Distribution network point cloud segmentation method based on environment fusion

Abstract

The invention discloses a power distribution network point cloud segmentation method based on environment fusion, and relates to the technical field of target segmentation, and the method comprises the steps: collecting point cloud data, including fused environment point cloud data; the hierarchical edge sensing network comprises a local embedding module, a global embedding module, an attention mechanism module and a multi-level down-sampling and up-sampling module. And the attention mechanism module and the jump connection between the multi-level down-sampling form a multi-level network. Point cloud data are input into a hierarchical edge sensing network, a local embedding module captures local features and neighborhood features of each point, a global embedding module identifies global features, an attention mechanism module is used for fusing the local features and the global features, and the features of a segmentation task are focused. And through multiple times of down-sampling and up-sampling of the multi-level network, segmenting the environment characteristics and the distribution network characteristics, and outputting a segmentation result of the distribution network. Therefore, the technical effect of improving the distribution network point cloud segmentation efficiency and accuracy is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of target segmentation, and in particular to a distribution network point cloud segmentation method based on environment fusion. Background Art

[0002] Point cloud data has become an important tool in the maintenance and management of smart grids and distribution networks. Point cloud data can accurately describe the three-dimensional structure of the distribution network and provide rich information for the design, monitoring and maintenance of the power grid. However, with the increasing complexity of the distribution network, especially the structure of the distribution network is more complex, and the types of distribution networks are different, and the environment in which different types of distribution networks are located is also complex and changeable. The existing automatic segmentation method of power point cloud is mainly suitable for high-voltage distribution networks with obvious features and small interference. The classification effect is poor for relatively low and complex distribution lines or some low-voltage transmission lines, and there are technical problems such as low segmentation efficiency and high segmentation cost. Summary of the invention

[0003] The present invention provides a distribution network point cloud segmentation method based on environment fusion to solve the technical problems of low segmentation efficiency and high segmentation cost in the prior art, and achieves the technical effect of improving the efficiency and accuracy of distribution network point cloud segmentation.

[0004] The distribution network point cloud segmentation method based on environment fusion provided by the present invention includes:

[0005] The point cloud data of the power distribution network is collected, wherein the point cloud data includes environmental point cloud data, and the environmental point cloud data is integrated with the point cloud data of the power distribution network.

[0006] A pre-built hierarchical edge perception network is obtained, wherein the hierarchical edge perception network includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling, and upsampling, wherein the attention mechanism module, global to local downsampling, and upsampling construct a multi-level network, and the multi-level network contains multiple jump connections, and the jump connection is a jump between the attention mechanism module and the global to local downsampling.

[0007] The distribution network point cloud data is input into the hierarchical edge perception network, the local features and local neighborhood features of each point in the distribution network point cloud data are captured by the local embedding module, and the global features are recognized by the global embedding module. The local features, local neighborhood features and global features are fused with the attention mechanism module to focus on the segmentation task features, and multiple downsampling and upsampling are performed through multiple jump connections contained in the multi-level network to segment the environmental features and the distribution network features, and the segmentation result of the distribution network is output.

[0008] In a feasible implementation, the local embedding module captures the local features and local neighborhood features of each point in the distribution network point cloud data, including:

[0009] Through distance measurement, the K neighbor points of each point in the point cloud data are determined.

[0010] The feature vectors of the K neighbor points are replicated K times to construct a feature matrix.

[0011] The K copied feature matrices are concatenated according to the channel dimension to obtain a stacked feature matrix, which includes the local features of the original point and the neighborhood.

[0012] The stacked feature matrix is ​​used as output to obtain the local features and local neighborhood features.

[0013] In a feasible implementation, before the K feature matrices obtained by copying are concatenated according to the channel dimension, the method further includes:

[0014] Configure the multi-scale K value.

[0015] Based on the multi-scale K value, multi-scale K neighbor points are determined.

[0016] Feature insertion is performed according to the feature vectors of the multi-scale K neighboring points to form multiple new feature dimensions.

[0017] Visually evaluate the multiple new feature dimensions, determine the relationship between the feature distribution and the target feature variable, and screen the new feature dimensions.

[0018] The new feature dimension is added to the concatenated feature.

