A 3D model layout method for substation electrical equipment

Abstract

The present invention discloses a three-dimensional model layout method for substation electrical equipment. This method obtains panoramic point clouds through a three-dimensional lidar, and performs clustering segmentation of the panoramic point clouds related to the point clouds of individual equipment according to the elevation information of the electrical equipment structure. Then, the matching relationships among the point clouds of individual equipment, three-dimensional models, and model point clouds are obtained to perform the attitude position layout of the three-dimensional models in the panoramic point clouds. The method of the present invention can achieve fast and accurate layout of three-dimensional models of substation electrical equipment.

Description

Technical field:

[0001] The invention relates to the field of digital auxiliary design of substations, and in particular to a three-dimensional model layout method for substation equipment. Background technology:

[0002] The design of substation engineering in power grid is undergoing a transformation from two-dimensional to three-dimensional, and three-dimensional digital design technology has been widely used in substation engineering.

[0003] Traditional substation modeling technology mainly uses professional software and manual marking to create the surface shape of objects. The massive amount of data leads to low efficiency and high labor intensity.

[0004] Some studies have directly generated twin space models through electrical topology diagrams, but this method is difficult to solve the problems of insufficient precision and spatial stacking of equipment, and its accuracy and precision are difficult to meet the requirements of analytical calculations.

[0005] Some studies have reconstructed substation scenes based on measured point cloud data. This method introduces a device model library with prior shape information, and reconstructs the power equipment monomer by registering the actual device point cloud with the model in the device model library, which can achieve rapid replication of the equipment layout in the actual scene. Two key steps are equipment shape retrieval and model-point cloud registration. Due to the large number of substation equipment models and complex shapes, and the defects of the original point cloud data (such as noise, occlusion, and uneven density), traditional 3D shape recognition algorithms based on shape descriptors such as SHOT and ESF are difficult to accurately match the device point cloud with the device model library, resulting in model retrieval errors. Due to the detailed shape differences between the device model and the scanned point cloud and the uneven distribution of surface points between the discretized model and the scanned point cloud, the corresponding points are prone to mismatching. Traditional point cloud registration algorithms based on corresponding points such as ICP and Super4PCS are prone to registration errors, resulting in problems such as inaccurate model posture and model position errors when using this method to reconstruct the substation scene. Therefore, the existing methods have not yet reached the quality level required for fully automated model creation. After using the existing methods to reconstruct the substation scene, a lot of manual verification and repair work is still required. Summary of the invention:

[0006] In order to solve the problems existing in the prior art, the present invention provides a three-dimensional model layout method for substation equipment.

[0007] The present invention adopts the following technical solution:

[0008] A three-dimensional model layout method for substation equipment, comprising:

[0009] Obtaining a panoramic point cloud of the substation, and preprocessing the panoramic point cloud;

[0010] Based on the octree connected grid labeling method, perform clustering segmentation on the preprocessed full-view point cloud with respect to the point cloud of a single device to obtain several point clouds of single devices in the preprocessed full-view point cloud;

[0011] Obtain the 3D models of substation equipment that match each single-device point cloud in the preset device model library respectively;

[0012] Obtain the model point cloud corresponding to the 3D model that matches each single-device point cloud, and complete the corresponding matching of the single-device point cloud, 3D model, and model point cloud;

[0013] In the preprocessed full-view point cloud, align the poses of each single-device point cloud and the model point cloud corresponding to it, and then replace the model point cloud with the 3D model, thereby completing the 3D model layout of the substation equipment.

[0014] Furthermore, the preprocessing includes removing the ground point cloud and power line point cloud in the full-view point cloud.

