A low-cost autonomous inspection method and device for overhead power lines by unmanned aerial vehicles

Abstract

The present invention provides a low-cost method and device for autonomous inspection of power overhead lines by unmanned aerial vehicles. The method includes: obtaining inspection images of the poles and towers to be inspected; identifying the types of poles and towers in the inspection images and extracting the three-dimensional coordinate information of the poles and towers; establishing a model library for various types of poles and towers and setting the UAV shooting points and flight paths for each type of pole and tower; based on the identified pole and tower types, the three-dimensional coordinate information of the poles and towers, and the model library, generating a parametric three-dimensional model of the identified poles and towers in a three-dimensional space map and obtaining the UAV shooting points and flight paths corresponding to the identified pole and tower types; analyzing whether there are channel tree hazards in each inspection image and optimizing the UAV shooting points and flight paths of the poles and towers with channel tree hazards; generating a final UAV autonomous inspection track file. The present invention can achieve one-time investment and construction and subsequent repeated application, greatly reducing the application cost of UAV autonomous inspection and realizing low-cost UAV autonomous inspection of distribution networks.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous inspection of unmanned aerial vehicles (UAVs), and particularly to a low-cost method and device for autonomous inspection of UAVs for overhead power lines. Background Art

[0002] In the field of power line inspection, the traditional manual inspection mode, due to its high dependence on human resources, long time consumption and limited efficiency, has been difficult to meet the requirements of efficient operation and maintenance of overhead power lines. In recent years, the leap-forward progress of UAV technology has brought a revolutionary change to this field. As an innovative alternative inspection method, autonomous inspection of UAVs is gradually emerging and showing strong application potential. The advantages of autonomous inspection of UAVs are as follows: First, using advanced three-dimensional laser point cloud camera technology to perform high-precision three-dimensional modeling of complex distribution network lines; subsequently, generating a flight path planning file based on these accurate three-dimensional coordinate information to guide the UAV to autonomously execute the inspection task according to the preset path, greatly improving the accuracy and efficiency of the inspection.

[0003] However, although autonomous inspection of UAVs shows great advantages in technology, its large-scale application faces severe challenges. Specifically, high-end three-dimensional laser point cloud camera equipment is expensive, often costing tens of thousands of yuan per unit, which constitutes an important threshold for initial investment; in addition, the cost of laser point cloud modeling service per kilometer is as high as one thousand yuan. This cost accumulation effect undoubtedly increases the economic burden for the inspection of large-scale overhead power lines. Therefore, the current high inspection cost has become a key factor restricting the in-depth application of UAV technology in the field of overhead power line inspection, thereby affecting the improvement of the overall inspection quality and efficiency.

[0004] At the same time, under complex mountain conditions, overhead power lines usually pass through a large number of tree areas. The growth of trees is uncontrollable, and after a long time, the trees may invade the power corridor; tree occlusion will bring great potential safety hazards to the autonomous inspection of UAVs, and there is a risk of UAV collision.

[0005] To sum up, although the technology of autonomous inspection of UAVs has broad prospects, to achieve large-scale application in the inspection of overhead power lines, it is necessary to continuously explore and optimize in terms of cost control and safety to overcome the existing bottlenecks and promote the industry to develop towards a more intelligent and economical direction. Therefore, there is an urgent need for a low-cost method and device for autonomous inspection of UAVs for overhead power lines to solve the problems existing in the prior art. Summary of the Invention

[0006] The object of the present invention is to provide a low-cost method for autonomous inspection of UAVs for overhead power lines, aiming to solve the problems of high cost and collision risk existing in the existing autonomous inspection methods of UAVs. The specific technical solutions are as follows:

[0007] A low-cost autonomous inspection method for unmanned aerial vehicles (UAVs) on overhead power lines, including:

[0008] Obtain inspection images of the poles and towers to be inspected;

[0009] Use a target detection model to identify the types of poles and towers in the inspection images and extract the three-dimensional coordinate information of the poles and towers;

[0010] Establish a model library for various types of poles and towers, and set the UAV shooting points and flight paths for each type of pole and tower;

