A photovoltaic defect identification method and system based on a UAV and an electronic device

Abstract

This application provides a method, system, and electronic device for photovoltaic defect identification based on unmanned aerial vehicles (UAVs), relating to the field of image processing technology. The method includes: acquiring an initial image of a photovoltaic array using a UAV; preprocessing the initial image to obtain a processed image; performing target detection on the processed image using a YOLOv11 model to obtain a first detection result, the first detection result including first position information of the target photovoltaic string; cropping the processed image based on the first position information to obtain a target image; and detecting the target image using a YOLOv11-OBB model to obtain a second detection result. When a target defect is determined to exist in the target photovoltaic string, the first position information, the defect type information of the target photovoltaic panel, and the target row and column information are output. Implementing the technical solution provided in this application improves the accuracy and efficiency of photovoltaic defect identification.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a photovoltaic defect identification method, system, and electronic device based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the rapid development of the photovoltaic industry and the continuous expansion of photovoltaic power plants, the demand for photovoltaic panel inspection and maintenance is increasing. Among related technologies, photovoltaic defect detection mainly relies on manual inspection or fixed camera monitoring. Manual inspection is inefficient, labor-intensive, and prone to missed or false detections due to human factors. Fixed camera monitoring is limited by installation location and angle, making it difficult to fully cover the photovoltaic array, and its defect identification capability in complex environments is limited. Furthermore, although there are methods using drones for photovoltaic inspection, the accuracy and efficiency of defect identification still need improvement. For example, existing drone inspection methods often use simple image acquisition followed by direct defect identification without effective image preprocessing and targeted target detection, resulting in low accuracy in defect identification under complex backgrounds and an inability to accurately output the specific location and type of defects. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides a photovoltaic defect identification method, system, and electronic device based on unmanned aerial vehicles (UAVs).

[0004] In a first aspect, this application provides a photovoltaic defect identification method based on a drone, comprising: acquiring an initial image of a photovoltaic array using a drone; preprocessing the initial image to obtain a processed image; performing target detection on the processed image using a YOLOv11 model to obtain a first detection result, the first detection result including first position information of a target photovoltaic string, wherein the target photovoltaic string is any photovoltaic string in the photovoltaic array; cropping the processed image according to the first position information to obtain a target image, wherein the target image includes the target photovoltaic string; detecting the target image using a YOLOv11-OBB model to obtain a second detection result; and when the second detection result indicates that a target defect exists in the target photovoltaic string, outputting the first position information, the defect type information of the target photovoltaic panel, and the target row and column information, wherein the target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect.

[0005] By adopting the above technical solutions, using drones to collect images can avoid the problems of low efficiency, easy omissions and false detections in manual inspections, and incomplete coverage of fixed camera monitoring. Preprocessing of the initial images can optimize image quality. Using the YOLOv11 model for target detection can obtain the first position information of the target photovoltaic string, so that the target image can be obtained by subsequent image cropping, and the target image can focus on the target photovoltaic string. Using the YOLOv11-OBB model to detect the target image, and outputting the first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel when a target defect is detected, the accurate identification and location of photovoltaic defects are realized, improving the accuracy and efficiency of photovoltaic defect identification, and accurately outputting the specific location and type information of defects.

[0006] Optionally, the YOLOv11-OBB model is used to detect the target image to obtain a second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, the first location information, the defect type information of the target photovoltaic panel, and the target row and column information are output. This includes: the YOLOv11-OBB model uses a rotating bounding box to detect a group of photovoltaic panels in the target image to obtain a set of location information, and identifies the local images corresponding to each rotating bounding box to obtain a set of identification results. The target photovoltaic string includes a group of photovoltaic panels, and the second detection result includes a set of location information and a set of identification results. When a set of identification results indicates that there is a target defect in the target photovoltaic panel, the target row and column information of the target photovoltaic panel is determined based on the set of location information, and the first location information, defect type information, and target row and column information are output. The set of identification results includes the defect type information.

[0007] By adopting the above technical solution, the YOLOv11-OBB model is used to detect a set of photovoltaic panels in the target image through a rotating frame to obtain a set of position information. The corresponding local image of the rotating frame is then identified to obtain a set of recognition results. This allows for more accurate positioning of photovoltaic panels and identification of defects. When a target photovoltaic panel is found to have a defect, the target row and column information of the target photovoltaic panel is determined based on this set of position information. The first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel can be output. This effectively improves the accuracy of photovoltaic defect identification and the completeness of the output information, solving the problems of low defect identification accuracy and inability to accurately output the specific location and type information of defects in related technologies.

[0008] Optionally, the target row and column information of the target photovoltaic panel is determined based on a set of location information, including: determining the center coordinates of each rotating frame based on a set of location information, wherein each photovoltaic panel in a set of photovoltaic panels corresponds to a rotating frame; determining the arrangement direction of each photovoltaic panel in a set of photovoltaic panels based on the long side direction of each rotating frame; and sorting according to the x or y values ​​of the center coordinates of each rotating frame to obtain the row and column distribution result of a set of photovoltaic panels, wherein the row and column distribution result includes the target row and column information.

[0009] By adopting the above technical solution, the center coordinates of each rotating frame are determined based on a set of position information, which can provide basic data for subsequent calculations; the arrangement direction of photovoltaic panels is determined based on the long side direction of the rotating frame, which can clearly show the arrangement of photovoltaic panels; the row and column distribution results are obtained by sorting according to the x or y values ​​of the center coordinates of the rotating frame to determine the target row and column information, which helps to accurately determine the position of the target photovoltaic panel in the photovoltaic string, thereby improving the accuracy and efficiency of photovoltaic defect identification and solving the problem that existing detection methods are difficult to accurately output the specific location information of defects.

[0010] Optionally, the first position information includes first center coordinate information, first length information, and first width information; a set of position information includes multiple second position information, and each photovoltaic panel in a set of photovoltaic panels corresponds to one second position information, the second position information including second center coordinate information, second length information, second width information, and rotation angle information.

[0011] By adopting the above technical solution, it is clear that the first position information includes the first center coordinate information, the first length information, and the first width information. A set of position information includes multiple second position information. Each second position information includes the second center coordinate information, the second length information, the second width information, and the rotation angle information. Each photovoltaic panel corresponds to one second position information. This can accurately characterize the position and size of the target photovoltaic string and photovoltaic panel, thereby more accurately locating and identifying the target photovoltaic panel and its defects, and improving the accuracy and precision of photovoltaic defect identification.

