Power corridor data adaptive classification method and system based on multiple scenes and multiple modes

Abstract

The invention discloses an adaptive classification method and system based on multi-scene and multi-mode power corridor data. The method comprises the steps of collecting multi-scene power corridor data and establishing a multi-scene data training set; the collected power corridor data are preprocessed; performing information supplementation on the acquired missing power corridor data by adopting multi-modal data, and extracting global features of multiple scenes and local features of single-shift power corridor data; and inputting the global features and the local features into a pre-constructed adaptive classification model for training to obtain a power corridor data classification result. According to the method, scene labels are marked on multi-period visible light data in single-file data, the scene occurrence probability is counted, and the maximum probability is the scene label of point cloud data of the file; then, for different scene types, different segmentation parameters are set through an adaptive method to perform classification processing on power corridors of different scenes; the multi-modal data multi-scene power corridor segmentation method has the advantages of high generalization and high classification precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network monitoring, and particularly relates to a method and system for adaptive classification of multi-scenario and multi-modal power corridor data. Background Art

[0002] The statements herein only provide background art related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of social economy, the power demand is continuously increasing, and the requirements for the power system are also getting higher and higher. As the main channel connecting the power source point and the load center, the safe, reliable and efficient operation of the power corridor is of great significance for ensuring the normal operation of the national economy and social life.

[0004] Currently, although many technologies have made great progress, there are still some problems to be solved urgently in the management and application of power corridor data:

[0005] (1) Due to the wide coverage and complex and changeable terrain of the power corridor, traditional data collection methods often rely on manual on-site measurement, which is time-consuming, laborious and error-prone.

[0006] (2) The data update speed is slow, and it is unable to reflect the latest situation of the power corridor in a timely manner, affecting the timeliness and accuracy of the data.

[0007] (3) The data analysis ability is limited. Although a large amount of power corridor data has been collected, due to the lack of efficient data processing and analysis tools, many potential valuable information has not been effectively mined and utilized.

[0008] The power corridor not only undertakes a huge power transmission task, but also is an important part of the grid structure safety. Therefore, accurately and timely mastering the high-precision power corridor category and data information is crucial for ensuring the efficient operation of the power system.

[0009] In recent years, algorithms based on machine learning and deep learning have been widely used in the industry for automatic classification of power corridor data. These algorithms can automatically identify and classify according to different types of ground object features, such as buildings, trees, ground, etc., greatly improving the work efficiency and classification accuracy. However, the generalization of these algorithms is relatively poor, and the classification accuracy of point cloud data for different scenarios is not high. Therefore, it is very important to improve the generalization of power corridor data classification. Summary of the Invention

[0010] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and provide a method for adaptive classification of multi-scenario and multi-modal power corridor data.

[0011] To achieve the above object, the present invention is implemented through the following technical solutions:

[0012] On the one hand, the technical solution of the present invention provides an adaptive classification method for multi-scenario and multi-modal power corridor data, including:

[0013] Collect multi-scenario power corridor data and establish a multi-scenario data training set;

[0014] Preprocess the collected power corridor data;

[0015] Use multi-modal data to supplement the missing information of the power corridor data, and extract the global features of multi-scenarios and the local features of single-span power corridor data;

[0016] Input the global features and local features into a pre-constructed adaptive classification model for training to obtain the classification result of the power corridor data.

[0017] In at least one embodiment, the types of power corridor data acquisition scenarios include: rural and urban areas, forests and deserts, mountains and plains; the multi-modal data includes visible light data and point cloud data;

[0018] The establishment of the multi-scenario data training set is specifically as follows: Define scene labels for the visible light data collected in multiple time periods of single-span point cloud data, and count the probabilities of the labels, and use the label with the highest probability as the scene label of the single-span point cloud power corridor, so as to establish a multi-scenario data training set.

[0019] In at least one embodiment, the preprocessing of the collected power corridor data includes: preprocessing the visible light data and point cloud data respectively, removing irrelevant noise points, performing binary processing on the visible light data, and performing coordinate transformation processing on the data to convert the point cloud data and visible light data to the same coordinate system.