[0019] In a feasible implementation, the fusing of local features, local neighborhood features and global features through an attention mechanism module includes:

[0020] According to the local features and global features, they are mapped into a query vector, a key vector, and a value vector through a weight matrix.

[0021] The similarity scores between the query vector and all key vectors are calculated to obtain an attention score.

[0022] The similarity scores between each query vector and all key vectors are normalized by a normalization function, and a normalized weight distribution relationship between each query vector and key vector is established.

[0023] The normalized weight is multiplied by the corresponding value vector to obtain the attention mechanism output.

[0024] In a feasible implementation, the normalized weight is multiplied by the corresponding value vector to obtain the attention mechanism output, including:

[0025] By formula: Calculate to obtain the output of the attention mechanism, where Q is the query vector, K is the key vector, V is the value vector, is the scaling factor.

[0026] In a feasible implementation manner, after outputting the segmentation result of the power distribution network, the method further includes:

[0027] The distribution line is identified based on the segmentation result of the distribution network, and the distribution location of the distribution line is extracted.

[0028] A line smoothness test is performed according to the distribution line positioning to obtain a smoothness test result.

[0029] According to the smoothness detection result, abnormal nodes are identified.

[0030] When the abnormal node does not exist, a segmentation result of the power distribution network is determined.

[0031] In a feasible implementation, after identifying the abnormal node, the method further includes:

[0032] When the abnormal node exists, the abnormal power distribution line and adjacent power distribution lines are obtained.

[0033] Abnormal node alignment compensation is performed on the abnormal distribution line to determine compensation characteristics.

[0034] Based on the compensation feature, the abnormal environment feature is deducted.

[0035] Environmental features are extracted based on the adjacent power distribution lines, and correlation analysis is performed using the environmental features and abnormal environmental features to determine correlation identification features.

[0036] The abnormal environment feature is re-segmented using the associated identification feature, and the distribution line feature is restored and re-segmented based on the compensation feature.

[0037] In a feasible implementation, extracting environmental features from the adjacent power distribution lines, performing correlation analysis using the environmental features and abnormal environmental features, and determining correlation identification features include:

[0038] According to the environmental features extracted from the adjacent power distribution lines, a feature initial point is set, and the feature initial point is a cluster of pixel points whose distance from the abnormal node meets the distance requirement.

[0039] Determine the neighborhood pixels of the feature initial point.

[0040] Pixel growth information is fitted based on the feature initial point, where the pixel growth information includes grayscale value, color, texture, and gradient.

[0041] The feature initial point is taken as a starting point, and growth coverage judgment is performed on neighboring pixels based on the pixel growth information, and a pixel area that meets the pixel growth information is taken as a new feature initial point.

[0042] New pixel growth information is fitted based on the new feature initial point, and the pixel growth information is iteratively used to determine the growth coverage of neighboring pixels until the abnormal environment feature is reached, and the associated identification feature is determined based on the current pixel growth information.

[0043] The present invention discloses a distribution network point cloud segmentation method based on environmental fusion, including: collecting distribution network point cloud data and environmental point cloud data, and fusing the two. Obtain a pre-constructed hierarchical edge perception network, which includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling and upsampling in a multi-level network, and contains multiple jump connections. Input the distribution network point cloud data into the hierarchical edge perception network, capture the local features and local neighborhood features of each point through the local embedding module, perform global feature recognition through the global embedding module, fuse local and global features through the attention mechanism module, and focus on the segmentation task features. Through multiple downsampling and upsampling in the multi-level network, segment the environmental features and the distribution network features, and output the segmentation results of the distribution network. The distribution network point cloud segmentation method based on environmental fusion disclosed by the present invention solves the technical problems of low segmentation efficiency and high segmentation cost, and achieves the technical effect of improving the efficiency and accuracy of distribution network point cloud segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the distribution network point cloud segmentation method based on environment fusion of the present invention.

[0045] Figure 2 It is a structural schematic diagram of the hierarchical edge perception network in the distribution network point cloud segmentation method based on environment fusion of the present invention.

[0046] Explanation of the accompanying drawings: local embedding module LE, global embedding module GE, attention mechanism module TA, global to local downsampling GLD, upsampling UPS, hierarchical edge perception network HEA. DETAILED DESCRIPTION

[0047] The technical solution provided in the embodiments of the present invention is to solve the technical problems of high latency, lack of real-time performance, incomplete monitoring, and insufficient predictability in the prior art. The overall idea adopted is as follows:

[0048] First, the point cloud data of the distribution network are collected, which includes the environmental point cloud data, and these environmental point cloud data are fused with the distribution network point cloud data.