[0012] Optionally, the yolov11-obb model uses InceptionNeXt as the backbone structure and employs a channel-enhanced feature pyramid network. The channel-enhanced feature pyramid network dynamically adjusts the weights of different channels during feature fusion through an adaptive channel weighting mechanism to enhance defect-related features.

[0013] By adopting the above technical solution, the yolov11-obb model uses InceptionNeXt as the backbone structure and employs a channel-enhanced feature pyramid network that dynamically adjusts the weights of different channels during feature fusion through an adaptive channel weighting mechanism. This enhances defect-related features and thus improves the accuracy and precision of defect identification.

[0014] Optionally, the yolov11-obb model introduces a coordinate attention mechanism in the detection head, so that image features correspond one-to-one with positional information. Also, the yolov11-obb model uses the KFIoU loss function instead of RotatedBboxLoss. The KFIoU loss function includes a scale-insensitive center point loss and a distance-independent term.

[0015] By adopting the above technical solutions, the introduction of a coordinate attention mechanism in the detection head can make image features correspond one-to-one with position information, thereby improving the accuracy of target position judgment during detection. The KFIoU loss function is used to replace RotatedBboxLoss, which is based on Gaussian modeling and Gaussian product approximation SkewIoU, and includes scale-insensitive center point loss and distance-independent terms, which can more accurately reflect the degree of overlap between target boxes and improve the accuracy of the model in target detection and localization.

[0016] Optionally, the yolov11-obb model is trained using defect sample data, which is obtained through one of the following methods: generating defect sample data using a generative adversarial network (GAN); obtaining defect sample data by simulating wind, sand, snow, and crack defects through image synthesis; or obtaining defect sample data by augmenting a small number of defect samples using transfer learning.

[0017] By adopting the above technical solution, the YOLOv11-OBB model is trained by using Generative Adversarial Networks (GANs) to generate defect sample data, simulating wind-blown sand and snow cracks through image synthesis, and augmenting the few-sample defects with transfer learning. This increases the amount of defect sample data used for training, improves the situation with few samples, and enables the model to better learn various photovoltaic defect features. As a result, the accuracy and generalization ability of the YOLOv11-OBB model for photovoltaic defect identification in UAV-based photovoltaic defect identification methods are improved.

[0018] Optionally, the initial image can be preprocessed to obtain a processed image, including: adjusting the brightness and enhancing the contrast of the initial image to obtain the processed image.

[0019] By adopting the above technical solution, the initial image of the acquired photovoltaic array is processed with brightness adjustment and contrast enhancement, which can improve the image quality and provide clearer and more accurate images for subsequent target detection using the YOLOv11 model and target image detection using the YOLOv11-OBB model. This improves the accuracy and efficiency of photovoltaic defect identification and reduces misjudgments and omissions caused by poor image quality.

[0020] Optionally, the second detection result also includes the area information of the target defect. The above method further includes: determining the target level of the target defect based on the defect type information and area information; and recommending the corresponding target maintenance strategy based on the target level.

[0021] By adopting the above technical solution, when using drones to identify photovoltaic defects, the area information of the target defect can be obtained. The target level of the target defect can be determined based on the defect type information and area information, and the corresponding maintenance strategy can be recommended based on the target level. This helps to more accurately assess the defect status, take appropriate maintenance measures, and improve the maintenance efficiency and effectiveness of photovoltaic panels.

[0022] In a second aspect of this application, a photovoltaic defect identification system based on a drone is also provided, comprising: an acquisition module for acquiring an initial image of a photovoltaic array using a drone; a processing module for preprocessing the initial image to obtain a processed image; a first detection module for performing target detection on the processed image using a YOLOv11 model to obtain a first detection result, the first detection result including first position information of a target photovoltaic string, wherein the target photovoltaic string is any photovoltaic string in the photovoltaic array; an acquisition module for cropping the processed image according to the first position information to obtain a target image, wherein the target image includes a target photovoltaic string; and a second detection module for detecting the target image using a YOLOv11-OBB model to obtain a second detection result, wherein when the second detection result indicates that a target defect exists in the target photovoltaic string, the module outputs the first position information, the defect type information of the target photovoltaic panel, and the target row and column information, wherein the target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect.

[0023] In a third aspect of this application, an electronic device is also provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the program to implement the method steps of any of the above claims.

[0024] In a fourth aspect of this application, a computer-readable storage medium is also provided, which stores instructions that, when executed by a processor, perform the method steps described above.

[0025] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: 1. Using drones to collect images can avoid the problems of low efficiency, easy omissions and false detections in manual inspections, and incomplete coverage of fixed camera monitoring. Using the YOLOv11 model for target detection can obtain the first position information of the target photovoltaic string. The YOLOv11-OBB model is used to detect the target image, and when a target defect is detected, it outputs the first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel. This achieves accurate identification and location of photovoltaic defects, improves the accuracy and efficiency of photovoltaic defect identification, and can accurately output the specific location and type information of defects. 2. When it is determined that the target photovoltaic panel has a target defect, the target row and column information of the target photovoltaic panel can be determined based on the set of position information. The first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel can be output, which effectively improves the accuracy of photovoltaic defect identification and the completeness of output information, and solves the problems of low defect identification accuracy and inability to accurately output the specific location and type information of defects in related technologies. 3. The yolov11-obb model uses InceptionNeXt as the backbone structure and employs a channel-enhanced feature pyramid network that dynamically adjusts the weights of different channels during feature fusion through an adaptive channel weighting mechanism. This enhances defect-related features and thus improves the accuracy and precision of defect identification. 4. By using Generative Adversarial Networks (GANs) to generate defect sample data, simulating wind-blown sand and snow-covered cracks in images to obtain defect sample data, and using transfer learning to augment the few sample defects to obtain defect sample data for training the YOLOv11-OBB model, the amount of defect sample data used for training can be increased, the situation of few sample data can be improved, and the model can better learn various photovoltaic defect features, thereby improving the accuracy and generalization ability of the YOLOv11-OBB model in photovoltaic defect identification methods based on UAVs. Attached Figure Description

