Power distribution network self-adaptive inspection method and system based on image recognition and medium

Abstract

The invention provides a power distribution network adaptive inspection method and system based on image recognition, and a medium, and belongs to the technical field of power distribution network line inspection. The self-adaptive routing inspection method comprises the following steps: S1, driving a routing inspection unmanned aerial vehicle to perform self-adaptive routing inspection along a power distribution network line, and obtaining a routing inspection image in real time; s2, preprocessing the inspection image, and obtaining an identification result of the inspection image based on an inspection image defect identification model; s3, judging whether the identification result is within a defect warning range, if so, sending warning information and marking a corresponding inspection image; s4, obtaining a prediction result of the inspection image based on the inspection image defect prediction model; and S5, judging whether the prediction result is within the potential defect range, and if so, sending early warning information and marking a corresponding inspection image. According to the method, the convenience and efficiency of power distribution network line inspection and maintenance can be effectively improved in a mode of combining self-adaptive inspection, defect identification and prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid line inspection, and in particular relates to a distribution network adaptive inspection method, system and medium based on image recognition. Background Art

[0002] As a crucial component of the power grid, the distribution network serves as a vital link between large-scale transmission networks and users, and its operational stability is crucial. Therefore, regular inspections of distribution network lines are necessary to keep abreast of the operational status of the distribution lines, as well as changes in the surrounding environment and protection zones, to ensure power supply security.

[0003] At present, for the maintenance of distribution network lines, inspection drones are generally used to inspect the distribution network lines. However, as described in the Chinese invention patent with application number "202110737182.0", during the drone inspection process, manual verification of the inspection image is usually required, and the distribution network line maintenance is performed after defects are found. The whole process is relatively cumbersome, making it very difficult to implement rapid inspection and maintenance of the distribution network lines. Summary of the Invention

[0004] The technical problem to be solved by the present invention is how to improve the convenience and efficiency of inspection and maintenance. In view of the shortcomings of the existing technology, a distribution network adaptive inspection method, system and storage medium based on image recognition are provided.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] In a first aspect, the present invention provides a distribution network adaptive inspection method based on image recognition, comprising:

[0007] S1: Drive the inspection drone to conduct adaptive inspections along the distribution network lines and obtain inspection images in real time;

[0008] S2. Preprocessing the inspection image and obtaining a recognition result of the inspection image based on an inspection image defect recognition model;

[0009] S3. Determine whether the recognition result is within the defect warning range. If so, send a warning message and mark the corresponding inspection image;

[0010] S4. Obtaining a prediction result of the inspection image based on the inspection image defect prediction model;

[0011] S5. Determine whether the prediction result is within the potential defect range. If so, send a warning message and mark the corresponding inspection image.

[0012] Compared with the prior art, the beneficial effects of the distribution network adaptive inspection method based on image recognition of the present invention include: first, the inspection drone is driven into the air through the mobile terminal of the distribution network adaptive inspection system based on image recognition, such as a remote control, and the adaptation of the inspection drone is turned on, so that the inspection drone moves along the distribution network line and collects inspection images in real time. The inspection images can be transmitted to the mobile terminal via the inspection drone, which is convenient for subsequent defect identification based on the inspection images, and the entire process does not require human participation, effectively improving the convenience of inspection; after the mobile terminal receives the inspection image in real time, the inspection image can be transmitted to the distribution station client, and the distribution station client has a built-in inspection image defect recognition model. At this time, the inspection image can be pre-processed first to ensure that the inspection image for subsequent defect identification meets the requirements, and then the inspection image is input into the inspection image defect recognition model. The inspection image can be identified by the inspection image defect recognition model to generate an identification result. At this time, the identification result can be compared with the defect warning range. If it is located within the defect warning range, If it is within the range of potential defects, it means that there is a defect in the distribution network line part in the inspection image, and a warning message is sent, so that the inspection and maintenance personnel can quickly receive the information and go for processing. At the same time, the inspection image will be marked for easy reference by the inspection and maintenance personnel, helping them to quickly locate the defect, thereby effectively improving the inspection and maintenance efficiency. On this basis, the platform master station has a built-in inspection image defect prediction model, and the distribution station client can synchronously transmit the inspection image to the platform master station. After the inspection image is input into the inspection image defect prediction model, a prediction result will be generated. At this time, the prediction result can be compared with the potential defect range. If it is within the potential defect range, it means that the distribution network line part in the inspection image will have a defect within a predictable time, and a warning message is sent, so that the inspection and maintenance personnel can quickly receive the information and go for processing, prevent problems before they occur, and effectively ensure the operational stability of the distribution network line. At the same time, the inspection image will be marked for easy reference by the inspection and maintenance personnel, helping them to quickly locate the defect, thereby effectively improving the inspection and maintenance efficiency.