[0049] Next, a pre-built hierarchical edge perception network is obtained, which includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling and upsampling. Among them, the attention mechanism module, global to local downsampling and upsampling jointly construct a multi-level network, and there are multiple jump connections in the multi-level network. These jump connections are mainly the connections between the attention mechanism module and the global to local downsampling.

[0050] Then, the point cloud data of the distribution network is input into this hierarchical edge perception network. Through the local embedding module, the network can capture the local features and local neighborhood features of each point; the global embedding module is used to identify global features. The attention mechanism module is responsible for fusing local features, local neighborhood features and global features to focus on the features in the segmentation task.

[0051] Finally, through multiple downsampling and upsampling in the multi-level network, the network can segment the environmental features and distribution network features, and output the segmentation results of the distribution network.

[0052] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention are shown in the drawings, rather than all of them.

[0053] Figure 1 The figure is a flow chart of a method for segmenting a distribution network point cloud based on environment fusion according to the present invention, wherein the method comprises:

[0054] Collecting distribution network point cloud data, wherein the point cloud data includes environmental point cloud data, and the environmental point cloud data is integrated with the distribution network point cloud data;

[0055] Optionally, laser radar, drones or ground mobile scanning systems are used to collect point cloud data of the distribution network. The collected point cloud data includes point cloud data related to the distribution network and environmental point cloud data around the distribution network. Among them, the point cloud data related to the distribution network includes the three-dimensional coordinate information of the poles, wires, transformers and other ancillary equipment of the distribution network; the environmental point cloud data includes the three-dimensional coordinate information of the terrain, buildings, trees and other natural or artificial obstacles.

[0056] Specifically, the environmental point cloud data and the distribution network point cloud data are mutually integrated point cloud data. In other words, the collected distribution network point cloud data includes the point cloud data of the distribution network neighborhood environment objects collected synchronously, and the data points in the collected distribution network point cloud data have not been clearly classified and are in a mixed state.

[0057] Obtain a pre-constructed hierarchical edge perception network, wherein the hierarchical edge perception network includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling, and upsampling, wherein the attention mechanism module, global to local downsampling, and upsampling construct a multi-level network, and the multi-level network contains multiple jump connections, and the jump connection is a jump between the attention mechanism module and the global to local downsampling;

[0058] Specifically, Figure 2 As shown in the figure, the pre-built hierarchical edge perception network HEA consists of multiple modules, including the local embedding module LE, the global embedding module GE, the attention mechanism module TA, the global to local downsampling GLD, and the upsampling USP. Among them, the local embedding module LE is responsible for extracting local features of the input image. Through convolution operations, this module can capture the detailed information of the image, such as edges and textures. The global embedding module GE is responsible for extracting global features of the input image. Unlike local embedding, this part focuses on the overall contextual information of the image to ensure that the model can understand the overall picture of the image. The attention mechanism module TA is used to highlight the features of important areas in the image and ignore unimportant parts. By calculating the attention weights of the input features, the model can focus more on key areas and improve the effectiveness of feature expression. The global to local downsampling GLD converts high-resolution feature maps to low-resolution feature maps by downsampling the features. This helps to reduce the amount of computation while retaining important global and local information. The upsampling USP restores the low-resolution feature maps to high resolution by upsampling the features to restore the point cloud resolution and increase the clarity of the point cloud. The upsampling operation can be combined with the deconvolution operation so that the restored feature map has better detail preservation.

[0059] The hierarchical edge awareness network HEA divides the environmental point cloud features and the distribution network point cloud features by focusing on the edge features of the point cloud, especially combining the global features with the local features in a small range, thereby improving the efficiency and accuracy of the distribution network point cloud segmentation, thereby solving the problem of deep fusion of the distribution network point cloud and the environmental point cloud.