[0026] Figure 1 This is a flowchart of a photovoltaic defect identification method based on an unmanned aerial vehicle (UAV) provided in an embodiment of this application; Figure 2 This is an example diagram of photovoltaic defect detection provided in an embodiment of this application; Figure 3 This is an example of a photovoltaic defect identification result provided in an embodiment of this application. Figure 1 ; Figure 4 This is an example of a photovoltaic defect identification result provided in an embodiment of this application. Figure 2 ; Figure 5 This is a structural block diagram of a photovoltaic defect identification system based on an unmanned aerial vehicle (UAV) according to an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0027] Explanation of reference numerals in the attached figures: 600 - Electronic device; 601 - Processor; 602 - Communication bus; 603 - User interface; 604 - Network interface; 605 - Memory. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0029] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0030] In the description of the embodiments of this application, the term "multiple" means two or more. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0031] This application provides a photovoltaic defect identification method based on unmanned aerial vehicles (UAVs), referring to... Figure 1 , Figure 1 This is a flowchart of a photovoltaic defect identification method based on an unmanned aerial vehicle (UAV) according to an embodiment of this application. The method includes: Step S101: Use a drone to acquire initial images of the photovoltaic array; Step S102: Preprocess the initial image to obtain the processed image; Step S103: Use the YOLOv11 model to perform target detection on the processed image to obtain a first detection result. The first detection result includes the first position information of the target photovoltaic string, wherein the target photovoltaic string is any photovoltaic string in the photovoltaic array. Step S104: The processed image is cropped according to the first location information to obtain the target image, wherein the target image includes the target photovoltaic string; Step S105: Use the yolov11-obb model to detect the target image and obtain the second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, output the first position information, the defect type information of the target photovoltaic panel, and the target row and column information, where the target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect.

[0032] Through the above steps, using drones to collect images can avoid the problems of low efficiency, easy omissions and false detections in manual inspections, and incomplete coverage of fixed camera monitoring. Preprocessing the initial images can optimize image quality. Using the YOLOv11 model for target detection can obtain the first position information of the target photovoltaic string, so that the target image can be obtained by subsequent image cropping, and the target image can focus on the target photovoltaic string. Using the YOLOv11-OBB model to detect the target image, and outputting the first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel when a target defect is detected, the accurate identification and location of photovoltaic defects is achieved, improving the accuracy and efficiency of photovoltaic defect identification, and accurately outputting the specific location and type information of the defect.

[0033] This embodiment uses a drone to acquire images of a photovoltaic array. For example, an image acquisition device mounted on the drone is used to acquire initial images of the photovoltaic array in a photovoltaic power station. The initial images are then processed and detected in stages. First, the acquired initial images are preprocessed to improve image quality and reduce noise and other interference factors, providing a better image foundation for subsequent target detection. Next, the YOLOv11 model is used to perform target detection on the processed images, determining the first position information of the target photovoltaic strings. That is, the YOLOv11 model is used to perform preliminary target detection on the panoramic image of the photovoltaic array acquired by the drone, locating the boundary positions of each photovoltaic string, including the first position information of the target photovoltaic strings. Finally, based on this first position information, the processed images are further processed... The image is cropped to obtain a target image containing only the target photovoltaic (PV) string. This allows focus on the region of interest, i.e., the specific target PV string, which can be any single PV string in the PV array. Then, the cropped target image is subjected to more refined detection using the YOLOv11-OBB model to determine the presence of defects. Specifically, an improved YOLOv11-OBB model (supporting rotated bounding box detection) is used to identify defects in the cropped image and output detailed defect information, including the first location information of the target PV string containing the defect, the defect type information of the target PV panel, and the target row and column information of the target PV panel within the target PV string. The YOLOv11-OBB model has better adaptability to tilted / rotated PV panels. The initial image can include visible light images and infrared thermal images. In related technologies, manual inspection of photovoltaic panels is time-consuming and labor-intensive, and is prone to missed or false detections due to fatigue, negligence, and other factors. Fixed cameras are limited by their installation position and angle, making it difficult to fully cover the photovoltaic array, and their ability to identify defects in complex environments is limited. Some related technologies also use single-model end-to-end detection methods, but these are prone to missed detections. This embodiment decomposes photovoltaic defect detection into three organically combined steps: "target location → area focusing → defect identification". In particular, it introduces an improved OBB (oriented bounding box) model, which solves the problem of inaccurate detection box matching caused by the regular tilted arrangement of photovoltaic panels. This embodiment utilizes drones to automatically acquire images and perform inspections, significantly improving inspection efficiency. The flexibility of drones allows them to acquire images from multiple angles, overcoming the limitations of fixed cameras. Through image preprocessing and a phased inspection model, especially the combination of YOLOv11 and YOLOv11-OBB models, defects on photovoltaic panels can be identified more accurately, reducing false detections and missed detections. The system can output the first location information of the target photovoltaic module with defects, the defect type information of the target photovoltaic panel, and the row and column information of the target photovoltaic panel, facilitating maintenance personnel to quickly locate and handle defects, thereby improving the maintenance efficiency and reliability of photovoltaic power plants.

[0034] In an optional embodiment, the YOLOv11-OBB model is used to detect the target image to obtain a second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, the first location information, the defect type information of the target photovoltaic panel, and the target row and column information are output. This includes: the YOLOv11-OBB model uses a rotating bounding box to detect a group of photovoltaic panels in the target image to obtain a set of location information, and identifies the local images corresponding to each rotating bounding box to obtain a set of identification results. The target photovoltaic string includes a group of photovoltaic panels, and the second detection result includes a set of location information and a set of identification results. When the set of identification results indicates that there is a target defect in the target photovoltaic panel, the target row and column information of the target photovoltaic panel is determined based on the set of location information, and the first location information, defect type information, and target row and column information are output. The set of identification results includes the defect type information.

[0035] In the above embodiments, the YOLOv11-OBB model is used to detect a group of photovoltaic panels in the target image by rotating the frame to obtain a set of position information. The corresponding local image of the rotating frame is identified to obtain a set of recognition results. This can more accurately locate the position of the photovoltaic panel and identify the defect situation. When it is determined that there is a target defect in the target photovoltaic panel, the target row and column information of the target photovoltaic panel is determined according to the set of position information. The first position information of the target photovoltaic string, the defect type information of the target photovoltaic panel, and the target row and column information of the target photovoltaic panel can be output, which can effectively improve the accuracy of photovoltaic defect identification and the completeness of output information. This solves the problems of low defect identification accuracy and inability to accurately output the specific location and type information of defects in related technologies.

