Transmission line identification and positioning method, device and storage device for live distribution network operations

Abstract

The present invention provides a method, device, and storage device for identifying and locating transmission lines for live operations on a distribution network. The method comprises: obtaining a transmission line point cloud through laser radar scanning; preprocessing the point cloud, wherein the preprocessing includes multiple filtering and point cloud clustering and segmentation; extracting transmission line features from the preprocessed point cloud; and identifying transmission line and lead targets using the extracted features. A transmission line identification and positioning device and storage device for live operations on a distribution network are provided, which are used to implement a method for identifying and locating transmission lines for live operations on a distribution network. The beneficial effects of the present invention are: the method of the present invention provides high real-time performance for live operations on a distribution network, can identify and locate online, and only takes 2 seconds to identify and locate a live operation scene on a distribution network, with excellent detection speed performance.

Description

Technical Field

[0001] The present invention relates to the field of live-line operations on distribution networks, and in particular to a method, device, and storage device for identifying and locating transmission lines for live-line operations on distribution networks based on machine learning. Background Art

[0002] Responding to the national call for intelligentization, robots are replacing humans. Using robots to complete high-risk live distribution network operations is crucial, significantly reducing the risks associated with manual labor. Correctly, quickly, and efficiently guiding robots is a crucial step in the process. Currently, live distribution network operations face environmental factors such as strong outdoor lighting, complex and unstable work environments, and a small and uneven sample size of transmission lines (main lines and lead lines). Consequently, machine learning and other methods have limited accuracy in identifying and positioning robots for live distribution network operations.

[0003] Existing methods for identifying and locating transmission lines during live distribution network operations fall into two main categories: depth camera-based methods and 3D point cloud-based methods. Depth camera-based methods use depth maps of the main and lead lines for stereo matching, then utilize deep learning and other methods to generate a 3D model of the transmission line. These methods ignore the complex working environment of live distribution network operations. Under strong outdoor sunlight, the depth camera's detection range is extremely short, and depth information is incomplete, making it difficult to accurately depict the transmission line. While they work better indoors, the methods are less adaptable. Methods based on three-dimensional point clouds are divided into traditional target detection methods and detection methods based on machine learning or deep learning. Traditional target detection methods mainly manually describe and extract three-dimensional point cloud features, which belong to the underlying global or local feature extraction. For target detection, feature description is the most important. The effects of different feature descriptors of objects are also different, and the robustness is poor. Methods based on deep learning use cropped data sets to train models. Existing methods mostly use neural networks to extract features during detection, and perform target recognition and detection on certain layers of output. However, these methods do not consider the importance of different features for recognition and detection tasks well, so that the feature information extracted by the neural network cannot well describe the unique features of the transmission line, and is relatively general. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a method, device and storage device for identifying and locating transmission lines for live operations on distribution networks. First, machine learning is suitable for the current situation where there are few data sets and lack of samples in the background of three-dimensional point cloud recognition of transmission lines. Classification and labeling are completed based on the data sets actually collected on site. Then, point cloud feature extraction is completed based on the characteristics of the transmission lines. On the basis of feature descriptors with strong adaptability and good effects, the shape and physical significance features of the transmission lines are added to complete the training of the machine learning model. On the basis of the classification results, combined with the correlation analysis algorithm, the transmission line identification and positioning is finally realized, and the recognition accuracy is improved. A method for identifying and locating transmission lines for live operations on distribution networks mainly includes:

[0005] S1: Obtaining the transmission line point cloud through LiDAR scanning;

[0006] S2: Preprocessing the point cloud, including multiple filtering and point cloud clustering and segmentation;

[0007] S3: Extract transmission line features from the preprocessed point cloud;

[0008] S4: Identify transmission lines and leads through the extracted features.

[0009] Furthermore, multiple three-dimensional lidars are used to complete the transmission line point cloud collection in different scenarios and perform data cropping.

