Intelligent detection and positioning system and method for defects of photovoltaic modules

Abstract

This invention relates to the field of photovoltaic power generation system inspection technology, and in particular to an intelligent detection and location system and method for photovoltaic module defects. The intelligent detection and location method for photovoltaic module defects includes the following steps: S1, acquiring image data; S2, classifying the image data obtained in step S1; S3, classifying and labeling the data obtained in step S2 and generating a dataset; S4, training a recognition network on the dataset obtained in step S3 and generating a network model; S5, establishing a recognition and location system based on the network model obtained in step S4. The intelligent detection and location system for photovoltaic module defects includes: a data acquisition terminal, a data transmission and storage layer, an intelligent recognition and processing layer, and an application layer. This invention trains a neural network model to identify various defects in photovoltaic panels, including shading, wiring problems, inherent defects, and missing parts. The system provides functions such as one-click data import, one-click recognition, manual verification, and report viewing for user convenience.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation system testing technology, and in particular to an intelligent detection and location system and method for photovoltaic module defects. Background Technology

[0002] Photovoltaic power generation, as a clean and renewable energy source, has experienced rapid development in recent years. However, during long-term operation, photovoltaic panels are affected by various factors, such as sunlight, temperature, humidity, dust, and bird droppings, leading to various defects such as shading, wiring problems, and inherent defects (e.g., microcracks / damage). These defects reduce the power generation efficiency of the photovoltaic panels and may even affect the normal operation of the photovoltaic power generation system. Manual inspection is inefficient, costly, and susceptible to subjective factors; infrared thermal imaging technology can detect hot spots on photovoltaic panels, but it cannot distinguish the causes of these hot spots, such as shading, wiring problems, or inherent defects. Drone inspection can quickly acquire image data of photovoltaic panels, but it requires professional personnel, is costly, and is easily affected by weather conditions. There is an urgent need for an intelligent detection and location system for photovoltaic panel defects that combines multi-source data acquisition from visible light and infrared thermal imaging by drones. This system should be able to efficiently and accurately identify various defects in photovoltaic panels and precisely locate their positions, providing crucial technical support for the operation and maintenance management of photovoltaic power generation systems. Among the existing technologies for identifying hot spots on photovoltaic modules, one method is based on the electrical characteristics of solar cells, which involves monitoring the output voltage, output current, and output power of the power generation module. The other method is based on image analysis of the power generation module, which mainly involves analyzing the temperature distribution of the photovoltaic panel under different operating conditions through thermal imaging and using machine learning or image processing methods to identify hot spots.

[0003] For example, prior art invention 202210209242.6 discloses a method for detecting defects in photovoltaic panels based on visible light and infrared vision. It uses a YOLOv3 deep learning network model for surface stain detection and an adaptive minimum threshold segmentation algorithm to detect hot spot effect areas in photovoltaic panels based on infrared photovoltaic panel images. However, it does not form a comprehensive intelligent multi-source data classification, detection, and localization recognition system using visible light and infrared thermal imaging.

[0004] For example, invention 202110046360.5 discloses an intelligent inspection method and system for photovoltaic power plants based on UAV images, which realizes intelligent detection of photovoltaic panel defects, builds an inspection playback function, and facilitates intuitive debugging and inspection management by operation and maintenance personnel through visualization of the inspection process. However, this invention does not perform artificial intelligence classification and recognition of photovoltaic panel defect data, nor does it generate defect location reports by comparing visible light images.

[0005] For example, invention 202311492717.8 discloses a method that uses an identification device to collect image data, temperature data, and current data of photovoltaic panels under different lighting conditions, different angle conditions, and different sizes. This enables real-time image recognition of photovoltaic panels without manual intervention, reducing labor costs. However, it only recognizes single components and is not suitable for identifying hot spots on a large number of components in an entire photovoltaic power station.

[0006] For example, invention patent 202411389826.1 discloses a method for using drones to collect images of photovoltaic power plants, which solves the problem that the terrain of photovoltaic power plants is rugged and varied, the string arrangement is complex, and map software cannot provide path navigation. However, it does not integrate visible light image data and deep learning algorithms for intelligent identification and localization of problems with different components.

[0007] For example, invention 202411389826.1 discloses a method for identifying hot spots in solar photovoltaic modules based on infrared images. This method requires converting infrared images into grayscale images and does not introduce intelligent recognition algorithms. It is not suitable for the characteristics of photovoltaic fields with many sites, wide coverage, and large quantity, and is more suitable for the recognition of single images.