[0036] The YOLOv11-OBB model uses a rotating bounding box to detect a group of photovoltaic (PV) panels in a target image, obtaining a set of positional information for each PV panel. This information includes the position of each PV panel within the group, such as the center coordinates, length, width, and rotation angle of the rotating bounding box. Compared to a regular rectangular bounding box, the rotating bounding box better fits tilted or irregularly placed PV panels, thus more accurately locating their positions in the image. The model then identifies the local image corresponding to each rotating bounding box, obtaining a set of recognition results. These results include information such as whether the PV panel has defects and the type of defect. If a set of recognition results indicates a defect in the target PV panel, the model determines the target row and column information of the target PV panel within the target PV string based on the positional information, i.e., determining the row and column number of the target PV panel in the target PV string, such as row 2, column 3. Combining this with the first positional information of the target PV string from the first detection result, the model outputs the first positional information of the target PV string, the defect type information of the target PV panel, and the target row and column information, providing detailed and accurate data for subsequent maintenance work. In actual photovoltaic power plants, photovoltaic panels may be irregularly placed due to factors such as terrain and installation methods. Ordinary rectangular frame detection in related technologies is difficult to accurately locate these irregular photovoltaic panels. This embodiment uses rotating frame detection. By recognizing the local image corresponding to the rotating frame, it can accurately determine whether there are defects in the photovoltaic panel and the type of defects. For example, the defect type can be hot spots, cracks, dirt (such as wind sand or snow accumulation), etc. At the same time, after detecting defects, it can further determine the target row and column information of the photovoltaic panel where the defect is located, solving the problem of inaccurate positioning of irregular photovoltaic panels. In addition, related technologies may only be able to detect the general existence of defects, but cannot accurately give the specific row and column information of the photovoltaic panel where the defect is located. This embodiment can not only identify the defect type, but also determine the target row and column information of the target photovoltaic panel based on a set of location information, and output the relevant information completely, solving the problems of accurate defect positioning and detailed information output.

[0037] In an optional embodiment, determining the target row and column information of the target photovoltaic panel based on a set of position information includes: determining the center coordinates of each rotating frame based on a set of position information, wherein each photovoltaic panel in the set of photovoltaic panels corresponds to a rotating frame; determining the arrangement direction of each photovoltaic panel in the set of photovoltaic panels based on the long side direction of each rotating frame; and sorting according to the x or y values ​​of the center coordinates of each rotating frame to obtain the row and column distribution result of the set of photovoltaic panels, wherein the row and column distribution result includes the target row and column information.

[0038] In the above embodiments, determining the center coordinates of each rotating frame based on a set of position information provides basic data for subsequent calculations; determining the photovoltaic panel arrangement direction based on the long side direction of the rotating frame clearly shows the arrangement of the photovoltaic panels; sorting by the x or y values ​​of the center coordinates of the rotating frame to obtain the row and column distribution results to determine the target row and column information helps to accurately determine the position of the target photovoltaic panel in the photovoltaic string, thereby improving the accuracy and efficiency of photovoltaic defect identification and solving the problem that existing detection methods are unable to accurately output the specific location information of defects.

[0039] This embodiment analyzes the center coordinates and long side direction of each rotating frame (corresponding to each photovoltaic panel) to infer the arrangement direction of the photovoltaic panels, and sorts them accordingly to determine their row and column distribution in the string. Specifically, firstly, the center coordinates of each rotating frame are determined based on a set of position information (position information of each rotating frame). Each photovoltaic panel corresponds to one rotating frame, so the center coordinates of each rotating frame can represent the center position of the corresponding photovoltaic panel. The arrangement direction of each photovoltaic panel in a string is determined based on the long side direction of each rotating frame. The long side direction of the rotating frame reflects the main extension direction of the photovoltaic panel in the image, thus inferring the arrangement direction of the photovoltaic panels (e.g., whether they are horizontally or vertically arranged). The panels are then sorted according to the x or y values ​​of the center coordinates of each rotating frame to obtain the row and column distribution result of the photovoltaic panel string. For example, if the photovoltaic panels are horizontally arranged, the row information can be determined by sorting according to the y-value of the center coordinates; if they are vertically arranged, the column information can be determined by sorting according to the x-value of the center coordinates. The row and column distribution result includes the target row and column information, that is, the specific row and column position of the target photovoltaic panel in the photovoltaic string. In different installation environments, photovoltaic panels may be arranged at different angles or in different directions, which poses a challenge to defect detection. The method in this embodiment can adapt to various arrangements by rotating the frame, ensuring that the position of the photovoltaic panels can be accurately identified even in complex layouts.

[0040] For example, assuming the target photovoltaic string consists of 6 photovoltaic panels, the yolov11-obb model is used to detect these 6 photovoltaic panels and provides the following information: Board 1: (100, 200), 10°; Board 2: (200, 200), 10°; Board 3: (300, 200), 10°; Board 4: (100, 300), 10°; Board 5: (200, 300), 10°; Board 6: (300, 300), 10°; The system groups and sorts the photovoltaic panels. First, it clusters the center points according to the Y-coordinate, identifying two rows: the first row (Y=200): panel 1, panel 2, panel 3; the second row (Y=300): panel 4, panel 5, panel 6. Then, within each row, the panels are sorted in ascending order of the X-coordinate: the first row: panel 1 → panel 2 → panel 3 → corresponding columns 1, 2, and 3; the second row: panel 4 → panel 5 → panel 6 → corresponding columns 1, 2, and 3. Finally, the system labels each photovoltaic panel with its specific row and column information in the target photovoltaic string, including the row and column information of the target photovoltaic panel.

[0041] In an optional embodiment, the first position information includes first center coordinate information, first length information, and first width information; a set of position information includes multiple second position information, and each photovoltaic panel in a set of photovoltaic panels corresponds to one second position information, the second position information including second center coordinate information, second length information, second width information, and rotation angle information.

[0042] In the above embodiments, the first position information includes first center coordinate information, first length information, and first width information. A set of position information includes multiple second position information. Each second position information includes second center coordinate information, second length information, second width information, and rotation angle information. Each photovoltaic panel corresponds to one second position information, which can accurately characterize the position and size of the target photovoltaic string and photovoltaic panel, thereby more accurately locating and identifying the target photovoltaic panel and its defects, and improving the accuracy and precision of photovoltaic defect identification.

[0043] The first location information describes the position of the target photovoltaic string, including first center coordinates, first length, and first width. The first center coordinates determine the center position of the target photovoltaic string in the image, while the first length and first width define its size range in the image. These three parameters can completely describe the position and scale of the target photovoltaic string in the initial image. A second set of location information describes the position of each photovoltaic panel in the target photovoltaic string, with each panel corresponding to a second set of location information. In addition to the center coordinates, length, and width (i.e., second center coordinates, second length, and second width) similar to the first location information, the second location information also includes rotation angle information. The rotation angle information takes into account the possible tilt or rotation of the photovoltaic panels in the actual scene, and can more accurately describe the posture of the photovoltaic panels in the image. Detailed location information enables the system to more accurately locate the position of the target photovoltaic panels, especially in complex environments (such as when photovoltaic panels are tilted or irregularly arranged). The rotation angle information can help to more accurately determine the orientation of the photovoltaic panels, thereby improving the accuracy of defect location. For example, the first position information includes the first center coordinates (250, 250), the first length (300), and the first width (150); for example, the second position information corresponding to the first photovoltaic panel includes the second center coordinates (100, 200), the second length (150), the second width (50), and the rotation angle of 10°. This is just one example.