[0020] In at least one embodiment, the use of multi-modal data to supplement the missing information of the power corridor data is specifically as follows: Use the visible light data of the power corridor to supplement the missing point cloud data; or, use the point cloud data of the power corridor to supplement the missing visible light data;

[0021] The local features of single-span power corridor data include: the density of point cloud data, the up and down relationship structure, texture information, and geometric features.

[0022] In at least one embodiment, the global features include tower structure features, conductor sag and span features, vegetation features, terrain features, and environmental features;

[0023] For the conductor sag and span features, use the asymmetric catenary equation to fit the shape of the conductor, and the asymmetric catenary equation is:

[0024]

[0025] In the formula, a is the catenary scale parameter; x represents the coordinate of any point on the catenary; c is used to adjust the position of the entire catenary in the vertical direction;

[0026] For vegetation characteristics, using the texture information of visible light data, through the analysis of visible light spectral information, rough classification of ground vegetation is carried out; classification of vegetation types is carried out by setting the point cloud density threshold;

[0027] Specifically, the density of the three-dimensional point set of the point cloud neighbors is represented by the number of three-dimensional points within a sphere centered on this three-dimensional point with a radius of R; the curvature calculation is as follows:

[0028]

[0029] Among them, fabs(·) represents calculating the absolute value, and λ1, λ2, and λ3 are the eigenvalues arranged from large to small of the covariance matrix formed by the three-dimensional point neighbor point set;

[0030] For terrain characteristics, a digital surface model and a digital elevation model are constructed for multi-scene point cloud data. The digital surface model obtains the height information of all objects on the ground surface, reflecting the actual elevation of the surface objects, and the digital elevation model represents the elevation of the ground, that is, the elevation of the bare ground;

[0031] The process of obtaining terrain characteristics is as follows: First, preprocess the point cloud data to remove non-ground points; second, interpolate the ground points to construct a digital elevation model; then, optimize the elevation and fill the holes; finally, smooth the terrain data and evaluate the accuracy.

[0032] In at least one embodiment, the adaptive classification model is constructed by splicing and weighting global features and local features, specifically including:

[0033] The input layer is used to receive the original point cloud data and visible light data, obtain the scene label of the data, count the scene types according to the scene label, and obtain the scene label of the point cloud;

[0034] The feature extraction layer is used to extract global features and local features;

[0035] The fusion layer is used to perform weighted fusion on the extracted global features and local features;

[0036] The fully connected layer is used to perform non-linear transformation on the feature vector to learn more complex patterns;

[0037] The output layer is used to output the classification result.

[0038] In at least one embodiment, after obtaining the power corridor data classification results, tree barrier detection and analysis are performed. The specific process is as follows:

[0039] First, determine the shape of the wire and use different wire fitting methods according to different scenarios;

[0040] Secondly, based on spectral analysis of visible light data, tree species are distinguished and a tree growth model is established. This growth model is used to predict tree growth in the next few years and assess their potential threat to power lines.

[0041] Next, based on the digital surface model information, the minimum distance between tree points and power lines is accurately calculated. Based on the safety distance standards for power lines, it is assessed whether the trees pose a safety hazard and a tree barrier hazard analysis report is generated.

[0042] In at least one embodiment, the tree obstacle hazard analysis report includes information on the location, distance, and risk level of the potential hazard.

[0043] On the other hand, the technical solution of the present invention also provides an adaptive classification system for multi-scenario and multi-modal power corridor data, including:

[0044] The data acquisition and processing module is configured to: collect multi-scenario power corridor data and establish a multi-scenario data training set; pre-process the collected power corridor data; and use multi-modal data to supplement the missing power corridor data;

[0045] The multi-scene global feature extraction module is configured to: determine the point cloud data scene based on visible light data, extract the tower structure, conductor distance and sag information in the scene, calculate the point cloud density, and scene characteristics such as season;

[0046] The single-stage power corridor data local feature extraction module is configured to: for the collected 3D point cloud data, extract the geometric structure, density information, and upper and lower structural relationship characteristics of the point cloud data through layered processing, and establish local feature information;

[0047] The adaptive training dataset module is configured to determine the global characteristics of the scene through a single-file power corridor scene type, and then classify the point cloud data type according to the local characteristics of the power corridor to obtain a high-precision classification result of the tower, conductor and vegetation types.

