Unmanned aerial vehicle nest based on fault image enhancement and optimization recognition

Abstract

The invention discloses an unmanned aerial vehicle nest based on fault image enhancement and optimization recognition, and relates to the technical field of overhead power line fault inspection, the unmanned aerial vehicle nest based on fault image enhancement and optimization recognition comprises an unmanned aerial vehicle nest main body, and a landing platform is arranged in the unmanned aerial vehicle nest main body. An unmanned aerial vehicle is arranged at the upper end of the landing platform, a millimeter-wave radar is installed on the lower end face of the unmanned aerial vehicle, a three-dimensional laser scanner is installed on one side of the millimeter-wave radar, the input end of the unmanned aerial vehicle nest body is electrically connected with an edge algorithm all-in-one machine, and the output end of the unmanned aerial vehicle nest body is electrically connected with a holder. According to the scheme, the problems that route planning is not accurate, collision damage is easily caused, cruising ability is poor, inspection efficiency is low, data management is difficult, and manual fault judgment is needed in actual application of an existing unmanned aerial vehicle are solved, the maintenance efficiency is greatly reduced, and unnecessary money and resource waste is caused.

Description

Technical Field

[0001] The present invention relates to the technical field of patrol inspection of overhead power line faults, and in particular to an unmanned aerial vehicle nest based on fault image enhancement and optimized recognition. Background Art

[0002] Inspection of overhead power line faults mainly involves inspection of the line conditions, as well as timely discovery and recording of equipment defects and fault points. This information will serve as the basis for line maintenance, and drones are often used for inspections.

[0003] For example, the Chinese authorized patent with announcement number CN112960132B (a distributed shared machine nest and its power line drone inspection method): includes a variety of inspection equipment and a mechanical arm that can replace the inspection equipment and drone batteries; the inspection method includes: selecting a distributed shared machine nest according to the power line inspection task and generating a three-dimensional inspection path; using the inspection drone in the distributed shared machine nest to perform the power line inspection task according to the three-dimensional inspection path to obtain inspection data; analyzing the inspection data based on image recognition technology and point cloud data processing technology to obtain the power line inspection results. The present invention can achieve efficient multi-functional, full-factor, and full-coverage autonomous inspection.

[0004] However, in actual applications, existing drones have exposed problems such as inaccurate route planning, easy collision damage, poor endurance, low inspection efficiency, difficult data management, and the need for manpower for fault identification. This not only greatly reduces the maintenance efficiency, but also causes unnecessary waste of money and resources; therefore, it does not meet the existing needs. In this regard, we propose a drone nest based on fault image enhancement and optimized recognition. Summary of the invention

[0005] The purpose of the present invention is to provide a drone nest based on fault image enhancement and optimized identification, so as to solve the problems of existing drones proposed in the above background technology, such as inaccurate route planning, easy collision damage, poor endurance, low inspection efficiency, difficult data management, and need for manpower for fault identification in actual applications, which not only greatly reduces the maintenance efficiency, but also causes unnecessary waste of money and resources.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a drone nest based on fault image enhancement and optimized identification, comprising: a drone nest main body, a landing platform is arranged inside the drone nest main body, a drone is arranged on the upper end of the landing platform, a millimeter-wave radar is installed on the lower end face of the drone, a three-dimensional laser scanner is installed on one side of the millimeter-wave radar, the input end of the drone nest main body is electrically connected to an edge algorithm all-in-one machine, and the output end of the drone nest main body is electrically connected to a gimbal.

[0007] Preferably, a protective cover is provided on the top of the drone nest body, a screw mounting groove is provided on the upper surface of the drone nest body, and a ball screw slide rail is installed inside the screw mounting groove, and the protective cover and the drone nest body are both semicircular.

[0008] Preferably, a fixed base is provided at the lower end of the drone nest body, a fixing frame is welded to the rear surface of the fixed base, and one end of the fixing frame is threadedly fixed to the drone nest body by a threaded fastener.