[0013] Optionally, the acquiring of the inspection image defect recognition model in S2 includes:

[0014] S21, establishing an original sample library of images of the distribution network line, wherein the original sample library includes at least typical component images and defect images;

[0015] S22, performing image augmentation on the defect image to generate a first artificial sample, and adding the first artificial sample to the original image sample library to generate an image sample library;

[0016] S23. Construct an analysis and judgment model based on the FPN network and the Faster-RCNN network, train the analysis and judgment model based on the image sample library, and obtain the inspection image defect recognition model.

[0017] Optionally, the pre-processing of the inspection image in S2 specifically includes:

[0018] S24, decomposing the inspection image using a fast adaptive two-dimensional empirical mode decomposition algorithm to obtain multiple IMF components and a residual;

[0019] S25, distinguishing the noise IMF component from the signal IMF component, performing adaptive threshold processing on the noise IMF component using an optimal linear interpolation threshold function algorithm, and obtaining the denoised noise IMF component;

[0020] S26 , combining the signal IMF component and the denoised noise IMF component to obtain the reconstructed inspection image.

[0021] Optionally, before performing adaptive threshold processing on the noise IMF component by using the optimal linear interpolation threshold function algorithm in S25, the step further includes: optimizing threshold parameters by using a QPSO algorithm.

[0022] Optionally, the S1 includes:

[0023] S11, acquiring the inspection images of the inspection starting point area at multiple angles, and generating the inspection direction and initial target coordinates of the inspection drone according to the inspection images;

[0024] S12: driving the inspection drone to move along the inspection direction, and updating the inspection direction according to the inspection image acquired in real time.

[0025] Optionally, the S4 includes:

[0026] S41, obtaining historical inspection images of defective parts of the distribution network line;

[0027] S42, classifying the historical inspection images at preset time intervals to form a defect prediction sample library;

[0028] S43, inputting the defect prediction sample library into the learning model for training to generate the inspection image defect prediction model;

[0029] S44 , inputting the inspection image into the inspection image defect prediction model to obtain a prediction result of the inspection image.

[0030] Optionally, before S23 and after S22, the acquisition of the inspection image defect recognition model in S2 further includes: performing image augmentation on the marked inspection image to generate a second artificial sample, and adding the second artificial sample to the image sample library.

[0031] Optionally, before driving the inspection drone to adaptively inspect along the distribution network line in S1, the distribution network adaptive inspection based on image recognition further includes: obtaining inspection items, and selecting the corresponding inspection drone from the inspection drone group according to the inspection items.

[0032] In a second aspect, the present invention further provides a distribution network adaptive inspection system based on image recognition, characterized in that it includes:

[0033] A mobile terminal, the mobile terminal being communicatively connected to the inspection drone;

[0034] A power distribution station client, the power distribution station client being communicatively connected with the mobile terminal;

[0035] The platform master station is communicatively connected with the power distribution station client and the mobile terminal respectively.

[0036] Compared with the prior art, the beneficial effects of the distribution network adaptive inspection system based on image recognition of the present invention are the same as the beneficial effects of the distribution network adaptive inspection method based on image recognition as described above, and will not be repeated here.

[0037] In a third aspect, the present invention further provides a computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the above-mentioned distribution network adaptive inspection method based on image recognition.