[0011] Based on the identified pole and tower types, the three-dimensional coordinate information of the poles and towers, and the model library, generate a parametric three-dimensional model of the identified poles and towers in a three-dimensional space map, and obtain the UAV shooting points and flight paths corresponding to the identified pole and tower types;

[0012] Analyze whether there are channel tree hazards in each inspection image. If there are channel tree hazards in the inspection image, optimize the UAV shooting points and flight paths of the poles and towers in the inspection image;

[0013] Generate the final UAV autonomous inspection track file based on the optimized UAV shooting points and flight paths and the UAV shooting points and flight paths that do not need to be optimized .

[0014] Preferably, the specific method for analyzing whether there are channel tree hazards in the inspection image is:

[0015] According to formula (2), convert the th inspection image from the RGB color channel to the HSV brightness space:

[0016] (2),

[0017] wherein, in formula (2) HSV represents the preprocessing operation of converting the RGB color channel to the HSV brightness space, represents the hue component, represents the saturation component, represents the value component, , represents the total number of inspection images;

[0018] According to formula (3) and formula (4), count the number of green plant tree pixel points in the inspection image ; if is greater than or equal to the set threshold, it is considered that there are green plant trees in the inspection image, that is, there are channel tree hazards; if ​If it is less than the set threshold, it is considered that there are no green plants or trees in the inspection image, that is, there are no channel tree problems;

[0019] (3),

[0020] (4),

[0021] Among them, is the number of pixel points belonging to green plants or trees counted in the inspection image ; represents the pixel points in the inspection image whose , and values satisfy formula (4).

[0022] Preferably, another way to analyze whether there are channel tree problems in the inspection image is:

[0023] Extract the spectral features of the trees in the inspection image , perform multi-scale tree feature-preserving filtering on the spectral features of the trees G , and combine the multi-scale filtering results to obtain G ; Use the principal component analysis method to perform decision fusion on and retain the first principal component ; Perform threshold processing on to generate a binary image , count the number of pixel points belonging to green plants or trees in the binary image , and if is greater than or equal to the determination threshold, it is considered that there are channel tree problems in the inspection image N ; Among them, N , represents the total number of inspection images. , represents the total number of inspection images.

[0024] Preferably, perform tree feature-preserving filtering on the spectral features of the trees according to formula (9): G (9),

[0025] (9),

[0026] represents domain transformation recursive filtering; are all filtering parameters under the th set of parameter conditions; represents the filtering result of the spectral features of the trees under the G th set of parameter conditions;

[0027] Combine the multi-scale filtering results according to formula (10) to obtain :[[]]

[0028] (10),

[0029] wherein, represents the total number of groups of parameter conditions.

[0030] Preferably, generate the final UAV autonomous inspection track file , specifically:[[]]

[0031] (5),

[0032] wherein, , and both have a value range of , represents the total number of inspection images, represents the number of inspection images with green plants and trees, represents the number of inspection images without green plants and trees, represents the th UAV shooting point and flight path of the pole corresponding to the inspection image with green plants and trees, represents the th UAV shooting point and flight path of the pole corresponding to the inspection image without green plants and trees, represents the type of pole in the inspection image, , represents the total number of pole types.

[0033] Preferably, use a UAV to take pictures directly above the pole to obtain inspection images of the pole; wherein, when taking pictures, ensure that the pole target is at the center position of the inspection image.

[0034] Preferably, the target detection model is obtained by improving the yolov8 model. Specifically: add a Transformer layer behind the SPPF layer in the backbone network of the yolov8 model, add a BLRA layer in front of the SPPF layer in the backbone network of the yolov8 model, and replace the Conv convolutional layer in the backbone network and the feature fusion network with a DSConv convolutional layer.

[0035] Preferably, input the power pole training dataset into the target detection model to complete model training, and use the trained target detection model to identify the type of pole in the inspection image and extract the three-dimensional coordinate information of the pole; wherein: respectively represent a type of pole, Indicates the total number of tower types.