[0015] Furthermore, the specific steps of performing clustering segmentation on the preprocessed full-view point cloud with respect to the single-device point cloud of substation equipment based on the octree connected grid labeling method to obtain several single-device point clouds in the preprocessed full-view point cloud are as follows:

[0016] First, traverse the preprocessed full-view point cloud, respectively obtain the maximum and minimum values in the x, y, and z-axis directions, use the leaf node grid where the minimum value is located as the origin of the grid coordinates, obtain the coordinates of each leaf node grid, and thus calculate the coordinates of the root node grid and each leaf node grid;

[0017] Second, perform Morton encoding on the coordinates of each leaf node grid;

[0018] Then, calculate the grid coordinates and Morton codes of each point cloud data point in the preprocessed full-view point cloud, thereby constructing a linear octree corresponding to the preprocessed full-view point cloud;

[0019] Finally, using the constructed linear octree, set the octree resolution n according to the interval distance between single substation equipment, and complete the clustering segmentation of the single-device point cloud in the preprocessed full-view point cloud according to the principle that the point cloud in the leaf node grid and the adjacent grid in the octree is marked as the same cluster.

[0020] Furthermore, the specific steps of obtaining the 3D models of substation equipment that match each single-device point cloud in the preset device model library respectively are as follows:

[0021] Obtain the local invariant descriptors of each single-device point cloud,

[0022] For each single-device point cloud, compare the local invariant descriptor thereof with the local invariant descriptors corresponding to each 3D model in the preset device model library, and use the 3D model with the highest similarity as the 3D model matching the single-device point cloud.

[0023] Further, the local invariant descriptor is obtained by extraction using a deep neural network, and the network structure of the deep neural network includes

[0024] a point cloud transformation layer, a grouped convolutional layer, two average pooling layers, and two fully connected layers.

[0025] Further, the specific steps for comparing the local invariant descriptor of each single-device point cloud with the local invariant descriptors corresponding to each 3D model in the preset device model library are as follows:

[0026] Perform discretization processing on each 3D model in the preset device model library respectively to obtain the model point cloud corresponding to each 3D model, denoted as D j j = 1,..., M, where M is the number of models; and denote the single-device point cloud as P i ; i = 1,..., N, where N is the number of single-device point clouds segmented from the panoramic point cloud;

[0027] Input each single-device point cloud and each model point cloud into the deep neural network respectively to obtain the feature vector f(P i , θ) extracted from each single-device point cloud by the deep neural network, and the feature vector f(D j , θ) extracted from each model point cloud by the deep neural network;

[0028] Calculate the similarity distance Dist between the feature vector f(P i , θ) extracted from each single-device point cloud by the deep neural network and the feature vector f(D j , θ) extracted from each model point cloud by the deep neural network;

[0029] Dist = ||f(D j , θ) - f(D i , θ)||

[0030] In the formula, ||*|| is the Euclidean distance operator.

[0031] Further, the specific steps for obtaining the model point cloud corresponding to the 3D model are as follows:

[0032] Perform discretization processing on the 3D model using a two-stage grid sampling method based on Poisson-Disk distribution to obtain the model point cloud corresponding to the 3D model.

[0033] Further, the specific steps for aligning the poses of each individual device point cloud and the corresponding and matched model point cloud are as follows:

[0034] Perform model registration on the model point cloud and the individual device point cloud using the Teaser++ algorithm to obtain the pose transformation matrix between the model point cloud and the individual device point cloud;

[0035] Use the pose transformation matrix to align the pose of the model point cloud to the coordinate system where the individual device point cloud is located.

[0036] Further, the individual device facilities include main transformers, lightning arresters, instrument transformers, capacitors, disconnectors, and gantries.

[0037] Further, the panoramic point cloud of the substation is obtained by a 3D lidar.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention provides a method for three-dimensional model layout of substation electrical equipment. The panoramic point cloud is obtained by a 3D lidar, and the panoramic point cloud is subjected to clustering segmentation related to the individual device point cloud according to the elevation information of the electrical equipment structure. Then, the matching relationships among the individual device point cloud, the three-dimensional model, and the model point cloud are obtained, and the pose position layout of the three-dimensional model in the panoramic point cloud is carried out. The method of the present invention can achieve fast and accurate layout of the three-dimensional model of substation electrical equipment.