[0038] Compared with the prior art, the beneficial effects of the computer storage medium of the present invention are the same as the beneficial effects of the distribution network adaptive inspection method based on image recognition as described above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention will be described in further detail below with reference to the accompanying drawings.

[0040] Figure 1 : Flowchart of the distribution network adaptive inspection method based on image recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to better understand the present invention, the content of the present invention is further clearly set forth below in conjunction with the examples, but the protection content of the present invention is not limited to the following examples. In the following description, a large number of specific details are provided in order to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details.

[0042] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0043] It should be noted that the modifications of "one" and "multiple" mentioned in the present invention are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0044] In a first aspect, an embodiment of the present invention provides a distribution network adaptive inspection method based on image recognition, including: S1, driving an inspection drone to adaptively inspect along the distribution network line and obtain inspection images in real time; S2, preprocessing the inspection image, and obtaining the recognition result of the inspection image based on the inspection image defect recognition model; S3, determining whether the recognition result is within the defect warning range, and if so, sending a warning message and marking the corresponding inspection image; S4, obtaining the prediction result of the inspection image based on the inspection image defect prediction model; S5, determining whether the prediction result is within the potential defect range, and if so, sending an early warning message and marking the corresponding inspection image.

[0045] Specifically, driving an inspection drone for adaptive inspection along distribution network lines involves adaptively planning the drone's location and inspection direction based on inspection images captured in real time by the drone, such as images of towers from various angles. The drone then hovers at a fixed position and captures images of fixed locations, such as tower insulators, to achieve adaptive cruising and automatic acquisition of inspection images. Inspection image preprocessing can include noise removal or brightness and color adjustment to meet the requirements of subsequent defect identification. Warning information can be text or image information sent to a mobile terminal to remind inspection and maintenance personnel to promptly repair the defect location, or it can be text or image information sent to a distribution station client to facilitate timely assignment of tasks to inspection and maintenance personnel based on the content. The recognition result can be a defect category, such as damaged conductor, broken insulator, or tilted tower. The defect warning range is preset with the defect category and corresponding image information to facilitate determination of whether the recognition result falls within the defect warning range. Correspondingly, the prediction result can also be a defect category, and the potential defect range is also preset with the defect category and corresponding image information.

[0046] In this embodiment, a mobile terminal, a power distribution station client, and a platform master station are set up to form a distribution network adaptive inspection system based on image recognition, so as to realize a distribution network adaptive inspection method based on image recognition by cooperating with an inspection drone, wherein the mobile terminal is connected to the inspection drone in communication, the power distribution station client is connected to the mobile terminal in communication, and the platform master station is connected to the power distribution station client and the mobile terminal in communication respectively. When performing the distribution network adaptive inspection, first, as shown in FIG. Figure 1 As shown in S1 in the figure, the mobile terminal of the distribution network adaptive inspection system based on image recognition, such as a remote control, drives the inspection drone to take off, turns on the self-adaptation of the inspection drone, and makes the inspection drone move along the distribution network line to collect inspection images in real time. The inspection images can be transmitted to the mobile terminal via the inspection drone, so as to facilitate the subsequent defect identification based on the inspection images. The whole process does not require human participation, which effectively improves the convenience of inspection. After the mobile terminal receives the inspection image in real time, the inspection image can be transmitted to the distribution station client. The distribution station client has a built-in defect recognition model based on the inspection image. At this time, Figure 1 As shown in S2 and S3 in the figure, the inspection image can be pre-processed first to ensure that the inspection image for subsequent defect identification meets the requirements, and then the inspection image is input into the inspection image defect recognition model. The inspection image can be identified by the inspection image defect recognition model to generate an identification result. At this time, the identification result can be compared with the defect warning range. If it is within the defect warning range, it means that there is a defect in the distribution network line part in the inspection image, so as to send a warning message, so that the inspection and maintenance personnel can quickly receive the information and go for processing. At the same time, the inspection image will also be marked for the convenience of inspection and maintenance personnel to review, helping inspection and maintenance personnel to quickly locate defects, thereby effectively improving the inspection and maintenance efficiency. On this basis The platform master station has a built-in inspection image defect prediction model. The distribution station client can synchronously transmit the inspection image to the platform master station. After the inspection image is input into the inspection image defect prediction model, a prediction result will be generated. At this time, the prediction result can be compared with the potential defect range. If it is within the potential defect range, it means that the distribution network line part in the inspection image will have defects within the predictable time, so as to send early warning information, so that the inspection and maintenance personnel can quickly receive the information and deal with it, prevent problems before they occur, and effectively ensure the operational stability of the distribution network line. At the same time, the inspection image will also be marked for the inspection and maintenance personnel to check and help them quickly locate the location, thereby effectively improving the inspection and maintenance efficiency.