[0036] Preferably, the th inspection image is input into the trained object detection model, and the recognition result of the th inspection image is expressed as:

[0037] (1),

[0038] wherein, is the recognition result of the th inspection image, represents the type of the tower in the th inspection image , , indicates the total number of tower types, represents the longitude and latitude of the tower in the th inspection image , represents the altitude information of the tower in the th inspection image .

[0039] Preferably, the drone shooting points and flight paths corresponding to the towers with channel tree problems are optimized. Specifically: on the basis of the set drone shooting points and flight paths, both the drone shooting points and flight paths of the towers are increased by H meters, where H > 0.

[0040] The present invention also provides a low-cost unmanned aerial vehicle (UAV) autonomous inspection device for overhead power lines, including a memory and a processor. When the processor runs the computer instructions stored in the memory, it executes the low-cost UAV autonomous inspection method for overhead power lines.

[0041] The present invention also provides another low-cost UAV autonomous inspection device for overhead power lines, including a path planning unit and a UAV. The UAV is used to collect inspection images and inspect the overhead lines; the path planning unit executes the low-cost UAV autonomous inspection method for overhead power lines to provide the final UAV autonomous inspection track file for the inspection of the UAV. .

[0042] Applying the technical solution of the present invention has the following beneficial effects:

[0043] The present invention presets models of different types of power transmission towers, as well as the shooting points and flight paths of unmanned aerial vehicles (i.e., the model library). According to the identified tower type, it matches the tower model, shooting points and flight paths of the unmanned aerial vehicle in the model library, and adaptively generates the final autonomous inspection flight track file of the unmanned aerial vehicle in combination with the three-dimensional coordinate information of the identified tower. The present invention can achieve one-time investment and construction, and subsequent repeated applications, greatly reducing the application cost of autonomous inspection of unmanned aerial vehicles and realizing low-cost autonomous inspection of unmanned aerial vehicles for distribution networks. In addition, the present invention qualitatively analyzes whether there are channel tree problems in the inspection image according to the number of green vegetation pixels in the inspection image, and then optimizes the flight track file of the unmanned aerial vehicle inspection, avoiding the risk of the unmanned aerial vehicle hitting a tree, improving the safety of the unmanned aerial vehicle inspection operation, and enhancing the quality and efficiency of distribution network operation and maintenance. The present invention can realize low-cost autonomous inspection of unmanned aerial vehicles for overhead power lines under complex mountain conditions, which has great significance and practical value in power intelligent operation and maintenance.

[0044] In addition to the purposes, features and advantages described above, the present invention has other purposes, features and advantages. The following will refer to the drawings for a more detailed description of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0046] Figure 1 is a flowchart of the method for low-cost autonomous inspection of unmanned aerial vehicles for overhead power lines in the present invention;

[0047] Figure 2 is a structural diagram of the target detection model in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below, and preferred embodiments of the present invention are given. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0050] Embodiment 1:

[0051] See Figure 1, this embodiment provides a low-cost autonomous inspection method for overhead power lines, especially applicable to complex mountainous conditions. There are many and dense trees in complex mountainous areas, and the risk of the drone colliding during the inspection process is high. The inspection method includes the following steps:

[0052] Step S1, obtain the inspection image of the tower to be inspected;

[0053] Specifically, in step S1, a drone is used to take pictures directly above the tower to obtain the inspection image of the tower (i.e., visible light image); further, when taking pictures, it should be ensured that the tower target is at the center position of the inspection image.

[0054] Step S2, use the target detection model to identify the type of the tower in the inspection image, and extract the three-dimensional coordinate information of the tower (i.e., longitude, latitude, and altitude information);

[0055] Specifically, the target detection model used in this embodiment is obtained by improving the yolov8 model. The improvement of the existing yolov8 model in this embodiment is as follows: a Transformer layer (i.e., encoder-decoder network module) is newly added behind the SPPF layer (i.e., spatial pyramid pooling layer) in the backbone network of the yolov8 model, and a BLRA layer (i.e., Bi-level Routing Attention, double-layer routing attention mechanism module) is added in front of the SPPF layer in the backbone network of the yolov8 model. At the same time, the Conv convolutional layer in the backbone network and the feature fusion network is replaced with a DSConv convolutional layer (i.e., serpentine convolutional layer). The improved yolov8 model improves the model calculation efficiency and performance through the BLRA layer, and enables self-attention calculation and the use of windowed attention mechanism in the local area of the image through the Transformer layer, improving the efficiency of the model for image data processing and maintaining a strong feature extraction ability. At the same time, the application of the DSConv convolutional layer can extract the detailed information of the tower target at different spatial scales, thereby improving the accuracy of AI recognition.

