Photovoltaic panel defect detection and classification method based on thermal infrared image

Abstract

The invention discloses a photovoltaic panel defect detection and classification method based on a thermal infrared image, and relates to the technical field of photovoltaic panel thermal infrared detection.The method comprises the steps that S1, a thermal infrared camera carried by an unmanned aerial vehicle is used for obtaining thermal infrared photos of a photovoltaic power station with the longitude, the latitude and the attitude angle; the unmanned aerial vehicle is used for carrying a thermal infrared camera to obtain pictures with longitude, latitude and attitude angle information, and photogrammetry software is combined to generate an image, so that the dependence of electroluminescence detection on special illumination and an external power supply is eliminated, detection work can be carried out in a conventional environment, a photovoltaic panel can be accurately positioned, and the detection efficiency is improved. The problem of difficult positioning caused by a large number of panels and repeated textures is solved, meanwhile, a deep learning algorithm is applied, multiple fault types can be accurately recognized, key information output of fault panels is integrated, detailed basis is provided for operation and maintenance personnel, maintenance work is effectively guided, the overall efficiency and accuracy of operation and maintenance of the photovoltaic power station are improved, and the working efficiency of the photovoltaic power station is improved. And safe and efficient operation of the photovoltaic power station is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal infrared detection of photovoltaic panels, and in particular to a photovoltaic panel defect detection and classification method based on thermal infrared images. Background Art

[0002] With the vigorous development of the photovoltaic industry, the scale of photovoltaic power stations continues to expand. During the operation cycle of a photovoltaic power station, various faults will inevitably occur in photovoltaic panels. These faults will not only reduce the power generation and shorten the battery life, but in severe cases, they may even cause the panels to burn out, greatly affecting the safe and efficient operation of the photovoltaic power station. Therefore, timely and accurate detection of photovoltaic panel faults is crucial. At present, photovoltaic panel fault detection methods are mainly divided into two categories: electrical signal-based detection and image-based detection. Image-based detection methods have attracted widespread attention due to the development of sensor and computer vision technology. Common methods use electroluminescence, visible light, and thermal infrared images for detection. Electroluminescence detection applies a forward bias current to the photovoltaic panel to make it emit light, thereby detecting tiny internal defects. However, it needs to be performed in a dark room or at night under no light conditions and requires an external power supply, which is greatly limited in application. Operation and maintenance personnel often use drones to take visible light photos for detection, but this method requires extremely high photo resolution, which is extremely unfriendly for rapid inspections.

[0003] Although some existing thermal image-based detection methods can separate faulty panels from thermal images, they have many shortcomings. On the one hand, they cannot effectively distinguish the defect type of the faulty panel. On the other hand, even if the fault type is distinguished, it is often not clear enough to reflect the cause of the fault and provide specific and effective solutions for operation and maintenance personnel. In addition, due to the large number of photovoltaic panels in photovoltaic power stations and their similar textures, it is extremely difficult to obtain the precise location of the faulty panel from panoramic images or photos of the electric field, which brings huge challenges to the operation and maintenance of photovoltaic power stations. Therefore, it is of great significance to develop a photovoltaic panel defect detection and classification method based on thermal infrared images. Summary of the Invention

[0004] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide a photovoltaic panel defect detection and classification method based on thermal infrared images. It can use a thermal infrared camera equipped with a drone to obtain photos with longitude, latitude, and attitude angle information and generate images. It can carry out detection work in a conventional environment, accurately locate photovoltaic panels, and overcome the positioning difficulties caused by a large number of panels and repeated textures. At the same time, it can accurately identify multiple fault types and integrate the key information output of the fault panel to provide detailed basis for operation and maintenance personnel and effectively guide maintenance work.

[0005] To solve the above technical problems, the present invention provides the following technical solution: a photovoltaic panel defect detection and classification method based on thermal infrared images, the method comprising the following steps: S1: Use the thermal infrared camera carried by the drone to obtain thermal infrared photos of the photovoltaic power station with longitude, latitude, and attitude angle; S2: Use photogrammetry software to process thermal infrared photos with longitude, latitude, and attitude angles to generate thermal infrared orthophotos (TIRDOM) of the photovoltaic power station; S3: Use the trained segmentation neural network to identify the pixels of the photovoltaic panel array in the TIRDOM and generate a binary image; S4: Based on the GDAL module, the photovoltaic panels in TIRDOM are divided into individual photovoltaic panels, saved in image format and their geographic coordinates are stored; S5: Use the trained classification neural network to detect faults in individual photovoltaic panels, identify the fault type if a fault exists, and classify it as a normal working panel if not. S6: Integrate the fault panel information and output the fault type and geographic coordinate information.