[0047] It should be noted that in this embodiment, a defect review module can also be built into the platform master station to perform secondary defect detection on the received inspection images through manual or AI, thereby reducing the probability of defects being missed and further improving the comprehensiveness and stability of inspection and maintenance.

[0048] It should be noted that in this embodiment, the distribution station customer has a built-in drone management module, an inspection task management module and a statistical analysis module. The drone management module can perform drone inventory management, drone status monitoring, drone service tracking and cumulative flight data storage. The inspection task management module can perform inspection work order management, equipment inventory management and inspection report generation, and the statistical analysis module can perform statistics on inspection operation data and generate analysis reports.

[0049] It should be noted that, in this embodiment, the platform main station also has a built-in situation display module, a UAV monitoring module and a defect identification management module. The situation display module can display the overall data of the UAV and inspection operations as well as warning and early warning information. The UAV monitoring module can obtain and store the UAV flight log and status data. The defect identification management module can conduct defect review, model evaluation, defect retrieval, distribution statistics, model training and testing based on defect identification and prediction.

[0050] Optionally, the acquisition of the inspection image defect recognition model in S2 includes: S21, establishing an image original sample library of the distribution network line, the image original sample library at least includes typical component images and defect images; S22, performing image augmentation on the defect image to generate a first artificial sample, adding the first artificial sample to the image original sample library, and generating an image sample library; S23, constructing an analysis and judgment model based on the FPN network and the Faster-RCNN network, training the analysis and judgment model based on the image sample library, and obtaining the inspection image defect recognition model.

[0051] In this optional embodiment, an original sample library of images of distribution network lines is first constructed, which includes at least typical component images and defect images. Specifically, the images in the original sample library are generally inspection images of distribution network lines collected by drones, infrared cameras, visible light cameras and other equipment, covering typical components such as conductors, insulators, hardware, and towers. Infrared thermal imaging images (to detect overheating defects) or laser point cloud data (to detect structural deformation) can also be added. For historical images of the same component, a time series sample library can be constructed to detect progressive defects. At the same time, the image annotation tool LabelImg or CVAT is used to annotate components (such as insulators, crossarms) and defects (such as cracks, rust) in the inspection images. , fracture) are annotated with bounding boxes. After the annotation is completed, the inspection images are classified. The component images in normal state (such as intact insulators and non-rusted hardware) are typical component images, and the images with typical defects (such as insulator bursts, wire breakage, and tower tilt) are defect images. The number of typical component images and defect images is adjusted to ensure that the ratio of defect samples to normal samples is reasonable (such as 1:3) to avoid the model being biased towards the majority class. Then, by augmenting the defect images, the diversity of samples is increased, the recall rate of the inspection image defect recognition model finally constructed is improved, overfitting is avoided, and the robustness of the inspection image defect recognition model to illumination changes and occlusion is enhanced. Specifically, Image augmentation can be achieved through geometric transformations such as rotation, color perturbations such as brightness adjustment, and generative adversarial networks (GANs). 3D modeling software (such as Blender) can also be used to simulate defects under extreme weather conditions (icing, strong winds), and style transfer (such as background changes from plains to mountainous areas) can be added during augmentation to account for the differences in power grids in different regions. At the same time, low-quality augmented samples can be filtered manually or through automated tools such as pre-trained models to achieve the purpose of augmentation. Finally, an analysis and judgment model is constructed based on the FPN network and the Faster-RCNN network, and the analysis and judgment model is trained based on the image sample library. Through the learning and training of the analysis and judgment model, an inspection image defect recognition model can be finally obtained. Specifically, For the FPN network, multi-scale features are extracted through the backbone network. Through a top-down path, high-level semantic features are fused with low-level detail features to improve the detection ability of small objects (such as missing pins). For the Faster-RCNN network, candidate boxes are generated through the RPN (Region Proposal Network). The candidate regions are accurately pooled through ROIAlign to preserve spatial information. At the same time, the classification and regression heads output defect categories and bounding box coordinates. During the learning and training process, loss functions (classification loss and regression loss), optimizers, and data loading are used for training. At the same time, the CBAM module is added to the FPN to enhance the feature response of the defect area. The Swin Transformer is used to replace the ResNet to capture long-range dependencies. The defect classification, segmentation (such as Mask R-CNN) and severity assessment are jointly trained for multi-task learning.With this setup, the inspection image defect recognition model can effectively improve the detection and recognition accuracy of small targets (such as insulator cracks) by utilizing FPN's multi-scale feature fusion, while also reducing the false positive rate by utilizing Faster R-CNN's precise candidate box regression.

