Distribution line defect management method and system based on multi-source data application

Abstract

The invention relates to the technical field of power distribution operation and maintenance management, in particular to a power distribution line defect management method and system based on multi-source data application, and the method comprises the following steps: obtaining inspection data when an inspection unmanned aerial vehicle advances along a predetermined route; performing data augmentation processing on the image data, and preprocessing the laser radar point cloud data; key equipment in the image data after data augmentation processing is identified and positioned; a Point Conv algorithm and a Hough algorithm are adopted to carry out classification and identification on the preprocessed point cloud data; the method comprises the following steps: carrying out insulator and hardware defect detection by adopting an OfficientNet classification network, and detecting conductor defects by introducing multi-stage down-sampling and a defect detection technology based on a U-Net segmentation algorithm; and establishing an overhead power transmission and distribution line tree obstacle real-time measurement and calculation model, and carrying out online detection on tree obstacle hidden dangers. The invention provides a distribution line defect management method based on multi-source data application, and aims to solve problems existing in traditional distribution line detection.

Description

Technical Field

[0001] This application relates to the technical field of distribution operation and maintenance management, and particularly to a distribution line defect management method and system based on multi-source data application. Background Art

[0002] In distribution operation and maintenance management, the inspection work is an important link to ensure the safe and stable operation of the power grid. However, there are currently common problems such as a shortage of inspection manpower, low efficiency of manual inspection, low accuracy of defect identification, and difficulty in accurately quantifying tree obstacle hazards. The traditional inspection plan formulation method relies on manual experience and judgment, lacking scientificity and precision. Currently, there is a coexistence of over-inspection and under-inspection in distribution lines, and no effective solution has been proposed yet. Summary of the Invention

[0003] Embodiments of this application provide a distribution line defect management method and system based on multi-source data application, aiming to solve the problems existing in traditional distribution line detection, such as low efficiency of manual inspection, low accuracy of defect identification, and difficulty in accurately quantifying tree obstacle hazards.

[0004] To solve the above technical problems, in a first aspect, embodiments of this application provide a distribution line defect management method based on multi-source data application, including the following steps: First, when the inspection unmanned aerial vehicle travels along a predetermined route, inspection data is acquired; the inspection data includes image data and lidar point cloud data; then, data augmentation processing is performed on the acquired image data, and preprocessing is performed on the lidar point cloud data; next, a key equipment in the image data after data augmentation processing is identified and located by using an improved YOLOv5 model based on an attention mechanism; and the preprocessed point cloud data is classified and identified by using a Point Conv algorithm and a Hough algorithm; then, an EfficientNet classification network is used for insulator and fitting defect detection, and a classification network improved by FocalLoss is used to improve the classification effect; by introducing multi-level downsampling and a defect detection technology based on the U-Net segmentation algorithm to fuse low-order features and high-order features, wire defects are detected; finally, a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines is established to perform online detection of tree obstacle hazards.

[0005] In some exemplary embodiments, after performing online detection of tree obstacle hazards, it further includes: using a database based on MySQL database technology to store, retrieve, and analyze the inspection data, and visually display the results.

[0006] In some exemplary embodiments, when the inspection UAV travels along a predetermined route, inspection data is acquired, including: when the inspection UAV travels along a predetermined route, a visible light camera is used to capture images of the equipment along the line, and a lidar is used to scan the surrounding environment, obtaining image data and lidar point cloud data respectively.

[0007] In some exemplary embodiments, data augmentation processing is performed on the acquired image data, including: performing operations such as brightness and saturation transformation, cropping and mixing, mosaic processing, rotation, scaling, and flipping on the acquired image data in sequence to expand the image training set and improve the model training effect; preprocessing the lidar point cloud data, including: defining the system coordinate system and generating a digital terrain model for the lidar point cloud data in sequence.

[0008] In some exemplary embodiments, the process of defining the system coordinate system includes: using IMU / GNSS as the combined system navigation and positioning coordinate system to transform the position, attitude, and coordinates of the lidar point cloud into the geographic coordinate system; the process of generating the digital terrain model for the lidar point cloud includes: performing denoising and filtering processing on the lidar point cloud data to generate the digital terrain model DEM results for subsequent data classification and target recognition; in order to obtain the digital terrain model, the ground data and non-ground data are separated, and the point cloud data is automatically filtered and classified to obtain the surface point cloud.

[0009] In some exemplary embodiments, a modified YOLOv5 model based on the attention mechanism is used to identify and locate key equipment in the image data after data augmentation processing, including: inserting a channel attention mechanism into each residual module of YOLOv5 to obtain a modified YOLOv5 model based on the attention mechanism; using the modified YOLOv5 model based on the attention mechanism to identify and locate transformers, switches, and insulators in the image data after data augmentation processing.

[0010] In some exemplary embodiments, Point Conv is defined as shown in the following formula:

[0011]

[0012] In the formula, S(δx, δy, δz) is the inverse density estimate of the point (δx, δy, δz); W(δx, δy, δz) is the approximate weight function; (δx, δy, δz) are the three-dimensional coordinates; F is the feature of the point;

[0013] The continuous version of the three-dimensional convolution is shown in the following formula:

[0014] Conv(W, F) xyz = ∫∫∫ (δx，δy，δz) F(x + δx, y + δy, z + δz)dδxδyδz

[0015] Regarding the point cloud as non-uniform sampling samples in a continuous space, in each local area, (δx, δy, δz) is any possible position in the local area.

[0016] In some exemplary embodiments, the Hough algorithm is used to classify and identify the preprocessed point cloud data, including: through sample training, extracting the features of the power transmission and distribution lines and their channel environment targets; selecting training samples to construct a point cloud classifier to classify the three-dimensional point cloud data, and dividing the original point cloud into ground points, vegetation points, building points, conductors, and towers according to categories; and combining with unmanned aerial vehicle lidar to realize the identification of three-dimensional scene targets in the unmanned aerial vehicle inspection of power transmission and distribution lines.

