q2, and q3. Finally, the two matrices are multiplied (translation matrix multiplied by rotation matrix) to calculate the rotation and translation matrix.

[0193] S531: Use x, y, and z to construct a translation matrix, which is:

[0194]

[0195] S532: Use q0, q1, q2, and q3 to construct a rotation matrix, which is:

[0196]

[0197] S533: Multiply the translation matrix and the rotation matrix to obtain a rotation-translation matrix, whose rotation-translation matrix A j for:

[0198]

[0199] Among them, the right-hand rule: hold the rotation axis with your right hand, with your thumb pointing to the positive direction of the rotation axis, then the direction pointed by the other four fingers is the rotation direction when the rotation angle is positive;

[0200] Positive angle: When the rotation angle θ is positive, the direction of rotation is determined by the right-hand rule; that is, if you point your right thumb in the direction of the rotation axis, the direction in which your fingers curl is the direction of rotation; in this case, the rotation is counterclockwise (from the positive direction of the rotation axis);

[0201] Negative angle: When the rotation angle θ is negative, the rotation direction is opposite to that of the positive angle, that is, clockwise (from the positive direction of the rotation axis);

[0202] S54: Based on A j, get the new source point cloud data S'; for each point in the point cloud data<a,b,c> Perform rotation and translation calculations to obtain the coordinates of the corresponding points<a',b',c'> , and its calculation formula is:

[0203]

[0204] After all points in S are rotated and translated, a new source point cloud data S' with all corresponding points is obtained, that is, the point cloud model of the corresponding equipment in the substation equipment model library is rotated and translated;

[0205] S55: Calculate the average Euclidean distance of the corresponding key points;

[0206] First, the key point extraction module KPSM (Key Points election Module) is used to obtain the three-dimensional coordinates of h ordered key points of T and S', respectively. A key point refers to a point with the strongest feature representation in point cloud data, that is, a point with a strong ability to describe surface shape. In KPSM, importance scores are used to describe the strength of feature representation. The higher the importance score, the stronger the representativeness of the point, and vice versa. Then, the average Euclidean distance d of the corresponding key points of T and S' is calculated. The corresponding key point refers to the key point with the highest importance score in T corresponding to the key point with the highest importance score in S', and so on. Finally, the h key points of T and S' are made to correspond to each other.

[0207] S551: Use the key point extraction module KPSM to extract the key points of S' and T; input the point cloud data T or S' into the key point extraction module in turn, use MLP to extract features for each point v, that is, use Conv1d(3,64) to extract features for v, and expand the feature representation of v from 3 dimensions to 64 dimensions; then input v into FC, the number of FC layers is (64,32,1), where the activation function of each layer is LeakyRelu, and each layer is normalized, and finally the importance score of v is obtained; the higher the importance score, the stronger the representativeness of the point, and vice versa, the weaker the representativeness, select the h points with higher rankings, and sort them, and then find the corresponding h points in turn, and finally get the three-dimensional coordinates of the h points.<x,y,z> ;

[0208] S552: Calculate the average Euclidean distance of the corresponding key points; for T and S', the key points in T are expressed as follows: <x 1i ,y 1i ,z 1i >, the key points of S' are expressed as <x 2i ,y 2i ,z 2i>, i = 1, 2, 3...h; each key point of T has a unique key point of S' corresponding to it, that is, key points with the same i correspond to each other;

[0209] Calculate the Euclidean distance d of the key points corresponding to S' and T in turn i , the formula is as follows:

[0210]

[0211] Calculate the average Euclidean distance d of the corresponding key points, the formula is as follows:

[0212]

[0213] S56: Set the threshold σ. If d>σ, T and S' are used as inputs and S52-S55 are repeated. Otherwise, the iteration ends.