[0006] Furthermore, the step S1 specifically includes the following steps: S1.1: The weather was clear and cloudless when the thermal infrared image was taken, and the batteries in the photovoltaic power station were operating normally and outputting current; S1.2: The PV power station photos taken by the drone must have a heading overlap of no less than 60% and a lateral overlap of no less than 30%. The photos must be taken perpendicular to the ground.

[0007] Furthermore, in step S3, the trained segmentation neural network is used to identify the pixels of the photovoltaic panel array in the TIRDOM to generate a binary image. Specifically, a semantic segmentation network is used to remove the TIRDOM background information and segment the photovoltaic panels. The specific method is as follows: S3.1: Select images from different PV power plants as training data. Generate corresponding binary masks by manually vectorizing the PV panel boundaries. After matching the original images with the masks, crop them into images of consistent size and with a certain degree of overlap as training data. S3.2: Randomly input the training data into the semantic segmentation neural network model in small batches for training. Since panel segmentation is a binary semantic segmentation task, the loss function during network training is set as: ,in is the predicted value of the network, For labels, is a smoothing term used to prevent the denominator from being 0; S3.3: Input the TIRDOM of the photovoltaic power station into the trained semantic segmentation neural network model. The model divides the pixels into two categories: photovoltaic panels and background environment, where the photovoltaic panel value is 1 and the background environment value is 0.

[0008] Furthermore, in step S4, the photovoltaic panels in the TIRDOM are segmented into individual photovoltaic panels and stored in image format based on the GDAL module, and their geographic coordinates are stored. The specific steps of the single panel image segmentation and coordinate extraction algorithm based on the GDAL library are as follows: S4.1: Convert the binary image predicted by the neural network in step S3.3 into a photovoltaic panel vector graphic; S4.2: Write an algorithm based on the GDAL library, input the PV panel vector graphics and the PV power plant TIRDOM into the program, and output a single PV panel image with geographic coordinates; S4.3: Write an algorithm based on the GDAL library to extract the central geographic coordinates of a single photovoltaic panel to implement the panel positioning function. The formula for calculating the panel center coordinates is: , ,in is the coordinate of the center point of the photovoltaic panel image, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image The maximum value in the coordinate direction, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image Maximum value in the coordinate direction.

[0009] Furthermore, the algorithm steps for classifying photovoltaic panel faults using a classification neural network in step S5 are as follows: S5.1: Create a classification dataset and label the segmented individual panels into six defect types (junction box overheating, substring open circuit, debris obstruction, panel breakage, dust, common hot spot, and normal panel) and one normal type. Set the PV panel image size to 64 × 64 pixels. S5.2: During model training, the classification data set is input into the neural network training model, and the loss function is the cross entropy function in the multi-classification task. The formula is: ,in represents the number of categories, is a symbolic function, sample The true category is equal to 1 if yes, 0 otherwise. For samples Belong to category The probability of the training process adopts the learning rate decay strategy, which gradually reduces the learning rate as the number of training rounds increases; S5.3: During model prediction, a single photovoltaic panel image is input into the trained classification neural network model to judge the working conditions of the photovoltaic panel and classify them into normal operation and various fault types.

[0010] Furthermore, in step S6, the faulty photovoltaic panel information is integrated, including the photovoltaic panel serial number, geographic coordinates, normal working status, faulty working status, fault type and fault area mask.