[0052] Optionally, the preprocessing of the inspection image in S2 specifically includes: S24, decomposing the inspection image by a fast adaptive two-dimensional empirical mode decomposition algorithm to obtain multiple IMF components and a residual;

[0053] S25. Distinguish the noise IMF component from the signal IMF component, perform adaptive threshold processing on the noise IMF component through the optimal linear interpolation threshold function algorithm, and obtain the denoised noise IMF component; S26. Merge the signal IMF component and the denoised noise IMF component to obtain a reconstructed inspection image.

[0054] Specifically, the fast adaptive two-dimensional empirical mode decomposition (BEMD) algorithm, based on BEMD, decomposes an image into multiple IMF components and a residual. Each IMF component represents spatial frequency information at a different scale. It optimizes the boundary effects and computational efficiency issues of traditional BEMD through local extreme point detection and surface interpolation (such as RBF or thin plate splines). It can also employ a pyramid decomposition strategy or GPU acceleration to effectively reduce the number of iterations. Furthermore, the fast adaptive two-dimensional empirical mode decomposition algorithm can also combine wavelet transforms to pre-decompose the image, effectively improving the accuracy of high-frequency component separation. It also adds white noise (CEEMDAN algorithm) to suppress modal aliasing.

[0055] It should be noted that the optimal linear interpolation threshold function dynamically adjusts the threshold based on the noise variance (such as the Donoho threshold) and the IMF energy to achieve threshold design. At the same time, the coefficients exceeding the threshold are linearly attenuated (soft threshold) or retained (hard threshold), and interpolation is performed between the two to balance fidelity and smoothness. A local window (such as 3×3) is introduced to calculate the spatially varying threshold to adapt to non-uniform noise.

[0056] It should be noted that when reconstructing the inspection image, the denoised noise IMF can be directly superimposed with the unprocessed signal IMF, and the residual retains the low-frequency information. In this process, the reconstructed image can be contrast enhanced (such as CLAHE) or sharpened (such as Laplace filtering).

[0057] In this optional embodiment, the inspection image is first decomposed by a fast adaptive two-dimensional empirical mode decomposition algorithm to obtain multiple IMF components and a residual, and the noise can be separated into high-frequency IMFs. The adaptive threshold suppresses the noise while retaining the edge (high gradient area), which facilitates subsequent denoising; then, the noise IMF component and the signal IMF component are distinguished, and the noise IMF component is adaptively thresholded by the optimal linear interpolation threshold function algorithm to obtain the denoised noise IMF component; finally, the signal IMF component and the denoised noise IMF component are merged to obtain a reconstructed inspection image, so that the inspection image for defect identification can meet the identification requirements and ensure the accuracy of defect identification.

[0058] Optionally, before performing adaptive threshold processing on the noise IMF component by using the optimal linear interpolation threshold function algorithm in S25, the method further includes: optimizing the threshold parameter by using a QPSO algorithm.