[0010] Furthermore, a multiple filtering method is used to filter out points outside the actual operation scene as well as invalid points such as miscellaneous points and outliers. Specifically, it includes straight-through filtering, voxel filtering and outlier filtering. Straight-through filtering is used to simply filter out points outside the specified range. Voxel filtering is used to complete point cloud downsampling. The outlier detection method based on the statistical Gaussian filtering method is used to complete the removal of outliers. The neighborhood average distance x can be obtained from the Gaussian distribution. i The probability density function of is shown in the following formula (1):

[0011]

[0012] Where: x i represents the average distance of any nearby point, μ represents the mean of the neighborhood average distance, and σ represents the standard deviation;

[0013] The threshold is set by the calculated probability density. Points smaller than the threshold indicate scattered points or sparse points with few neighboring points and need to be removed.

[0014] Furthermore, the single point cloud after clustering segmentation is divided into main lines, lead lines, and other objects, where the main lines and lead lines are straight-line and curved-shaped wires, respectively.

[0015] Furthermore, in step S3, the descriptor based on the viewpoint feature histogram is used to extract the features of the transmission line. The specific process is as follows:

[0016] (1) Find the geometric center point of a single point cloud cluster after point cloud clustering and segmentation, expand the fast point feature histogram so that it can use the entire point cloud object for calculation and estimation. When calculating FPFH, the point pair between the center point of the object and all other points on the object surface is used as the calculation unit to complete the calculation of FPFH parameters α, θ, and φ, which are recorded as SPFH values. On this basis, the angle between the viewpoint direction and the estimated normal of each point is added as a histogram to calculate the viewpoint component, that is, the viewpoint direction variable is directly integrated into the relative normal angle calculation in the FPFH calculation;

[0017] The calculation formula of SPFH value is shown in formula (2):

[0018]

[0019] Among them, u, v, w are the origin p s The three components of n t is the neighbor point p t vector, d is the Euclidean distance between two points, α, θ, φ are n t The angle between the vectors of u, v, and w components;

[0020] Formula (2) is the calculation of a single query point. The k-neighborhood of each point is re-determined and the adjacent SPFH values ​​are used to calculate. The final histogram is called FPFH, as shown in formula (3):

[0021]

[0022] Among them, w k is the weight, indicating the query point P q and the neighboring point P in a given metric space k The distance between q represents the query point, P k Represents the nearest neighbor point, SPFH(P q ) represents the query point P q SPFH value, SPFH(P k ) represents point P k SPFH value, k is a positive integer greater than 1;

[0023] (2) According to step (1), the VFH feature descriptor is obtained. Then, according to the working environment and the actual position of the main line and the lead line targets, the features of the number of point clouds, the distance between the center points of the point clouds, and the three-dimensional size of the point cloud rectangular envelope box are added to the feature description to obtain the final transmission line feature.

[0024] Furthermore, the correlation analysis method based on the SVM classifier is used to complete the identification of the main line and the lead line. The identification process is as follows:

[0025] 1) Process the scanned point cloud and segment it into individual point clouds through point cloud preprocessing;

[0026] 2) Extract features from the i-th single target point cloud;

[0027] 3) Complete SVM classification and recognition based on composite features for the i-th single target point cloud: Input the i-th single target point cloud into the trained SVM classifier and obtain the output result. The result 0 indicates other objects, the result 1 indicates the main line, and the result 2 indicates the lead line.

[0028] The Pearson correlation coefficient method is used to calculate and analyze the correlation between the i-th single target point cloud and the template point cloud. The result value is -1 to 1, 0 to 1 represents positive correlation, and -1 to 0 represents negative correlation. The closer to 1, the greater the correlation, and the closer to -1, the less relevant. The template point cloud is a lead template point cloud close to the live operation site of the distribution network, and the point cloud quality is good.

[0029] 4) Complete steps 2) to 4) for all segmented single target point clouds;

[0030] 5) Mark the point clouds with the greatest correlation in the identified main line and lead line results respectively to complete the identification and positioning of the target point cloud.

[0031] A storage device stores instructions and data for implementing a method for identifying and locating transmission lines for live operations in a distribution network.