[0048] In at least one embodiment, it further includes a tree obstacle detection and analysis module, which is configured to: first, determine the shape of the wire, and adopt different wire fitting methods according to different scenarios; second, distinguish the types of trees based on the spectral analysis of visible light data, establish a growth model of the trees, and use the growth model to predict the growth of the trees in the next few years and evaluate their potential threats to the power line; then, accurately calculate the minimum distance between the tree points and the power line according to the digital surface model information, and evaluate whether the trees pose a safety hazard according to the safety distance standard of the power line, and generate a tree obstacle hazard analysis report.

[0049] The beneficial effects of the above technical solutions of the present invention are as follows:

[0050] A multi-scenario and multi-modal power corridor data adaptive classification method of the present invention first labels the multi-temporal visible light data in a single-span data with scene labels, determines the scene type of the visible light data in each time period, and counts the probability of the scene occurrence. The maximum probability is the scene label of the point cloud data of this span; then, for different scene types, different segmentation parameters are set by an adaptive method to classify and process the power corridors of different scenes; on this basis, tree obstacle analysis and detection are performed for inspection and maintenance output. The multi-modal data multi-scene power corridor segmentation method has high generalization and high classification accuracy. Brief Description of the Drawings

[0051] The specification drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0052] Figure 1 It is a flow schematic diagram of a multi-scenario and multi-modal power corridor data adaptive classification method of the present invention;

[0053] Figure 2 It is a structural schematic diagram of a multi-scenario and multi-modal power corridor data adaptive classification system of the present invention;

[0054] Figure 3 It is a flow schematic diagram of constructing an adaptive model by splicing and weighting global and local features in Embodiment 1 of the present invention. Detailed Description of the Specific Embodiment

[0055] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0056] As introduced in the background art, the object of the present invention is to overcome the deficiencies existing in the above-mentioned prior art, and provide a method and system for adaptively classifying multi-scenario and multi-modal power corridor data.

[0057] Embodiment 1

[0058] In a typical embodiment of the present invention, as Figure 1 shown, this embodiment discloses a method for adaptively classifying multi-scenario and multi-modal power corridor data, including:

[0059] S100: Collect multi-scenario power corridor data and establish a multi-scenario data training set;

[0060] S200: Preprocess the collected power corridor data;

[0061] S300: Use multi-modal data to supplement the information of the power corridor data with missing collection;

[0062] S400: Extract the global features of multiple scenarios;

[0063] S500: Extract the local features of single-span power corridor data;

[0064] S600: Input the global features and local features into a pre-constructed adaptive classification model for training to obtain the classification result of the power corridor data.

[0065] The following is a detailed description.

[0066] S100: Collect multi-scenario power corridor data and establish a multi-scenario data training set.

[0067] Since the collected data is becoming more and more extensive, in order to make full use of different modal data and increase the accuracy and reliability of the data, in recent years, many multi-modal data training algorithms have emerged, but more are based on the fusion of multi-modal data, and there are almost no algorithms for adaptively classifying multi-scenario data using multi-modal data. Different scenarios have different fine structures of the poles and towers, different sag structures of the power lines, and there are also differences in terrain, point cloud density, and environment. Therefore, for different scenarios, the ability of adaptive classification is required.

[0068] S101. Collection of multi-scenario and multi-modal data.

[0069] Collect multi-scenario and multi-modal power corridor data through drones, robots or handheld power inspection equipment, and use this as the basic training data.