[0011] Furthermore, according to the above method, a photovoltaic panel fault location and detection system based on thermal infrared images is used, which includes: Image acquisition module, used to obtain thermal infrared photos of photovoltaic power plants with longitude, latitude, and attitude angles; Photogrammetry module, used to perform aerial triangulation on thermal infrared photos of photovoltaic power plants with longitude, latitude, and attitude angles to generate TIRDOM of the photovoltaic power plant; The panel recognition module is used to use a semantic segmentation neural network to identify pixels belonging to photovoltaic panels in the photovoltaic power station TIRDOM and separate the photovoltaic panels from the background; Single panel segmentation and positioning module, used to segment the photovoltaic panel part in the TIRDOM of the entire photovoltaic power station into single panel images and locate the geographical coordinates of single photovoltaic panels; The fault detection and classification module is used to classify a single photovoltaic panel image using a classification neural network to identify its panel working status and fault type; The data integration module is used to integrate the output results, including the photovoltaic panel serial number, geographic coordinates, and fault type information.

[0012] Furthermore, the image acquisition module includes an image storage unit, which is used to store thermal infrared photos of the photovoltaic power station with longitude, latitude, and attitude angle acquired by the thermal infrared camera carried by the drone in a local storage device and a cloud storage server, and the stored image data is encrypted using an encryption algorithm. The image acquisition module has an automatic calibration function, which automatically calibrates the parameters of the thermal infrared camera before shooting.

[0013] Furthermore, when performing aerial triangulation, the photogrammetry module automatically adjusts measurement parameters, including the distribution and number of image control points, according to the different terrains of the photovoltaic power station and the layout characteristics of the photovoltaic panels.

[0014] Compared with existing technologies, this photovoltaic panel defect detection and classification method based on thermal infrared images has the following beneficial effects: 1. The present invention uses a thermal infrared camera mounted on a drone to acquire photos with longitude, latitude, and attitude angle information, and combines it with photogrammetry software to generate images. This eliminates the reliance of electroluminescence detection on special lighting and an external power supply, and can carry out detection work in conventional environments. It uses a deep learning algorithm to accurately locate photovoltaic panels, overcoming the positioning difficulties caused by a large number of panels and repeated textures. At the same time, it can accurately identify multiple fault types and integrate key information output from faulty panels, providing detailed information for operation and maintenance personnel, effectively guiding maintenance work, improving the overall efficiency and accuracy of photovoltaic power station operation and maintenance, and ensuring the safe and efficient operation of photovoltaic power stations.

[0015] 2. The present invention trains a segmentation neural network to identify photovoltaic panel pixels in TIRDOM, segments the panels and extracts geographic coordinates, thereby achieving precise positioning of a large number of photovoltaic panels with repeated textures, solving the problem of difficulty in obtaining the specific location of faulty panels from panoramic images or photos. By training a classification neural network to detect faults and identify types, the present invention integrates the serial number, coordinates, fault type and other information outputs of the faulty panel to accurately distinguish between multiple fault types, provide detailed fault information to operation and maintenance personnel, effectively guide the formulation of maintenance measures, and improve the efficiency and accuracy of photovoltaic power station operation and maintenance.

[0016] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0018] Figure 1 Flowchart of photovoltaic panel defect detection and classification method based on thermal infrared images; Figure 2 Schematic diagram of the structure of the photovoltaic panel defect detection and classification system based on thermal infrared images. DETAILED DESCRIPTION

[0019] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0020] Example 1

[0021] In a large photovoltaic power station located on the edge of the desert, photovoltaic panels frequently malfunction due to long-term erosion by wind and sand and high temperature environment. In order to detect and handle the malfunctions in a timely manner and ensure the stable power generation efficiency of the power station, Figure 1 and Figure 2 , decided to use the photovoltaic panel defect detection and classification method based on thermal infrared images for daily inspections. The power station covers a large area, with a large number of photovoltaic panels and a complex layout. The traditional manual inspection method is inefficient and difficult to accurately locate faults. Therefore, an efficient and accurate detection technology is urgently needed.

[0022] A professional drone equipped with a high-precision thermal infrared camera is used for shooting. On clear and cloudless days, ensure that the batteries in the photovoltaic power station are operating normally and outputting current to obtain thermal infrared images that accurately reflect the working status of the panels. According to the shooting requirements, the drone's heading overlap is set to 65% and the lateral overlap is set to 35%. Keep the camera vertical to the ground during the flight. During the flight, the drone's thermal infrared camera continuously acquires thermal infrared photos with longitude, latitude, and attitude angles, and uses the image storage unit in the image acquisition module to store the photos in the local storage device and cloud storage server using an encryption algorithm for subsequent processing and analysis. Before each shooting, the automatic calibration function of the image acquisition module will automatically calibrate the parameters of the thermal infrared camera to ensure that the quality of the thermal infrared photos taken meets the standards.