[0214] Among them: σ: represents the average distance between the key points of the two point cloud models, that is, the average Euclidean distance of the key points is greater than σ, indicating that the registration effect is poor; it is used to determine the number of iterations of the registration network, which needs to be set according to the actual scenario of the substation, for example, it can be initially set to 0.1m;

[0215] S57: After the iteration is completed, all the rotation and translation matrices A generated during the iteration are j Multiply them together to get a final rotation and translation matrix A, that is j is the number of iterations, and then the three-dimensional model in the substation equipment model library is rotated and translated according to A, and finally the three-dimensional model of the target equipment is obtained. The target equipment is a substation equipment with an absolute position in a unified coordinate system after a specific type of equipment in the substation equipment model library is rotated and translated;

[0216] in:

[0217] The registration network and key point extraction network need to be pre-trained;

[0218] 1. Registration Network SERN

[0219] (1) Constructing a training set

[0220] Taking the device point cloud model constructed in Step 1 as the sample set, for each sample template in the training set, a rotation matrix R and a translation vector t are randomly generated, where R represents the rotation around the x-axis. Rotate around the y axis Rotate around the z axis The matrix t is generated by uniform sampling from [-1,1]. Apply R,t to the template point cloud to generate the source point cloud, thus obtaining the point cloud model to be registered.

[0221] (2) Training Model

[0222] First, randomly initialize the network parameters and set hyperparameters, such as learning rate and optimizer; here, set the learning rate to 0.001 and use the Adam optimizer;

[0223] Then, the template point cloud model and the source point cloud model in the sample set are input into the registration network, and the multi-layer perceptron (MLP) is used to extract features from the point cloud. For each point in each point cloud model, Conv1d(3,64), Conv1d(64,64), Conv1d(64,128) and Conv1d(128,512) are used in turn for feature extraction, and then the features are aggregated into a 512-dimensional global feature through a mixed pooling layer (MixedPool mixed pooling layer means selecting the maximum value and the average value in each dimension and adding them up), thereby retaining the most significant features of the input point cloud. Then, the The Concatenate layer aggregates the two 512-dimensional global features of the Template point cloud and the Source point cloud into a 1024-dimensional feature. Finally, the fully connected layer (FC) converts the 1024-dimensional feature into a 7-dimensional feature. The specific hierarchy is: FC(1024,512), FC(512,512), FC(512,256), FC(256,7). Finally, a seven-tuple is output to represent the transformation parameters of the point cloud registration. The first three are the translation amounts, and the last four are the coefficients of the unit quaternion. The Source point cloud is transformed according to the predicted rotation and translation matrix, and ideally should coincide with the Template point cloud.

[0224] During the training process, the loss function is defined, including geometric distance loss and feature matching loss, and the error is calculated based on the loss function; the error is back-propagated to update the model parameters, so that the performance of the network on the training set is continuously improved and the loss value is continuously reduced;

[0225] Finally, a trained SERN model is obtained, which can accurately register the input point cloud model and output a seven-tuple;

[0226] 2 Key point extraction module KPSM:

[0227] (1) Constructing a training set

[0228] Use the training set of the PointNet network, and for each substation equipment point cloud data set in the training set, select h points with the strongest representative features as key points and assign importance scores. The remaining points are randomly set with scores lower than the minimum importance score. The key points of the same type of equipment are the same. When all samples are successfully marked, they are used as the training sample set of the key point extraction module.

[0229] Training key point extraction module

[0230] The input point cloud data is first processed by a multi-layer perceptron (MLP). In this process, assume that we have n points and the data dimension of each point is 3 (usually 3D coordinates: x, y, z). Through a layer of MLP, the data of each point is converted into a 64-dimensional feature vector. The structure of this layer of MLP can be expressed as (3, 64), where the input dimension is 3 and the output dimension is 64.

[0231] After extracting the 64-dimensional features of each point, these feature vectors are input into a fully connected layer network; this fully connected layer network consists of two layers with a structure of (64, 32, 1); the first layer reduces the 64-dimensional features to 32 dimensions, and the second layer further reduces the dimensions to 1 dimension; each layer uses LeakyReLU as an activation function to introduce nonlinearity so that the model can capture more complex features; in addition, batch normalization (BN) is applied after each layer to accelerate training and stabilize model performance;

[0232] After being processed by the fully connected layer network, each point will receive a score; the higher the score, the more representative the point is in the point cloud; the lower the score, the less representative the point is; all points are sorted according to the score, and h points with higher scores are selected. For the same device point cloud model, ideally, the h points are the same;

[0233] During the training process, define a suitable loss function (such as mean square error loss) and optimizer (such as Adam), and perform model training. In each training cycle, the model performs forward propagation and back propagation on the training set, calculates the error through the loss function, and updates the model parameters.