[0060] Optionally, the hierarchical edge perception network HEA adopts a multi-level network structure, including multiple attention mechanism modules TA, global to local downsampling GLD and upsampling USP grouped and connected in a certain order. At the same time, there are multiple skip connections (Skip Connections) in the hierarchical edge perception network HEA. Among them, the skip connection refers to the direct connection between the attention mechanism module TA and the global to local downsampling GLD. Through the skip connection, feature information can be transmitted between different layers of the network, the multi-scale information of the feature map can be retained, and the overall performance of the model can be improved.

[0061] The above method, through the combination of local embedding module and global embedding module, enables the hierarchical edge perception network HEA to capture local details and global context information at the same time, improving the accuracy of feature extraction. The resolution of the feature map is reduced by global to local downsampling GLD, which reduces the amount of calculation and improves the computational efficiency of the model. The feature map is weighted and processed by the attention mechanism module TA to highlight important areas and enhance the effectiveness of feature expression. Configuring jump connections to retain feature information at different levels helps to improve the overall performance of the model.

[0062] The distribution network point cloud data is input into the hierarchical edge perception network, the local features and local neighborhood features of each point in the distribution network point cloud data are captured by the local embedding module, and the global features are recognized by the global embedding module. The local features, local neighborhood features and global features are fused with the attention mechanism module to focus on the segmentation task features, and multiple downsampling and upsampling are performed through multiple jump connections contained in the multi-level network to segment the environmental features and the distribution network features, and the segmentation result of the distribution network is output.

[0063] Specifically, first, the distribution network point cloud data (including environmental point cloud data) is input into the input end (Input) of the hierarchical edge perception network HEA; then, the local embedding module LE and the global embedding module GE respectively obtain the local neighborhood features of multiple points, the local neighborhood features and the global features of the entire point cloud data; then, the local features, local neighborhood features and global features are input into the attention mechanism module TA, focusing on the key features of the segmentation task for feature fusion; then, multiple downsampling operations and upsampling operations are performed through the global to local downsampling GLD and upsampling USP at different levels in the multi-level network, and the features are transferred between different levels based on multiple jump connections to realize the segmentation of environmental features and distribution network features; finally, the segmented distribution network point cloud data is output from the output end (HEA Output) of the hierarchical edge perception network HEA to obtain the segmentation result of the distribution network and environmental features.

[0064] Through the above steps and modules, the hierarchical edge perception network HEA can efficiently and accurately segment the point cloud data of the distribution network and the environment, solve the problem of deep fusion of the distribution network point cloud and the environmental point cloud, and improve the efficiency and accuracy of the distribution network point cloud segmentation.

[0065] In some embodiments, capturing the local features and local neighborhood features of each point in the distribution network point cloud data by the local embedding module includes:

[0066] Through distance measurement, the K neighbor points of each point in the point cloud data are determined; the feature vectors of the K neighbor points are copied K times to construct a feature matrix; the K copied feature matrices are spliced ​​according to the channel dimension to obtain a stacked feature matrix, which includes the local features of the original point and the neighborhood; the stacked feature matrix is ​​used as output to obtain the local features and local neighborhood features.

[0067] Optionally, the local features and local neighborhood features of each point are captured by a local embedding module LE. First, the K nearest neighboring points of each point in the point cloud data are determined by a distance measurement method (such as Euclidean distance) to form the neighborhood of the point.

[0068] Furthermore, the feature vectors of the K nearest neighbor points are copied K times, that is, the feature vectors of each point in the K nearest neighbor points are copied to construct a feature matrix. Exemplarily, the feature matrix is ​​a KxC matrix, where the K rows of the feature matrix correspond to the feature vectors of the K nearest neighbor points, and the C columns correspond to the feature dimension C of the feature vectors of the K nearest neighbor points.

[0069] Optionally, the K copied feature matrices are concatenated according to the channel dimension to obtain a stacked feature matrix. This stacked feature matrix includes the local features of the K original points and the K neighborhoods. Each row or column of the stacked feature matrix contains the features of the original point and its neighborhood points. The K layers of feature matrices in the stacked feature matrix correspond to the K nearest neighbor points, forming a feature representation that integrates local information. The stacked feature matrix is ​​used as the output of the local embedding module LE to obtain the local features and local neighborhood features.

[0070] Through the above steps, the local embedding module LE can effectively capture the local features of each point and its neighborhood features, so that the local information of the point cloud data is fully utilized, the local representation ability of the point cloud data is enhanced, and the basis for subsequent global feature extraction and feature fusion is provided.