[0032] A transmission line identification and positioning device for live operations on a distribution network comprises: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a transmission line identification and positioning method for live operations on a distribution network.

[0033] The beneficial effects of the technical solution provided by the present invention are as follows: the present invention builds a transmission line classification and identification model for transmission lines in live distribution network operations to meet the real-time requirements of industrial sites; designs a target recognition algorithm based on machine learning and correlation analysis, and performs target recognition and detection by fusing the VFH features of transmission line objects with actual physical features, and based on the recognition results and correlation analysis of the SVM model trained with a small sample, which can effectively complete target recognition and improve the descriptiveness of the identified objects. In addition to extracting global features of the geometric domain such as VFH, the present invention also adds other feature descriptions related to the main line, lead line, operation scene, etc., and completes the main line and lead line target identification and positioning under multi-feature composite criteria. The feature description ability is stronger, more suitable for distribution network live operation, and the recognition performance is better; compared with the method based on traditional detection methods and deep learning, the method of the present invention not only introduces machine learning into the traditional method to improve the recognition performance, but also selects a more suitable machine learning algorithm for the small sample and few classification conditions of the distribution network live operation background, which can effectively enhance the recognition and positioning of target objects under such conditions; the method of the present invention provides high real-time performance for distribution network live operation, can be identified and positioned online, and it only takes 2 seconds to identify and locate a distribution network live operation scene, with good detection speed performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0035] Figure 1 The present invention is a flowchart of a method for identifying and locating transmission lines during live operations on a distribution network.

[0036] Figure 2 Schematic diagram of the original scanned point cloud image (with RGB information) in an embodiment of the present invention.

[0037] Figure 3 This is a result diagram after multiple filtering in an embodiment of the present invention.

[0038] Figure 4 Schematic diagram of FPFH calculation content in an embodiment of the present invention.

[0039] Figure 5 2 is a diagram showing the principle of calculating the viewpoint direction component in an embodiment of the present invention.

[0040] Figure 6 This is a block diagram of the target recognition algorithm implementation in an embodiment of the present invention.

[0041] Figure 7 This is a flowchart of point cloud clustering segmentation based on the region growing algorithm in an embodiment of the present invention.

[0042] Figure 8 2 is a target recognition result diagram in an embodiment of the present invention.

[0043] Figure 9 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0045] Embodiments of the present invention provide a method, device, and storage device for identifying and locating transmission lines during live operations on a distribution network.

[0046] Please refer to Figure 1 , Figure 1 This is a flow chart of a method for identifying and locating transmission lines during live operations on a distribution network according to an embodiment of the present invention, comprising:

[0047] S1: Obtaining the transmission line point cloud through LiDAR scanning;

[0048] S2: Preprocessing the point cloud, including multiple filtering and point cloud clustering and segmentation;

[0049] S3: Extract transmission line features from the preprocessed point cloud;

[0050] S4: Identify transmission lines and leads through the extracted features.

[0051] The above steps can be specifically divided into: data set collection and production, feature description and extraction, model training, and classification and recognition.

[0052] 1. Dataset Collection, Production and Preprocessing

[0053] This method uses multiple 3D LiDARs to collect data on power lines in different scenarios. A 3D LiDAR with a horizontal field of view of 70° and a vertical field of view of 77° is used to scan the scene. Each point cloud data point, used as input for the model, undergoes preprocessing using multiple filtering methods to remove noise and other artifacts, improving point cloud quality and subsequent calculation speed. The point cloud data acquired through scanning is quite large, containing not only the spatial coordinates of the scanned object but also geometric information reflecting the object's spatial dimensions and structure. The point cloud of the transmission line is collected in the laboratory, substation, and other outdoor scenes to create a data set. The obtained point cloud data is divided into a training set, a validation set, and a test set, and the data is cropped. The point cloud of each scene is subdivided into three categories: main line, lead line, and other objects. The feature values ​​of the three types of point clouds are labeled and classified to lay the foundation for the subsequent feature extraction of the transmission line and the training of the support vector machine (SVM) classifier. After training according to the general process of the support vector machine, the svm.xml model file is obtained. The single point cloud feature after segmentation is input to it to obtain the classification results of "0" (other objects), "1" (main line), and "2" (lead line). The present invention explores the preprocessing and clustering segmentation of point cloud data suitable for live distribution network operations, paving the way for subsequent target identification.