[0044] In an optional embodiment, the yolov11-obb model uses InceptionNeXt as the backbone structure and employs a channel-enhanced feature pyramid network. The channel-enhanced feature pyramid network dynamically adjusts the weights of different channels during feature fusion through an adaptive channel weighting mechanism to enhance defect-related features.

[0045] In the above embodiments, the yolov11-obb model uses InceptionNeXt as the backbone structure and employs a channel-enhanced feature pyramid network that dynamically adjusts the weights of different channels during feature fusion through an adaptive channel weighting mechanism. This enhances defect-related features and thus improves the accuracy and precision of defect identification.

[0046] InceptionNeXt is a novel convolutional neural network architecture that combines the design philosophy of Inception modules with the advantages of modern Transformer-style networks (such as ConvNeXt). As the backbone structure of the YOLOv11-OBB model, InceptionNeXt can efficiently extract multi-scale features from images, providing rich feature representations for subsequent object detection. While maintaining high efficiency, it has stronger feature extraction capabilities, making it particularly suitable for object detection tasks in complex scenes. The Channel Enhancement Feature Pyramid Network introduces an adaptive channel weighting mechanism to dynamically adjust the importance weights of different channels during multi-scale feature fusion. During feature fusion, the importance of features from different channels for defect detection may vary. Through the adaptive channel weighting mechanism, the model can dynamically adjust the weights of different channels according to the needs of the current task, enhancing feature channels relevant to defect detection and suppressing irrelevant feature channels. This mechanism can highlight feature channels related to defect recognition and suppress irrelevant or interfering information, thereby improving the recognition ability of small targets and low-contrast defects. This dynamic adjustment mechanism enables the model to more effectively extract and utilize features related to defect detection, improving the accuracy and robustness of defect detection.

[0047] In an optional embodiment, the yolov11-obb model introduces a coordinate attention mechanism in the detection head, so that image features correspond one-to-one with positional information. In addition, the yolov11-obb model uses the KFIoU loss function instead of RotatedBboxLoss. The KFIoU loss function includes a scale-insensitive center point loss and a distance-independent term.

[0048] In the above embodiments, introducing a coordinate attention mechanism into the detection head can make image features correspond one-to-one with position information, improving the accuracy of target position judgment during detection; using the KFIoU loss function instead of RotatedBboxLoss, based on Gaussian modeling and Gaussian product approximation SkewIoU, and including scale-insensitive center point loss and distance-independent terms, can more accurately reflect the degree of overlap between target boxes and improve the accuracy of the model in target detection and localization.

[0049] A position-aware feature enhancement module is introduced into the detection head to dynamically associate image features with spatial coordinates (x, y), enabling the model to focus more on the precise boundary position of the photovoltaic panel and reduce the positioning error of the rotated box. The YOLOv11-OBB model introduces a coordinate attention mechanism into the detection head, which establishes a one-to-one correspondence between image features and positional information. Specifically, it processes the feature map and learns the relationship between the features at each position and the actual target position by focusing on the feature responses at different positions in the feature map. In this way, the model can more accurately locate the target position based on features when performing target detection, especially for rotated box target detection (OBB), and can better adapt to different angles and directions of the target. The traditional rotated box loss (RotatedBboxLoss) has certain limitations in handling rotated box target detection. The KFIoU loss function is designed based on Gaussian modeling and the Gaussian product approximation SkewIoU (Slant IoU). It includes a scale-insensitive center point loss and a distance-independent term. The scale-insensitive center point loss means that the loss function is more stable in locating the center point when dealing with targets of different sizes, and will not produce large errors due to changes in target scale. The distance-independent term reduces unnecessary interference from distance factors during the calculation process, focusing more on the actual overlap of targets. The use of the KFIoU loss function allows the model to optimize parameters more effectively during training. The scale-insensitive center point loss and distance-independent term reduce unnecessary interference, allowing the model to focus more on the actual overlap of targets, improving the model's convergence speed and stability. Ultimately, the model's performance in the rotated bounding box target detection task is significantly improved, enabling more accurate detection of defects in photovoltaic panels and reducing false positives and false negatives. In this way, the KFIoU loss function can more accurately measure the difference between the predicted bounding box and the ground truth bounding box, guiding the model to learn more accurate detection results.

[0050] In an optional embodiment, the yolov11-obb model is trained using defect sample data, which is obtained through one of the following methods: generating defect sample data using a generative adversarial network (GAN); obtaining defect sample data by simulating sandstorms, snow accumulation, and crack defects through image synthesis; or obtaining defect sample data by augmenting a small number of defect samples using transfer learning.

[0051] In the above embodiments, the YOLOv11-OBB model is trained by using Generative Adversarial Networks (GANs) to generate defect sample data, using image synthesis to simulate wind-blown sand and snow cracks, and using transfer learning to augment the data of few-sample defects. This increases the amount of defect sample data used for training, improves the situation of few-sample data, and enables the model to better learn various photovoltaic defect features, thereby improving the accuracy and generalization ability of the YOLOv11-OBB model in photovoltaic defect identification methods based on UAVs.

[0052] Generative Adversarial Networks (GANs) are used to generate defect sample data. A GAN consists of a generator and a discriminator. The generator attempts to generate data similar to real defect samples, while the discriminator distinguishes between the generated data and real defect samples. Through continuous adversarial training, the generator gradually learns to generate increasingly realistic defect samples. This generated data can be used to expand the training set, allowing the model to encounter more defect samples of different types and features, thus improving its generalization ability. Defect sample data is obtained by simulating sandstorms, snow accumulation, and cracks through image synthesis. This method is based on existing photovoltaic panel images and uses image processing techniques to synthesize the effects of sandstorms, snow accumulation, and cracks into the images. For example, image filtering and texture addition can be used to simulate sandstorm effects, pixel modification can simulate snow cover, and line drawing can simulate crack defects. In this way, a large amount of sample data with simulated defects can be quickly generated, providing rich material for model training. Transfer learning is used to expand the data for fewer defect samples: Transfer learning leverages models pre-trained on other relevant datasets, transferring the learned features and knowledge to the current defect detection task. For defect types with few samples, features can be extracted using pre-trained models, and fine-tuned on these limited samples. This approach uncovers potential information within the few-sample defect data, expanding the model's capabilities and allowing it to learn more features about these defects during training, thus enhancing its ability to identify them. In real-world photovoltaic defect detection scenarios, collecting a large amount of sample data encompassing various defect types is challenging, and there may be situations where the number of samples for certain defect types is extremely limited. This can lead to the model failing to fully learn the features of various defects during training, resulting in overfitting and impacting the model's generalization ability and detection accuracy. The three methods mentioned above for acquiring defect sample data effectively expand the training set and address the problem of insufficient sample data. Real-world photovoltaic power plant environments are complex and variable, with various environmental disturbances such as wind, sand, and snow accumulation, as well as various types of cracks and defects. Collecting sample data covering all situations in real-world environments is costly and time-consuming. Image synthesis methods can simulate these complex real-world scenarios, allowing the model to access sample data closer to actual conditions during training, thereby improving its detection performance in real-world scenarios.