[0056] See Figure 2 , the target detection model in this embodiment will be described in detail below:

[0057] The target detection model in this embodiment includes a backbone network, a feature fusion network, and a detection network. The backbone network is used to extract multi-scale deep visual features from the inspection image, the feature fusion network is used to fuse features of different scales, and the detection network is responsible for outputting detection information.

[0058] The backbone network sequentially includes DSConv convolutional layers, DSConv convolutional layers, 3×C2F layers (where C2f is a feature fusion layer), DSConv convolutional layers, 6×C2F layers, DSConv convolutional layers, 6×C2F layers, DSConv convolutional layers, 3×C2F layers, BLRA layer, SPPF layer, and Transformer layer; where: The DSConv convolutional layer is used to dynamically adjust the shape and weights of the convolutional kernel, which can better adapt to the diversity and complexity in the image, thereby extracting richer and more accurate features; The 3×C2F layer and 6×C2F layer are both used to splice the features of different branches in the channel dimension to achieve feature fusion; The SPPF layer is used to enhance the model's detection ability for targets of different scales; The Transformer layer is used to perform self-attention calculation and use a windowed attention mechanism to improve the efficiency of the model in processing image data; The BLRA layer is used to improve the model's calculation efficiency and performance. Specifically, the BLRA layer contains two attention mechanism channels. The upper-branch attention mechanism channel interacts with all image patches through the global self-attention mechanism and generates a global image representation; The lower-branch attention mechanism channel uses the local self-attention mechanism to interact with each image patch and its neighboring image patches and generates a local image representation; Through this double-layer routing attention mechanism, the improved yolov8 model can simultaneously capture global and local feature information, thereby improving the model's performance in visual tasks.

[0059] The feature fusion network sequentially includes UP layers (i.e., upsampling layers, Upsample layers), Concat layers (i.e., splicing layers), 3×C2F layers, UP layers, Concat layers, 3×C2F layers, DSConv convolutional layers, Concat layers, 3×C2F layers, DSConv convolutional layers, Concat layers, and 3×C2F layers; where: The UP layer is an upsampling operation used to fuse the information extracted from the deep feature map onto the shallower feature map so that it can match the higher-resolution feature map; The Concat layer is used to merge feature maps from different depths to form a more comprehensive representation; The 3×C2F layer is used to splice the features of different branches in the channel dimension to achieve feature fusion.

[0060] The detection network includes three Detect layers (i.e., detection layers). The last three 3×C2F layers of the feature fusion network are respectively connected to a Detect layer, and the Detect layer is used to determine the objects present in the image and their locations.

[0061] Among them, the first 6×C2F layer in the backbone network is connected to the second Concat layer in the feature fusion network. The DSConv convolutional layer between the 6×C2F layer and the 3×C2F layer in the backbone network (where the 6×C2F layer is in front of the DSConv convolutional layer and the 3×C2F layer is behind the DSConv convolutional layer) is connected to the first Concat layer in the feature fusion network. The Transformer layer in the backbone network is connected to the first UP layer and the last Concat layer in the feature fusion network. The first 3×C2F layer in the feature fusion network is connected to the penultimate Concat layer in the feature fusion network.