[0054] The point cloud data generated by LiDAR scanning is large, and its quality is affected by complex environmental conditions. Therefore, a series of preprocessing operations must be performed on the point cloud to filter out points outside the actual working scene, as well as invalid points such as clutter and outliers. These invalid points can be eliminated using multiple filtering methods, namely, through-filtering, voxel filtering, and outlier filtering.

[0055] Through-filtering is a very simple and effective filtering algorithm that uses parameter settings to remove points outside the scope of the operating scene. Through-filtering simply removes points outside the specified range and cannot filter out sparse, outlier noise points. These noise points can be further filtered using an outlier filtering algorithm. However, due to the large amount of data in the point cloud obtained by radar scanning and the fact that the point cloud quality is affected by the lidar itself and complex environmental factors, some non-sparse noise points may be present. Therefore, voxel filtering is required before outlier filtering to complete point cloud downsampling. Point cloud downsampling can reduce the number of points by tens of thousands to hundreds of thousands, depending on the number of original point clouds. It can also process non-sparse noise points, paving the way for outlier filtering to remove non-sparse noise points.

[0056] The present invention uses an outlier detection method based on the statistical Gaussian filtering method to remove outliers. The neighborhood average distance x can be obtained from the Gaussian distribution. i The probability density function of is shown in the following formula (1):

[0057]

[0058] Where: x i represents the average distance of any nearby point, μ represents the mean of the neighborhood average distance, and σ represents the standard deviation.

[0059] Outlier filtering assumes that all points in space conform to a certain distribution, and uses discordance check to identify outliers. The specific process is to perform statistical analysis on the neighborhood of each point, and remove points that do not conform to the Gaussian distribution as outliers. The present invention is based on the judgment standard that the average distance from a point to its neighboring points conforms to the Gaussian distribution, and filters out points whose average distance is outside the standard range as outliers. Assume that the parameter neighborhood k = 30, and take 2.4 times the standard deviation, that is, for each point in the point cloud, calculate its average distance from the 30 neighboring points. When the value is greater than 2.4 times the standard deviation, it is recorded as an outlier and needs to be filtered out. Experiments have shown that when the total number of neighboring points is too large and the standard deviation multiple is too small, excessive denoising will result; otherwise, the noise cannot be completely removed. Therefore, the values ​​of the two require multiple experiments and debugging to complete, to prevent incomplete or excessive denoising of the point cloud data. The original scene scan is as follows Figure 2 As shown, in order to improve the computer processing speed, the RGB information of the scanned point cloud is discarded, as shown in Figure 3 Shown are the results after multiple filtering.

[0060] Point cloud segmentation is a prerequisite for subsequent feature extraction, target classification, and recognition. In the subsequent point cloud feature extraction and target recognition based on the multi-feature conformance criterion of the SVM classifier, it is necessary to separate the target objects (transmission lines, leads) in the test scene from the debris, and then perform feature extraction to pave the way for identifying the transmission lines and leads. The segmentation algorithm based on Euclidean clustering cannot make good use of other features of the point cloud besides spatial information, and cannot complete the segmentation well. The present invention uses point cloud clustering based on the region growing algorithm to complete the segmentation. The basic idea of ​​the region growing algorithm is to group points with similar properties to form regions. First, for each region to be segmented, a seed is found as the starting point for growth. Then, points in the neighborhood around the seed with the same or similar properties as the seed (determined according to a pre-determined growth or similarity criterion, mostly normal vectors and curvature) are merged into the region where the seed is located. The new points continue to act as seeds and grow in all directions until no more points that meet the conditions can be included, and a region is grown.