[0053] In an optional embodiment, the initial image is preprocessed to obtain a processed image, including: performing brightness adjustment and contrast enhancement processing on the initial image to obtain the processed image.

[0054] In the above embodiments, brightness adjustment and contrast enhancement processing are performed on the initial image of the acquired photovoltaic array to improve image quality. This provides clearer and more accurate images for subsequent target detection using the YOLOv11 model and target image detection using the YOLOv11-OBB model, thereby improving the accuracy and efficiency of photovoltaic defect identification and reducing misjudgments and omissions caused by poor image quality.

[0055] Image brightness and contrast directly affect the recognizability of information within an image. In photovoltaic defect recognition scenarios, initial images captured by drones may be too bright, too dark, or lack sufficient contrast due to varying lighting conditions (such as strong light, weak light, and shadows), making it difficult to clearly present the features and potential defects of the photovoltaic panel. Adjusting the image brightness makes the pixel values ​​in the image more uniform, avoiding interference from overly bright or dark areas in subsequent detection. Contrast enhancement processing can amplify the brightness differences between different pixels in the image, making the distinction between the target and the background more obvious. This is crucial for subsequent target detection and defect recognition, as a clearer image improves the model's recognition accuracy. Brightness adjustment enhances the overall visibility of the image, making details clearer. Brightness adjustment involves linear or non-linear transformation of the brightness value of each pixel in the image to achieve an appropriate overall brightness level, avoiding the loss of important information due to excessive brightness or darkness. Contrast enhancement, on the other hand, amplifies the brightness differences between different pixels in the image, highlighting edges and details, making the outline, texture, and potential defects of the photovoltaic panel more apparent, facilitating accurate identification and analysis by subsequent target detection models. Brightness adjustment and contrast enhancement processes can significantly improve the quality of the initial image, making the details in the image clearer and providing a better foundation for subsequent defect detection.

[0056] In an optional embodiment, the second detection result further includes area information of the target defect, and the method further includes: determining the target level of the target defect based on the defect type information and area information; and recommending a corresponding target maintenance strategy based on the target level.

[0057] In the above embodiments, when using drones to identify photovoltaic defects, the area information of the target defect can be obtained. The target level of the target defect can be determined based on the defect type information and area information, and the corresponding maintenance strategy can be recommended based on the target level. This helps to more accurately assess the defect status, take appropriate maintenance measures, and improve the maintenance efficiency and effectiveness of photovoltaic panels.

[0058] The second detection result includes not only defect type information but also the area information of the target defect. This can be achieved by counting pixels or performing geometric calculations on the detected defect area to obtain the defect size. Based on the defect type and area information, the target defect level is determined. This is typically achieved through pre-defined rules or models. For example, defects are classified into different levels (e.g., minor, moderate, severe) based on their type and size. Different types and sizes of defects may have varying degrees of impact on the performance of the photovoltaic system. For instance, a small but severe crack may require more immediate attention than a large, minor stain. Based on the determined target defect level (or target defect grade), a corresponding target maintenance strategy is recommended. For example, for low-level defects, regular inspections are conducted to observe whether the defects show signs of expansion or worsening. For instance, for low-priority defects (such as small areas of dust cover), regular cleaning can be scheduled. For medium-level defects, professional personnel conduct on-site assessments to determine whether repairs are necessary. For example, for medium-level defects (such as large hot spots or cracks), immediate repairs may be recommended. For high-level defects, maintenance personnel are immediately dispatched to replace or repair the photovoltaic panels to minimize the impact on the photovoltaic system's power generation efficiency. This strategy can be customized according to the actual application scenario and maintenance standards.

[0059] Taking hot spot defects (local aging caused by high temperature) as an example, when the hot spot area accounts for less than 5% of the total area of ​​a single photovoltaic panel, it is considered a mild defect; when the proportion is 5%-15%, it is considered a moderate defect; and when the proportion is greater than 15%, it is considered a moderate defect. The proportion range here is only an example and can be set as needed in actual applications.

[0060] In an optional embodiment, when using the yolov11-obb model to detect the target image, multi-scale feature extraction is performed on the target image. Convolution operations are performed on the target image using convolution kernels of different scales to obtain feature information of each photovoltaic panel and defect at different scales.

[0061] In the above embodiments, when using the YOLOv11-OBB model to detect target images, multi-scale feature extraction is performed on the target images. This means that the model performs convolution operations on the target image using convolution kernels of different scales, thereby obtaining feature information of each photovoltaic panel and defect at different scales; the size of the convolution kernel determines the scale of image features that the model can capture. Smaller convolution kernels can capture detailed information in the image, which is very effective for detecting defects in small-sized photovoltaic panels (such as micro-cracks, local damage, etc.); while larger convolution kernels can extract overall features and contextual information of the image, which helps to identify large-sized photovoltaic panels or defects over a large area (such as large-area shading, severe deformation, etc.). By integrating feature information extracted at different scales, the model can more comprehensively and accurately identify each photovoltaic panel and its defects. By using convolution kernels of different scales, the model can simultaneously capture local details and global structure in the image, thereby gaining a more comprehensive understanding of the image content. Through multi-scale feature extraction, the model can simultaneously capture local details and global structure in the image, thereby more accurately detecting photovoltaic panels and defects of different sizes and improving detection accuracy.

[0062] In an optional embodiment, determining the target row and column information of the target photovoltaic panel based on a set of location information includes: determining the center coordinates of each rotating frame based on a set of location information, wherein each photovoltaic panel in a set of photovoltaic panels corresponds to a rotating frame; grouping the center coordinates of each rotating frame by Y coordinate and sorting them by X coordinate to obtain the row and column distribution result, wherein the row and column distribution result includes the target row and column information.