[0017] In some exemplary embodiments, when using the EfficientNet classification network for insulator and fitting defect detection, the FocalLoss is used as the loss function to improve the classification effect;

[0018] The loss function L sum has the following expression:

[0019] L sum = α1 × L easy + α2 × L hard

[0020] Among them, FocalLoss defines weights according to the ease of sample discrimination, with a smaller weight of α1, and a larger weight of α2 is added to samples that are difficult to distinguish; in the formula, samples with a classification confidence close to 1 or close to 0 are called easily distinguishable samples, and the rest are called difficult-to-distinguish samples; for features such as long and thin conductors, U-Net extracts feature information from different levels through stacking multiple convolutional and pooling operations to achieve accurate positioning and segmentation of the target; for tree obstacle distance monitoring, by establishing a real-time measurement and calculation model for tree obstacles in overhead power transmission and distribution lines, using the real-time IMU data of the inspection unmanned aerial vehicle to correct the measured distance to obtain an accurate tree obstacle distance; and calculating the clearance distance between the tree obstacle and the conductor through the distance calculation formula;

[0021] The distance calculation formula is as follows:

[0022] H(α0, ω) = L(α0, ω) * cos(ω) * sin(α0)

[0023] H(α0, ω) ∈ S, H(α0, ω) < β

[0024] Wherein, S is the distance set of tree obstacle detection points; L is the set of tree obstacle point positions; the distance set H(α0, ω) of the detection point from the small tower is the clearance distance between the tree detection point and the wire; L(α0, ω) is the spatial distance between the UAV and the detected tree; α0 is the horizontal angle, and ω is the vertical angle; α0 ∈ Q, L0 = L(α0, ω)*sin(ω) + D is the distance from the tree obstacle to the small tower, and D is the distance from the UAV to the small tower.

[0025] In a second aspect, the embodiment of the present application also provides a distribution line defect management system based on multi-source data application, which performs distribution line defect management by using the distribution line defect management method based on multi-source data application described in the above embodiment, including a data acquisition module, a data preprocessing module, an equipment identification module, a defect detection module, and a tree obstacle analysis module connected in sequence; wherein, the data acquisition module is used to take images of the equipment along the line with a visible light camera and scan the surrounding environment with lidar when the inspection UAV travels along a predetermined route, respectively obtaining image data and lidar point cloud data; the data preprocessing module is used to perform data augmentation processing on the acquired image data and preprocess the lidar point cloud data; the equipment identification module is used to identify and locate key equipment in the image data after data augmentation processing by using an improved YOLOv5 model based on the attention mechanism; and classify and identify the preprocessed point cloud data by using the Point Conv algorithm and the Hough algorithm; the defect detection module is used to detect insulator and fitting defects by using an EfficientNet classification network, and use a classification network improved by FocalLoss to improve the classification effect; aiming at the characteristics of long and thin wire-like targets, by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, to fuse low-order features and high-order features and improve the detection ability of wire-like defects; the tree obstacle analysis module is used to establish a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines and perform online detection of tree obstacle hazards.

[0026] The technical solutions provided by the embodiments of the present application have at least the following advantages:

[0027] The embodiments of the present application provide a distribution line defect management method and system based on multi-source data application. The method includes the following steps: First, when the inspection UAV travels along a predetermined route, inspection data is acquired; the inspection data includes image data and lidar point cloud data; Then, data augmentation processing is performed on the acquired image data, and preprocessing is performed on the lidar point cloud data; Next, a key equipment in the image data after data augmentation processing is identified and located by using an improved YOLOv5 model based on the attention mechanism; and the Point Conv algorithm and the Hough algorithm are used to classify and identify the preprocessed point cloud data; Then, an EfficientNet classification network is used for insulator and fitting defect detection, and a classification network improved by FocalLoss is used to improve the classification effect; By introducing multi-level downsampling and a defect detection technology based on the U-Net segmentation algorithm, low-order features and high-order features are fused to detect conductor defects; Finally, a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines is established to perform online detection of tree obstacle hazards.

[0028] The distribution line defect management method and system based on multi-source data application provided by the present application combine visible light image technology and lidar technology to achieve intelligent identification of key equipment in distribution lines, automatic defect detection, accurate ranging of surrounding tree obstacles, and hazard assessment. The purpose of the present application is to improve the comprehensiveness, accuracy, and efficiency of distribution line defect management, reduce the risk of human misjudgment, achieve early discovery and timely handling of line safety hazards, thereby significantly improving the reliability and safety of the distribution system, while optimizing resource allocation and reducing operation and maintenance costs. Through this method of multi-source data fusion and intelligent analysis, the system can comprehensively and accurately manage various defects and hazards of distribution lines, greatly improve the detection efficiency and accuracy, reduce labor costs and safety risks, thereby ensuring the safe and stable operation of the power grid and adapting to the changing power grid environment and requirements. Description of the Drawings

[0029] One or more embodiments are exemplarily illustrated by the pictures in the corresponding drawings. These exemplary illustrations do not constitute a limitation on the embodiments. Unless otherwise stated, the figures in the drawings do not constitute a scale limitation.

[0030] Figure 1 It is a schematic flowchart of a distribution line defect management method based on multi-source data application provided by an embodiment of the present application.

[0031] Figure 2 It is a schematic structural diagram of a distribution line defect management system based on multi-source data application provided by an embodiment of the present application.

[0032] Figure 3Schematic diagram of the three-dimensional lidar coordinate system provided by an embodiment of the present application.

[0033] Figure 4 Schematic diagram of enhancing the YOLOv5 residual module using the channel attention mechanism provided by an embodiment of the present application.

[0034] Figure 5 Schematic diagram of enhancing the YOLOv5 backbone using the multi-head attention mechanism provided by an embodiment of the present application.

[0035] Figure 6 Diagram of the point cloud classification method provided by an embodiment of the present application.

[0036] Figure 7 Schematic diagram of the online calculation process for the tree obstacle distance based on lidar provided by an embodiment of the present application.

[0037] Figure 8 Real-time measurement and calculation model for tree obstacles on overhead power transmission and distribution lines provided by an embodiment of the present application. Detailed implementation manners

[0038] As can be seen from the background art, there are currently common problems such as a shortage of inspection manpower, low efficiency of manual inspection, low accuracy of defect identification, and difficulty in accurately quantifying tree obstacle hazards.

[0039] To solve the above technical problems, the embodiments of the present application provide a distribution line defect management method based on multi-source data application, including the following steps: First, when the inspection unmanned aerial vehicle travels along a predetermined route, inspection data is acquired; the inspection data includes image data and lidar point cloud data; then, data augmentation processing is performed on the acquired image data, and preprocessing is performed on the lidar point cloud data; next, a YOLOv5 model improved based on the attention mechanism is used to identify and locate key devices in the image data after data augmentation processing; and the Point Conv algorithm and the Hough algorithm are used to classify and identify the preprocessed point cloud data; then, an EfficientNet classification network is used for insulator and fitting defect detection, and a classification network improved using FocalLoss is used to improve the classification effect; by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, low-order features and high-order features are fused to detect conductor defects; finally, a real-time measurement and calculation model for tree obstacles on overhead power transmission and distribution lines is established to perform online detection of tree obstacle hazards. The embodiments of the present application provide a distribution line defect management method and system based on multi-source data application, aiming to solve the problems existing in traditional distribution line detection, such as low efficiency of manual inspection, low accuracy of defect identification, and difficulty in accurately quantifying tree obstacle hazards.