[0234] S6: placing the three-dimensional models of all target devices to obtain a three-dimensional model of the substation;

[0235] The specific steps of S6 are as follows: placing the three-dimensional models of all target devices to obtain a three-dimensional model of the substation; placing the three-dimensional models of all target devices in sequence according to the absolute position information of the three-dimensional models of the target devices in a unified coordinate system to finally obtain the three-dimensional model of the substation, that is, in a professional modeling software, setting a three-dimensional space and determining the origin position, placing the three-dimensional models of all target devices according to the absolute coordinates of the three-dimensional models of the target devices to finally obtain the three-dimensional model of the substation;

[0236] The details are as follows:

[0237] (1) In a unified coordinate system, determine the location of the origin (0, 0, 0);

[0238] (2) Query the absolute position of the three-dimensional model of the target device;

[0239] (3) placing the three-dimensional model of the target device at the corresponding position in the unified coordinate system;

[0240] (4) For example, if the absolute position of the transformer is (1, 1, 1), then the transformer is placed at (1, 1, 1), and the lightning arrester is at (2, 1, 1), then the lightning arrester should be placed at (2, 1, 1);

[0241] (5) After all target equipment is placed, the three-dimensional model of the substation is obtained.

[0242] Based on the above, the present invention first uses LiDAR technology to collect high-precision point cloud data of the substation. Then, by collecting and integrating detailed information of various types of equipment, a comprehensive substation equipment model library is constructed, thereby significantly improving the construction accuracy of the substation 3D model. Then, the DBSCAN and RANSAC algorithms are used to pre-process the substation site cloud data.

[0243] Aiming at the limitations of the traditional Euclidean clustering algorithm, this paper proposes a point cloud spatial clustering method PCSCK. This method accelerates the clustering of substation cloud data by only calculating each point in the point cloud data once and using Kd-Tree, which greatly reduces the demand for computing resources and significantly improves the segmentation efficiency, thus achieving fast and accurate segmentation of substation equipment.

[0244] Furthermore, the present invention trains and optimizes a deep learning recognition network PointNet, which can quickly identify substation equipment and has a high degree of reusability, effectively reducing costs.

[0245] In view of the shortcomings of the ICP registration algorithm, the present invention proposes a point cloud model registration method, which iteratively uses the deep learning registration network SERN for registration. During the iteration, the key points are extracted by the key point extraction module KPSM, and the average Euclidean distance of the corresponding key points is calculated to determine the number of SERN iterations. This method improves the efficiency and robustness of substation equipment registration and avoids the problem of the registration algorithm falling into the local optimum.

[0246] Finally, in a unified coordinate system, the 3D models of all substation equipment are accurately placed to form a complete 3D model of the substation.

[0247] In practical applications, the present invention can quickly reconstruct the three-dimensional model of the substation, which not only improves the accuracy of the model, but also reduces the cost of making the three-dimensional model of the substation, providing great convenience for subsequent operation, maintenance and management.

[0248] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for constructing a three-dimensional model of a substation based on deep learning, characterized in that: The specific steps are as follows: S1: Collect substation cloud data and build substation equipment model library; S2: pre-processing the substation cloud data; S3: segment the pre-processed substation cloud data to obtain a single point cloud set; S4: Identify the single point cloud set, and then search for the point cloud model of the corresponding equipment in the substation equipment model library; S5: registering the single point cloud set and the point cloud model of the corresponding device to obtain a three-dimensional model of the target device; S6: Place the three-dimensional models of all target devices to obtain a three-dimensional model of the substation.

2. A method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S1 are as follows: S11, collect substation cloud data: according to the substation area, understand the substation layout and equipment distribution in detail, select the location to install the lidar equipment to ensure that the substation target area can be fully covered, and set the corresponding lidar layout plan based on the field of view occlusion, and arrange the lidar. At the same time, set the parameters of the lidar scanning range, scanning rate, and angular resolution to collect the substation cloud data N1. Each point cloud data in N1 is a triple<x,y,z> , where x is the x-axis coordinate in the unified coordinate system, y is the y-axis coordinate in the unified coordinate system, and z is the z-axis coordinate in the unified coordinate system; S12, build a substation equipment model library: collect relevant information of various equipment in the substation, including drawings, photos and measurement data, use professional modeling software for manual modeling, and build three-dimensional models of various equipment in the substation; at the same time, use the constructed three-dimensional model to perform triangular meshing, and obtain a point cloud model after uniform density sampling; then save the models of various equipment into the substation equipment model library, where each type of model contains a three-dimensional model and a point cloud model, and the size of each type of model is consistent with the size of the actual equipment. Finally, a model library containing various substation equipment models is obtained.