[0008] For example, invention patent 202410903620.X discloses a method for detecting hot spot faults in photovoltaic panels, which solves the technical problem of time-varying hot spot distribution caused by changes in the position and shape of hot spots due to changes in the sun's position and the presence of shading objects at different times. Invention CN202410487334.X discloses a method and system for detecting faults in new energy photovoltaic modules. By constructing a single-scale significance value that can effectively reflect the difference between weak hot spot areas and normal areas, the significance detection effect of the module saliency map is improved, and the accuracy of fault detection results is increased.

[0009] For example, invention 202010298498.X discloses a method for detecting hot spots on photovoltaic panels. It constructs an experimental platform and detects whether there are hot spots on photovoltaic panels based on experimental data, experimental images, measured data, and measured images. However, this invention focuses on detecting the presence of hot spots on photovoltaic panels under different levels of dirt.

[0010] For example, invention 202010740350.7 discloses a method and system for locating hot spots on photovoltaic panels, which processes infrared images of photovoltaic panels to determine whether each grid of infrared images of the photovoltaic panel contains hot spots. However, this invention cannot achieve intelligent classification, identification and location of different hot spots.

[0011] For example, invention 202011605219.6 discloses a method for detecting hot spots on photovoltaic panels, including a data acquisition device, a detection device, a cleaning mechanism, a control console, etc. It detects the presence of hot spots by comparing the separately acquired photovoltaic panel temperature with thermal imaging images. This invention mainly targets the detection of faults through infrared images, but does not combine visible light image information, which has limitations in stain location.

[0012] While existing technologies offer various detection methods as described above, the challenges remain for photovoltaic (PV) projects, especially those located in mountainous or offshore areas. Various defects in PV panels, such as shading, wiring problems, and inherent defects (like microcracks / damage), can reduce power generation efficiency and even disrupt the normal operation of the power generation system. Manual inspection is inefficient, costly, and susceptible to subjective factors. Although numerous identification methods combining drone inspection and artificial intelligence algorithms exist, they suffer from limitations such as limited data sources, location difficulties, limited applicability, and low accuracy. Summary of the Invention

[0013] To overcome the shortcomings of existing technologies, this invention provides a photovoltaic module defect intelligent detection and location system and method that can import diverse data, perform one-click intelligent identification, manual defect verification, precise location, and report generation, thereby significantly improving the operation and maintenance efficiency and accuracy of photovoltaic power plants. To solve the above technical problems, this invention provides the following technical solution: The intelligent detection and location method for photovoltaic module defects includes the following steps: S1. Collect image data: Use a drone equipped with a dual-mode infrared camera to conduct aerial inspections of photovoltaic power stations and collect zoom and infrared image data of photovoltaic panels. S2. Classify the image data obtained in step S1, and classify the infrared data according to the different types of the field station. S3. Classify and label the data obtained in step S2 and generate a dataset. Use image labeling software to classify and label the hotspots according to their types, including at least occlusion, short circuit, no load, damage, missing, and water accumulation. After classification and labeling, divide the data into training set, test set, and validation set in a ratio of 80:10:10. S4. Train the recognition network and generate a network model using the dataset obtained in step S3. The YOLOX model is used for training. S5. Based on the network model in step S4, establish an identification and positioning system to realize data import, one-click intelligent identification, manual defect verification and report generation functions.

[0014] Based on the above technical solutions, the present invention can also adopt the following further technical solutions: Step S3 further includes reading all the original images and labeled data of the training set, validation set, and test set respectively, associating the image data and labeled data, checking whether the label boxes are correct, that is, the label boxes cannot exceed the image edges, the width and height must be positive values, and displaying the labeling effect in the training set.

[0015] Step S4 includes: S41. Construct the YOLOX model; S42. Adjust the image size, perform data augmentation, and display the data augmentation effect in the visualization window; S43. Training parameter settings: Select adam as the optimization function, set the initial learning rate to 0.0001, set the minibatchsize to 16, and the maximum number of iterations to 100. Specifically, a cosine annealing learning rate decay strategy is adopted:

[0016] in The current learning rate, The initial learning rate, To minimize the learning rate, For the current round, This refers to the total number of rounds.

[0017] S44. Train the YOLOX detector, evaluate the model using the test set, start detecting targets, evaluate the difference between the detected targets and the real targets, and calculate precision, recall, and average accuracy (AP). Plot the PR curve and calculate the average precision. Specifically, the calculation formula is as follows:

[0018]

[0019] S45. Use the YOLOX detector to detect targets, load the trained network model, read the image to be detected, perform target detection, and display the detection results; Specifically, for the input image, the model outputs a set of detection results. ,in For category labels, For the coordinates of the predicted bounding box, This represents the confidence score. A confidence threshold is set. Only retain The test results.

[0020] S46. Design a batch target detection system based on MATLAB appdesigner.