[0023] Photogrammetry software was used to process the acquired thermal infrared photos with longitude, latitude, and attitude angles. When performing aerial triangulation, the photogrammetry module automatically adjusted measurement parameters based on the desert terrain of the photovoltaic power station and the irregular layout of the photovoltaic panels. It rationally increased the number of image control points and optimized their distribution, thereby improving the accuracy and quality of the generated thermal infrared orthophoto TIRDOM. After processing, a clear and accurate TIRDOM of the photovoltaic power station was successfully generated, laying the foundation for subsequent panel identification and fault detection.

[0024] To ensure that the training data is broadly representative, training data was selected from images of photovoltaic power stations in different terrains (including mountains, plains, hills, etc.), different climatic conditions (high temperature, low temperature, high humidity, low humidity, etc.), and different manufacturers and models. During the data selection process, the impact of different environmental factors on the thermal infrared imaging of photovoltaic panels was fully considered to ensure that all possible situations were covered. For example, in a high-temperature environment, the heating characteristics of photovoltaic panels may be different from those in a normal temperature environment. Photovoltaic panels produced by different manufacturers also have different thermal infrared radiation characteristics due to differences in materials and processes.

[0025] For the selected images, the photovoltaic panel boundaries were carefully vectorized manually using professional image editing software to generate corresponding binary masks. Then, the original images were precisely aligned with the masks and, using image cropping tools, they were cropped into images of 256×256 pixels in size and 40% overlap as training data. This cropping method not only ensures the diversity of the training data, but also ensures that each image contains sufficient panel feature information, which is conducive to improving the training effect of the neural network model.

[0026] The training data is randomly input into the semantic segmentation neural network model in small batches for training. Since panel segmentation belongs to the binary semantic segmentation task, the loss function is set as follows during network training: ,in is the predicted value (output value) of the network, For labels, It is a smoothing term used to prevent the denominator from being zero. During the training process, a dynamic learning rate adjustment strategy is adopted. The learning rate is automatically adjusted according to the number of training rounds and the convergence of the model. A larger learning rate is set in the initial stage to speed up the convergence of the model. As the training progresses, when the convergence speed of the model gradually slows down, the learning rate is gradually reduced to avoid the model oscillation near the optimal solution, thereby improving the training accuracy of the model. After hundreds of rounds of training and optimization, the performance of the model has gradually stabilized, and the recognition accuracy of photovoltaic panels and background environment has reached a high level.

[0027] After training is complete, the TIRDOM of the photovoltaic power station is input into the trained semantic segmentation neural network model. The model analyzes and judges each pixel in the image, accurately dividing it into two categories: photovoltaic panels (value 1) and background environment (value 0), thereby obtaining a clear binary image. Through this process, the photovoltaic panels and the background are accurately separated, providing a clear image foundation for subsequent panel segmentation and fault detection.

[0028] The photovoltaic power station TIRDOM is input into the trained semantic segmentation neural network model. The model divides the pixels into two categories: photovoltaic panels (value 1) and background environment (value 0), successfully generating a binary image and effectively separating the photovoltaic panels from the background environment.

[0029] Using image processing algorithms, the binary image predicted by the neural network is converted into a photovoltaic panel vector graphic. This conversion process uses edge detection, contour extraction and other technologies to convert the photovoltaic panel boundaries in the binary image into vector form, facilitating subsequent precise processing and analysis.

[0030] Based on the powerful geospatial data processing capabilities of the GDAL library, a special algorithm program is written, and the generated photovoltaic panel vector graphics and photovoltaic power station TIRDOM are input into the program at the same time. After a series of calculations and processing, a single photovoltaic panel image with precise geographic coordinates is output. These images not only contain the shape and texture information of the photovoltaic panel, but also accurately mark its position coordinates in the actual geographic space. Based on the GDAL library, a special algorithm is written, and the photovoltaic panel vector graphics and photovoltaic power station TIRDOM are input into the program, and a single photovoltaic panel image with geographic coordinates is output.