[0070] In this embodiment, the multi-scenarios include a total of 6 scenarios, and their types include rural and urban areas, forests and deserts, mountains and plains. The multi-modal data includes two-dimensional visible light data and three-dimensional point cloud data. By collecting power corridor data of multiple scenarios and multiple modalities, the generalization of the power corridor data classification algorithm can be improved.

[0071] S102. Establish a multi-scenario data training set.

[0072] After collecting the multi-scenario and multi-modal power corridor data, for the collected multi-scenario data, define scene labels for the visible light data collected in multiple time periods of the single-span point cloud data, and count the probabilities of the labels. Use the label with the highest probability as the scene label of the single-span point cloud power corridor, and establish a multi-scenario data training set in this way, which can enrich the data scenarios and improve the generalization of the model.

[0073] S200. Preprocess the collected power corridor data.

[0074] First, clean the data and select data with rich scenarios and diverse tower types for training.

[0075] Secondly, in order to improve the accuracy of the data in the training set, preprocess the collected two-dimensional visible light data and three-dimensional point cloud data of the power corridor.

[0076] Specifically, perform denoising and registration processing on the visible light data and the point cloud data respectively, remove irrelevant noise points, perform binarization processing on the visible light data, and perform coordinate transformation processing on the data to convert the point cloud data and the visible light data to the same coordinate system.

[0077] S300. Use multi-modal data to supplement the information of the power corridor data with missing acquisitions.

[0078] During the data acquisition process by drones, robots or handheld power inspection equipment, there may be reasons such as occlusion, resulting in missing information in the collected visible light data or point cloud data. Therefore, it is necessary to use multi-modal data to supplement the data information.

[0079] Specifically, when there is missing point cloud data and the visible light data has collected relevant information at the missing part of the point cloud data, at this time, use the visible light data of the power corridor to supplement the missing point cloud data with information; when there is missing visible light data and the point cloud data has collected relevant information at the missing part of the visible light data, at this time, use the point cloud data of the power corridor to supplement the missing visible light data with information. By making full use of data of different modalities, the accuracy and integrity of the collected power corridor data are increased, and the reliability of the data classification result is further ensured.

[0080] S400. Extract global features for multiple scenarios.

[0081] In this embodiment, the data is divided into four types: vegetation, buildings, poles, and wires, that is, the data classification results are vegetation, buildings, poles, and wires respectively. Due to different scenarios, there are obvious differences in the structure of poles, the sag of wires, the density of vegetation, the undulation of terrain, and the change of seasons. For example, in mountainous scenarios, poles with large floor areas and hard building materials are generally used, the spacing between poles is large, and the shape and sag of wires also vary greatly. Secondly, the vegetation coverage density is different, and the terrain slope is also different.

[0082] Therefore, in order to improve the generalization of the point cloud data category classification algorithm, it is crucial to extract global features. The global features mainly include five aspects: pole structure features, wire sag and span features, vegetation features, terrain features, and environmental features, which are specifically as follows:

[0083] 1) Pole structure feature information: For poles with different structures and heights in different scenarios, their materials, design, and installation methods are also different. For example, in urban areas, the pole structure is relatively low and is a steel pipe tower structure; for mountainous areas, the pole height is generally higher and is made of high-strength steel. The pole structure is generally a cup-shaped tower or a cat-head tower.

[0084] 2) Sag and span feature information: Due to the large terrain undulation and complex natural environment in mountainous areas, in order to ensure power safety, the span is generally small and the sag is stable; while in cities, the span is large, and the sag design mainly considers factors such as pedestrian safety. Therefore, the threshold of sag and the span are used as global features for wire extraction.

[0085] In order to fit the shape of the wire, the centroid coordinates of the poles and the sag parameters can be obtained, the distance between the poles can be calculated, and the power wire can be fitted through the distance and sag. The fitting function involves parameters such as elevation, sag, and span. For mountainous areas, more natural factors need to be considered, and the sag control of the power line is more strict. Therefore, the standard catenary equation is no longer applicable. Therefore, in this embodiment, an asymmetric catenary equation is used to fit the shape of the wire, and the formula is as follows:

[0086]

[0087] In the formula, a is the catenary scale parameter; x represents the coordinate of any point on the catenary; c is used to adjust the position of the entire catenary in the vertical direction.