[0009] Preferably, a maintenance opening is provided on the front surface of the drone nest body, and a protective door is provided inside the maintenance opening.

[0010] Preferably, the edge algorithm integrated machine includes COSRTK real-time line patrol and error correction technology, T-FAS drone image enhancement and optimization recognition algorithm, and electrical patrol fault image data set.

[0011] Preferably, the COSRTK real-time line patrol and error correction technology is implemented based on the Kalman algorithm.

[0012] Preferably, the electrical inspection fault image dataset is based on open source datasets at home and abroad, coupled with over 20,000 adversarial network algorithms, and is formed by combining deep learning with self-collected datasets.

[0013] Preferably, the COSRTK real-time line patrol and error correction technology includes RTK technology and three-dimensional laser point cloud accurate hazard level judgment technology, and the UAV adopts RTK technology for positioning.

[0014] Preferably, the UAV is composed of three parts: a UAV carrier system, an automatic flight system and a ground control system.

[0015] Preferably, the T-FAS UAV image enhancement and optimization recognition algorithm performs enhancement processing on the decomposed low-illuminance illumination image based on the Encoder-Decoder structure.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] 1. The present invention realizes "active early warning, intelligent decision-making, automatic inspection, and intelligent management and control" by carrying COSRTK real-time line inspection and error correction technology based on Kalman algorithm and self-created T-FAS UAV image enhancement and optimization recognition algorithm. Compared with the existing UAV that imports video into computer and recognizes it frame by frame by backstage personnel, it improves the equipment perception ability, defect discovery ability, status control ability, active early warning ability and emergency externalization ability of operation and maintenance personnel, realizes the intelligent operation and maintenance of power grid, accelerates the implementation process of intelligent inspection, intelligent operation and intelligent safety, and solves the problems of existing UAVs in actual application, such as inaccurate route planning, easy collision damage, poor endurance, low inspection efficiency, difficult data management, and need for manpower for fault identification, which not only greatly reduces the maintenance efficiency, but also causes unnecessary waste of money and resources.

[0018] 2. A protective cover is provided on the upper end of the main body of the drone nest, and an electric push rod for driving the landing platform to move up and down is installed in the main body of the drone nest. Both the protective cover and the main body of the drone nest are semicircular. After the drone lands on the landing platform, the electric push rod drives the landing platform to move down. After closing the protective cover, it becomes a complete circle. The electric push rod drives the landing platform to move up and contact the protective cover. The lower surface of the protective cover has a sponge bottom plate, which has a limiting and squeezing effect on the top of the drone to prevent the drone from shaking when the drone nest moves. The drone nest with a circular structure only needs to roll the whole when moving, and does not require multiple people to lift it, which improves the convenience of movement. Compared with installing rollers, the ground contact area is small, and the rolling method will be more labor-saving and improve convenience. In addition, with a circular structure, rainwater can fall along the arc surface on rainy days. Compared with the existing square shape, it is not easy for rainwater to accumulate, and there is no gap in the middle, so it is not easy to get water in, which improves the waterproof effect and safety. Through the circular design, it can be moved with less effort and the waterproof effect can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a schematic diagram of the main structure of the drone nest of the present invention;

[0020] Figure 2 This is a schematic diagram of the main structure of the drone nest after the protective cover of the present invention is closed;

[0021] Figure 3 This is a schematic diagram of the rear view structure of the main body of the drone nest of the present invention;

[0022] Figure 4 It is a flowchart of the overall technology and data transmission of the drone of the present invention;

[0023] Figure 5 A schematic diagram of the classification of target detection methods by combining the DCGAN algorithm of the present invention with deep learning expansion technology;

[0024] Figure 6 The core calculation control diagram of the UAV automatic flight system of the present invention;

[0025] Figure 7 A diagram of the ground control system of the present invention;

[0026] Figure 8 The COSRTK measurement technology principle diagram of the present invention;

[0027] Fig. 9 The flowchart of the reconstruction of the three-dimensional laser point cloud model of the present invention;

[0028] Fig.10 A diagram of the steps affecting geometric correction of the present invention;