[0061] The point cloud clustering segmentation flow chart based on region growing algorithm is as follows Figure 7 As shown, the specific process is as follows:

[0062] 1. Select the first seed point and put it into the seed point sequence;

[0063] 2. Search for the neighborhood point of the current seed point and calculate the angle between the normal of the neighborhood point and the normal of the current seed point;

[0064] 3. Determine whether the angle is less than the smoothing threshold. If so, go to step 4; if not, return to step 2.

[0065] 4. Add this field point to the current area;

[0066] 5. Calculate the curvature of each field point;

[0067] 6. Determine whether the curvature is less than the curvature threshold. If so, go to step 7. If not, do not put the seed point into the seed point sequence.

[0068] 7. Put the seed point into the seed point sequence;

[0069] 8. Determine whether the seed point sequence is empty. If so, the traversal ends and the region segmentation is completed. If not, delete the current seed point, select a new seed point, and then go to step 2.

[0070] A curvature-based region growing algorithm results in one or more cluster sets, each of which is considered to be part of the same smooth surface.

[0071] 2. Feature Description and Extraction

[0072] The feature descriptor of the method of the present invention uses a descriptor based on the viewpoint feature histogram (VFH) to extract the features of the transmission line, and combines the physical and spatial features of the transmission line itself with it to improve the descriptiveness of the features. The VFH feature is improved by adding the viewpoint vector to the fast point feature histogram (FPFH) feature. FPFH is a feature descriptor for describing three-dimensional point cloud features, which is used to represent features. It is derived from the FPFH descriptor and is a global feature descriptor used for cluster recognition and posture estimation. Because FPFH has good acquisition speed and recognition power, on its basis, the constructed features maintain the property of scaling invariance, and at the same time, different postures must be distinguished, and viewpoint variables are added during calculation. The features of the present invention are composed of several features of the object added on the basis of the VFH feature.

[0073] Figure 4 The advantage of FPFH is that it only calculates the query point P. q and the neighboring points ( Figure 4 The characteristic elements α, θ, and φ of the point feature histogram (the central gray line) are recorded as the simplified point feature histogram (SPFH) value, which reduces the interconnection between adjacent points and reduces the computational complexity. The SPFH value is calculated as shown in formula (2):

[0074]

[0075] This is the calculation of a single query point. The k-neighborhood of each point is re-determined. The final histogram calculated using the neighboring SPFH values ​​is called FPFH. The FPFH calculation formula is extended from SPFH, as shown in formula (3):

[0076]

[0077] where w k is the weight, which represents the distance between the query point and its neighboring points in the given metric space.

[0078] The VFH feature consists of the following two parts: 1. A component related to the viewpoint direction, such as Figure 5 As shown; 2. A component describing the surface shape containing an extended FPFH.

[0079] The VFH feature is calculated as follows:

[0080] The geometric center of the object is found and the FPFH is expanded to use the entire point cloud object for calculation and estimation. When calculating FPFH, the point pairs between the object center and all other points on the object surface are used as calculation units to complete the calculation of FPFH parameters α, θ, and φ. On this basis, the angle between the viewpoint direction and the estimated normal of each point is added as a histogram to calculate the viewpoint component. In other words, the viewpoint direction variable is directly integrated into the relative normal angle calculation in the FPFH calculation.

[0081] This approach preserves both the recognition capabilities of the FPFH descriptor and its scaling-invariant properties, further facilitating subsequent target identification. However, VFH feature descriptors alone are insufficient. In the complex environment of live distribution network operations, in addition to extracting global geometric features like VFH, other feature descriptions related to the main line, lead wires, and operational scenarios should be incorporated to identify and locate main and lead wire targets using a multi-feature composite criterion.