[0062] The processing process of the inspection image in the target detection model is as follows:

[0063] The inspection image is input into the backbone network. First, it passes through two DSConv convolutional layers in sequence to extract more abundant and accurate feature data A. Then, the feature data A is input into the 3×C2F layer (i.e., three C2F layers) for processing. The feature B output by the 3×C2F layer enters the DSConv convolutional layer and the 6×C2F layer (i.e., six C2F layers) for processing in sequence, and outputs feature Y1 and feature Y2. After feature Y1 passes through the DSConv convolutional layer, the 6×C2F layer, and the DSConv convolutional layer for processing, it outputs feature B1 and feature B2. Feature B1 passes through the 3×C2F layer, the BLRA layer, the SPPF layer, and the Transformer layer for processing, and outputs feature D1 and feature D2.

[0064] Furthermore, feature Y2, feature B2, feature D1, and feature D2 are all input into the feature fusion network. Feature D1 is input into the UP layer for processing and outputs feature E. Feature E and feature B2 are spliced through the Concat layer to output feature F. Feature F passes through the 3×C2F layer for processing and outputs feature G1 and feature G2. Feature G1 passes through the UP layer and is transported to the Concat layer to be spliced with feature Y2, and then input into the 3×C2F layer for processing to output feature K1 and feature K2. Feature K1 passes through the DSConv convolutional layer and enters the Concat layer to be spliced with feature G2, and then input into the 3×C2F layer for processing to output feature L1 and feature L2. Feature L1 passes through the DSConv convolutional layer and enters the Concat layer to be spliced with feature D2, and then input into the 3×C2F layer for processing to output feature M.

[0065] Among them, feature K2 is transported to the first Detect layer to determine the objects existing in the image and their positions. Feature L2 is transported to the second Detect layer to determine the objects existing in the image and their positions. Feature M is transported to the third Detect layer to determine the objects existing in the image and their positions.

[0066] During the processing of the model, the C2F layer performs feature transformation, branch processing, and feature fusion operations on the input data through two convolutional layers (cv1 and cv2), extracts and transforms the features of the input data, and generates a more representative feature output.

[0067] In this embodiment, in order to adapt to the spatial morphological characteristics of the pole tower, on the one hand, the target detection model introduces a serpentine convolutional layer (i.e., a serpentine convolution module), which has good feature representation in the case of processing slender tubular structures (such as slender targets like pole towers). Secondly, in the complex power inspection background, the ordinary YOLO algorithm is difficult to handle the problem of small target recognition (for example, after the drone raises its height to take pictures, the pole tower target will appear in the form of a small target). Therefore, by introducing a transformer module, self-attention calculation is performed in the local area and a windowed attention mechanism is used, which improves the efficiency of the model in processing image data and maintains a strong feature extraction ability for small target pole towers. Finally, since the spatial resolution of the drone inspection images is very high, for example, the size of the pictures taken by DJI is 5000×4000 pixels, and a picture is more than 20 megabytes, therefore, how to accelerate the training and inference efficiency of the model is crucial for improving the inspection efficiency. In this embodiment, by introducing the BLRA attention mechanism module and using a dynamic and query-aware sparse attention mechanism to solve this problem, the target detection model of this embodiment can well perform the task of power inspection image recognition.

[0068] Further, the power pole tower training data set is input into the target detection model to complete the model training, where: respectively represent a type of pole tower, represents the total number of pole tower types. In this embodiment, five types of pole towers are set, represents a single-circuit straight pole tower, represents a portal straight pole tower, represents a distribution transformer pole tower, represents a double-circuit straight pole tower, represents a double-circuit portal tension pole tower, and each type of pole tower contains 2000 samples; during training, the core parameters are set as follows: the batch size is set to 32, the initial learning rate is set to 10 -3 , the optimizer is the Adam algorithm, the label-smoothing is set to 0.1, and the epoch is set to 120.

[0069] After the model training is completed, the detection and recognition of the pole tower types in the inspection images can be realized. Assume that the th inspection image input is , where , represents the total number of inspection images input, and the recognition result of the -th inspection image is expressed as:

[0070] (1),

[0071] Among them, is the recognition result of the -th inspection image, represents the type of the tower in the -th inspection image , , represents the longitude and latitude of the tower in the -th inspection image , represents the altitude information of the tower in the -th inspection image .