[0029] Fig.11 It is a schematic diagram of the Relight-Ne network structure of the present invention;

[0030] Fig.12 This is a flowchart of drone image recognition of the present invention;

[0031] Fig.13 The formula of the present invention is to align point cloud data to the same coordinate system through a six-parameter model;

[0032] Fig.14 is the Retinex theoretical image formula of the present invention;

[0033] Fig.15 This is a schematic diagram of the cross-sectional structure of the main body of the drone nest of the present invention;

[0034] In the figure: 1. UAV nest body; 2. Protective cover; 3. Landing platform; 4. UAV; 5. Fixed base; 6. Maintenance opening; 7. Protective door; 8. Screw mounting slot; 9. Fixed bracket; 10. Threaded fasteners; 11. Millimeter wave radar; 12. 3D laser scanner; 13. Edge algorithm all-in-one machine. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0036] See also Figure 1-11An embodiment of the present invention is as follows: a drone nest based on fault image enhancement and optimization recognition, comprising: a drone nest body 1, a landing platform 3 is arranged inside the drone nest body 1, a drone 4 is arranged at the upper end of the landing platform 3, a millimeter wave radar 11 is installed on the lower end surface of the drone 4, a three-dimensional laser scanner 12 is installed on one side of the millimeter wave radar 11, an input end of the drone nest body 1 is electrically connected to an edge algorithm integrated machine 13, an output end of the drone nest body 1 is electrically connected to a gimbal, a protective cover 2 is arranged on the top of the drone nest body 1, a lead screw mounting groove 8 is arranged on the upper surface of the drone nest body 1, and a ball screw slide is installed inside the lead screw mounting groove 8, the protective cover 2 and the drone nest body 1 are both semicircular, a fixed base 5 is arranged at the lower end of the drone nest body 1, a fixing frame 9 is welded on the rear surface of the fixing base 5, one end of the fixing frame 9 is threadedly fixed to the drone nest body 1 through a threaded fastener 10, an inspection opening 6 is arranged on the front surface of the drone nest body 1, and a protective door 7 is arranged inside the inspection opening 6.

[0037] An electric push rod is installed in the main body 1 of the drone nest to drive the landing platform 3 to move up and down. The protective cover 2 and the main body 1 of the drone nest are both semicircular. After closing the protective cover 2, it becomes a complete circle. The drone nest with a circular structure only needs to be rolled as a whole when moving, and does not require multiple people to lift it, which improves the convenience of mobility. In addition, due to the circular structure, rainwater can fall along the arc surface on rainy days. Compared with the existing square structure, it is not easy for rainwater to accumulate, and there is no gap in the middle, so it is not easy for water to enter, which improves the waterproof effect and safety.

[0038] By equipping the COSRTK real-time line patrol and error correction technology based on the Kalman algorithm and the self-created T-FAS drone image enhancement and optimization recognition algorithm, we can achieve "active early warning, intelligent decision-making, automatic inspection, and intelligent management and control", improve the equipment perception, defect detection, status control, active early warning and emergency externalization capabilities of operation and maintenance personnel, realize the intelligent operation and maintenance of the power grid, and accelerate the implementation of intelligent inspection, intelligent operation and intelligent safety.

[0039] It is further introduced that the edge algorithm all-in-one machine 13 includes COSRTK real-time line patrol and error correction technology, T-FAS drone image enhancement and optimization recognition algorithm and electric patrol fault image dataset. COSRTK real-time line patrol and error correction technology is based on Kalman algorithm, and the electric patrol fault image dataset is based on open source datasets at home and abroad, plus more than 20,000 adversarial network algorithms, which are formed by combining deep learning self-collected datasets. The image dataset has doubled, improving the quality and accuracy of the generated images. The above-mentioned electric patrol fault image dataset is the existing technology.