[0040] The following will elaborate on the embodiments of the present application in conjunction with the accompanying drawings. However, those of ordinary skill in the art can understand that in the embodiments of the present application, many technical details are presented to help readers better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present application can still be implemented.

[0041] Referring to Figure 1 , an embodiment of the present application provides a distribution line defect management method based on multi-source data application, including the following steps:

[0042] Step S1: When the inspection drone travels along a predetermined route, obtain inspection data; the inspection data includes image data and lidar point cloud data.

[0043] Step S2: Perform data augmentation processing on the collected image data and preprocess the lidar point cloud data.

[0044] Step S3: Use an improved YOLOv5 model based on the attention mechanism to identify and locate key devices in the image data after data augmentation processing; and use the Point Conv algorithm and the Hough algorithm to classify and identify the preprocessed point cloud data.

[0045] Step S4: Use the EfficientNet classification network for insulator and fitting defect detection, and use a classification network improved by FocalLoss to enhance the classification effect; by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, fuse low-order features and high-order features to detect wire defects.

[0046] Step S5: Establish a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines to detect tree obstacle hazards online.

[0047] In some embodiments, after step S5 for online detection of tree obstacle hazards, it further includes: Step S6: Use a database based on MySQL (relational database management system) database technology to store, retrieve, and analyze the inspection data, and visually display the results.

[0048] Referring to Figure 2, the embodiment of the present application also provides a distribution line defect management system based on multi-source data application. The distribution line defect management method described in the above embodiment is used for distribution line defect management, including a data acquisition module, a data preprocessing module, an equipment identification module, a defect detection module, and a tree obstacle analysis module connected in sequence; among them, the data acquisition module is used to use a visible light camera to capture images of equipment along the line and a lidar to scan the surrounding environment when the inspection drone travels along a predetermined route, respectively obtaining image data and lidar point cloud data; the data preprocessing module is used to perform data augmentation processing on the acquired image data and preprocess the lidar point cloud data; the equipment identification module is used to use an improved YOLOv5 model based on the attention mechanism to identify and locate key equipment in the image data after data augmentation processing; and use the Point Conv algorithm and the Hough algorithm to classify and identify the preprocessed point cloud data; the defect detection module is used to use the EfficientNet classification network to detect insulator and fitting defects, and use a classification network improved by FocalLoss to improve the classification effect; aiming at the characteristics of long and thin wire-like targets, by introducing multi-level downsampling and a defect detection technology based on the U-Net segmentation algorithm, to fuse low-order features and high-order features and improve the detection ability of wire-like defects; the tree obstacle analysis module is used to establish a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines and perform online detection of tree obstacle hazards.

[0049] The overall structure diagram of the distribution line defect management system based on multi-source data application provided by the present application is as Figure 2 shown. The system operation steps include: First, data acquisition is performed: the inspection drone travels along a predetermined route, and a visible light camera is used to capture images of equipment along the line, and a lidar is used to scan the surrounding environment. Then, data preprocessing is performed: data augmentation is performed on the acquired images, and the obtained lidar point cloud data is preprocessed. Next, equipment identification is performed: an improved YOLOv5 model based on the attention mechanism is used to identify and locate key equipment (such as transformers, switches, insulators, etc.) in the images; the Point Conv algorithm and the Hough algorithm are used to classify and identify the point cloud data. Then, defect detection and tree obstacle analysis are performed: the EfficientNet classification network is used to detect insulator and fitting defects, and a classification network improved by FocalLoss is used to further improve the classification effect. Aiming at the characteristics of long and thin wire-like targets, by introducing multi-level downsampling and a defect detection technology based on the U-Net segmentation algorithm, to fuse low-order features and high-order features and improve the detection ability of wire-like defects. The present application also proposes a method for online calculation of tree obstacle distance based on lidar, establishes a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines, and realizes online detection of tree obstacle hazards.

[0050] In some embodiments, when the inspection UAV travels along a predetermined route in step S1, inspection data is acquired, including: when the inspection UAV travels along the predetermined route, a visible light camera is used to capture images of the equipment along the line, and a lidar is used to scan the surrounding environment, respectively obtaining image data and lidar point cloud data.

[0051] In some embodiments, in step S2, data augmentation processing is performed on the acquired image data, including: successively performing operations such as brightness and saturation transformation, cropping and mixing, mosaic processing, rotation, scaling, and flipping transformation on the acquired image data to expand the image training set and improve the model training effect.

[0052] It should be noted that data augmentation is an important means to improve the robustness of the model in dealing with complex environments. By introducing noise into the training set images, the model is forced to learn noisy samples to reduce the overfitting phenomenon of the model on the original data set, assist in model training, and improve model performance.

[0053] Specifically, due to differences in inspection time, shooting angle, and exposure time, UAV aerial images usually have different brightness and saturation levels. By modifying the brightness and saturation of a single image, the shooting results under high-light (or upward shooting) at noon and low-light (or downward shooting) in the evening are effectively simulated, enhancing the network's adaptability. The Mosaic method can achieve a stable performance improvement for the object detection network only by randomly splicing 4 image samples. The specific process is to first randomly crop a part of each image, then randomly scale the cropped image, and finally splice all the images into a new sample. The Mosaic method can generate more complex and diverse backgrounds without adding data, and the cropping operation can also generate undetected and partially missing detection targets.

[0054] The CutMix augmentation method generates new training samples by pasting the randomly cropped sub-images of a certain category of images onto another image, effectively improving the performance of the classification network. The pasting operation in CutMix can simulate the occlusion effect of multiple objects on some key features in real images, improving the model's detection ability for overlapping and crowded objects. On the other hand, it can prevent the object detection network from over-associating the object to be detected with the background, thereby avoiding overfitting, and then improving the generalization ability of the network. For rotation, scaling, and flipping transformation, since the key components of the power transmission and distribution lines are often installed on the poles at different angles, and the flight attitude, shooting angle, and distance of the inspection UAV are diverse, even the same type of equipment will appear in different poses and sizes in the aerial images. The rotation, scaling, and flipping operations simulate the changes in shooting angle and distance, prompting the network to learn more key features of the equipment and improving the sample utilization rate.

[0055] In step S2, preprocessing is performed on the lidar point cloud data, including: successively performing system coordinate system definition and generation processing of the lidar point cloud digital terrain model on the lidar point cloud data.

[0056] In some embodiments, the process of system coordinate system definition includes: using IMU / GNSS as a combined system navigation and positioning coordinate system to transform the position, attitude, and coordinates of the lidar point cloud into the geographic coordinate system.