[0031] In order to achieve accurate positioning of the panel, an algorithm is written again based on the GDAL library to extract the central geographic coordinates of a single photovoltaic panel. , ,in is the coordinate of the center point of the photovoltaic panel image, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image The maximum value in the coordinate direction, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image The maximum value in the coordinate direction accurately calculates the center coordinates of each photovoltaic panel and completes the positioning of the panel.

[0032] A classification dataset was created, and the segmented individual cell panels were labeled into six defect types, including junction box overheating, substring open circuit, debris obstruction, panel breakage, dust, common hot spot, and normal panel, as well as one normal type. Considering that the resolution of most photovoltaic panels after TIRDOM segmentation in photovoltaic power plants does not exceed 70 pixels, the image size of the photovoltaic panels was uniformly set to 64×64 pixels.

[0033] During model training, the classification data set is input into the neural network training model, and the loss function uses the cross entropy function in the multi-classification task. The formula is: ,in represents the number of categories, is a symbolic function, sample The true category is equal to 1 if yes, 0 otherwise. For samples Belong to category The training process adopts a learning rate decay strategy. As the number of training rounds increases, the learning rate is gradually reduced to improve the convergence speed and stability of the model.

[0034] During model prediction, a single photovoltaic panel image is input into the trained classification neural network model to judge the working condition of the photovoltaic panel and accurately classify it into normal operation and various fault types.

[0035] Through the data integration module, the detected fault panel information is integrated, including the photovoltaic panel serial number, geographic coordinates, normal working status, fault working status, fault type, fault area mask, etc. This information is organized into a detailed table data to clearly present the specific situation of each fault panel, which is convenient for operation and maintenance personnel to view and use.

[0036] In summary, through this implementation, the photovoltaic panel defect detection and classification method based on thermal infrared images was successfully applied in this large-scale desert photovoltaic power station. During the actual detection process, the faulty photovoltaic panel was located quickly and accurately, and the positioning accuracy reached a high level, effectively solving the positioning problem caused by the large number and complex layout of photovoltaic panels. At the same time, it can accurately identify various fault types, such as heating of the junction box and panel breakage, and provide detailed fault information to operation and maintenance personnel, helping them to formulate targeted maintenance plans in a timely manner, greatly improving the operation and maintenance efficiency of the photovoltaic power station. Compared with traditional detection methods, it saves a lot of manpower and time costs, ensures the stable power generation of the photovoltaic power station, and improves the economic benefits and reliability of the power station.

[0037] Example 2 There is a photovoltaic power station in a mountainous area. It is built on a hilly terrain. Due to the complex terrain, manual inspection is not only difficult but also inefficient. In addition, the changing climate conditions and frequent natural disasters have a high risk of photovoltaic panel failure. To ensure the stable operation of the power station, see Figure 1 and Figure 2 , it was decided to adopt a photovoltaic panel defect detection and classification method based on thermal infrared images for monitoring.

[0038] Select a drone that can adapt to complex environments and equip it with a high-resolution thermal infrared camera. Choose a sunny and cloudless day when the photovoltaic power station batteries are operating normally and outputting current to carry out shooting. According to the characteristics of mountainous terrain, the drone's heading overlap is set to 70% and the lateral overlap is set to 32% to ensure that the photos can completely cover the photovoltaic area. Keep the camera perpendicular to the ground during shooting, and adjust the flight altitude according to terrain changes. After the drone obtains thermal infrared photos with longitude, latitude, and attitude angle information, it uses the image storage unit of the image acquisition module and uses an encryption algorithm to store them in local storage devices and cloud storage servers. Before shooting, the automatic calibration function of the image acquisition module calibrates the thermal infrared camera parameters to ensure shooting quality.

[0039] Professional photogrammetry software is used to process thermal infrared photos with longitude, latitude, and attitude angles. Due to the complex mountainous terrain, the photogrammetry module intelligently adjusts the measurement parameters according to the terrain undulations and the distribution of photovoltaic panels when performing aerial triangulation. The distribution of image control points in areas with large terrain changes is encrypted, and the number of image control points is increased to generate high-precision thermal infrared orthophotos TIRDOM, providing accurate basic data for subsequent analysis.

[0040] Training data was selected from images of photovoltaic power stations in different terrains, such as mountains and plains. Images under different lighting, temperature, and humidity environments, as well as images of photovoltaic panels from different manufacturers and models, were taken into account. The boundaries of the photovoltaic panels were manually vectorized to generate a binary mask. The original image was aligned with the mask and cropped into 256×256 pixel images with a 40% overlap as training data.