[0040] The three-dimensional model layout method of the present invention is based on the octree connected grid labeling method to perform clustering segmentation of the panoramic point cloud related to the individual device point cloud of the electrical equipment. This design can retain the panorama of the substation while ignoring the noisy point cloud data in the substation scene, making the layout of the three-dimensional model of the present invention have high accuracy.

[0041] The three-dimensional model layout method of the present invention makes the original "messy" substation scene "neat" by removing the ground point cloud and power line point cloud in the panoramic point cloud, making the layout of the three-dimensional model of the present invention have high accuracy.

[0042] The three-dimensional model layout method of the present invention uses the similarity of local invariant descriptors between point clouds to obtain the three-dimensional model corresponding to the individual device point cloud in the scanned panoramic point cloud. This design can weaken the sensitivity to the uneven density distribution and local shape differences of the point cloud. Compared with the matching of traditional models and scanned point clouds, the present invention can better adapt to cross-source point cloud shape retrieval.

[0043] The Teaser++ registration algorithm adopted by the 3D model layout method of the present invention can adapt to the wrong matching point pairs caused by the local shape differences between the model and the point cloud, and improve the accuracy of the attitude alignment between the device model and the device point cloud. Description of the Drawings:

[0044] Figure 1 It is a flowchart of the 3D model layout method of substation electrical equipment in Embodiment 1;

[0045] Figure 2 It is a schematic diagram of the power line removal result in the application embodiment;

[0046] Figure 3 It is a schematic diagram of the clustering segmentation result of the single-device point cloud in the application embodiment;

[0047] Figure 4(a) shows the gantry identified in the panoramic point cloud of the substation in the application embodiment;

[0048] Figure 4(b) shows the gantry model after automatic layout in the application embodiment. Detailed Implementation Modes:

[0049] Embodiment 1:

[0050] A 3D model layout method of substation electrical equipment of the present invention includes:

[0051] s1. Obtain the panoramic point cloud of the substation through a 3D lidar, and preprocess the panoramic point cloud;

[0052] s2. Based on the octree connected grid labeling method, perform clustering segmentation on the preprocessed panoramic point cloud for the single-device point cloud related to the electrical equipment to obtain several single-device point clouds in the preprocessed panoramic point cloud; The single devices include typical electrical equipment and facilities such as main transformers, lightning arresters, instrument transformers, capacitors, disconnectors, and gantries.

[0053] s3. Obtain the 3D models of the electrical equipment that match each single-device point cloud in a preset device model library respectively;

[0054] s4. Obtain the model point cloud corresponding to the 3D model that matches each single-device point cloud, and complete the corresponding matching of the single-device point cloud, the 3D model, and the model point cloud;

[0055] s5. In the preprocessed panoramic point cloud, align the attitudes of each single-device point cloud and the model point cloud corresponding to it, and then replace the model point cloud with the 3D model, thereby completing the 3D model layout of the electrical equipment in the substation.

[0056] Embodiment 2:

[0057] On the basis of the first embodiment, the further design of this embodiment lies in that the preprocessing in this example includes removing the ground point cloud and power line point cloud from the panoramic point cloud. A large amount of ground point cloud and power line point cloud are contained in the panoramic point cloud obtained by 3D lidar scanning, which are interferences for subsequent point cloud clustering segmentation and model matching. Therefore, before that, it is necessary to remove the ground point cloud and power line point cloud from the panoramic point cloud.