[0057] It should be noted that the Inertial Measurement Unit (IMU) is a sensor system that can measure and output the triaxial acceleration and angular velocity of an object. It is a core technology in many application fields such as modern navigation systems, unmanned aerial vehicles, autonomous driving vehicles, virtual reality, and augmented reality devices. The Global Navigation Satellite System (GNSS) refers to a series of space constellations orbiting the Earth that provide positioning, navigation, and timing services by transmitting radio signals to the Earth's surface and near-Earth space regions.

[0058] Specifically, for requirements such as multi-sensor fusion positioning solution, point cloud compensation, and tree obstacle detection, it is necessary to transform the position, attitude, and coordinates of the lidar point cloud into the geographic coordinate system to achieve the purpose of unifying all information into the global coordinate system. Using IMU / GNSS as a combined system navigation and positioning coordinate system, the defined coordinate system of its solution result is the geographic coordinate system, without conversion. The original lidar data consists of several components such as the vertical angle ω, azimuth and distance R corresponding to the point cloud. The X, Y, and Z coordinates of each lidar point in the radar body coordinate system are calculated by polar coordinates. The three-dimensional lidar coordinate system is as Figure 3 shown. Based on this method, during the point cloud compensation process, the position and attitude information of the original point cloud via IMU / GNSS combined navigation is directly transformed into the three-dimensional coordinates of the geographic coordinate system, and the transformation process is realized by a 6-degree-of-freedom pose, including a 3-degree-of-freedom attitude matrix and a 3-degree-of-freedom position offset.

[0059] In some embodiments, the process of generating the lidar point cloud digital terrain model includes: performing denoising and filtering processing on the lidar point cloud data to generate the digital terrain model DEM results for subsequent data classification and target recognition; in order to obtain the digital terrain model, separating the ground data and non-ground data, and automatically filtering and classifying the point cloud data to obtain the surface point cloud.

[0060] It should be noted that a Digital Elevation Model (DEM) is a digital simulation of the ground terrain (i.e., a digital representation of the terrain surface morphology) achieved through limited terrain elevation data. It is a physical ground model that represents ground elevation in the form of an ordered numerical array.

[0061] The process of generating a laser point cloud digital ground model includes: First, the inspection UAV collects laser point cloud data based on a three-dimensional lidar sensor, and then the data is denoised and filtered to generate ground digital elevation model (DEM) data for subsequent data classification and target recognition. Second, to obtain the digital ground model, it is necessary to separate the ground data and non-ground data, and automatically filter and classify the point cloud data to obtain the surface point cloud.

[0062] Then, the irregular triangulation network (TIN) filtering algorithm is used to filter the laser point cloud data. This algorithm updates the triangulation network by calculating the distance and angle of the laser points to the triangles they belong to. It has the advantage of directly processing the original point cloud data, avoiding the accuracy loss caused by the interpolation process, and thus has a certain universality for data with different point densities. The specific processing process is as follows: First, the laser point cloud data is segmented, and the point with the lowest height Z value in each block is selected as the seed point to construct an initial sparse triangulation network. Then, the angles and vertical distances between the points and vertices in the triangulation network are calculated and inserted into the triangulation network. Finally, the triangulation network is optimized to more accurately represent the terrain of the surveyed area.

[0063] When selecting the classified point cloud, it is necessary to consider the point cloud data and categories participating in the ground point classification, and set parameters such as the maximum building size and the maximum terrain slope during automatic filtering. To ensure the accuracy of the classification results, tools such as translation and rotation are also needed to check the classification results. If obvious classification errors are found, manual fine classification operations need to be carried out again to improve the algorithm accuracy.

[0064] The denoising process includes: By setting the distance between each point and its neighboring points, if it exceeds the average distance deviation range between points, it is determined as a noise point and deleted, eliminating the discrete noise points in the point cloud data.

[0065] Finally, the DEM and DSM are generated. The process includes: First, it is necessary to interpolate and regular grid the ground points to generate the digital elevation model DEM results. Based on the point cloud data, through automatic ground point detection and classification, an appropriate grid size and interpolation method are selected to obtain high-quality digital surface model DSM (Digital Surface Model, DSM) data. Further classify the DSM data to remove information such as ground buildings and vegetation to generate the DEM. Use GNSS positioning and DEM to calculate the distance of the UAV from the ground for ground part laser point cloud classification.

[0066] In some embodiments, in step S3, the improved YOLOv5 model based on the attention mechanism is used to identify and locate key devices in the image data after data augmentation processing, including: inserting a channel attention mechanism into each residual module of YOLOv5 to obtain the improved YOLOv5 model based on the attention mechanism; using the improved YOLOv5 model based on the attention mechanism to identify and locate transformers, switches, and insulators in the image data after data augmentation processing.

[0067] In step S3, an online identification technology for key devices of distribution lines based on the YOLOv5 algorithm is adopted. Considering that devices such as insulators, fittings, and conductors of distribution lines are arranged compactly, occlusion occurs frequently when taking pictures. In order to improve the detection ability of the target detection network in case of occlusion, etc., an attention mechanism is adopted to improve the feature extraction and recognition ability for non-obvious targets.

[0068] Insert a channel attention mechanism into each residual module of YOLOv5, as Figure 4 shown.

[0069] The channel attention mechanism can effectively improve the fitting ability of each residual module, thereby improving the network performance. The computational cost of the spatial attention mechanism is highly correlated with the size of the feature map. Existing research generally uses the spatial attention mechanism for small-size and high-dimensional feature maps.

[0070] Considering that there is already a spatial pyramid pooling structure (SPP) in the YOLOv5 backbone to enhance the receptive field, a multi-head attention mechanism MHAM is used to replace the SPP, as Figure 5 shown. MHAM can more effectively enhance the global feature modeling ability of the network and improve the detection ability for occluded and non-significant targets.

[0071] It should be noted that the channel attention mechanism is a technique that focuses on modeling the correlation between each feature channel and enhances or suppresses different channels according to different tasks. The core of this mechanism lies in assigning different weights to each channel of the input image in order to better extract and utilize the information in the image. The multi-head attention mechanism is a technique widely used in the fields of natural language processing and machine learning, aiming to improve the expressive ability and generalization ability of the model. It calculates the attention weights by introducing multiple independent attention heads and splices or weights and sums the results of these heads to obtain a richer representation. SPP is a technique for convolutional neural networks (CNNs). Its main role is to convert feature maps of different sizes into feature vectors of a fixed length for input into the fully connected layer. The essence of SPP is multi-layer maxpooling. By adaptively adjusting the size and stride of the pooling window, a feature vector of a fixed dimension can be output regardless of the size change of the input feature map.