[0082] Based on the existing 308-dimensional VFH feature, the feature description is supplemented with features such as the number of point clouds, the distance between point cloud centers, and the three-dimensional dimensions of the point cloud rectangular envelope, based on the operating environment and the actual locations of the main and lead lines. This completes the multi-feature composite judgment criteria. The number of point clouds and the distance between point cloud centers reflect the specific range of the main and lead lines in actual operation, which is determined by the operational and process requirements. The three-dimensional dimensions of the envelope are determined by the morphological characteristics of the main and lead lines. The result of VFH feature extraction is a one-dimensional matrix of 308 values. These 308 numbers represent the VFH characteristics of an object. Based on an existing mathematical formula, object characteristics are added to this matrix. After the 308 numbers are supplemented with data such as the number of point clouds and distance, the composite feature is constructed.

[0083] The composite features selected in this embodiment are shown in Table 1:

[0084] Table 1 Multi-feature description

[0085]

[0086] After completing feature extraction of 3D point cloud data, the target can be identified and classified into the three categories 0, 1, and 2 mentioned above.

[0087] 3. Model Training

[0088] The method of the present invention uses support vector machine (SVM) classification based on composite features and correlation analysis between the point cloud to be identified and the template point cloud to identify the main line and the lead line. The optimized extracted features are input to complete the training of the SVM classifier. The SVM classifier is a well-known classifier in the field of machine learning. Because it can find the global optimal solution when there are few training samples, the SVM classifier is widely used in fields such as target classification and nonlinear regression.

[0089] This paper uses a Gaussian kernel function to design an SVM classifier to solve the three-class classification problem. After clustering and segmentation, the individual point clouds are divided into "main line," "lead line," and "other objects," where the main line and lead line are straight and curved wires, respectively. When the number of training sample sets for the two categories is small and mixed, the SVM strives for generalization while requiring a low cost of misclassification at the intersection of the two categories, which can lead to complex classification surfaces or over-learning, thereby reducing generalization ability. When the sample data on the classification surface contains noise, it may lead to incorrect classification.

[0090] The training of the three types of point clouds, namely "main line", "lead line" and "other objects", which have been cropped, are classified and stored. The features of each type of point cloud (the values ​​obtained from the feature description in the previous article, i.e., VFH values) are extracted, labeled and put into classifier training. The trained model is used to complete the classification and recognition of point cloud objects in the test set.

[0091] The trained model satisfies the formula: i (ωx i +b)=1

[0092]

[0093]

[0094] Among them, ω represents the normal vector of the hyperplane, b represents the displacement term of the hyperplane, and x i Represents the vector representation of the i-th sample, y i represents the classification label of the i-th sample, α i represents the coefficient in the constraint problem, x jrepresents the vector representation of the jth sample, and N represents the number of samples.

[0095] 4. Classification and Identification

[0096] The present invention first pre-processes the input point cloud to improve the point cloud quality and subsequent calculation speed, and then clusters and segments the processed point cloud to achieve feature extraction of a single point cloud cluster, which is input into the trained model for classification. Based on the classification results, combined with correlation analysis, the final classification and recognition results are obtained. The present invention uses an SVM classifier based on composite features and a correlation analysis based on template matching combined with an improved target recognition method (referred to as a correlation analysis method based on an SVM classifier) ​​to complete the recognition of main lines and leads. The improvements are: ① Integrate SVM and correlation analysis to complete comprehensive judgment; ② The feature description part combines existing feature descriptors with specific engineering object features to form a composite feature criterion. The method of the present invention increases the recognition rate by 10% compared to the separate SVM method and the separate correlation method. The block diagram of the target recognition algorithm is as follows Figure 6 shown.

[0097] The SVM classifier is described above, while correlation analysis solves the following two problems: determining the statistical association between two or more variables; and if a correlation exists, further analyzing the strength and direction of the association. The present invention mainly uses the Pearson correlation coefficient to perform correlation analysis to determine the similarity between the object to be tested and the template object.

[0098] For the same type of samples, let the given sample pairs be {(x1,y1),(x2,y2),…,(x n ,y n )}, then the sample Pearson correlation coefficient ρ X,Y The definition of is formula (5):

[0099]

[0100] Where n is the number of samples, which in this embodiment refers to the dimension of a single point cloud feature after segmentation; X i , Y i is the i-th dimension data of a single point cloud feature, Represents the data corresponding to the template point cloud.