[0088] 3) Vegetation feature information: Different vegetation has different reflectivity to light. Therefore, the texture information of visible light data can be used to analyze the visible light spectrum information to perform a coarse classification of ground vegetation. At the same time, since the density of higher vegetation in forests is high and the point cloud density of plain farmland is low, vegetation type can be classified by setting a point cloud density threshold. The density of a point cloud neighboring 3D point set is represented by the number of 3D points within a sphere with a radius R centered at the 3D point. The curvature is calculated as:

[0089]

[0090] Where fabs(·) represents the calculated absolute value, and λ1, λ2, and λ3 are the eigenvalues of the covariance matrix of the three-dimensional point neighboring point set, arranged from large to small.

[0091] 4) Global terrain feature information: Digital surface models and elevation modeling are performed on multi-scene point cloud data. The digital surface model obtains the height information of all objects on the surface (including the ground, buildings, vegetation, etc.), reflecting the actual elevation of surface objects. The digital elevation model represents the elevation of the ground, that is, the elevation of the bare ground. The digital elevation model excludes the height of non-ground objects such as buildings and vegetation.

[0092] Specifically, the process of obtaining the global features of the terrain is as follows:

[0093] First, the point cloud data is preprocessed, including denoising and registration processing, to remove non-ground points;

[0094] Secondly, interpolate the ground points to construct a digital elevation model; the interpolation method uses Kriging interpolation:

[0095]

[0096] Among them, h0 is the height of the interpolation point, γ i is the weight coefficient, h i is the height of the known point.

[0097] Therefore, when calculating vegetation height, you need to calculate:

[0098] H 植被 =H 表面 -H 地面

[0099] Among them, H 植被 is the vegetation height, H 表面 Represents the height of the digital surface model, H 地面 Indicates the height of the digital elevation model.

[0100] Next, perform elevation optimization and void filling; finally, perform smoothing processing and accuracy evaluation on the terrain data. The purpose of smoothing is to reduce noise and make the terrain surface more natural and smooth. A Gaussian filtering method based on a linear filtering technique of weighted average can be used. The weights are determined according to the Gaussian distribution, which can effectively remove high-frequency noise while retaining the main features of the terrain. Since a larger standard deviation σ value will make the image smoother, but it may also blur some details, a suitable standard deviation, that is, the σ value, needs to be selected during the processing.

[0101] In this way, extracting the terrain information of the scene can enrich the spatial resolution and vertical accuracy of the data, facilitate the extraction of each point cloud type and vegetation type information, which is crucial for subsequent tree obstacle calculation and analysis and can facilitate tree growth simulation; at the same time, at the visual level, the point cloud data can be directly used to generate a 3D model, which makes the tree obstacle analysis more intuitive and accurate. Through the 3D model, the relative positions between the trees and the power lines can be clearly seen, and the relationships between the trees and the lines can be viewed from multiple perspectives, which helps to comprehensively evaluate potential tree obstacle risks.

[0102] 5) Environmental feature information: When collecting power corridor data, seasonal information will have significant distinguishing features for data collection and extraction. For example, in the summer scene, the ground vegetation coverage density is large, while in the winter, the vegetation coverage density is small. In winter, it is easier to extract power lines and tower structures due to the lack of tree occlusion. Therefore, in this embodiment, sensor data is used to detect the current environmental conditions, such as season, weather, vegetation status, etc.

[0103] S500. Extract the local features of the single-span power corridor data.