[0040] Further introduction, COSRTK real-time line inspection and error correction technology includes RTK technology and 3D laser point cloud accurate judgment of hidden danger level technology, and UAV 4 uses RTK technology for positioning. UAV 4 consists of three parts: UAV carrier system, automatic flight system and ground control system. The UAV carrier system is used to provide power for the UAV and perform high-precision and refined automatic inspection of high-voltage power line poles and overhead towers. The automatic flight system allows the UAV to fly automatically according to the pre-planned route, and controls the onboard camera to take pictures of the required location at each waypoint. The flight of the UAV according to the planned route mainly depends on path planning and navigation control technology. These technologies include global path planning and local path planning, as well as related algorithms such as A* algorithm, Dijkstra algorithm, artificial potential field method, fast random tree (RRT) algorithm, ant colony algorithm, genetic algorithm, etc. These algorithms analyze the environmental map to generate an optimized path, which is then converted into heading and altitude control instructions to guide the UAV to fly along the predetermined path. This is an existing technology. The RTK system provides real-time positioning to avoid collisions due to position errors during automatic flight of the UAV. The ground control system receives the position signal in real time to ensure that the UAV has fixed reference coordinates during flight. The position autopilot system receives the RTK signal to control the UAV flight and improve accuracy performance.

[0041] It is further introduced that the T-FAS UAV image enhancement and optimization recognition algorithm enhances the decomposed low-light illumination image based on the Encoder-Decoder structure to improve the image recognition capability.

[0042] Working principle: When using it, we create an overhead power line fault image database based on open source data sets at home and abroad, plus more than 20,000 adversarial network (DCGAN) algorithms by combining deep learning self-collected data sets. In the future inspection process, the inspected overhead power line faults will be uploaded to the database for processing and storage, and the model training of overhead power line faults will be carried out to iterate a more accurate model. The Feitongxunchang project team uses deep convolution to strengthen the adversarial network algorithm. DCGAN uses a fully convolutional network to replace the multi-layer perceptron in GAN, which improves the quality of synthetic images and training speed. The data expansion methods based on deep learning mainly include feature space expansion, adversarial training, generative adversarial networks, neural style transfer and meta-learning, such as Figure 5As shown in the figure, all the expansion methods are applied to images in the input space, and deep learning has a powerful function in mapping high-dimensional input to low-dimensional representation. Therefore, the input can be converted into a low-dimensional intermediate representation through the network, and data expansion processing is performed on it. Among them, we divide the various defects that occur into three categories according to the standards: nine facility defects, six ancillary facility defects, and four external hidden dangers. The specific process is that the drone collects real-time data of the tower and transmits it to the aircraft carrier. The aircraft carrier summarizes and analyzes the obtained data, generates a report and sends it to the cloud for the staff. The DCGAN algorithm doubles the image data set by combining the data expansion technology of deep learning, thereby improving the quality and accuracy of the generated images.

[0043] COSRTK real-time line patrol and error correction technology includes RTK technology and 3D laser point cloud accurate judgment of hidden danger level technology, and UAV 4 uses RTK technology for positioning. RTK technology is essentially a real-time dynamic high-precision positioning technology. Positioning accuracy is very important for multi-rotor UAV autonomous line patrol. The 3D coordinates of the measuring point can be assigned to the coordinate system in real time, and the composition, implementation principle, RTK measurement requirements and base station construction of multi-rotor UAV RTK are carried out; UAV 4 consists of three parts: UAV carrier system, automatic flight system and ground control system. The UAV carrier system is used to provide power for the UAV and perform high-precision and refined automatic inspection of high-voltage power line poles and overhead towers. The automatic flight system allows the UAV to fly automatically according to a pre-planned route, and controls the onboard camera at each waypoint to take pictures of the required position. The RTK system provides real-time positioning to avoid collisions due to position errors during automatic flight of the UAV, and the ground control system Receive position signals in real time to ensure that the UAV has fixed reference coordinates during flight. The position autopilot system receives RTK signals to control the flight of the UAV. COSRTK technology is a technology that integrates the advantages of high efficiency, automation, high integration, and strong data processing capabilities. The extended Kalman filter method can greatly improve the measurement accuracy of RTK technology. One of the characteristics of this calculation method is that as a huge permanent base station, several or dozens of small base stations have been set up, which can cover a larger measurement area at any time. There is no need to set up base stations at any time. The mobile platform with a gasoline generator drives the machine nest to move, which can cover a larger measurement area at any time. The UAV 4 is equipped with a three-dimensional laser scanner 12 to fly above the line to obtain the line body and channel data information of the current flight position. Only a high-precision three-dimensional transmission line model can meet the needs of inspection. It is an indispensable and important part of the automatic inspection of UAVs, and its accuracy will directly affect the safety of the entire operation.