[0072] In step S3, the Point Conv algorithm and the Hough algorithm are mainly used to classify and identify the point cloud data.

[0073] For discrete point cloud data, it is necessary to classify it into point cloud sets with different object attributes for subsequent screening and classification of target object recognition. The point cloud data is represented as a set of 3D points {p i | i = 1,..., n}, and each point contains features such as its position vector (X, Y, Z), color, and surface normal. Since the relative positions of the points in each local area are different, the point coordinates p = (x, y, z) in the point cloud do not necessarily lie on a fixed grid but adopt arbitrary continuous values.

[0074] To make the convolution compatible with the point set, a permutation-invariant convolution operation called the Point Conv algorithm is proposed. Point Conv approximates the weight function W(δx, δy, δz) from the three-dimensional coordinates (δx, δy, δz) and the inverse density S(δx, δy, δz) obtained through kernel density estimation and MIP-based non-linear transformation using a multi-layer perceptron (MLP). Since the weight function highly depends on the distribution of the input point cloud, the weights of the MLP in Point Conv are consistent across all points. To calculate the inverse density scale S(δx, δy, δz), first, the kernel density estimation (KDE) is used to evaluate the density at different positions of the point cloud offline, and one-dimensional non-linear transformation is adaptively performed through the MLP.

[0075] It should be noted that MLP is an artificial neural network model, which is a feedforward neural network composed of an input layer, a hidden layer, and an output layer. KDE is a non-parametric statistical method used to estimate the probability density function behind data samples. This method is not restricted by the form of the data distribution and can adapt to probability density functions of various shapes. Kernel density estimation makes inferences based on data samples and does not require too many assumptions about the overall distribution, so it can better reflect the actual situation of the data.

[0076] In some embodiments, Point Conv is defined as shown in the following formula:

[0077]

[0078] In the formula, S(δx, δy, δz) is the inverse density estimate of the point (δx, δy, δz) (the greater the density, the smaller the inverse density estimate); W(δx, δy, δz) is the approximate weight function; (δx, δy, δz) are three-dimensional coordinates; F is the feature of the point;

[0079] The continuous version of the three-dimensional convolution is shown in the following formula:

[0080] Conv(W, F) xyz = ∫∫∫ (δx，δy，δz) F(x + δx, y + δy, z + δz) dδxδyδz

[0081] Regarding the point cloud as a non-uniform sampling sample in the continuous space, in each local region, (δx, δy, δz) is any possible position in the local region.

[0082] Let C in , C out be the number of channels of the input feature and the output feature, k, C in , C out be the index of the k-th neighborhood, the C in -th channel of the input feature, and the C out -th channel of the output feature. The input point is the 3D local position Plocal ∈ RK×3, and the decentralized coordinates of the points within the 3D local region are calculated by subtracting the centroid of the local region and the coordinates of the feature. A 1×1 convolution is used to implement MLP. The output of the weight function is W. After convolution, the feature F in of the local region with K adjacent points is encoded into the output feature F out as shown in the following formula:

[0083]

[0084] To aggregate features in the entire point cloud set, the algorithm adopts a hierarchical structure that can combine detailed features of small regions into abstract features covering a larger spatial range. This hierarchical structure consists of multiple feature encoding modules, each module being similar to a layer in a convolutional neural network (CNN). The key layers of each feature encoding module include a sampling layer and a grouping layer.

[0085] Based on the above three-dimensional point cloud data classification algorithm, a key object point cloud classification method based on the Hough algorithm is as Figure 6 shown. In step S3, the Hough algorithm is used to classify and identify the preprocessed point cloud data, including: through sample training, extracting the features of transmission and distribution line and its channel environment objects; selecting training samples to construct a point cloud classifier to classify the three-dimensional point cloud data, dividing the original point cloud into ground points, vegetation points, building points, conductors, and towers according to categories; and combining with unmanned aerial vehicle (UAV) lidar to realize the recognition of three-dimensional scene objects in UAV patrol of transmission and distribution lines.

[0086] As Figure 6 shown, from the original three-dimensional point cloud data, training samples are manually selected. To ensure the accuracy of classification, different tower types and line types need to be considered during selection. At the same time, the ground points are separated using a filtering algorithm to further improve the accuracy of classification. In the recognition of conductors and ground wires, the Hough line detection algorithm is adopted because this algorithm can effectively overcome the discreteness of conductor and ground wire data. In the recognition of towers, erosion operation is used to eliminate the boundaries of conductors and ground wires, and dilation operation is used to obtain the connected regions in order to identify the aggregation point groups of tower objects. In this algorithm, when calculating the features by selecting the neighborhood points of each point, the following four methods of selecting neighborhood points are used: K-nearest neighbor points (K points around each point), spherical neighborhood (points within a sphere with radius R), cylindrical neighborhood (points within a cylinder with radius R), and grid neighborhood (points within each grid when the point cloud data is divided into grids).

[0087] During the clustering operation of this algorithm, the covariance matrix needs to be calculated. The calculation method is: constructing a covariance matrix for each point and its K-nearest neighbor points, and calculating the eigenvalues γ1, γ2, γ3 according to the covariance matrix, where γ1 >> γ2 >> γ3 >> 0. The features calculated based on the eigenvalues are shown in the following formula:

[0088] Sum = γ1 + γ2 + γ3

[0089]

[0090] In the formula, Sum is the sum of eigenvalues, Omnivariance is the total variance, Eigenentropy is the feature, Anisotropy is the anisotropy, Planarity is the flatness, and Linearity is the linearity.

[0091] The above content is the point cloud feature extraction algorithm. Combining with the manually marked point cloud classification labels, a point cloud classifier is constructed. This classifier is small in scale and can be deployed on the inspection UAV side and the server side at the same time to achieve automatic classification of the point cloud. The classification results include ground points, conductors, towers, and trees and buildings, etc.

[0092] In some embodiments, when using the EfficientNet classification network for insulator and fitting defect detection in step S4, the Focal Loss is used as the loss function to improve the classification effect.

[0093] Focal Loss defines weights according to the ease of sample discrimination. The smaller weight is α1, and a larger weight α2 is added to the samples that are difficult to distinguish; the loss function L sum The expression is shown as follows:

[0094] L sum = α1 × L easy + α2 × L hard In the formula, samples with classification confidence close to 1 or close to 0 are called easily distinguishable samples, and the rest are called difficult-to-distinguish samples. That is, samples with a sure grasp of confirming the attribute are called easily distinguishable samples, and samples without a sure grasp of confirming the attribute are called difficult-to-distinguish samples.