[0063] In the above embodiments, a set of position information is used to determine the center coordinates of each rotating frame. The center coordinates of each rotating frame are grouped by the Y coordinate and sorted by the X coordinate to obtain the row and column distribution result containing the target row and column information. When there is a target defect in the target photovoltaic string, the target row and column information, defect type information and the first position information of the target photovoltaic string can be accurately output. This solves the problems of low efficiency of manual inspection, incomplete coverage of fixed camera monitoring and low accuracy and efficiency of existing drone detection, and improves the accuracy and efficiency of photovoltaic defect identification in complex backgrounds.

[0064] First, the center coordinates of each rotating frame are determined based on a set of position information. Each photovoltaic panel corresponds to one rotating frame, so the center coordinates of each rotating frame can represent the center position of the corresponding photovoltaic panel. The center coordinates of each rotating frame are grouped according to the Y coordinate, that is, the center coordinates with similar Y coordinates are grouped together. This can determine the row information of the photovoltaic panel. Within each group, the center coordinates are sorted according to the X coordinate, which can determine the column information of the photovoltaic panel. In this way, the row and column distribution result of the photovoltaic panel is finally obtained, which includes the target row and column information of the target photovoltaic panel.

[0065] Obviously, the embodiments described above are only some embodiments of this application, and not all embodiments. The present application will be specifically described below with reference to specific embodiments.

[0066] This application provides a method for photovoltaic defect identification based on the YOLOv11 and YOLOv11-OBB models, relating to the field of artificial intelligence. This method enables intelligent analysis of photovoltaic defects. The embodiments of this application are described below.

[0067] The method in this embodiment mainly employs target detection and rotating frame positioning to locate defects; The yolov11 and yolov11-obb models were used for photovoltaic string detection and photovoltaic panel defect identification, respectively. Data synthesis and image fusion are performed, such as downloading images of sandstorms, snow, and rocks from the internet and synthesizing images of specified regions, i.e., synthesizing images with few defective samples; GANs can also be used to generate defective samples; and multi-sampling techniques can be used to collect few samples; in order to solve the problem of few defective samples in related technologies. The YOLOv11-OBB model backbone was modified to use InceptionNeXt, achieving high throughput while maintaining competitive performance. Channel Enhancement Feature Pyramid Network (CE-FPN) was employed to mitigate feature loss caused by upsampling. Finally, a CoordAttention mechanism was introduced into the detection head to ensure a one-to-one correspondence between image features and location information, improving the algorithm's localization accuracy. This resulted in performance improvements of 1.1%, 1.5%, and 1.9% compared to YOLOv11-OBB-M. KFIoU replaces yolov11-obb's RotatedBboxLoss. KFIoU is based on Gaussian modeling and Gaussian product to effectively approximate the SkewIoU loss. It includes two terms: a scale-insensitive center point loss, used to quickly shorten the distance between the center points of two bounding boxes; and a distance-independent term, which uses a Gaussian distribution product to mimic the mechanism of SkewIoU. Within a certain distance (e.g., within 9 pixels), it aligns with the SkewIoU loss at a trend level, improving the recognition map by 0.7%.

[0068] The overall steps of the photovoltaic defect identification method in this application embodiment include: S1, Image Acquisition, uses a drone to acquire and label images; the image acquisition module can also perform some simple processing on the acquired images, such as adjusting the brightness and contrast of the images to improve the image quality and clarity; S2, object detection and rotated bounding box recognition; S3 outputs the position information of the photovoltaic string detection frame, the position information of the photovoltaic panel, and the defect information; S4, Photovoltaic panel position row and column information, uses the center position of the rotating frame as the arrangement basis, calculates the tilt angle of each frame, for example, grouping by Y coordinate (row): grouped by vertical coordinate (Y value), each group represents a row, divided into 2 rows. Based on the center of the smallest coordinate value frame, a y distance threshold is determined, dividing into 2 rows; sorted by X coordinate (column) within the group: points in each row are arranged in ascending order by horizontal coordinate (X value).

[0069] Considering the lack of defective sample data in related technologies, the image data collected by drones can be labeled for further training of subsequent models.

[0070] Figure 2 This is an example diagram of photovoltaic defect detection provided in an embodiment of this application. Figure 2 It includes a detection frame for a single photovoltaic string and a rotating frame for a single photovoltaic panel. Figure 2 The individual photovoltaic panel in the rotating frame shown is defective. Figure 3 , Figure 4 These are examples of photovoltaic defect identification results. Figure 1 , two , Figure 3 The defect identification result was bird droppings, with a confidence level of 96.1%. Figure 4 The defect identification result was an infrared hot spot, with confidence levels of 95% and 92.03%. It should be noted that... Figure 3 The corresponding original image is a visible light image. Figure 2 , Figure 4 The corresponding original image is an infrared thermal image.

[0071] This application also provides a photovoltaic defect identification system based on unmanned aerial vehicles (UAVs), such as... Figure 5 As shown, Figure 5 This is a structural block diagram of a photovoltaic defect identification system based on an unmanned aerial vehicle (UAV) according to an embodiment of this application. The system includes: Acquisition module 51 is used to acquire initial images of the photovoltaic array using a drone; Processing module 52 is used to preprocess the initial image to obtain the processed image; The first detection module 53 is used to perform target detection on the processed image using the YOLOv11 model to obtain a first detection result. The first detection result includes the first position information of the target photovoltaic string, wherein the target photovoltaic string is any photovoltaic string in the photovoltaic array. The acquisition module 54 is used to crop the processed image according to the first position information to obtain a target image, wherein the target image includes a target photovoltaic string; The second detection module 55 is used to detect the target image using the yolov11-obb model and obtain a second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, it outputs the first position information, the defect type information of the target photovoltaic panel, and the target row and column information, wherein the target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect.

[0072] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0073] This application also provides a computer-readable storage medium storing instructions that, when executed by a processor, perform the steps of any of the methods described above.

[0074] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0075] This application also discloses an electronic device. For example... Figure 6 As shown, Figure 6 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 600 may include: at least one processor 601, at least one communication bus 602, a user interface 603, at least one network interface 604, and a memory 605.

[0076] The communication bus 602 is used to enable communication between these components.

[0077] The user interface 603 may include a display screen and a camera. Optionally, the user interface 603 may also include a standard wired interface and a wireless interface.

[0078] The network interface 604 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0079] The processor 601 may include one or more processing cores. The processor 601 connects to various parts of the electronic device (such as a server) using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 605, and by calling data stored in memory 605. Optionally, the processor 601 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 601 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 601 and may be implemented as a separate chip.