[0058] Among them, the ground point cloud is removed by the following method:

[0059] Considering that the substation ground is generally flat, and the existence of noise points will affect the accuracy of the least squares method for fitting the plane. Therefore, the random sample consensus (RANSAC) algorithm is used to extract the ground, and the interference of noise is eliminated by iterative fitting, which greatly improves the fitting accuracy. The specific steps are as follows

[0060] Find the ground points of the panoramic point cloud through the RANSAC algorithm, fit the found ground points to form a plane, calculate the normal vector of the plane, and make the normal vector parallel to the z-axis of the panoramic point cloud through rotation transformation. At this time, the plane formed by fitting is flush with the plane corresponding to the xy-axis. The point cloud data points in the plane where the lowest point cloud data point in the panoramic point cloud is located and the space above this plane by h1 are used as the ground point cloud segmentation. h1 is generally taken as 0.2m.

[0061] Power lines are important components of power stations. Each power equipment is connected together through power lines, which are usually erected above the insulators of power equipment and on the sides of poles. The connection between each equipment makes the clustering segmentation of the equipment very difficult. Therefore, the power lines can be extracted first before segmenting the power equipment. Power lines usually have the following characteristics: (1) The projection of the power line in the XOY plane is linearly distributed. (2) The shape of the power line is a linear structure and is thinner than other equipment. The power line point cloud is removed by the following method:

[0062] Use the method of line fitting to obtain the candidate points of the panoramic point cloud, and use the principal component analysis (PCA) method to screen out the power line point cloud data points in the candidate points as the power line point cloud and remove them. Among them, the method for obtaining candidate points is as follows:

[0063] The panoramic point cloud is equally divided into several horizontal slices at the same vertical height interval h2 to obtain several layers with a thickness of h2; h2 is generally taken as 0.3 meters.

[0064] Project the point cloud data points in the layers with a height greater than h3 onto a two-dimensional horizontal plane, and extract the line segments therein through the Hough transform respectively; h3 can be selected according to the height range of the overhead power lines in the substation;

[0065] In the horizontal projection plane of this layer, calculate the straight-line equation for the straight lines with a calculated length greater than the threshold h4, and extend 0.5 m on each side of the straight-line equation, so as to determine the spatial region where the power lines exist in this layer. The point cloud data points within this spatial region are candidate points; generally, h4 is taken as 1 m.

[0066] 1.2.2) The screening method for power line point cloud data points is as follows:

[0067] Use the PCA principal component analysis method to obtain the PCA eigenvalues and corresponding eigenvectors of each candidate point. Since the power lines are thin and linear, there is only one main direction within the neighborhood of the candidate points. Also, because the power lines are horizontally arranged, the eigenvector v1 corresponding to λ1 is approximately perpendicular to the Z-axis. That is, if the PCA eigenvalues and corresponding eigenvectors of the candidate point satisfy the following formula, it means that the candidate point is a power line point cloud data point; otherwise, the candidate point is not a power line point cloud data point.

[0068]

[0069] In the formula, λ1, λ2, and λ3 are the first, second, and third PCA eigenvalues of the candidate point respectively; v1 is the eigenvector corresponding to the first PCA eigenvalue λ1; ||*|| is the modulus operator; ζ is the power line main direction threshold; θ is the power line orientation threshold. As a preferred solution, ζ = 1e3 and θ = 0.8, and better results can be obtained for power line extraction.

[0070] Example 2:

[0071] The further design of this example is based on Example 1. In this example, the octree connectivity grid marking method is used, that is, the point clouds in the octree grid and adjacent grids are marked as the same cluster. Thus, the clustering segmentation of the single-device point clouds related to substation equipment is performed on the preprocessed panoramic point cloud. The specific steps to obtain several single-device point clouds in the preprocessed panoramic point cloud are as follows:

[0072] First, traverse the preprocessed panoramic point cloud, and respectively obtain the maximum and minimum values in the x, y, and z-axis directions. Use the leaf node grid where the minimum value is located as the origin of the grid coordinates to obtain the coordinates of each leaf node grid, so as to calculate the coordinates of the root node grid and each leaf node grid.

[0073] Second, perform Morton encoding on the coordinates of each leaf node grid.

[0074] Then, calculate the grid coordinates and Morton codes of each point cloud data point in the preprocessed panoramic point cloud, so as to construct a linear octree corresponding to the preprocessed panoramic point cloud.