[0059] In this optional embodiment, in order to further improve the image denoising effect, the quantum-behaved particle swarm optimization (QPSO) algorithm can be used to adaptively optimize the threshold parameters of the IMF component (such as the Donoho threshold, the soft / hard threshold interpolation ratio, etc.), so as to achieve an optimal balance between noise suppression and detail preservation.

[0060] Optionally, S1 includes: S11, obtaining inspection images of the inspection starting point area at multiple angles, and generating the inspection direction and initial target coordinates of the inspection drone based on the inspection images; S12, driving the inspection drone to move along the inspection direction, and updating the inspection direction based on the inspection images obtained in real time.

[0061] Specifically, before conducting adaptive inspections, digital twin modeling is first performed. Using LiDAR or historical inspection data, a 3D model of the distribution network is constructed. The coordinates of key components (conductors, insulators, and towers) are annotated. This effectively reduces the initial positioning error of the inspection drone and improves path planning efficiency. Sensor calibration is also performed, calibrating the relative positions of multiple cameras (wide-angle and telephoto), LiDAR, and IMU to ensure data alignment in time and space.

[0062] In this optional embodiment, after the inspection drone is launched, multi-angle image acquisition is first performed, and the inspection drone hovers in the inspection starting point area. The camera is controlled by the gimbal to shoot at least 3 sets of images from different angles (looking down 30°, horizontal, and looking up 30°), and visible light (RGB) and infrared (IR) images are synchronously acquired for subsequent temperature anomaly detection. At the same time, the YOLOv7+Transformer model is used to detect the wires and towers in the image, and the pixel coordinates are output. Then, the pixel coordinates are converted into the initial target coordinates in the world coordinate system (such as the first base tower position) through stereo vision triangulation (binocular camera) or LiDAR point cloud matching. Finally, the inspection direction is calculated, and the initial heading angle (yaw angle ψ) is generated according to the starting point and the target coordinates. At the same time, the digital twin model is combined to verify whether the path avoids known obstacles (such as trees under the high-voltage corridor); on this basis, dynamic path update is performed, first real-time image stream processing is performed, and the airborne JetsonAGX Orin runs lightweight models (such as MobileNetV3+DeepSORT), tracks wires at 30fps and calculates deviation distances (Δx, Δy), dynamically adjusts the drone's roll and yaw angles based on the deviation distance, and adapts the weight coefficients to wind speed (Kalman filtering estimates wind disturbances). If a temporary obstacle (such as a flying bird) is detected, local replanning (RRT* algorithm) is triggered to generate a flight path. Finally, the target coordinates are updated every 200 meters, and the visual and IMU data are integrated through SLAM (ORB-SLAM3) to correct the accumulated positioning error, ultimately achieving precise adaptive cruising for the inspection drone.

[0063] It should be noted that in other embodiments of the present invention, the traditional artificial potential field method can be combined with the AI algorithm, and the jitter suppression adjustment coefficient can be set to obtain an improved artificial potential field method, and a global obstacle avoidance path can be collaboratively searched. Local targets are set on this path to divide the complex area into simple local areas, and local path planning is implemented. The final planned path is obtained by integrating the results of each local path planning.

[0064] Optionally, S4 includes: S41, obtaining historical inspection images of defective parts of the distribution network lines; S42, classifying the historical inspection images at preset time intervals to form a defect prediction sample library; S43, inputting the defect prediction sample library into the learning model for training to generate an inspection image defect prediction model; S44, inputting the inspection image into the inspection image defect prediction model to obtain the prediction result of the inspection image.