[0044] 3D laser scanning technology uses the information of line towers and ground objects to construct models of power grid lines and power line channels. The constructed model needs to meet the requirements of 3D scene reconstruction of transmission lines. Subsequent route planning and automatic flight will use this model as an important reference. The 3D laser model reconstruction of high-voltage power lines mainly includes: classification of laser point cloud data, laser point cloud detection of each power line, segmentation of laser point cloud and 3D building model of power lines; point cloud data registration is to unify the laser data obtained by scanning multiple different measuring stations into the same coordinate system, which is conducive to the splicing of point cloud data and the splicing of measuring stations. The most commonly used method is the directional splicing method based on the same-name points. The same-name points can be specific targets or the foundation points of towers with other obvious features and other object points on the ground. There must be more than four same-name points in the splicing process. The following calculation process takes a certain measuring station as the base station, and the three rotation parameters and three translation parameters of other measuring stations relative to the base station are calculated based on the same same-name point. The following formula is to align the point cloud data to the same coordinate system for analysis through the six-parameter model, see Fig.13 .

[0045] Fig.13 In the formula, (X, Y, Z) is the coordinate after the point cloud is aligned, (x, y, z) is the original coordinate of the point cloud, α, β, γ are three rotation parameters, Δx, Δy, Δz are three translation parameters. The important purpose of line and channel model construction is to convert the collected point cloud data into a three-dimensional surface model. Point cloud data is a large collection of data points, but the construction of the transmission line model only requires some accurate feature lines.

[0046] The images taken by the drone have a small phase, so the images need to be stitched before they can be used for subsequent geometric correction. The geometric correction of drone images is essentially to project the collected original image data onto a plane, and at the same time add the original image to the map projection system to make it meet the requirements of geometric correction, thus forming a new image.

[0047] Timely registration of power line point cloud data and images is actually to ensure the reliability of data fusion. Before the 3D laser scanning data and drone images are used to reconstruct the 3D scene, the laser point cloud and drone image systems must be registered. This allows the data in the sky and the data on the earth to be aggregated into the same coordinate system. By using the maximum inter-class variance method to segment the point cloud of a single tree into a forest point cloud and a ground point cloud, the tree height can be calculated. The laser coordinate system is used when collecting laser point cloud data. This is an independent coordinate system. Analysts must scan The target control point information in the scene determines more than four feature points, such as the top of the tower, pole tower, three-phase conductor, ground wire, trees under the channel, etc., and measures their coordinate values ​​in the local coordinate system. It is also necessary to find the coordinate values ​​of the corresponding measuring points in the laser point cloud system in the point cloud data. There is a large rotation angle between the local coordinate system and the coordinate system of the laser scanning instrument, so the Rodriguez matrix method is used to transform the coordinates and calculate the rotation and translation parameters of the two coordinate systems. After the point cloud alignment is completed, the uniform sampling method and human-computer interaction filtering method are used to filter the noise and resample the data.