[0072] Step S3: Establish a model library for various types of towers, and set the UAV shooting points and flight paths for each type of tower;

[0073] Specifically, those skilled in the art can set the UAV shooting points and UAV flight paths for various types of towers according to the actual situation, and the detailed settings of the UAV shooting points and flight paths for various types of towers will not be introduced in detail in this embodiment. In this embodiment, the UAV shooting points and flight paths of the -th type of tower are recorded as , , represents the total number of tower types.

[0074] Step S4: Based on the tower type identified in step S2, the three-dimensional coordinate information of the tower, and the model library in step S3, generate a parametric three-dimensional model of the identified tower in the three-dimensional space map, and obtain the UAV shooting points and flight paths corresponding to the identified tower type;

[0075] Preferably, step S4 is specifically to obtain the three-dimensional model of the tower from the model library according to the identified tower type, determine the position of the tower in the three-dimensional space map based on the three-dimensional coordinate information of the identified tower, and after obtaining the parametric three-dimensional model, the UAV flight track files for each type of tower can be preset, that is, set the flight path and shooting points when each type of tower is photographed, etc.; among them, the flight path plans the shooting order between each shooting point.

[0076] Step S5: Analyze whether there are channel tree hazards in each inspection image. If there are channel tree hazards in the inspection image, optimize the UAV shooting points and flight paths of the tower in the inspection image;

[0077] Preferably, the specific method for analyzing whether there are channel tree problems in the inspection images is as follows:

[0078] First, convert the inspection images from the RGB color channels to HSV the luminance space for better processing of the colors in the inspection images:

[0079] (2),

[0080] wherein, in formula (2) HSV represents the preprocessing operation of converting the RGB color channels to HSV the luminance space, represents the hue component, represents the saturation component, represents the value component.

[0081] Then, count the number of pixel points belonging to green plants and trees in the inspection images according to formulas (3) and (4); if is greater than or equal to the set threshold, it is considered that there are green plants and trees in the inspection images, that is, there are channel tree problems; if is less than the set threshold, it is considered that there are no green plants and trees in the inspection images, that is, there are no channel tree problems;

[0082] (3),

[0083]

[0084] wherein, is the number of pixel points belonging to green plants and trees counted in the inspection image , represents detecting the pixel points in the inspection image whose , and values satisfy formula (4).

[0085] Preferably, in this embodiment, the set threshold is taken as 500 according to experience; of course, in other embodiments, the set threshold may be set to other values, and those skilled in the art can set it according to the actual situation, and the possible value ranges of the set threshold will not be described in detail in this embodiment.

[0086] According to the standards of distribution network operation and maintenance, there should be no trees under power poles and corridors. Therefore, if green plants or trees are identified in the inspection images (such as when trees grow under power poles and corridors), in order to avoid unnecessary drone collisions, the drone shooting points and flight paths of the poles in the inspection images with green plants or trees need to be raised by H meters. In this embodiment, H is taken as 5 meters. Those skilled in the art can determine the value of H according to the actual terrain, experience, etc., and no detailed description is given in this embodiment.

[0087] Step S6: Generate the final autonomous drone inspection flight path file ;

[0088] In step S6, it is possible that the drone shooting points and flight paths of some poles have been optimized, while those of the remaining poles do not need to be optimized. Therefore, the final autonomous drone inspection flight path file consists of two parts:

[0089] (5),

[0090] Among them, , and both have a value range of , represents the total number of input inspection images, represents the number of inspection images with green plants or trees, represents the number of inspection images without green plants or trees, represents the drone shooting points and flight paths of the pole corresponding to the th inspection image with green plants or trees, represents the drone shooting points and flight paths of the pole corresponding to the th inspection image without green plants or trees, represents the type of the pole in the inspection image, , represents the total number of pole types.

[0091] At this point, any drone only needs to import the defined final autonomous drone inspection flight path file , and it can complete autonomous inspection without any further manual intervention or processing.

[0092] This embodiment also provides a low-cost autonomous drone inspection device for overhead power lines, including a memory and a processor. The memory and the processor are connected. When the processor runs the computer instructions stored in the memory, it executes the low-cost autonomous drone inspection method for overhead power lines.