3. The method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S2 are as follows: S21, use DBSCAN algorithm to denoise the point cloud data: use DBSCAN algorithm to denoise N1 and obtain the denoised substation cloud data N2; S22, use the RANSAC algorithm to remove the ground point cloud: use the RANSAC algorithm to remove the ground point cloud from N2, and obtain the substation cloud data N3 after the ground point cloud is removed and denoised.

4. The method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S3 are as follows: S31: Establish an empty unvisited point set Q, the elements of which include all unvisited points, add each point in N3 to Q, and establish an empty visited point set P, the elements of which include all visited points, where the points with category information are visited points; S32: For any point q in Q, a radius search method is used to search a sphere with q as the center and r as the radius, and all points within the sphere are added to a set R, which contains all points within the sphere with q as the center and r as the radius; S33: Use the Kd-Tree algorithm to search all points in R, return the distance from each point to q, and sort them in ascending order, then select the k points farthest from q and form a set K; S34: Determine whether the points in the above set K have been visited: If there is no visited point in K, add a category information C to each point in R i , where i starts at 1 and is automatically increased by 1 each time a category is added; add each point in R to P, and then delete all visited points in Q; If there are visited points in K, add all visited points to set B, which only contains the visited points queried in the current sphere, and find the earliest added category information C from B min , that is, the subscript min is the smallest of all category information, and the category information C of each point is queried in P p , if C p The category information C of any point in B b The same, that is, the subscript p = b, then C p Set to C min , add all unvisited points in R to C min , then add to P, and finally delete all visited points in Q; S35: Repeat S32 to S34 until Q is empty; S36: For P, points with the same category information are clustered into a single point cloud set, and finally n single point cloud sets are obtained.

5. The method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S4 are as follows: S41: input a single point cloud set; S42, use the PointNet network to identify the monomer point cloud set: use the multi-layer perceptron MLP to extract features from each point in the monomer point cloud set, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128), Conv1d(128,512) and Conv1d(512,024) to extract features from each point in turn, expand its feature representation from 3 dimensions to 1024 dimensions, and then use the maximum pooling layer to obtain the global 1024-dimensional features of the monomer point cloud set. The global 1024-dimensional features of the monomer point cloud set are used as data input into several fully connected layers FC, and the FC converts the global features into an e-dimensional feature through linear transformation and nonlinear activation function. Finally, the e-dimensional features are normalized using the Softmax activation function to obtain the probability distribution of each category and determine what type of device it is; S43: According to the identified equipment type, search for the corresponding type of substation equipment in the substation equipment model library, and finally obtain the point cloud model of the corresponding equipment.