[0080] The memory 605 may include random access memory (RAM) or read-only memory. Optionally, the memory 605 may include a non-transitory computer-readable storage medium. The memory 605 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 605 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 605 may also be at least one storage device located remotely from the aforementioned processor 601. (Refer to...) Figure 6 The memory 605, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a photovoltaic defect identification method based on unmanned aerial vehicles.

[0081] exist Figure 6In the illustrated electronic device 600, the user interface 603 is mainly used to provide an input interface for the user and acquire user input data; while the processor 601 can be used to call an application program of a photovoltaic defect identification method based on a drone stored in the memory 605. When executed by one or more processors 601, the electronic device 600 performs one or more of the methods described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0082] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0083] In the various embodiments provided in this application, it should be understood that the disclosed apparatus or system can be implemented in other ways. For example, the apparatus or system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0084] The above description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will be readily apparent to those skilled in the art upon consideration of the disclosure herein.

[0085] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art that are not described in this disclosure.

Claims

1. A photovoltaic defect identification method based on unmanned aerial vehicles (UAVs), characterized in that, include: Use drones to acquire initial images of the photovoltaic array; The initial image is preprocessed to obtain the processed image; The processed image is subjected to target detection using the YOLOv11 model to obtain a first detection result. The first detection result includes the first location information of the target photovoltaic string, wherein the target photovoltaic string is any one of the photovoltaic strings in the photovoltaic array. The processed image is cropped according to the first location information to obtain a target image, wherein the target image includes the target photovoltaic string; The target image is detected using the yolov11-obb model to obtain a second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, the first location information, the defect type information of the target photovoltaic panel, and the target row and column information are output. The target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect. The target image is detected using the YOLOv11-OBB model to obtain a second detection result. When the second detection result indicates that a target defect exists in the target photovoltaic string, the first location information, the defect type information of the target photovoltaic panel, and the target row and column information are output. This includes: the YOLOv11-OBB model uses a rotating frame to detect a group of photovoltaic panels in the target image to obtain a set of location information, and identifies the local images corresponding to each rotating frame to obtain a set of identification results. The target photovoltaic string includes the group of photovoltaic panels, and the second detection result includes the set of location information and the set of identification results. When the set of identification results indicates that the target photovoltaic panel has the target defect, the target row and column information of the target photovoltaic panel is determined based on the set of location information, and the first location information, the defect type information, and the target row and column information are output. The set of identification results includes the defect type information. Determining the target row and column information of the target photovoltaic panel based on the set of position information includes: determining the center coordinates of each rotating frame based on the set of position information, wherein each photovoltaic panel in the set of photovoltaic panels corresponds to a rotating frame; determining the arrangement direction of each photovoltaic panel in the set of photovoltaic panels based on the long side direction of each rotating frame; and sorting the photovoltaic panels according to the x or y values ​​of the center coordinates of each rotating frame to obtain the row and column distribution result of the set of photovoltaic panels, wherein the row and column distribution result includes the target row and column information.

2. The method according to claim 1, characterized in that, The first position information includes first center coordinate information, first length information, and first width information; the set of position information includes multiple second position information, and each photovoltaic panel in the set of photovoltaic panels corresponds to one second position information, the second position information including second center coordinate information, second length information, second width information, and rotation angle information.

3. The method according to claim 1, characterized in that, The yolov11-obb model uses InceptionNeXt as its backbone structure and employs a channel-enhanced feature pyramid network. This network uses an adaptive channel weighting mechanism to dynamically adjust the weights of different channels during feature fusion, thereby enhancing defect-related features.

4. The method according to claim 1, characterized in that, The YOLOv11-OBB model introduces a coordinate attention mechanism in the detection head, which makes image features correspond one-to-one with position information. In addition, the YOLOv11-OBB model uses the KFIoU loss function instead of RotatedBboxLoss. The KFIoU loss function includes a scale-insensitive center point loss and a distance-independent term.

5. The method according to claim 1, characterized in that, The yolov11-obb model is trained using defect sample data, which is obtained through one of the following methods: The defect sample data is generated using a Generative Adversarial Network (GAN). The defect sample data was obtained by simulating wind, sand, snow accumulation, and crack defects through image synthesis. The defect sample data is obtained by using transfer learning to augment the data of a small number of defects.

6. A photovoltaic defect identification system based on unmanned aerial vehicles (UAVs), characterized in that, include: The acquisition module is used to acquire initial images of the photovoltaic array using a drone; The processing module is used to preprocess the initial image to obtain the processed image; The first detection module is used to perform target detection on the processed image using the YOLOv11 model to obtain a first detection result. The first detection result includes the first location information of the target photovoltaic string, wherein the target photovoltaic string is any one of the photovoltaic strings in the photovoltaic array. The acquisition module is used to crop the processed image according to the first location information to obtain a target image, wherein the target image includes the target photovoltaic string; The second detection module is used to detect the target image using the yolov11-obb model to obtain a second detection result. When the second detection result indicates that there is a target defect in the target photovoltaic string, the module outputs the first location information, the defect type information of the target photovoltaic panel, and the target row and column information, wherein the target photovoltaic panel refers to the photovoltaic panel in the target photovoltaic string that has the target defect. The second detection module is used to obtain the second detection result and output the first location information, the defect type information of the target photovoltaic panel, and the target row and column information in the following manner: The yolov11-obb model uses a rotating frame to detect a group of photovoltaic panels in the target image to obtain a set of location information, and identifies the local images corresponding to each rotating frame to obtain a set of identification results. The target photovoltaic string includes the group of photovoltaic panels, and the second detection result includes the set of location information and the set of identification results. When the set of identification results indicates that the target photovoltaic panel has the target defect, the target row and column information of the target photovoltaic panel is determined according to the set of location information, and the first location information, the defect type information, and the target row and column information are output. The set of identification results includes the defect type information. The process of determining the target row and column information of the target photovoltaic panel based on the set of position information includes: determining the center coordinates of each rotating frame based on the set of position information, wherein each photovoltaic panel in the set of photovoltaic panels corresponds to a rotating frame; determining the arrangement direction of each photovoltaic panel in the set of photovoltaic panels based on the long side direction of each rotating frame; and sorting the photovoltaic panels according to the x or y values ​​of the center coordinates of each rotating frame to obtain the row and column distribution result of the set of photovoltaic panels, wherein the row and column distribution result includes the target row and column information.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, perform the method as described in any one of claims 1 to 5.

Images