[0071] In some implementations, before concatenating the K feature matrices obtained by copying according to the channel dimension, the method further includes:

[0072] Configure a multi-scale K value; determine multi-scale K neighbor points based on the multi-scale K value; perform feature insertion according to the feature vectors of the multi-scale K neighbor points to form multiple new feature dimensions; perform visual evaluation on the multiple new feature dimensions, determine the relationship between feature distribution and target feature variables, and screen new feature dimensions; and add the new feature dimensions to the splicing features.

[0073] Optionally, before splicing the K feature matrices obtained by copying according to the channel dimension, a multi-scale K value is configured, that is, multiple different K values ​​are selected to capture neighborhood features at different scales. Specifically, the larger the K value, the larger the range of captured local features and local neighborhood features, and the larger the corresponding data processing amount.

[0074] Optionally, based on the multi-scale K value, multi-scale K neighbor points are determined. That is, for each point, its neighbor point set under different K values ​​is determined respectively, wherein the feature dimensions of the feature vectors corresponding to the multi-scale K neighbor points are different. Further, feature insertion is performed based on the feature vectors of the multi-scale K neighbor points to form multiple new feature dimensions. These new feature dimensions reflect the differences in neighborhood features at different scales. Exemplarily, a smaller scale may have feature dimensions for identifying subtle textures or color changes of fabrics that are missing on a larger scale.

[0075] Optionally, visually evaluate the feature distribution and its relationship with the target variable through visualization methods such as scatter plots, heat maps, and box plots, and determine the relationship between the feature distribution and the target feature variable based on the visualization results to screen new feature dimensions. Specifically, select those feature dimensions that have a significant correlation with the target variable (such as classification labels or regression values) as new feature dimensions.

[0076] Furthermore, the filtered new feature dimensions are added to the concatenated features to obtain the final stacked feature matrix, ensuring that the stacked feature matrix can represent richer and more diverse local features of the original points and their neighborhoods, making more comprehensive use of the local information of the point cloud data.

[0077] In some embodiments, the fusing of local features, local neighborhood features and global features through an attention mechanism module includes:

[0078] According to the local features and global features, they are mapped into query vectors, health vectors, and value vectors through a weight matrix; the similarity scores between the query vector and all health vectors are calculated to obtain an attention score; the similarity scores between each query vector and all health vectors are normalized through a normalization function to establish a normalized weight distribution relationship between each query vector and the health vector; the normalized weight is multiplied by the corresponding value vector to obtain an attention mechanism output.

[0079] Specifically, local features, local neighborhood features and global features are integrated. First, the local features and global features are linearly transformed by multiplying the weight matrix with the local features and the weight matrix with the global features, and the local features and global features are mapped to their corresponding query vectors Q, key vectors K and value vectors V. The dimension and format of the features are unified to facilitate subsequent calculations. Among them, the query vector represents the target features that need to be paid attention to; the key vector represents the content of all features; and the value vector represents the feature content that needs to be extracted.

[0080] Furthermore, the similarity scores between the query vector and all key vectors are calculated through a dot product operation, and the similarity scores are stored as attention scores. Then, the similarity scores between each query vector and all key vectors are normalized using a normalization function, and the similarity scores are converted into normalized weight distributions, so that the sum of all weights is 1.

[0081] Optionally, the normalized weight distribution relationship is constructed based on the normalized probability distribution, reflecting the attention of the hierarchical edge perception network HEA to each position. Then, multiple normalized weights in the normalized weight distribution relationship are multiplied and summed with the corresponding value vector to obtain the output of the attention mechanism. The output of the attention mechanism represents the fused feature information.

[0082] Through the attention mechanism module, TA focuses on important features and ignores irrelevant information. Local features, local neighborhood features and global features are integrated, making the feature representation more comprehensive and accurate, thereby improving the effect of feature extraction and classification.

[0083] In some implementations, the normalized weight is multiplied by the corresponding value vector to obtain the attention mechanism output, including:

[0084] By formula: Calculate to obtain the output of the attention mechanism, where Q is the query vector, K is the key vector, V is the value vector, is the scaling factor.

[0085] Optionally, Q×K represents a similarity score, It is used to perform dimension division on the similarity score to prevent the dot product value from being too large, which would lead to gradient vanishing when calculating the softmax.