[0048] The use of three-dimensional laser radar measurement technology can be used to make up for the shortcomings of low spatial positioning judgment accuracy of manual inspection technology and difficulty in accurately measuring the distance between the cross-over and the conductor and ground wire in the power line corridor. The drone measurement system has high efficiency and huge advantages in rapid line inspection, close-range observation and fault repair of power grid transmission lines. It is a comprehensive system with advanced technology. Most laser points in the corridor of high-voltage overhead transmission line are surface vegetation points and crossed power line points, with few building points. After removing other object points, basically only vegetation points are left. The vegetation growth characteristics show complex and irregular shape distribution, which leads to the difficulty of data collection and accurate measurement. It is the biggest safety impact of transmission lines. In view of the shortcomings of the current inspection method, the three-dimensional laser point cloud data can solve the high-precision measurement and refined three-dimensional modeling problems of corridor terrain, surface vegetation, cross-over objects, line sags, etc. in the three-dimensional channel of the transmission line, greatly improving the operation efficiency of drones in transmission line inspection. At the same time, after the point cloud three-dimensional model is established, the dangerous points of the power grid line are analyzed.

[0049] We use the above method to perform 3D laser scanning on some line sections, process and analyze the collected real-time point clouds, and finally form a real-time working condition report.

[0050] Then, the image is enhanced and optimized by using the T-FAS UAV image enhancement and optimization recognition algorithm. First, the Retinex-Net low-light image enhancement network is analyzed and constructed. The network integrates image decomposition and subsequent enhancement operations. The Retinex-Net network introduces the traditional human color perception Retinex theory into deep learning. The theory believes that if the observed image can be decomposed into two parts, namely the reflected image and the illuminated image, let S represent the source image, then it can be expressed as: Fig.14 .

[0051] Fig.14 Where R is the reflected image;

[0052] I——Photographic image.

[0053] In the decomposition network, T-FAS decomposes the training image into a reflection image and an illumination image; in the adjustment step, the Relight-Net network based on the Encoder-Decoder structure enhances the illumination image; finally, the adjusted illumination image and reflection image are reconstructed to obtain the enhanced result. Under the constraint that the low / normal illumination images share the same reflection image and illumination image smoothness, the T-FAS algorithm extracts the same reflection image from different illumination images through network training. The Relight-Net network structure is as follows: Fig.11 shown.

[0054] Traditional image decomposition algorithms usually use images to convolve with various filter operators to obtain illumination images, such as Gaussian filtering, Kalman filtering, bilateral filtering, etc. Although the method is simple to operate, it cannot adapt to various complex scenes due to the small number of filter operator parameters. The T-FAS algorithm uses a convolutional neural network to learn the relationship between low illumination and normal illumination in a data-driven manner. The decomposition network first extracts features by a convolutional layer with a convolution kernel size of 3×3. The feature map is mapped to a reflection image R and an illumination image I through a mapping structure composed of six convolutional layers. Finally, the Sigmoid function is used to constrain the reflection image and the illumination image to be between [0, 1].

[0055] The T-FAS algorithm enhances the low-light illumination image obtained by decomposition based on the Encoder-Decoder structure. This structure introduces skip connections, establishes channels between feature maps of the same scale, and performs feature superposition at the pixel level. This can not only solve the gradient disappearance caused by the superposition of convolutional layers, but also facilitate the back propagation of gradients and accelerate network convergence. T-FAS converts the low-light illumination image I lowAs input, it first passes through the downsampling layer in sequence to obtain the illumination information contained in the large-scale feature map, and then reconstructs the local illumination information through the upsampling operation. Each convolution layer sets 64 convolution kernels of size 3×3, and finally passes through a convolution kernel of size 1×1 to compress the number of channels and simplify the features to obtain the final reconstructed illumination image.

[0056] In order to better extract the target features and speed up the network training process, we first borrow the idea of ​​transfer learning, that is, use the weight file obtained through training with the dataset image as input as the initial value of the model weight, such as Fig.12 As shown in the figure, the data set contains many types of targets, and the initial weight file already has a good target feature extraction capability. On this basis, the network training can converge quickly. At the same time, in order to avoid the damage to the front-layer weight file caused by the difference in target features during the initial training process, a two-stage training method is adopted. That is, when training the first five generations, the first 226 layers of the network, that is, the feature extraction layer, are frozen. In the subsequent training process, the feature extraction layer is unfrozen and the entire network is trained.