[0041] The training data is randomly input into the semantic segmentation neural network model in small batches for training, and the loss function is set to After multiple rounds of training and optimization, the model performance is stable. When the photovoltaic power station TIRDOM is input into the trained model, the pixels are divided into two categories: photovoltaic panels (value 1) and background environment (value 0), and a clear binary image is obtained, realizing the accurate separation of photovoltaic panels and background.

[0042] Convert the binary image predicted by the neural network into a photovoltaic panel vector graphic. Write an algorithm based on the GDAL library, input the vectorized graphics and TIRDOM into the program, and output a single photovoltaic panel image with geographic coordinates. Use the GDAL library algorithm according to the formula , Calculate the center coordinates of the panel to achieve precise positioning of the panel, overcoming the problem of difficult positioning of photovoltaic panels in mountainous areas.

[0043] A classification dataset was created, and the segmented individual battery panels were marked as seven types: junction box heating, substring open circuit, debris occlusion, panel breakage, dust, common hot spot, and normal panel. The image size was set to 64×64 pixels.

[0044] When training the model, the data set is input into the neural network and the cross entropy function is selected As the loss function, a learning rate decay strategy is adopted in the training. After the model training is completed, a single photovoltaic panel image is input to determine whether its working status is normal or there is a fault, and the fault type is determined.

[0045] The data integration module collects information such as the fault panel's serial number, geographic coordinates, working status, fault type, fault area mask, etc., and outputs it in a table format. Operations and maintenance personnel can quickly obtain detailed information about the fault panel through the table, providing strong support for maintenance decisions.

[0046] In summary, after implementing this method in this mountain photovoltaic power station, the inspection problem caused by the complex terrain was successfully solved, the faulty photovoltaic panel was quickly located, the positioning error was within the allowable range, the positioning efficiency was greatly improved, and various types of faults were accurately identified with high classification accuracy, enabling operation and maintenance personnel to quickly formulate maintenance plans and reduce the impact of faults on power generation. Overall, the operation and maintenance costs were reduced, the power generation stability and economic benefits of the photovoltaic power station were improved, and the reliable operation of the mountain photovoltaic power station was effectively guaranteed.

[0047] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A photovoltaic panel defect detection and classification method based on thermal infrared images, characterized in that: The method comprises the following steps: S1: Use the thermal infrared camera carried by the drone to obtain thermal infrared photos of the photovoltaic power station with longitude, latitude, and attitude angle; S2: Use photogrammetry software to process thermal infrared photos with longitude, latitude, and attitude angles to generate thermal infrared orthophotos (TIRDOM) of the photovoltaic power station; S3: Use the trained segmentation neural network to identify the pixels of the photovoltaic panel array in the TIRDOM and generate a binary image; S4: Based on the GDAL module, the photovoltaic panels in TIRDOM are divided into individual photovoltaic panels, saved in image format and their geographic coordinates are stored; S5: Use the trained classification neural network to detect faults in individual photovoltaic panels, identify the fault type if a fault exists, and classify it as a normal working panel if not. S6: Integrate the fault panel information and output the fault type and geographic coordinate information.

2. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 1 is characterized in that: The step S1 specifically includes the following steps: S1.1: The weather was clear and cloudless when the thermal infrared image was taken, and the batteries in the photovoltaic power station were operating normally and outputting current; S1.2: The PV power station photos taken by the drone must have a heading overlap of no less than 60% and a lateral overlap of no less than 30%. The photos must be taken perpendicular to the ground.

3. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 1, characterized in that: In step S3, the trained segmentation neural network is used to identify the pixels of the photovoltaic panel array in the TIRDOM and generate a binary image. Specifically, a semantic segmentation network is used to remove the TIRDOM background information and segment the photovoltaic panels. The specific method is as follows: S3.1: Select images from different PV power plants as training data. Generate corresponding binary masks by manually vectorizing the PV panel boundaries. After matching the original images with the masks, crop them into images of consistent size and with a certain degree of overlap as training data. S3.2: Randomly input the training data into the semantic segmentation neural network model in small batches for training. Since panel segmentation is a binary semantic segmentation task, the loss function during network training is set as: ,in is the predicted value of the network, For labels, is a smoothing term used to prevent the denominator from being 0; S3.3: Input the TIRDOM of the photovoltaic power station into the trained semantic segmentation neural network model. The model divides the pixels into two categories: photovoltaic panels and background environment, where the photovoltaic panel value is 1 and the background environment value is 0.

4. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 3 is characterized in that: In step S4, the photovoltaic panels in the TIRDOM are segmented into individual photovoltaic panels and stored in image format based on the GDAL module, and their geographic coordinates are stored. The specific steps of the single panel image segmentation and coordinate extraction algorithm based on the GDAL library are as follows: S4.1: Convert the binary image predicted by the neural network in step S3.3 into a photovoltaic panel vector graphic; S4.2: Write an algorithm based on the GDAL library, input the PV panel vector graphics and the PV power plant TIRDOM into the program, and output a single PV panel image with geographic coordinates; S4.3: Write an algorithm based on the GDAL library to extract the central geographic coordinates of a single photovoltaic panel to realize the panel positioning function. , in is the coordinate of the center point of the photovoltaic panel image, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image The maximum value in the coordinate direction, Photovoltaic panel image Minimum value in the coordinate direction, Photovoltaic panel image Maximum value in the coordinate direction.

5. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 1, characterized in that: The algorithm steps for using a classification neural network to classify photovoltaic panel faults in step S5 are as follows: S5.1: Create a classification dataset and label the segmented individual panels into six defect types (junction box overheating, substring open circuit, debris obstruction, panel breakage, dust, common hot spot, and normal panel) and one normal type. Set the PV panel image size to 64 × 64 pixels. S5.2: During model training, the classification data set is input into the neural network training model, and the loss function is the cross entropy function in the multi-classification task. The formula is: ,in represents the number of categories, is a symbolic function, sample The true category is equal to 1 if yes, 0 otherwise. For samples Belong to category The probability of the training process adopts the learning rate decay strategy, which gradually reduces the learning rate as the number of training rounds increases; S5.3: During model prediction, a single photovoltaic panel image is input into the trained classification neural network model to judge the working conditions of the photovoltaic panel and classify them into normal operation and various fault types.

6. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 1, characterized in that: In step S6, the faulty photovoltaic panel information is integrated, including the photovoltaic panel serial number, geographic coordinates, normal working status, faulty working status, fault type and fault area mask.

7. The photovoltaic panel defect detection and classification method based on thermal infrared images according to any one of claims 1 to 6, wherein a photovoltaic panel fault location and detection system based on thermal infrared images is used, characterized in that: The system includes: Image acquisition module, used to obtain thermal infrared photos of photovoltaic power plants with longitude, latitude, and attitude angles; Photogrammetry module, used to perform aerial triangulation on thermal infrared photos of photovoltaic power plants with longitude, latitude, and attitude angles to generate TIRDOM of the photovoltaic power plant; The panel recognition module is used to use a semantic segmentation neural network to identify pixels belonging to photovoltaic panels in the photovoltaic power station TIRDOM and separate the photovoltaic panels from the background; Single panel segmentation and positioning module, used to segment the photovoltaic panel part in the TIRDOM of the entire photovoltaic power station into single panel images and locate the geographical coordinates of single photovoltaic panels; Fault detection classification module, which is used to classify a single photovoltaic panel image using a classification neural network. Identify the panel's working status and fault type; The data integration module is used to integrate the output results, including the photovoltaic panel serial number, geographic coordinates, and fault type information.

8. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 7, characterized in that: The image acquisition module includes an image storage unit, which is used to store thermal infrared photos of the photovoltaic power station with longitude, latitude, and attitude angle acquired by the thermal infrared camera carried by the drone in a local storage device and a cloud storage server, and the stored image data is encrypted using an encryption algorithm. The image acquisition module has an automatic calibration function, which automatically calibrates the parameters of the thermal infrared camera before shooting.

9. The photovoltaic panel defect detection and classification method based on thermal infrared images according to claim 7, characterized in that: When performing aerial triangulation, the photogrammetry module automatically adjusts measurement parameters according to the different terrains of the photovoltaic power station and the layout characteristics of the photovoltaic panels, including adjusting the distribution and number of image control points.

Images