[0075] Finally, using the constructed linear octree, set the octree resolution n according to the interval distance between individual substation equipment. According to the principle that the point cloud in the leaf node grid and the adjacent grid in the octree are marked as the same cluster, complete the clustering segmentation of the point cloud of individual equipment in the preprocessed panoramic point cloud.

[0076] The above algorithm is an extension of the image connected region labeling algorithm in three-dimensional space. Compared with point cloud clustering algorithms based on point cloud distribution such as kmeans and DBSCAN, the point cloud clustering based on octree connected grid labeling rasterizes the three-dimensional space using the octree, avoiding the calculation for each point, having a significant advantage in the speed of point cloud clustering, and being able to ensure the correctness of clustering.

[0077] Example Three:

[0078] The further design of this example based on Example One is that the specific steps to obtain the three-dimensional model of the substation equipment matching each individual equipment point cloud in the preset equipment model library are as follows:

[0079] Obtain the local invariant descriptors of each individual equipment point cloud. For each individual equipment point cloud, compare the similarity of its local invariant descriptors with the local invariant descriptors corresponding to each three-dimensional model in the preset equipment model library, and take the three-dimensional model with the highest similarity as the three-dimensional model matching the individual equipment point cloud.

[0080] Example Four:

[0081] The further design of this example based on Example Three is that in this example, the local invariant descriptors are obtained by extracting with a deep neural network. The network structure of this deep neural network includes a point cloud transformation layer, a grouped convolutional layer, two average pooling layers, and two fully connected layers.

[0082] Example Five:

[0083] The further design of this example based on Example Three is that in this example, the local invariant descriptor (LocalInvariant Feature, LIF) is obtained by extracting with a deep neural network. In the problem of cross-source point cloud shape retrieval, the density distributions of the model point cloud obtained by discrete sampling of the model and the real lidar scan point cloud are uneven, and there are certain differences between the local shapes of the actual equipment and the model. The local invariant descriptor can weaken the sensitivity to the uneven density distribution of the point cloud and the local deformation difference, so it can adapt to cross-source point cloud shape retrieval. The network structure of this deep neural network includes four modules, specifically:

[0084] The first module is the point cloud transformation layer, which transforms the point cloud into a 3D tensor that can be processed by a convolutional kernel. The second module contains a group convolution layer, a BN+ReLU layer, and an AvgPooling layer. The third module contains a group deconvolution layer, a BN+ReLU layer, and an AvgPooling layer. Then there is a feature extraction layer, which utilizes the feature extraction ability of the convolutional layer to map the features contained in the 3D tensor to the hidden layer feature space, and then learns the feature representation of the distribution. Finally, there is a fully convolutional layer, which maps the learned distribution feature representation to the sample label space, acting as a classifier, and finally outputs a 33-dimensional feature vector.

[0085] For each single device point cloud, the specific steps for comparing the similarity between its local invariant descriptor and the local invariant descriptors corresponding to each 3D model in the preset device model library are as follows:

[0086] Discretize each 3D model in the preset device model library respectively to obtain the model point cloud corresponding to each 3D model, denoted as D j j = 1,..., M, where M is the number of models; and denote the single device point cloud as P i ; i = 1,..., N, where N is the number of single device point clouds segmented from the panoramic point cloud;

[0087] Input each single device point cloud and each model point cloud into the deep neural network respectively to obtain the feature vector f(P i , θ) extracted from each single device point cloud by the deep neural network, and the feature vector f(D j , θ) extracted from each model point cloud by the deep neural network;

[0088] Calculate the similarity distance Dist between the feature vector f(P i , θ) extracted from each single device point cloud by the deep neural network and the feature vector f(D j , θ) extracted from each model point cloud by the deep neural network respectively;

[0089] Dist = ||f(D j , θ) - f(D i , θ)||

[0090] In the formula, ||*|| is the Euclidean distance operator.