[0086] Furthermore, after processing Applying the softmax function, we can get the attention weight Finally, the attention weights are multiplied by the value vector V to obtain the output of the attention mechanism.

[0087] Furthermore, after outputting the segmentation result of the power distribution network, the method further includes:

[0088] The segmentation result of the distribution network is used to identify distribution lines and extract the distribution line distribution location; line smoothness detection is performed based on the distribution line distribution location to obtain a smoothness detection result; abnormal nodes are identified based on the smoothness detection result; when the abnormal nodes do not exist, the segmentation result of the distribution network is determined.

[0089] Optionally, after obtaining the segmentation results of the distribution network, the segmentation results are used to identify the distribution lines, wherein the segmentation results are point cloud data related to the distribution network, specifically, point cloud data of the poles, distribution lines, transformers and other ancillary equipment of the distribution network. Exemplarily, based on feature extraction algorithms (such as edge detection algorithms, Hough transforms, etc.) and target detection technologies, the distribution line point cloud in the segmentation results of the distribution network is extracted to generate the distribution line distribution location.

[0090] Optionally, according to the distribution location of the distribution lines, the shape or layout of the distribution lines is analyzed based on computational geometry methods to detect the smoothness of the lines.

[0091] Exemplarily, a smoothness calculation method (such as polynomial fitting, B-spline curve, etc.) is used to detect the smoothness of the distribution line, and based on the smoothness detection results, abnormal nodes in the distribution line are identified, where the abnormal nodes may be points that deviate significantly from the smooth curve due to noise or data errors.

[0092] Optionally, when there are no abnormal nodes, the segmentation result of the power distribution network can be considered valid. Correspondingly, if there are abnormal nodes, the segmentation result is reprocessed or manually checked and corrected.

[0093] In some embodiments, after identifying abnormal nodes, the method further includes:

[0094] When the abnormal node exists, the abnormal distribution line and the adjacent distribution lines are obtained; the line abnormal node alignment compensation is performed according to the abnormal distribution line to determine the compensation characteristics; based on the compensation characteristics, the abnormal environmental characteristics are deducted; the environmental characteristics are extracted according to the adjacent distribution line, and the correlation analysis is performed between the environmental characteristics and the abnormal environmental characteristics to determine the correlation identification characteristics; the abnormal environmental characteristics are re-segmented using the correlation identification characteristics, and the distribution line characteristics are restored and segmented based on the compensation characteristics.

[0095] Specifically, first, the distribution line containing the abnormal node is identified and extracted as the abnormal distribution line, and the distribution line directly adjacent to the abnormal distribution line is identified as the adjacent distribution line. Exemplarily, the distribution line adjacent to the abnormal node is found using a spatial neighborhood search algorithm.

[0096] Furthermore, alignment compensation is performed on the abnormal nodes in the abnormal distribution line to eliminate the deviation caused by the abnormal nodes. At the same time, based on the alignment compensation result, compensation features are determined, wherein the compensation features include position, direction and morphological features.

[0097] Optionally, an interpolation algorithm is used to align and compensate abnormal nodes, including linear interpolation and spline interpolation.

[0098] Optionally, use a smoothing algorithm (such as Gaussian smoothing) to process abnormal nodes to achieve alignment compensation

[0099] Furthermore, based on the compensation feature, the compensation vector is generated with the abnormal node as the starting point combined with the compensation feature, and then the neighborhood feature of the compensation vector is extracted as the abnormal environment feature. Among them, the compensation vector represents the vector from the abnormal node to the compensation position, which contains direction and amplitude information. Through the above method, the environmental features around the abnormal node can be accurately extracted to provide data support for subsequent correlation analysis.

[0100] Optionally, correlation analysis is performed using abnormal environmental characteristics and environmental characteristics of adjacent distribution lines, including calculating the correlation coefficient between the abnormal environmental characteristics and the environmental characteristics of adjacent distribution lines, determining their correlation, and screening out correlation identification features.

[0101] Optionally, the abnormal environment features are re-segmented based on the associated identification features, and the abnormal environment features are segmented into environment-related point cloud data and power supply data related to the distribution network. Then, the distribution line features are restored to the abnormal nodes based on the compensation features, and the neighborhood space of the abnormal nodes is re-segmented, so as to correct the abnormal distribution lines with abnormal nodes.