[0057] Firstly, the existing power transmission line image sample set is subjected to denoising and color correction. Then, the location and type of faults are marked in the power transmission line images. Finally, the power transmission line images are rotated and scaled to expand the data set and realize data set generation.

[0058] The dataset is then divided into training and testing sets.

[0059] The image data of the training set is first subjected to feature extraction by a VGG16 convolutional neural network; then, a priori box design and prior knowledge transfer learning are introduced into the RPN to generate a set of region proposals; finally, the output of the model is obtained through classification and regression to obtain the data model, and the operation is repeated and iterated to obtain the optimal hyperparameters. The optimal hyperparameter combination is input into the model to obtain the trained prior knowledge Faster R-CNN image recognition model, and then the test set is input into the trained model to verify the effectiveness of the model.

[0060] Then, drone equipment is used to capture images of the power transmission lines. The images are then transmitted to a cloud server for denoising and color correction. Finally, the processed images are input into the trained prior knowledge Faster R-CNN image recognition model to obtain recognition results, and the recognition algorithm is optimized.

[0061] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A drone nest based on fault image enhancement and optimization recognition, comprising a drone nest body (1), characterized in that: A landing platform (3) is arranged inside the drone nest body (1), a drone (4) is arranged at the upper end of the landing platform (3), a millimeter wave radar (11) is installed on the lower end surface of the drone (4), a three-dimensional laser scanner (12) is installed on one side of the millimeter wave radar (11), the input end of the drone nest body (1) is electrically connected to an edge algorithm integrated machine (13), and the output end of the drone nest body (1) is electrically connected to a gimbal.

2. The drone nest based on fault image enhancement and optimization recognition according to claim 1 is characterized by: A protective cover (2) is provided at the top of the drone nest body (1); a lead screw installation groove (8) is provided on the upper surface of the drone nest body (1); and a ball screw guide rail is installed inside the lead screw installation groove (8); and the protective cover (2) and the drone nest body (1) are both semicircular.

3. The drone nest based on fault image enhancement and optimization recognition according to claim 2 is characterized in that: A fixed base (5) is provided at the lower end of the drone nest body (1), a fixed frame (9) is welded to the rear surface of the fixed base (5), and one end of the fixed frame (9) is threadedly fixed to the drone nest body (1) via a threaded fastener (10).

4. The drone nest based on fault image enhancement and optimization recognition according to claim 3 is characterized by: The front surface of the drone nest body (1) is provided with an inspection opening (6), and a protective door (7) is provided inside the inspection opening (6).

5. The drone nest based on fault image enhancement and optimization recognition according to claim 1 is characterized by: The edge algorithm integrated machine (13) includes COSRTK real-time line patrol error correction technology, T-FAS unmanned aerial vehicle image enhancement and optimization recognition algorithm and power patrol fault image data set.

6. The drone nest based on fault image enhancement and optimization recognition according to claim 5 is characterized by: The COSRTK real-time line patrol and error correction technology is implemented based on the Kalman algorithm.

7. The drone nest based on fault image enhancement and optimization recognition according to claim 5 is characterized by: The electric patrol fault image dataset is based on open source datasets at home and abroad, plus more than 20,000 adversarial network algorithms, and is formed by combining deep learning self-collected datasets.

8. The drone nest based on fault image enhancement and optimization recognition according to claim 5 is characterized by: The COSRTK real-time line patrol and error correction technology includes RTK technology and three-dimensional laser point cloud accurate hazard level judgment technology, and the drone (4) uses RTK technology for positioning.

9. The drone nest based on fault image enhancement and optimization recognition according to claim 8 is characterized by: The unmanned aerial vehicle (4) consists of three parts: an unmanned aerial vehicle carrier system, an automatic flight system and a ground general control system.

10. The drone nest based on fault image enhancement and optimization recognition according to claim 5, characterized in that: The T-FAS UAV image enhancement and optimization recognition algorithm performs enhancement processing on the decomposed low-illuminance illumination image based on the Encoder-Decoder structure.

Images