[0091] Example 6:

[0092] The further design of this example is based on Example 1, Example 2, Example 3, Example 4 or Example 5. The specific steps for obtaining the model point cloud corresponding to the 3D model in this example are as follows:

[0093] The 3D model is discretized using a two-stage grid sampling method based on Poisson-Disk distribution to obtain the model point cloud corresponding to the 3D model.

[0094] Example Seven:

[0095] This example is further designed based on Example One. The specific steps for aligning the poses of each individual device point cloud and the corresponding and matching model point cloud are as follows:

[0096] Perform model registration on the model point cloud and the individual device point cloud using the Teaser++ algorithm to obtain the pose transformation matrix between the model point cloud and the individual device point cloud;

[0097] Use the pose transformation matrix to align the pose of the model point cloud to the coordinate system where the individual device point cloud is located.

[0098] Application Example:

[0099] This example applies the 3D model layout method of the present invention to a certain substation. The substation is equipped with devices and facilities such as main transformers, lightning arresters, instrument transformers, capacitors, disconnecting switches, and gantries. The specific steps include:

[0100] Obtain the panoramic point cloud of the substation and perform preprocessing on the panoramic point cloud to remove the ground point cloud and power line point cloud; The power line removal result is as Figure 2 shown.

[0101] Based on the octree connected grid labeling method, perform clustering segmentation on the preprocessed panoramic point cloud related to the individual device point cloud to obtain several individual device point clouds in the preprocessed panoramic point cloud; The segmentation and clustering results are as Figure 3 shown, Figure 3 The two parts framed in the figure are two types of individual device point clouds respectively. The lower type of individual device point cloud is the combined switch point cloud, and the point cloud with lightning arresters and shrubs in the upper window. Among them, the lightning arrester will be recognized in the next step, while the shrubs cannot be matched with any devices in the model library and are removed.

[0102] Obtain the 3D models of substation equipment that match each individual device point cloud in the preset device model library;

[0103] Obtain the model point cloud corresponding to the 3D model that matches each individual device point cloud to complete the corresponding matching of the individual device point cloud, 3D model, and model point cloud;

[0104] In the preprocessed panoramic point cloud, the point cloud with the recognition result of the gantry is shown in Figure 4(a). Align the pose of each gantry point cloud with the corresponding model point cloud to obtain the pose transformation matrix. Then use this matrix to replace the point cloud with a 3D model, thus completing the 3D model layout of the gantry in the substation, and the layout result is shown in Figure 4(b).

Claims

1. A three-dimensional model layout method for substation electrical equipment, characterized in that: Including: Obtain the panoramic point cloud of the substation and preprocess the panoramic point cloud; Based on the octree connected grid labeling method, perform clustering segmentation of the preprocessed panoramic point cloud related to the point cloud of individual equipment to obtain several point clouds of individual equipment in the preprocessed panoramic point cloud; The specific steps of performing clustering segmentation of the preprocessed panoramic point cloud related to the point cloud of substation equipment to obtain several point clouds of individual equipment in the preprocessed panoramic point cloud based on the octree connected grid labeling method are as follows: First, traverse the preprocessed panoramic point cloud, respectively obtain the maximum and minimum values in the x, y, and z axis directions, use the leaf node grid where the minimum value is located as the origin of the grid coordinates to obtain the coordinates of each leaf node grid, and thus calculate the coordinates of the root node grid and each leaf node grid; Secondly, perform Morton encoding on the coordinates of each leaf node grid; Then, calculate the grid coordinates and Morton codes where each point cloud data point in the preprocessed panoramic point cloud is located, thereby constructing a linear octree corresponding to the preprocessed panoramic point cloud; Finally, using the constructed linear octree, set the octree resolution n according to the interval distance between individual substation equipment, and complete the clustering segmentation of the point clouds of individual equipment in the preprocessed panoramic point cloud according to the principle that the point clouds in the leaf node grid and the adjacent grid in the octree are marked as the same cluster; Respectively obtain the three-dimensional models of substation equipment that match each point cloud of individual equipment in the preset equipment model library; Obtain the model point cloud corresponding to the three-dimensional model that matches each point cloud of individual equipment, and complete the corresponding matching of the point cloud of individual equipment, the three-dimensional model, and the model point cloud; In the preprocessed panoramic point cloud, align the postures of each point cloud of individual equipment and the corresponding model point cloud, and then replace the model point cloud with the three-dimensional model, thereby completing the three-dimensional model layout of the substation equipment.