[0095] The most classic loss function in traditional classification loss is the standard cross-entropy. Taking binary classification as an example, it is shown as follows:

[0096]

[0097] In the formula, y = 1 or 0 is the label of the positive and negative samples, and p is the probability predicted by the network. To unify p and 1 - p, the p function is set as follows:

[0098]

[0099] When the loss function deals with the class imbalance problem, the redundancy of a certain class will dominate the loss function, causing the model to lose its effectiveness. To solve the class imbalance problem, a weight factor is added, that is, the balanced cross-entropy. The balanced cross-entropy is used as the weight as follows:

[0100] CE(p t ) = -α t log(p t )

[0101] When dealing with the situation of a large number of negative samples and a small number of positive samples, even if the weight of the negative samples is set very low, the loss function of the negative samples may still dominate because of the large number of negative samples.

[0102] Samples with a very high confidence level also incur losses in the standard cross-entropy. Moreover, in practice, negative samples with a very high confidence level often account for the vast majority of the total samples. If this part of the loss is removed or reduced, the efficiency of the loss function will be higher. Therefore, to reduce the proportion of the loss of samples with a very high confidence level in the total loss, a weight factor can be added before the standard cross-entropy to improve the performance during training, as shown in the following formula:

[0103] FL(p t )=-(1-p t ) γ logp t

[0104] In step S4, for features such as long and thin wires, by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, low-level features and high-level features are fused to improve the detection ability of wire-like defects.

[0105] It should be noted that U-Net is an image segmentation network based on a fully convolutional network, and its network structure is designed with a symmetric U-shaped structure, which can effectively segment targets of different sizes and scales.

[0106] For features such as long and thin wires, U-Net can extract feature information from different levels by stacking multiple convolutional and pooling operations to achieve accurate positioning and segmentation of the target. Combining the characteristics of wire-like defects, a semantic segmentation network can be used as a refined post-processing method for the target detection network. The wire regions extracted by the YOLOv5 algorithm are finely segmented by semantic segmentation for abnormal protrusions, bifurcations, and bends of the wires, etc., to improve the diagnostic ability of wire-like defects.

[0107] To meet the requirements of lightweight and high precision, the U-Net algorithm with complete symmetry is selected as the semantic segmentation algorithm for wire-like target objects. The target is segmented through multi-level downsampling to identify the differences between the target and the surrounding environment, thereby improving the segmentation judgment: high-resolution features are extracted from the corresponding positions of the feature extraction part for feature fusion, which can provide more refined features for the segmented target. For example, gradient information helps to divide the target boundary, and its loss function is cross-entropy.

[0108] U-Net uses a structure similar to an encoder-decoder. In the encoding stage, in its computational process, the input image first passes through several convolutional modules to be encoded into multi-scale feature maps. Each convolutional module can be a direct convolution stack similar to the VGG network or multiple residual modules similar to ResNet. Each convolutional module contains a convolutional layer or pooling layer with a stride of 2 to downsample the input image or features, thereby weakening the high-frequency features of the input features or image and transforming the low-frequency global features into a higher-dimensional semantic space. Therefore, the multi-scale feature maps can fully represent the local details and global semantic features of the input image. In the decoding stage, U-Net starts from the highest-level features, uses convolutional modules to further transform the feature maps, and fuses them with the feature maps of the lower layer using downsampling operations. The downsampling operation can be bilinear interpolation or transposed convolution, and their actual effects are similar. In this chapter, the interpolation method is used to reduce the computational complexity. This chapter also introduces a fully connected structure similar to DenseNet to realize the fusion of multi-scale information contained in each layer of feature maps, further strengthening the model's ability to extract wires of various scales and different sizes.

[0109] To address the problem of easy breakage and discontinuity in the wire segmentation map, an elastic transformation augmentation method is used to improve the network segmentation performance. The augmentation process of elastic transformation is as follows:

[0110] (1) Generate a random displacement vector for each pixel position in the image to form a displacement field, where the length of the random vector is used to control the degree of the final deformation.

[0111] (2) Apply Gaussian filtering to the displacement field to ensure the smoothness of the displacement field

[0112] (3) Apply the filtered displacement field to the original image to offset its pixel positions

[0113] (4) Resample and grid the offset image to generate the augmented image. The result of elastic transformation augmentation is to make the original straight wire edges zigzag, increasing the network generalization ability, thereby improving the wire segmentation accuracy at the pixel level of the network.

[0114] In step S5 for tree obstacle distance monitoring, by establishing a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines, the measured distance is corrected using the real-time IMU data of the inspection UAV to obtain an accurate tree obstacle distance; and the clearance distance between the tree obstacle and the wire is calculated through the distance calculation formula.

[0115] The online calculation method for tree obstacle distance based on lidar adopted in step S5 has a process as Figure 7 shown.

[0116] As Figure 7As shown in the figure, the measured distance is corrected using the real-time IMU data of the inspection UAV to obtain the accurate tree obstacle distance. This method is similar to the method for detecting wires and requires corresponding pose compensation.

[0117] The inspection UAV detects real-time data through lidar and calculates the clearance distance between the tree obstacle hazard and the wire. When the system detects that the clearance distance between the obstacle and the wire exceeds the set threshold, it will automatically save the obstacle information and issue a real-time warning at the ground monitoring terminal to notify the inspection personnel.

[0118] Based on the above method, a real-time measurement model for tree obstacles in overhead transmission and distribution lines is proposed. During flight, the inspection UAV automatically measures the shortest distance from all trees in the span between two towers to the wire through lidar. The side view of the ideal measurement by 3D lidar is as follows. The horizontal azimuth angle of the 3D lidar is defined as 0 degrees, and the position of the UAV in space when the plane is in the same plane as the wire is the ideal ranging point.

[0119] Set the safe horizontal distance between the UAV and the wire as R1 (the value varies with different voltage levels), and the dangerous distance of the tree obstacle β < R2. Define the time period T for each scan cycle of the lidar data, calculate the tree obstacle distance once, L(α0, ω) is the spatial distance between the UAV and the detected tree, where α0 is the horizontal angle and ω is the vertical angle. Define S as the set of distances of the tree obstacle detection points, L as the set of positions of the tree obstacle points, that is, the set of distances of the detection points from the small tower. H(α0, ω) is the clearance distance from the tree detection point to the wire. The distance calculation formula is as follows:

[0120] H(α0, ω) = L(α0, ω) * cos(ω) * sin(α0)

[0121] H(α0, ω) ∈ S, H(α0, ω) < β, when

[0122] In the formula, S is the set of distances of the tree obstacle detection points; L is the set of positions of the tree obstacle points; the set of distances of the detection points from the small tower. H(α0, ω) is the clearance distance from the tree detection point to the wire; L(α0, ω) is the spatial distance between the UAV and the detected tree; α0 is the horizontal angle and ω is the vertical angle; α0 ∈ Q, L0 = L(α0, ω) * sin(ω) + D is the distance from the tree obstacle to the small tower, and D is the distance between the UAV and the small tower. There may be overlapping tree obstacle points in the scanned area during the radar scan interval T. To avoid duplicate recording of tree obstacles, according to experience, set the tree obstacles within ±2m of L as one point.