6. The method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S5 are as follows: S51: The point cloud model of the corresponding equipment in the substation equipment model library is used as the source point cloud data S, and the monomer point cloud set is used as the sample point cloud data T, which are input into the deep learning registration network SERN; S52: SERN extracts, connects and aligns S and T to obtain a seven-tuple<x,y,z,q0,q1,q2,q3> ; In the seven-tuple, the first element x represents the distance moved along the x-axis; the second element y represents the distance moved along the y-axis; the third element z represents the distance moved along the z-axis; the last four elements are q0, q1, q2, and q3, respectively, where q0 is the cosine component of the rotation angle θ, cos(θ / 2), i.e., q0 = cos(θ / 2), θ refers to the angle of rotation of a point around the rotation axis according to the right-hand rule. The rotation axis is a direction in three-dimensional space and is represented by a unit vector Q1 is The component u in the x-axis direction x The product of and sin(θ / 2); q2 is The component u in the y-axis direction y The product of and sin(θ / 2); q3 is The component u in the z-axis direction z The product of q1 and sin(θ / 2), that is, q1 = u x sin(θ / 2),q2=u y sin(θ / 2),q3=u z sin(θ / 2); Direction by u x 、u y 、u z Jointly determined, the calculation formula is: That is, the rotation angle is obtained from q0, the direction of the rotation axis is obtained from q1, q2, and q3, and q0 2 +q1 2 +q2 2 +q3 2 =1; S53: Using the seven-tuple, calculate the rotation and translation matrix; through the seven-tuple<x,y,z,q0,q1,q2,q3> The first element x, the second element y, and the third element z in the matrix are used to construct a translation matrix, and then the rotation matrix is ​​constructed according to the fourth to seventh elements q0, q1, q2, and q3. Finally, the two matrices are multiplied to calculate the rotation and translation matrix. S531: Use x, y, and z to construct a translation matrix, which is: S532: Use q0, q1, q2, and q3 to construct a rotation matrix, which is: S533: Multiply the translation matrix and the rotation matrix to obtain a rotation-translation matrix, whose rotation-translation matrix A j for: S54: Based on A j , get the new source point cloud data S'; for each point in the point cloud data<a,b,c> Perform rotation and translation calculations to obtain the coordinates of the corresponding points<a',b',c'> , and its calculation formula is: After all points in S are rotated and translated, a new source point cloud data S' with all corresponding points is obtained, that is, the point cloud model of the corresponding equipment in the substation equipment model library is rotated and translated; S55: Calculate the average Euclidean distance of the corresponding key points; S551: Use the key point extraction module KPSM to extract the key points of S' and T; input the point cloud data T or S' into the key point extraction module in turn, use MLP to extract features for each point v, that is, use Conv1d(3,64) to extract features for v, and expand the feature representation of v from 3 dimensions to 64 dimensions; then input v into FC, the number of FC layers is (64,32,1), where the activation function of each layer is LeakyRelu, and each layer is normalized, and finally the importance score of v is obtained; the higher the importance score, the stronger the representativeness of the point, and vice versa, the weaker the representativeness, select the h points with higher rankings, sort them, and then find the corresponding h points in turn, and finally get the three-dimensional coordinates of the h points<x,y,z> ; S552: Calculate the average Euclidean distance of the corresponding key points; for T and S', the key points in T are expressed as follows: <x 1i ,y 1i ,z 1i >, the key points of S' are expressed as <x 2i ,y 2i ,z 2i >, i = 1, 2, 3...h; each key point of T has a unique key point of S' corresponding to it, that is, key points with the same i correspond to each other; Calculate the Euclidean distance d of the key points corresponding to S' and T in turn i , the formula is as follows: Calculate the average Euclidean distance d of the corresponding key points, the formula is as follows: S56: Set the threshold σ. If d>σ, T and S' are used as inputs and S52-S55 are repeated. Otherwise, the iteration ends. S57: After the iteration is completed, all the rotation and translation matrices A generated during the iteration are j Multiply them together to get a final rotation and translation matrix A, that is j is the number of iterations, and then the three-dimensional model in the substation equipment model library is rotated and translated according to A, and finally the three-dimensional model of the target equipment is obtained. The target equipment is a substation equipment with an absolute position in a unified coordinate system after a specific type of equipment in the substation equipment model library is rotated and translated.

7. A method for constructing a three-dimensional substation model based on deep learning according to claim 6, characterized in that: The operation process of the deep learning registration network SERN is as follows: Process 1: Use MLP to extract features from each point in S and T, that is, use Conv1d(3,64), Conv1d(64,64), Conv1d(64,128) and Conv1d(128,512) to extract features from each point in turn, and expand its feature representation from 3 dimensions to 512 dimensions; Process 2: Aggregate the features into a 512-dimensional global feature through the hybrid pooling layer; Process 3: Use the Concatenate layer to aggregate the 512-dimensional global features of T and S into a 1024-dimensional feature, and convert the 1024-dimensional feature into a 7-dimensional feature through FC. The specific hierarchical structure is: FC(1024,512), FC(512,512), FC(512,256), FC(256,7); Process 4: Convert 7-dimensional features into 7-tuples<x,y,z,q0,q1,q2,q3> .

8. The method for constructing a three-dimensional substation model based on deep learning according to claim 1, characterized in that: The specific steps of S6 are as follows: placing the three-dimensional models of all target devices to obtain a three-dimensional model of the substation; in a unified coordinate system, placing the three-dimensional models of all target devices in sequence according to the absolute position information of the three-dimensional models of the target devices, and finally obtaining the three-dimensional model of the substation, that is, in a professional modeling software, setting a three-dimensional space and determining the origin position, placing all the three-dimensional models of the target devices according to the absolute coordinates of the three-dimensional models of the target devices, and finally obtaining the three-dimensional model of the substation.

Images