[0104] For a single-span tower, there are also obvious characteristic differences in its categories. Using the point cloud data, the following aspects of characteristic information can be obtained:

[0105] 1) In terms of height: Generally, the wires are the highest, followed by the towers and the ground. Therefore, the point cloud can be classified into two parts: the ground and above the ground. Among them, for the part above the ground: extract the towers and transmission wires; for the ground part: mainly divide it into trees, high-rise buildings, and low vegetation.

[0106] 2) In terms of geometric structure features: The outer contour of a building is a regular geometric figure, while the outer contour of a tree is an irregular closed curve figure. Therefore, trees and buildings can be distinguished by geometric shapes.

[0107] 3) The density of the point cloud data: The density feature reflects the distribution of the point cloud data in space and helps to distinguish different types of ground objects. For example, wires are usually sparser than vegetation and the ground.

[0108] 4) Texture information: Since visible light data has rich texture information, different categories have different spectral reflectivities, so spectral reflection information of different categories can be captured.

[0109] Therefore, when classifying single-pole tower data, information in four aspects, namely height, geometric structure, density, and color texture, can be fused into the power corridor data information as a multi-dimensional vector.

[0110] S600. Input the global features and local features into a pre-constructed adaptive classification model for training to obtain the classification result of the power corridor data.

[0111] The classification method in this embodiment adopts a segmentation method combining traditional and deep learning: for data in different scenarios, the deep learning model has generalization ability, and the model dynamically adjusts its own parameters and structure according to the changes and feedback of the data during the training process.

[0112] Specifically, as Figure 3 shown, an adaptive model is constructed by splicing and weighting global and local features as follows:

[0113] 1) Input layer: Receive the original point cloud and visible light data, obtain the scene label of the data, and count the scene types according to the scene label to obtain the scene label of the point cloud.

[0114] 2) Feature extraction layer: Extract global and local features. For example, when training the power corridor data of a summer forest, first extract the global information of the power corridor, such as terrain shape, climate information, density of the point cloud data, and height and type of the tower. Since the vegetation is lush in summer, the set vegetation density will be higher. At the same time, there are generally no high-rise buildings in the forest, but more tall vegetation. Therefore, when extracting local features, the high-rise building category can be classified as the tall vegetation type.

[0115] 3) Fusion layer: Weight and fuse the extracted global and local features.

[0116] 4) Fully connected layer: Perform non-linear transformation on the feature vector to learn more complex patterns.

[0117] 5) Output layer: Output the classification result.

[0118] In addition, by adopting the above method, through the extraction of global features and local features of the power corridor data, high-precision information on towers, conductors, vegetation, and terrain can be obtained. After obtaining the classification result of the power corridor data, tree obstacle detection and analysis can be carried out.

[0119] For tree obstacle detection and analysis, first determine the shape of the conductor, and adopt different methods to fit the conductor according to different scenarios;

[0120] Secondly, according to the spectral analysis of visible light data, classify the types of trees, establish a growth model of the trees, and use the growth model to predict the growth of the trees in the next few years and evaluate their potential threats to power lines;

[0121] Next, based on the digital surface model information, the minimum distance between the tree points and the power lines can be accurately calculated. According to the safety distance standard of the power lines, evaluate whether the trees pose a safety hazard and generate a tree obstacle hazard analysis report. Among them, the tree obstacle hazard analysis report includes information such as the location, distance, and risk level of potential hazards.

[0122] A method for adaptively classifying multi-scenario and multi-modal power corridor data in this embodiment first tags the scenes of multi-temporal visible light data within a single-span data, determines the scene types of the visible light data for each time period, and counts the probability of the scene occurrence. The maximum probability is the scene tag of the point cloud data for this span. Then, for different scene types, different segmentation parameters are set through an adaptive method to classify the power corridors of different scenes. On this basis, tree obstacle analysis and detection are performed for inspection and maintenance output. The multi-modal data multi-scene power corridor segmentation method has high generalization and relatively high classification accuracy.