[0101] A coefficient close to 1 indicates a positive correlation, and a coefficient close to -1 indicates a negative correlation. The larger the coefficient value, the better.

[0102] The recognition implementation process of the correlation analysis method based on the SVM classifier is as follows:

[0103] 1) Process the scanned point cloud and segment it into individual point clouds through point cloud preprocessing;

[0104] 2) Extract features from the i-th single target point cloud;

[0105] 3) Complete SVM classification and recognition based on composite features for the i-th single target point cloud;

[0106] 4) Complete the correlation calculation and analysis of the i-th single target point cloud and the template point cloud based on VFH features, and sort them;

[0107] 5) Complete the above steps 2) to 4) for all (n) segmented single target point clouds;

[0108] 6) Mark the point clouds with the greatest correlation in the identified “main line” and “lead line” categories respectively to complete the identification and positioning of the target point cloud.

[0109] The number of test sets used was 50, and the experimental results showed that the recognition rate of transmission lines was 90%, and the recognition rate of lead wires was 94%. Figure 8 As shown, the black frame is the transmission line and lead.

[0110] Table 2 Comparative analysis of experimental results

[0111]

[0112] See Figure 9 , Figure 9 4 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a transmission line identification and positioning device 401 for live operation of a distribution network, a processor 402 and a storage device 403.

[0113] A transmission line identification and positioning device 401 for live operations on a distribution network: The transmission line identification and positioning device 401 for live operations on a distribution network implements the transmission line identification and positioning method for live operations on a distribution network.

[0114] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the transmission line identification and positioning method for live distribution network operations.

[0115] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the transmission line identification and positioning method for live operations in a distribution network.

[0116] The beneficial effects of the present invention are as follows: the present invention builds an identification and detection model for transmission lines in live distribution network operations to meet the real-time requirements of industrial sites; designs a target recognition algorithm based on machine learning and correlation analysis, which integrates the VFH features of transmission line objects with actual physical features, and performs target recognition and detection based on the recognition results and correlation analysis of the SVM model trained with a small sample, which can effectively complete target recognition and improve the descriptiveness of the identified objects. In addition to extracting global features of the geometric domain such as VFH, the present invention also adds other feature descriptions related to the main line, lead line, operation scene, etc., and completes the main line and lead line target identification and positioning under multi-feature composite criteria. The feature description ability is stronger, more suitable for distribution network live operation, and the recognition performance is better; compared with the method based on traditional detection methods and deep learning, the method of the present invention not only introduces machine learning into the traditional method to improve the recognition performance, but also selects a more suitable machine learning algorithm for the small sample and few classification conditions of the distribution network live operation background, which can effectively enhance the recognition and positioning of target objects under such conditions; the method of the present invention provides high real-time performance for distribution network live operation, can be identified and positioned online, and it only takes 2 seconds to identify and locate a distribution network live operation scene, with good detection speed performance.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying and locating transmission lines during live operations on a distribution network, characterized by: include: S1: Obtaining the transmission line point cloud through LiDAR scanning; S2: Preprocessing the point cloud, including multiple filtering and point cloud clustering and segmentation; S3: Extract transmission line features from the preprocessed point cloud; In this step, the descriptor based on the viewpoint feature histogram is used to extract the transmission line features. The specific process is as follows: (1) Find the geometric center point of a single point cloud cluster after point cloud clustering and segmentation, expand the fast point feature histogram so that it can use the entire point cloud object for calculation and estimation. When calculating FPFH, the point pair between the center point of the object and all other points on the object surface is used as the calculation unit to complete the calculation of FPFH parameters α, θ, and φ, which are recorded as SPFH values. On this basis, the angle between the viewpoint direction and the estimated normal of each point is added as a histogram to calculate the viewpoint component, that is, the viewpoint direction variable is directly integrated into the relative normal angle calculation in the FPFH calculation; The calculation formula of SPFH value is shown in formula (2): Among them, u, v, w are the origin p s The three components of n t is the neighbor point p t vector, d is the Euclidean distance between two points, α, θ, φ are n t The angle between the vectors of u, v, and w components; Formula (2) is the calculation of a single query point. The k-neighborhood of each point is re-determined and the adjacent SPFH values ​​are used to calculate. The final histogram is called FPFH, as shown in formula (3): Among them, w k is the weight, indicating the query point P q and the neighboring point P in a given metric space k The distance between q represents the query point, P k Represents the nearest neighbor point, SPFH(P q ) represents the query point P q SPFH value, SPFH(P k ) represents point P k SPFH value, k is a positive integer greater than 1; (2) According to step (1), the VFH feature descriptor is obtained, and then according to the working environment and the actual position of the main line and the lead line target, the features of the number of point clouds, the distance between the center points of the point clouds, and the three-dimensional size of the point cloud rectangular envelope are added to the feature description to obtain the final transmission line feature; S4: Identify transmission lines and leads through the extracted features.