[0123] The lidar scans the line tree area, and all the points in this area are used as the point set J of the beam points on the two-dimensional plane coordinate system with the inspection drone as the coordinate origin. The scanned data points in other angular areas are used as the set K of tree obstacle points on the two-dimensional plane coordinate system with the inspection drone as the coordinate origin. Calculate the point set J and the point set K to calculate the shortest path between the two point sets.

[0124] Compared with the prior art, the advantages of the power distribution line defect management method and system based on multi-source data application provided by this application are as follows:

[0125] (1) Multi-source data fusion improves detection accuracy: This system combines visible light images and lidar technology to achieve a full range of detection of power distribution lines. Visible light images are used for equipment identification and defect detection, while lidar technology accurately measures the distance to tree obstacles. This method of multi-source data fusion significantly improves the accuracy and comprehensiveness of detection. Compared with the traditional method using a single data source, it can better identify and quantify various potential hazards.

[0126] (2) Intelligent algorithms improve efficiency: The system adopts a variety of advanced intelligent algorithms, such as the improved YOLOv5 model for equipment identification, EfficientNet and FocalLoss for defect detection, and the segmentation algorithm based on U-Net for conductor defect detection. These algorithms greatly improve the detection efficiency and accuracy, can quickly process a large amount of data, and reduce manual intervention.

[0127] (3) Real-time measurement and warning function: The online calculation method of tree obstacle distance based on lidar proposed by the system can measure the clearance distance between the tree obstacle and the conductor in real time, and automatically issue a warning within the dangerous distance. This function greatly improves the timeliness of potential hazard discovery and helps prevent potential safety accidents.

[0128] (4) Data augmentation improves model robustness: The system adopts a variety of data augmentation methods, such as brightness and saturation changes, mosaic processing, cropping and mixing, etc., which effectively improve the adaptability of the model to complex environments. This not only improves the detection accuracy but also enhances the stability of the system under different weather and lighting conditions.

[0129] (5) Visualization and data management: The system adopts MySQL database technology to realize the storage, retrieval and analysis of inspection data. This data management method helps to track the line condition for a long time, supports trend analysis and predictive maintenance, thereby optimizing resource allocation and reducing operating costs.

[0130] In summary, the distribution line defect management method and system based on multi-source data application provided by this application, on the one hand, combines the multi-source data fusion method of visible light images and lidar technology. Through steps such as data acquisition, preprocessing, and coordinate system conversion, it can simultaneously obtain the visual information of equipment and the three-dimensional spatial information of the surrounding environment, providing a basis for comprehensive and accurate defect detection and hidden danger assessment. On the other hand, this application adopts a series of intelligent recognition and detection algorithms, including the improved YOLOv5 model, EfficientNet classification network, loss function improved by FocalLoss, segmentation algorithm based on U-Net, etc. The specific improvement methods of these algorithms, such as the introduction of attention mechanism, application of multi-level downsampling, etc. In addition, this application also proposes an online calculation method for the distance between tree obstacles and power lines based on lidar. This method can measure the clearance distance between tree obstacles and power lines in real time and issue automatic warnings. The specific implementation process of this method includes IMU data correction, distance calculation formula, etc. In addition, this application also adopts a variety of data augmentation methods, such as brightness and saturation changes, mosaic processing, cropping and mixing, rotation and flipping changes, etc., which are the key to improving the model performance. Finally, the distribution line defect management system based on multi-source data application provided by this application, the overall system architecture includes the design and integration methods of modules such as data acquisition, preprocessing, equipment recognition, defect detection, tree obstacle analysis, and result visualization. This architecture can realize the full-process automatic processing from data acquisition to result output.

[0131] With the above technical solutions, the embodiments of this application provide a distribution line defect management method and system based on multi-source data application. The method includes the following steps: First, when the inspection drone travels along a predetermined route, inspection data is acquired; the inspection data includes image data and lidar point cloud data; then, the acquired image data is subjected to data augmentation processing, and the lidar point cloud data is preprocessed; next, the improved YOLOv5 model based on the attention mechanism is used to identify and locate key equipment in the data-augmented image data; and the Point Conv algorithm and Hough algorithm are used to classify and identify the preprocessed point cloud data; then, the EfficientNet classification network is used for insulator and fitting defect detection, and the classification network improved by FocalLoss is used to improve the classification effect; by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm to fuse low-order features and high-order features, wire-type defects are detected; finally, a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines is established to perform online detection of tree obstacle hidden dangers.

[0132] The distribution line defect management method and system based on multi-source data application provided by this application combine visible light image technology and lidar technology to achieve intelligent identification of key equipment on the distribution line, automatic defect detection, accurate ranging of surrounding tree obstacles, and hidden danger assessment. The purpose of this application is to improve the comprehensiveness, accuracy, and efficiency of distribution line defect management, reduce the risk of human misjudgment, achieve early detection and timely handling of line safety hazards, thereby significantly enhancing the reliability and safety of the distribution system, while optimizing resource allocation and reducing operation and maintenance costs. Through this method of multi-source data fusion and intelligent analysis, this system can comprehensively and accurately manage various defects and hidden dangers on the distribution line, greatly improve the detection efficiency and accuracy, reduce labor costs and safety risks, thereby ensuring the safe and stable operation of the power grid and adapting to the changing power grid environment and requirements.

[0133] Those of ordinary skill in the art can understand that the above-described embodiments are specific examples for implementing this application, and in actual applications, various changes can be made in form and details without departing from the spirit and scope of this application. Any person skilled in the art can make their own changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be determined by the scope defined in the claims.

Claims

1. A distribution line defect management method based on multi-source data application, characterized in that, Including the following steps: When the inspection UAV travels along a predetermined route, obtain inspection data; The inspection data includes image data and lidar point cloud data; Perform data augmentation processing on the collected image data and preprocess the lidar point cloud data; Use the improved YOLOv5 model based on the attention mechanism to identify and locate key devices in the image data after data augmentation processing; and use the Point Conv algorithm and the Hough algorithm to classify and identify the preprocessed point cloud data; Use the EfficientNet classification network for insulator and fitting defect detection, and use the classification network improved by FocalLoss to improve the classification effect; by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, to fuse low-order features and high-order features for conductor defect detection; Establish a real-time measurement and calculation model for tree obstacles on overhead transmission and distribution lines to perform online detection of tree obstacle hazards.