[0123] Embodiment 2

[0124] In a typical implementation manner of the present invention, as Figure 2 shown, this embodiment discloses a system for adaptively classifying multi-scenario and multi-modal power corridor data, including:

[0125] A data acquisition and processing module, configured to: collect multi-scenario power corridor data and establish a multi-scenario data training set; perform preprocessing on the collected power corridor data; and use multi-modal data to supplement information for the missing power corridor data in the collection;

[0126] A multi-scenario global feature extraction module, configured to: determine the point cloud data scene according to the visible light data, extract the tower structure, wire distance, and sag information in the scene, calculate the point cloud density, and features such as the season of the scene;

[0127] A local feature extraction module for single-span power corridor data, configured to: for the collected three-dimensional point cloud data, extract the geometric structure, density information, and upper and lower structural relationship features of the point cloud data by layer processing of the point cloud data, and establish local feature information;

[0128] An adaptive training data set module, configured to: determine the global features of the scene through the single-span power corridor scene type, and then classify the point cloud data types according to the local features of the power corridor to obtain classification results of tower poles, wires, and vegetation types with relatively high accuracy.

[0129] Further, it further includes a tree obstacle detection and analysis module, which is configured to: first, determine the shape of the wire and adopt different wire fitting methods according to different scenarios; second, distinguish the types of trees based on the spectral analysis of visible light data, establish a growth model of the trees, and use the growth model to predict the growth of the trees in the next few years and evaluate its potential threat to the power line; then, according to the digital surface model information, accurately calculate the minimum distance between the tree points and the power line, and evaluate whether the trees pose a safety hazard according to the safety distance standard of the power line, and generate a tree obstacle hazard analysis report.

[0130] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An adaptive classification method for multi-scenario and multi-modal power corridor data, characterized in that Including: Collecting multi-scenario power corridor data and establishing a multi-scenario data training set; Preprocessing the collected power corridor data; Using multi-modal data to supplement the missing information of the power corridor data, and extracting the global features of multi-scenarios and the local features of single-span power corridor data; Inputting the global features and local features into a pre-constructed adaptive classification model for training to obtain the classification results of the power corridor data.

2. The multi-scenario multi-modal power corridor data adaptive classification method according to claim 1, wherein The types of power corridor data collection scenarios include: rural and urban areas, forests and deserts, mountains and plains; multi-modal data includes visible light data and point cloud data; Specifically, establishing a multi-scenario data training set means: defining scene labels for the visible light data collected in multiple time periods of single-span point cloud data, and statistically calculating the probabilities of the labels, and taking the label with the highest probability as the scene label of the single-span point cloud power corridor, so as to establish a multi-scenario data training set.

3. The multi-scenario and multi-modal power corridor data adaptive classification method according to claim 1, characterized in that The preprocessing of the collected power corridor data includes: preprocessing the visible light data and point cloud data respectively, removing irrelevant noise points, performing binary processing on the visible light data, and performing coordinate transformation processing on the data to convert the point cloud data and visible light data into the same coordinate system.

4. The multi-scenario and multi-modal power corridor data adaptive classification method according to claim 2, wherein Specifically, using multi-modal data to supplement the missing information of the power corridor data means: using the visible light data of the power corridor to supplement the missing information of the point cloud data; or, using the point cloud data of the power corridor to supplement the missing visible light data. The local features of single-span power corridor data include: the density of point cloud data, the up-down relationship structure, texture information, and geometric features.