2. The three-dimensional model layout method of the power conversion equipment in the substation according to claim 1, wherein: The preprocessing includes removing the ground point cloud and the power line point cloud in the panoramic point cloud.

3. The three-dimensional model layout method of the power conversion equipment in the substation according to claim 1, wherein: The specific steps of respectively obtaining the three-dimensional models of substation equipment that match each point cloud of individual equipment in the preset equipment model library are as follows: Obtain the local invariant descriptors of each point cloud of individual equipment, For each point cloud of individual equipment, compare the similarity of its local invariant descriptor with the local invariant descriptors corresponding to each three-dimensional model in the preset equipment model library, and use the three-dimensional model with the highest similarity as the three-dimensional model that matches the point cloud of individual equipment.

4. The three-dimensional model layout method of substation electrical equipment according to claim 3, characterized in that: The local invariant descriptor is obtained by extracting with a deep neural network, and the network structure of the deep neural network includes a point cloud transformation layer, a grouped convolutional layer, two average pooling layers, and two fully connected layers.

5. The three-dimensional model layout method of the power conversion equipment in the substation according to claim 4, wherein: The specific steps of comparing the similarity of the local invariant descriptor of each point cloud of individual equipment with the local invariant descriptors corresponding to each three-dimensional model in the preset equipment model library are as follows: Discretize each 3D model in the preset device model library respectively to obtain the model point cloud corresponding to each 3D model, denoted as D j j = 1,..., M, where M is the number of models; and denote the single-device point cloud as P i ; i = 1,..., N, where N is the number of single-device point clouds segmented from the panoramic point cloud Respectively input each individual device point cloud and each model point cloud into the deep neural network to obtain the feature vector f(P i ,θ) extracted from each individual device point cloud by the deep neural network, and the feature vector f(D j ,θ) extracted from each model point cloud by the deep neural network; Calculate the similarity distance Dist between the feature vectors f(P i , θ) extracted from the point clouds of each individual device by the deep neural network and the feature vectors f(D j , θ) extracted from the point clouds of each model by the deep neural network; Dist=||f(D j ,θ)-f(D i ,θ)|| In the formula, ||*|| is the Euclidean distance operator.

6. The three-dimensional model layout method for substation electrical equipment according to claim 1, characterized in that: The specific steps of obtaining the model point cloud corresponding to the three-dimensional model are as follows: Discretize the three-dimensional model by using a two-stage grid sampling method based on the Poisson-Disk distribution to obtain the model point cloud corresponding to the three-dimensional model.

7. The three-dimensional model layout method of substation electrical equipment according to claim 1, characterized in that: The specific steps for aligning the poses of each single-device point cloud and the corresponding and matched model point cloud are as follows: Perform model registration on the model point cloud and the single-device point cloud using the Teaser++ algorithm to obtain the pose transformation matrix between the model point cloud and the single-device point cloud; Use the pose transformation matrix to align the pose of the model point cloud to the coordinate system where the single-device point cloud is located.

8. The three-dimensional model layout method of the power conversion equipment in the substation according to claim 1, wherein: The single-device facilities include main transformers, lightning arresters, instrument transformers, capacitors, disconnectors, and gantries.

9. The three-dimensional model layout method of the power conversion equipment in the substation according to claim 1, characterized in that: The panoramic point cloud of the substation is obtained by a 3D lidar.

Images