[0093] This embodiment also provides another low-cost autonomous inspection device for overhead power lines, including a path planning unit and a drone. The drone is used to collect inspection images and inspect the overhead lines. The path planning unit executes the low-cost autonomous inspection method for overhead power line drones to provide the final autonomous inspection flight track file for the drone's inspection. 。

[0094] Embodiment 2:

[0095] The difference between this embodiment and Embodiment 1 lies only in the different method of analyzing whether there are channel tree hazards in the inspection images. The specific method of analyzing whether there are channel tree hazards in the inspection images in this embodiment is as follows:

[0096] Extract the spectral features of the trees in the inspection images and perform multi-scale tree feature-preserving filtering on the spectral features of the trees G . Combine the multi-scale filtering results to obtain G ; Use the principal component analysis method to perform decision fusion on and retain the first principal component ; Perform threshold processing on to generate a binary image . Count the number of pixel points belonging to green plants in the binary image . If is greater than or equal to the determination threshold, it is considered that there are channel tree hazards in the inspection image N . N is greater than or equal to the determination threshold, it is considered that there are channel tree hazards in the inspection image .

[0097] Furthermore, the specific method of extracting the spectral features of the trees in the inspection images is as follows: G According to formula (6), convert the

[0098] th inspection image from the RGB color channel to the HSV brightness space:

[0099] (6),

[0100] wherein, in formula (6) HSV represents the preprocessing operation of converting the RGB color channel to HSV the brightness space, represents the hue component, represents the saturation component, represents the value component, , represents the total number of inspection images;

[0101] Obtain the inspection images according to Formula (7) and Formula (8). The spectral features of the trees G :

[0102] (7),

[0103] (8),

[0104] wherein, represents detecting the inspection images in , and the pixel points whose values satisfy Formula (8).

[0105] Although most of the target pixel points of green plants and trees can be detected through the function, the noise and isolated background in the image still interfere with the target. Therefore, to further improve the representation ability of the tree spectral features G , in this embodiment, a multi-scale tree feature-preserving filtering process is also performed on the tree spectral features G , and then the multi-scale filtering results are combined to obtain , specifically:

[0106] Perform tree feature-preserving filtering on the tree spectral features G according to Formula (9):

[0107] (9),

[0108] represents domain transformation recursive filtering; are all filtering parameters under the parameter conditions of the th group, set according to empirical values; represents the filtering result of the tree spectral features under the parameter conditions of the G th group;

[0109] Combine the multi-scale filtering results according to Formula (10) to obtain :

[0110] (10),

[0111] wherein, represents the total number of groups of parameter conditions (i.e., the number of types of scales).

[0112] Furthermore, in order to fully exploit the tree feature information under different filtering scales, in this embodiment, the principal component analysis method is used for Perform decision fusion processing and retain the first principal component :

[0113] (11),

[0114] Among them, represents the dimensionality reduction operation of principal component analysis.

[0115] Obtain the fusion result and perform threshold processing on it to generate a binary image , in , the pixels with a pixel value of 1 represent the pixels that are very likely to be green plants and trees; to reduce the misjudgment rate, a determination threshold is set in this embodiment. When the number of pixels with a pixel value of 1 N is greater than or equal to the determination threshold, it is considered that there are tree hazards in the inspection image; among them, the operation of counting the pixel points belonging to green plants and trees in the binary image is expressed as:

[0116] (12),

[0117] Among them, num is the operation of counting the number of pixel points.

[0118] In this embodiment, the determination threshold is set to 500; compared with the tree hazard recognition method in Embodiment 1, the recognition method in this embodiment can avoid the interference of noise and isolated backgrounds in the image, improve the accuracy of tree hazard recognition, and avoid misjudgment.