5. The multi-scenario and multi-modal power corridor data adaptive classification method according to claim 1, characterized in that, The global features include tower structure features, conductor sag and span features, vegetation features, terrain features, and environmental features; For the conductor sag and span features, an asymmetric catenary equation is used to fit the shape of the conductor, and the asymmetric catenary equation is: where a is the catenary scale parameter; x represents the coordinate of any point on the catenary; c is used to adjust the position of the entire catenary in the vertical direction; For the vegetation features, using the texture information of the visible light data, through the analysis of the visible light spectrum information, a rough classification of the ground vegetation is carried out; the vegetation type is classified by setting the point cloud density threshold; Specifically, the density of the three-dimensional point set of the point cloud neighbors is represented by the number of three-dimensional points in a sphere with a radius of R centered on the three-dimensional point; the curvature calculation is: where fabs(·) represents calculating the absolute value, and λ1, λ2, and λ3 are the eigenvalues arranged from large to small of the covariance matrix formed by the three-dimensional point neighbor point set; For the terrain features, a digital surface model and a digital elevation model are constructed for the multi-scenario point cloud data. The digital surface model obtains the height information of all objects on the ground surface, reflecting the actual elevation of the ground surface objects, and the digital elevation model represents the elevation of the ground, that is, the elevation of the bare ground; The process of obtaining the terrain features is: first, preprocessing the point cloud data to remove non-ground points; second, interpolating the ground points to construct a digital elevation model; then, optimizing the elevation and filling holes; finally, smoothing the terrain data and evaluating the accuracy.

6. The multi-scenario and multi-modal power corridor data adaptive classification method according to claim 1, wherein The adaptive classification model is constructed by splicing and weighting global features and local features, specifically including: An input layer, which is used to receive the original point cloud data and visible light data, obtain the scene labels of the data, count the scene types according to the scene labels, and obtain the scene labels of the point cloud. A feature extraction layer, which is used to extract global features and local features. A fusion layer, which is used to perform weighted fusion on the extracted global features and local features. A fully connected layer, which is used to perform non-linear transformation on the feature vectors and learn more complex patterns. An output layer, which is used to output the classification results.

7. The multi-scenario multi-modal power corridor data adaptive classification method according to claim 1, wherein After obtaining the classification results of the power corridor data, tree obstacle detection and analysis are carried out. The specific process is as follows: First, determine the shape of the conductor, and adopt different methods to fit the conductor according to different scenes. Secondly, according to the spectral analysis of the visible light data, distinguish the types of trees, establish a growth model of the trees, and use the growth model to predict the growth of the trees in the next few years and evaluate their potential threats to the power lines. Then, according to the digital surface model information, accurately calculate the minimum distance between the tree points and the power lines, and evaluate whether the trees pose a safety hazard according to the safety distance standard of the power lines, and generate a tree obstacle hazard analysis report.

8. The multi-scenario and multi-modal power corridor data adaptive classification method according to claim 7, wherein The tree obstacle hazard analysis report includes the location, distance, and risk level information of potential hazards.

9. An adaptive classification system for multi-scenario and multi-modal power corridor data, characterized in that, Including: A data acquisition and processing module, which is configured to: acquire multi-scene power corridor data and establish a multi-scene data training set; Preprocess the acquired power corridor data; use multi-modal data to supplement the information of the missing power corridor data. A multi-scene global feature extraction module, which is configured to: determine the point cloud data scene according to the visible light data, extract the tower structure, conductor distance, and sag information in the scene, calculate the point cloud density, and the seasonal characteristics of the scene. A local feature extraction module for single-span power corridor data, which is configured to: for the acquired three-dimensional point cloud data, extract the geometric structure, density information, and upper and lower structure relationship features of the point cloud data by layering the point cloud data, and establish local feature information. An adaptive training data set module, which is configured to: determine the global features of the scene through the single-span power corridor scene type, and then classify the point cloud data types according to the local features of the power corridor to obtain classification results of tower poles, conductors, and vegetation types with higher accuracy.

10. The multi-scenario and multi-modal power corridor data-based adaptive classification system according to claim 9, wherein It also includes a tree obstacle detection and analysis module, which is configured to: first determine the shape of the conductor, and adopt different methods to fit the conductor according to different scenes; secondly, according to the spectral analysis of the visible light data, distinguish the types of trees, establish a growth model of the trees, and use the growth model to predict the growth of the trees in the next few years and evaluate their potential threats to the power lines; then, according to the digital surface model information, accurately calculate the minimum distance between the tree points and the power lines, and evaluate whether the trees pose a safety hazard according to the safety distance standard of the power lines, and generate a tree obstacle hazard analysis report.

Images