[0065] In this optional embodiment, first, historical inspection images of defective parts of the distribution network line are obtained, wherein the historical inspection images of defective parts of the distribution network line (such as insulator damage, wire corrosion, etc.) are collected, and it is necessary to ensure that the images cover a variety of defect types, environmental conditions (lighting, weather) and equipment status. The data sources may include drone inspections, cameras or manual photography, etc., and can be combined with multimodal data such as infrared thermal imaging and lidar point clouds to effectively improve the comprehensiveness of subsequent defect predictions; then, the historical inspection images are classified at preset time intervals to form a defect prediction sample library, wherein the historical images are classified at preset time intervals (such as monthly, quarterly) to form a time-series defect sample library, based on the defect evolution At the same time, a label system (such as defect level and location label) and metadata (such as ambient temperature and load current) are introduced to build a structured sample library to improve the comprehensiveness of the defect prediction sample library. The defect prediction sample library is then input into the learning model for training to generate an inspection image defect prediction model. The defect prediction sample library can be used to train learning models such as CNN and Transformer. Lightweight models such as MobileNet are suitable for edge deployment and can be placed on the distribution station client, while complex models such as ResNet are suitable for high-precision cloud prediction and can be placed on the platform master station. During training, time series data can be combined with LSTM or 3D CNN to capture defect evolution characteristics. Transfer learning (pre-trained model + fine-tuning) or federated learning (multi-institutional data collaboration) can also be used to optimize training efficiency. Finally, the inspection image is input into the inspection image defect prediction model to obtain the prediction result of the inspection image. The new inspection image is input into the model to output the defect type, location and severity prediction. At the same time, it can be combined with real-time data streams (such as drone inspection video streams) to achieve online detection and alarm.

[0066] Optionally, before S23 and after S22 , the acquisition of the inspection image defect recognition model in S2 further includes: performing image augmentation on the marked inspection image to generate a second artificial sample, and adding the second artificial sample to the image sample library.

[0067] In this optional embodiment, in order to ensure the timeliness and diversity of the image sample library, while the defect image is augmented to generate a first artificial sample and added to the image sample library, the inspection image acquired in real time can also be augmented to generate a second artificial sample. The second artificial sample can be added to the image sample library, effectively improving the timeliness and diversity of the image sample library.

[0068] Optionally, before driving the inspection drone to adaptively inspect along the distribution network line in S1, the distribution network adaptive inspection based on image recognition further includes: obtaining inspection items, and selecting a corresponding inspection drone from the inspection drone group according to the inspection items.

[0069] In this optional embodiment, the inspection items can be divided into UAV power patrol, UAV defect patrol, UAV fault line detection and UAV tree barrier measurement, etc. For different inspection items, the same inspection UAV can be used to load different types of tools. For example, the UAV power patrol system is based on digital photos and uses photogrammetry as a technical means, while the UAV defect patrol controls the unmanned gimbal, takes aerial photos of various parts of the line, and provides real-time feedback of the corresponding video images, which helps to fully detect the ground conductors and hardware that are blocked on the ground. Therefore, inspection UAVs equipped with different types of tools or different types of inspection UAVs can also be organized into inspection UAV groups. Before conducting an inspection, you can select the appropriate inspection drone according to the inspection project to effectively improve the efficiency of inspection and maintenance; on this basis, in the drone inspection of overhead lines, you can use a combination of unmanned helicopters and fixed-wing aircraft, or you can choose other drone combinations according to actual conditions to effectively complete the overhead line inspection work. For example, if it is known that a fault has occurred somewhere on the overhead line, you can use a combination of small rotor-type drones and helicopter-type drones to complete the inspection work. In this way, the advantages and disadvantages of the two can complement each other. Therefore, for the same inspection project, different types of inspection drones can be selected for combined use.

[0070] In a third aspect, an embodiment of the present invention provides an adaptive inspection system for a distribution network based on image recognition, comprising: a mobile terminal, which is communicatively connected to an inspection drone; a distribution station client, which is communicatively connected to the mobile terminal; and a platform master station, which is communicatively connected to the distribution station client and the mobile terminal respectively.

[0071] The technical effects of the distribution network adaptive inspection system based on image recognition in this embodiment are similar to the technical effects of the above-mentioned distribution network adaptive inspection method based on image recognition, and will not be repeated here.

[0072] In a third aspect, an embodiment of the present invention provides a computer storage medium storing a computer program, which implements the above-mentioned distribution network adaptive inspection method based on image recognition when executed by a processor.