2. The method for managing distribution line defects based on multi-source data applications according to claim 1, characterized in that After performing online detection of tree obstacle hazards, it also includes: Use a database based on MySQL database technology to store, retrieve, and analyze the inspection data, and visually display the results.

3. The method for managing distribution line defects based on multi-source data applications according to claim 1, characterized in that When the inspection UAV travels along a predetermined route, obtain inspection data, including: When the inspection UAV travels along a predetermined route, use a visible light camera to capture images of the devices along the line, and use lidar to scan the surrounding environment to obtain image data and lidar point cloud data respectively.

4. The method for managing distribution line defects based on multi-source data applications according to claim 1, wherein, Perform data augmentation processing on the collected image data, including: Perform operations such as brightness and saturation transformation, cropping and mixing, mosaic processing, rotation, scaling, and flipping on the collected image data in sequence to expand the image training set and improve the model training effect; Preprocess the lidar point cloud data, including: Perform system coordinate system definition and lidar point cloud digital terrain model generation processing on the lidar point cloud data in sequence.

5. The method for managing distribution line defects based on multi-source data applications according to claim 1, wherein The process of the system coordinate system definition includes: Use IMU / GNSS as the combined system navigation and positioning coordinate system to transform the position, attitude, and coordinates of the lidar point cloud into the geographic coordinate system; The process of the lidar point cloud digital terrain model generation processing includes: Perform denoising and filtering processing on the lidar point cloud data to generate the digital terrain model DEM results for subsequent data classification and target recognition; In order to obtain the digital terrain model, separate the ground data and non-ground data, and automatically filter and classify the point cloud data to obtain the ground surface point cloud.

6. The method for managing distribution line defects based on multi-source data application according to claim 1, characterized in that, Use the improved YOLOv5 model based on the attention mechanism to identify and locate key devices in the image data after data augmentation processing, including: Insert a channel attention mechanism into each residual module of YOLOv5 to obtain the improved YOLOv5 model based on the attention mechanism; Use the improved YOLOv5 model based on the attention mechanism to identify and locate transformers, switches, and insulators in the image data after data augmentation processing.

7. The method for managing distribution line defects based on multi-source data applications according to claim 1, characterized in that, Point Conv is defined as shown in the following formula: Where S(δx, δy, δz) is the inverse density estimate of the point (δx, δy, δz); W(δx, δy, δz) is the approximate weight function; (δx, δy, δz) are three-dimensional coordinates; F is the feature of the point; The continuous version of the three-dimensional convolution is shown as follows: Conv(W, F) xyz = ∫∫∫ (δx，δy，δz) F(x + δx, y + δy, z + δz) dδxδyδz Regarding the point cloud as non-uniform sampling samples in a continuous space, in each local region, (δx, δy, δz) is any possible position in the local region.

8. The distribution line defect management method based on multi-source data application according to claim 1, characterized in that, The Hough algorithm is used to classify and identify the preprocessed point cloud data, including: Through sample training, extract the features of the power transmission and distribution lines and their channel environment targets; Select training samples to construct a point cloud classifier to classify the three-dimensional point cloud data, and divide the original point cloud into ground points, vegetation points, building points, conductors, and towers according to categories; and combine with the UAV lidar to realize the recognition of three-dimensional scene targets in the UAV inspection of power transmission and distribution lines.

9. The method for managing distribution line defects based on multi-source data applications according to claim 1, characterized in that When using the EfficientNet classification network for insulator and fitting defect detection, use FocalLoss as the loss function to improve the classification effect; Loss function L sum The expression of which is shown as follows: L sum = α1 × L easy + α2 × L hard Among them, FocalLoss defines weights according to the ease of sample discrimination, with a smaller weight of α1, and a larger weight of α2 is added to samples that are difficult to distinguish; in the formula, samples with a classification confidence close to 1 or close to 0 are called easily distinguishable samples, and the rest are called difficult-to-distinguish samples; For features such as long and thin conductors, U-Net extracts feature information from different levels by stacking multiple layers of convolution and pooling operations to achieve accurate positioning and segmentation of the target; For tree obstacle distance monitoring, by establishing a real-time measurement and calculation model for tree obstacles in overhead power transmission and distribution lines, use the real-time IMU data of the inspection UAV to correct the measured distance to obtain an accurate tree obstacle distance; and calculate the clearance distance between the tree obstacle and the conductor through the distance calculation formula; The said distance calculation formula is shown as follows: H(α0, ω) = L(α0, ω) * cos(ω) * sin(α0) H(α0, ω) ∈ S, H(α0, ω) < β In the formula, S is the set of distances of tree obstacle detection points; L is the set of positions of tree obstacle points; the set of distances H(α0, ω) of the detection points from the small tower number is the clearance distance between the tree detection points and the conductor; L(α0, ω) is the spatial distance between the UAV and the detected tree; α0 is the horizontal angle, ω is the vertical angle; α0 ∈ Q, L0 = L(α0, ω) * sin(ω) + D is the distance of the tree obstacle from the small tower number, and D is the distance of the UAV from the small tower.

10. A distribution line defect management system based on multi-source data application performs distribution line defect management by using the distribution line defect management method based on multi-source data application described in any one of claims 1 to 9, characterized in that, Including a data acquisition module, a data preprocessing module, a device identification module, a defect detection module, and a tree obstacle analysis module connected in sequence; The said data acquisition module is used to, when the inspection UAV travels along a predetermined route, use a visible light camera to take images of the equipment along the line and use lidar to scan the surrounding environment to obtain image data and lidar point cloud data respectively; The said data preprocessing module is used to perform data augmentation processing on the acquired image data and preprocess the lidar point cloud data; The device recognition module is used to identify and locate key devices in the image data after data augmentation processing by using the improved YOLOv5 model based on the attention mechanism; and classify and identify the preprocessed point cloud data by using the Point Conv algorithm and the Hough algorithm; The defect detection module is used to detect insulator and fitting defects by using the EfficientNet classification network, and use the classification network improved by FocalLoss to improve the classification effect; aiming at the characteristics of long and thin wire-like objects, by introducing multi-level downsampling and defect detection technology based on the U-Net segmentation algorithm, to fuse low-order features and high-order features and improve the detection ability of wire defects; The tree obstacle analysis module is used to establish a real-time measurement and calculation model for tree obstacles in overhead transmission and distribution lines to detect tree obstacle hazards online.

Images