2. A method for identifying and locating transmission lines during live operations on a distribution network according to claim 1, characterized in that: In step S1, the process of obtaining point cloud data is as follows: multiple three-dimensional laser radars are used to complete the transmission line point cloud collection in different scenes, and the data is cropped.

3. The method for identifying and locating a transmission line during live working on a distribution network according to claim 1, wherein: A multiple filtering method is used to filter out points outside the actual operation scene as well as invalid points such as miscellaneous points and outliers. Specifically, it includes straight-through filtering, voxel filtering and outlier filtering. Straight-through filtering is used to simply filter out points outside the specified range. Voxel filtering is used to complete point cloud downsampling. The outlier detection method based on the statistical Gaussian filtering method is used to complete the removal of outliers. The neighborhood average distance x is obtained from the Gaussian distribution. i The probability density function of is shown in the following formula (1): Where: x i represents the average distance of any nearby point, μ represents the mean of the neighborhood average distance, and σ represents the standard deviation; The threshold is set by the calculated probability density. Points smaller than the threshold indicate scattered points or sparse points with few neighboring points and need to be removed.

4. A method for identifying and locating transmission lines during live operations on a distribution network according to claim 2, characterized in that: After clustering and segmentation, the single point cloud is divided into main line, lead line, and other objects, where the main line and lead line are straight-line and curved-shaped wires respectively.

5. The method for identifying and locating transmission lines during live operations on a distribution network according to claim 1, wherein: The main line and the lead line are identified by the correlation analysis method based on the SVM classifier. The identification process is as follows: 1) Process the scanned point cloud and segment it into individual point clouds through point cloud preprocessing; 2) Extract features from the i-th single target point cloud; 3) Complete SVM classification and recognition based on composite features for the i-th single target point cloud: Input the i-th single target point cloud into the trained SVM classifier and obtain the output result. The result 0 indicates other objects, the result 1 indicates the main line, and the result 2 indicates the lead line. The Pearson correlation coefficient method is used to calculate and analyze the correlation between the i-th single target point cloud and the template point cloud. The result value is -1 to 1, 0 to 1 represents positive correlation, and -1 to 0 represents negative correlation. The closer to 1, the greater the correlation, and the closer to -1, the less relevant. The template point cloud is a lead template point cloud close to the live operation site of the distribution network, and the point cloud quality is good. 4) Complete steps 2) to 4) for all segmented single target point clouds; 5) Mark the point clouds with the greatest correlation in the identified main line and lead line results respectively to complete the identification and positioning of the target point cloud.

6. A storage device, characterized in that: The storage device stores instructions and data for implementing the method for identifying and locating transmission lines for live operations on a distribution network as described in any one of claims 1 to 5.

7. A transmission line identification and positioning device for live operations in a distribution network, characterized by: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the transmission line identification and positioning method for live distribution network operations as described in any one of claims 1 to 5.

Images