[0102] Through the above process, abnormal distribution lines can be effectively handled and the reliability and accuracy of distribution network point cloud segmentation can be improved.

[0103] In some implementations, environmental features are extracted based on the adjacent distribution lines, and correlation analysis is performed using the environmental features and abnormal environmental features to determine associated identification features, including: setting a feature initial point based on the environmental features extracted from the adjacent distribution lines, the feature initial point being a cluster of pixel points whose distance from the abnormal node meets the distance requirement; determining neighborhood pixels of the feature initial point; fitting pixel growth information based on the feature initial point, the pixel growth information including grayscale value, color, texture, and gradient; taking the feature initial point as the starting point, performing growth coverage judgment on neighborhood pixels based on the pixel growth information, and taking the pixel area that meets the pixel growth information as a new feature initial point; fitting new pixel growth information based on the new feature initial point, iteratively using the pixel growth information to perform growth coverage judgment on neighborhood pixels until the abnormal environmental features are reached, and determining the associated identification features based on the current pixel growth information.

[0104] Specifically, first, environmental features are extracted from adjacent distribution lines. The environmental features include information such as the grayscale value, color, texture, and gradient of the pixel points, which reflect the environmental features during normal segmentation. Then, according to the position of the abnormal node, the pixel points that meet certain requirements for the distance from the abnormal node are selected as the feature initial point. The feature initial point is the starting point for growth and analysis. It needs to maintain a certain distance from the abnormal node in order to distinguish normal and abnormal environmental features. Next, the pixels around the feature initial point are determined as neighborhood pixels, which contain pixel information related to the feature initial point. Then, based on the feature initial point, the pixel growth information, including grayscale value, color, texture, and gradient, is fitted, which reflects the growth trend and characteristics of the pixel points, which help to identify and distinguish normal and abnormal environmental features.

[0105] Furthermore, starting from the feature initial point, the growth coverage of the neighboring pixels is judged based on the pixel growth information to determine which pixel regions meet the requirements of the pixel growth information. These regions can be used as new feature initial points. In other words, the pixel regions that meet the pixel growth information are regions with continuous pixel types.

[0106] Furthermore, based on the new feature initial point, new pixel growth information is fitted, and iterative growth coverage judgment is performed. The iterative process continuously expands the feature initial point and grows to cover new pixel areas until the abnormal environmental features are covered. Based on the current pixel growth information, the associated recognition features are determined.

[0107] In summary, the distribution network point cloud segmentation method based on environment fusion provided by the present invention has the following technical effects:

[0108] By collecting distribution network point cloud data, wherein the point cloud data includes environmental point cloud data, and the environmental point cloud data is integrated with the distribution network point cloud data; obtaining a pre-constructed hierarchical edge perception network, wherein the hierarchical edge perception network includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling, and upsampling, wherein the attention mechanism module, global to local downsampling, and upsampling construct a multi-level network, and the multi-level network contains multiple jump connections, and the jump connection is the jump between the attention mechanism module and the global to local downsampling; inputting the distribution network point cloud data into the hierarchical edge perception network, using the local embedding module to capture the local features and local neighborhood features of each point in the distribution network point cloud data, and using the global embedding module to perform global feature recognition, and using the attention mechanism module to fuse the local features and local neighborhood features with the global features, focusing on the segmentation task features, and using the multi-level network to contain multiple jump connections to perform multiple downsampling and upsampling, segmenting the environmental features and the distribution network features, and outputting the segmentation results of the distribution network. Thus, the technical effect of improving the efficiency and accuracy of distribution network point cloud segmentation is achieved.

[0109] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the above-mentioned embodiments. It should be understood that those skilled in the art can still modify the technical solutions recorded in the above-mentioned embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A distribution network point cloud segmentation method based on environment fusion, characterized in that: The method comprises: Collecting distribution network point cloud data, wherein the point cloud data includes environmental point cloud data, and the environmental point cloud data is integrated with the distribution network point cloud data; Obtain a pre-constructed hierarchical edge perception network, wherein the hierarchical edge perception network includes a local embedding module, a global embedding module, an attention mechanism module, global to local downsampling, and upsampling, wherein the attention mechanism module, global to local downsampling, and upsampling construct a multi-level network, and the multi-level network contains multiple jump connections, and the jump connection is a jump between the attention mechanism module and the global to local downsampling; The distribution network point cloud data is input into the hierarchical edge perception network, the local features and local neighborhood features of each point in the distribution network point cloud data are captured by the local embedding module, and the global features are recognized by the global embedding module. The local features, local neighborhood features and global features are fused with the attention mechanism module to focus on the segmentation task features, and multiple downsampling and upsampling are performed through multiple jump connections contained in the multi-level network to segment the environmental features and the distribution network features, and the segmentation result of the distribution network is output.

2. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 1, characterized in that: The local embedding module captures the local features and local neighborhood features of each point in the distribution network point cloud data, including: Determine the K neighbor points of each point in the point cloud data through distance measurement; The feature vectors of the K neighboring points are copied K times to construct a feature matrix; The K copied feature matrices are spliced ​​according to the channel dimension to obtain a stacked feature matrix, where the stacked feature matrix includes local features of the original point and the neighborhood; The stacked feature matrix is ​​used as output to obtain the local features and local neighborhood features.

3. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 2, characterized in that: Before the K feature matrices obtained by copying are concatenated according to the channel dimension, the method further includes: Configure multi-scale K value; Based on the multi-scale K value, determining multi-scale K neighbor points; Perform feature insertion according to the feature vectors of the multi-scale K neighboring points to form multiple new feature dimensions; Performing visual evaluation on the multiple new feature dimensions, determining the relationship between feature distribution and target feature variables, and screening new feature dimensions; The new feature dimension is added to the concatenated feature.

4. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 1, characterized in that: The local features, local neighborhood features and global features are integrated through the attention mechanism module, including: According to the local features and the global features, mapping is performed into a query vector, a key vector, and a value vector through a weight matrix; Calculate the similarity scores between the query vector and all key vectors to obtain an attention score; The similarity scores between each query vector and all key vectors are normalized by a normalization function, and a normalized weight distribution relationship between each query vector and the key vector is established; The normalized weight is multiplied by the corresponding value vector to obtain the attention mechanism output.

5. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 4, characterized in that: The normalized weight is multiplied by the corresponding value vector to obtain the attention mechanism output, including: By formula: Calculate to obtain the output of the attention mechanism, where Q is the query vector, K is the key vector, V is the value vector, is the scaling factor.

6. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 1, characterized in that: After outputting the segmentation result of the power distribution network, the method further includes: Performing distribution line identification on the segmentation result of the distribution network, and extracting the distribution location of the distribution lines; Perform line smoothness detection according to the distribution line positioning to obtain a smoothness detection result; According to the smoothness detection result, identifying abnormal nodes; When the abnormal node does not exist, a segmentation result of the power distribution network is determined.

7. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 6, characterized in that: After identifying the abnormal node, the method further includes: When the abnormal node exists, obtaining the abnormal distribution line and adjacent distribution lines; Perform line abnormal node alignment compensation according to the abnormal distribution line and determine compensation characteristics; Based on the compensation feature, deducting abnormal environment features; Extracting environmental features according to the adjacent power distribution lines, performing correlation analysis using the environmental features and abnormal environmental features, and determining correlation identification features; The abnormal environment feature is re-segmented using the associated identification feature, and the distribution line feature is restored and re-segmented based on the compensation feature.

8. The method for point cloud segmentation of a power distribution network based on environment fusion according to claim 7, characterized in that: Extracting environmental features according to the adjacent power distribution lines, performing correlation analysis using the environmental features and abnormal environmental features, and determining correlation identification features, including: According to the environmental features extracted from the adjacent power distribution lines, a feature initial point is set, where the feature initial point is a cluster of pixel points whose distance from the abnormal node meets the distance requirement; Determine the neighborhood pixels of the feature initial point; Fitting pixel growth information based on the feature initial point, wherein the pixel growth information includes gray value, color, texture, and gradient; Taking the feature initial point as the starting point, performing growth coverage judgment on neighboring pixels based on the pixel growth information, and taking the pixel area that meets the pixel growth information as the new feature initial point; New pixel growth information is fitted based on the new feature initial point, and the pixel growth information is iteratively used to determine the growth coverage of neighboring pixels until the abnormal environment feature is reached, and the associated identification feature is determined based on the current pixel growth information.

Images