[0073] The technical effect of the computer storage medium in this embodiment is similar to the technical effect of the above-mentioned distribution network adaptive inspection method based on image recognition, and will not be repeated here.

[0074] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A distribution network adaptive inspection method based on image recognition, characterized in that: include: S1. Driving the inspection drone to adaptively inspect along the distribution network line and acquiring inspection images in real time; S2. Preprocessing the inspection image and obtaining a recognition result of the inspection image based on an inspection image defect recognition model; S3. Determine whether the recognition result is within the defect warning range. If so, send a warning message and mark the corresponding inspection image; S4. Obtaining a prediction result of the inspection image based on the inspection image defect prediction model; S5. Determine whether the prediction result is within the potential defect range. If so, send a warning message and mark the corresponding inspection image.

2. The method for adaptive inspection of distribution network based on image recognition according to claim 1, characterized in that: The acquisition of the inspection image defect recognition model in S2 includes: S21, establishing an original sample library of images of the distribution network line, wherein the original sample library includes at least typical component images and defect images; S22, performing image augmentation on the defect image to generate a first artificial sample, and adding the first artificial sample to the original image sample library to generate an image sample library; S23. Construct an analysis and judgment model based on the FPN network and the Faster-RCNN network, train the analysis and judgment model based on the image sample library, and obtain the inspection image defect recognition model.

3. The distribution network adaptive inspection method based on image recognition according to claim 2, characterized in that: The pre-processing of the inspection image in S2 specifically includes: S24, decomposing the inspection image using a fast adaptive two-dimensional empirical mode decomposition algorithm to obtain multiple IMF components and a residual; S25, distinguishing the noise IMF component from the signal IMF component, performing adaptive threshold processing on the noise IMF component using an optimal linear interpolation threshold function algorithm, and obtaining the denoised noise IMF component; S26 , combining the signal IMF component and the denoised noise IMF component to obtain the reconstructed inspection image.

4. The method for adaptive inspection of distribution network based on image recognition according to claim 3, characterized in that: In the S25 , before the adaptive threshold processing is performed on the noise IMF component by the optimal linear interpolation threshold function algorithm, the method further includes: optimizing the threshold parameters by the QPSO algorithm.

5. The distribution network adaptive inspection method based on image recognition according to claim 1, characterized in that: Said S1 comprises: S11, acquiring the inspection images of the inspection starting point area at multiple angles, and generating the inspection direction and initial target coordinates of the inspection drone according to the inspection images; S12: driving the inspection drone to move along the inspection direction, and updating the inspection direction according to the inspection image acquired in real time.

6. The method for adaptive inspection of distribution network based on image recognition according to claim 1, characterized in that: The S4 includes: S41, obtaining historical inspection images of defective parts of the distribution network line; S42, classifying the historical inspection images at preset time intervals to form a defect prediction sample library; S43, inputting the defect prediction sample library into the learning model for training to generate the inspection image defect prediction model; S44 , inputting the inspection image into the inspection image defect prediction model to obtain a prediction result of the inspection image.

7. The distribution network adaptive inspection method based on image recognition according to claim 2, characterized in that: Before S23 and after S22 , the acquisition of the inspection image defect recognition model in S2 further includes: performing image augmentation on the marked inspection image to generate a second artificial sample, and adding the second artificial sample to the image sample library.

8. The distribution network adaptive inspection method based on image recognition according to claim 1, characterized in that: Before driving the inspection drone to adaptively inspect along the distribution network line in S1, the distribution network adaptive inspection based on image recognition further includes: obtaining inspection items, and selecting the corresponding inspection drone from the inspection drone group according to the inspection items.

9. A distribution network adaptive inspection system based on image recognition, characterized in that: include: A mobile terminal, the mobile terminal being communicatively connected to the inspection drone; A power distribution station client, the power distribution station client being communicatively connected with the mobile terminal; The platform master station is communicatively connected with the power distribution station client and the mobile terminal respectively.

10. A computer storage medium, characterized in that It stores a computer program, which, when executed by a processor, implements the distribution network adaptive inspection method based on image recognition as described in any one of claims 1 to 8.

Images