[0119] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A low-cost autonomous inspection method for overhead power lines by drones, characterized in that: include: Obtain inspection images of the tower to be inspected; Use the target detection model to identify the type of tower in the inspection image and extract the 3D coordinate information of the tower; Establish a model library of various types of towers, and set the drone shooting points and flight paths for each type of tower; Based on the identified tower type, the tower's three-dimensional coordinate information and the model library, a parametric three-dimensional model of the identified tower is generated in the three-dimensional space map, and the drone shooting points and flight paths corresponding to the identified tower type are obtained; Analyze whether there is channel tree damage in each inspection image. If there is channel tree damage in the inspection image, optimize the drone shooting points and flight path of the tower in the inspection image; Generate the final drone autonomous inspection track file based on the optimized drone shooting points and flight paths and the drone shooting points and flight paths that do not need to be optimized ; The specific method to analyze whether there is channel tree problem in the inspection image is as follows: Extract inspection images Spectral characteristics of trees in G , for tree spectral characteristics G Perform multi-scale tree feature preservation filtering and combine the multi-scale filtering results to obtain ; Using principal component analysis Perform decision fusion and retain the first principal component ;right Perform threshold processing to generate a binary image , statistical binary image The number of pixels belonging to green plants and trees N ,like N If it is greater than or equal to the judgment threshold, the inspection image is considered There is a channel tree problem in , Indicates the total number of inspection images; Extract inspection images Spectral characteristics of trees G The specific method is: According to formula (6), Inspection image Convert from RGB color channels to HSV brightness space: （6）， Among them, in formula (6) HSV Represents the preprocessing operation of converting RGB color channels into HSV brightness space. represents the hue component, represents the saturation component, Indicates the brightness component; According to formula (7) and formula (8), the inspection image is obtained Spectral characteristics of trees in G : （7）， （8）， in, Indicates the inspection image middle , and The pixel whose value satisfies formula (8); According to formula (9), the spectral characteristics of trees G Perform tree feature preserving filtering: （9）， Representation domain transform recursive filtering; All are filtering parameters under the conditions of the zth group of parameters; Represents the spectral characteristics of trees under the zth group of parameters G The filtering result of According to formula (10), the multi-scale filtering results are combined to obtain : （10）， in, Indicates the total number of groups of parameter conditions.

2. The low-cost autonomous inspection method for overhead power lines by drones according to claim 1 is characterized in that: The principal component analysis method was used to Perform decision fusion and retain the first principal component , expressed as: （11）， in, Represents the principal component analysis dimensionality reduction operation.

3. The low-cost autonomous inspection method for overhead power lines by drone according to claim 1 is characterized in that: The target detection model is improved based on the yolov8 model, specifically: a Transformer layer is added after the SPPF layer in the yolov8 model backbone network, a BLRA layer is added before the SPPF layer in the yolov8 model backbone network, and the Conv convolution layer in the backbone network and the feature fusion network is replaced with a DSConv convolution layer.

4. The low-cost autonomous inspection method for overhead power lines by drone according to claim 1 is characterized in that: The first Inspection image Input the trained target detection model, then The recognition result of the inspection image is expressed as: （1）， in, For the The recognition result of the inspection image is Indicates Inspection image Type of tower, , Indicates the total number of tower types, Indicates Inspection image The longitude and latitude of the middle tower, Indicates Inspection image The altitude information of the middle tower.

5. The low-cost autonomous inspection method for overhead power lines by drone according to claim 1 is characterized in that: The drone shooting points and flight paths corresponding to the towers with channel tree problems are optimized. Specifically, based on the set drone shooting points and flight paths, the drone shooting points and flight paths of the towers are increased by H meters, where H is greater than 0.

6. A low-cost UAV autonomous inspection device for power overhead lines, characterized in that: It comprises a memory and a processor, and when the processor runs the computer instructions stored in the memory, it executes the low-cost autonomous inspection method of overhead power lines by unmanned aerial vehicles as described in any one of claims 1 to 5.

7. A low-cost UAV autonomous inspection device for power overhead lines, characterized in that: The invention comprises a path planning unit and a drone, wherein the drone is used to collect inspection images and inspect the overhead lines; the path planning unit executes the low-cost autonomous inspection method for overhead power lines by drones as described in any one of claims 1 to 5, and provides a final autonomous inspection track file for the drone inspection. .

Images