[0021] Step S41 includes: S411. The Anchor-Free method is adopted to transform target detection into a direct regression of the target's center point and its width and height. Specifically, for each position (x, y) on the input image, the model predicts the probability p that the target center point exists. ij The regression formula is as follows: (The formula is incomplete and requires further context to be fully translated.)

[0022] in, This is the grid cell offset. This is the offset predicted by the network. For the preset scale, This represents the scaling factor for network predictions. The Anchor-Free mechanism simplifies the model structure while improving detection accuracy and generalization ability.

[0023] S412. Decouple the classification task from the regression task and process them separately through two independent convolutional branches; Specifically, the classification branch is used to predict the target class probability P. class The regression branch is used to predict the bounding box coordinates and the confidence level P. obj Let the input feature map be... The output of the decoupling detection head can then be expressed as:

[0024] in For the number of categories, This represents the number of regression parameters (typically 4 coordinates plus 1 confidence score). This decoupled structure effectively avoids mutual interference between classification and regression tasks in feature learning, improving the model's target localization ability in complex contexts.

[0025] S413. A dynamic label allocation strategy is adopted, which dynamically allocates the optimal positive sample to each target by calculating the cost matrix between the predicted box and the ground truth box. Specifically, let the set of prediction boxes be... The set of real bounding boxes is For the first The first prediction box and the first prediction box A real frame, cost Defined as:

[0026] in For classifying losses, To regress the loss, For the prior loss at the center point, and The balancing coefficient is determined by solving the optimal transport problem, and is dynamically selected for each ground truth box. Using the optimal prediction box as a positive sample avoids the sample imbalance problem caused by the traditional fixed threshold allocation method, which is especially suitable for scenarios where the defect targets in photovoltaic modules are small and the categories are unbalanced.

[0027] S414. The total loss function consists of classification loss, regression loss, and confidence loss.

[0028] Specifically, its expression is:

[0029] in: Classification loss Using cross-entropy loss (BCE Loss):

[0030] Regression loss Using IoU Loss (GIoU Loss):

[0031] in , For inclusion and The smallest closure region.

[0032] Confidence loss Use binary cross-entropy loss to measure the probability of the target's existence:

[0033] Step S42 includes color dithering enhancement, random translation, and random scaling in the HSV space; The color dithering enhancement in the HSV space is achieved by adding random perturbations to the hue (H), saturation (S), and brightness (V) channels, respectively. The specific formula is as follows:

[0034] The random translation performs a horizontal or vertical translation on the image with probability p, and the translation amounts Δx and Δy follow a uniform distribution. The random scaling is based on probability. Scaling the image, scaling factor And adjust the coordinates of the annotation box accordingly.

[0035] Furthermore, the Mosaic data augmentation method is integrated, which randomly scales four images and stitches them into one, enhancing the model's robustness to defects in photovoltaic modules under different scales, lighting conditions, and shading conditions.

[0036] Step S46 includes: S4601. Add an attribute to the class definition to store the index of the currently displayed image, initialize it to 1, convert the degree, minute, and second array to a string in degree, minute, and second format, and set the selected folder as the working path; S4602. Prompt the user to select multiple image files, create a progress bar, ensure that the file name is a cell array, initialize a counter, iterate through each selected image file, update the progress bar, and read the image metadata; S4603. Obtain longitude, latitude, altitude, and modification date information from metadata; obtain the modification date; convert longitude and latitude into strings in degree, minute, and second format; add image information and date to the table data; update the number of successfully imported images; and close the progress bar. S4604, Update the text boxes for the total number of images imported successfully and those that failed to import, display the data in a table, load the YOLOX object detector, obtain the total number of images, initialize the total time, create and display the progress, iterate through each image, and update the progress bar; S4605. Obtain the image name, read the image and start timing, use the YOLOX model for object detection, set the confidence threshold, stop timing and accumulate the detection time. If the detection result is not empty, process the detection result, convert the detection label into a string array, and iterate through each detection result. S4606. Obtain the longitude, latitude, and sampling time of the image, add the image information to the table data, ensuring that it is a character and numeric type, close the progress bar dialog box, calculate the average frame rate and display it; S4607. Obtain the names of the infrared and visible light images, read and display the infrared image, read and display the visible light image, perform target detection, add a label to each detected target, and display a prompt on the image indicating that no target was detected; S4608. Obtain and display relevant defect information, the index of the currently displayed image, the name of the infrared image, construct the filename of the visible light image, read the infrared image and refresh the display; S4609. Read the visible light image and refresh the display to show the defect text information and detection results; S4610. Check if the last image has been reached. A pop-up prompt box asks whether to re-verify. Perform the corresponding operation according to the user's selection, re-verify, return to the first image, and end. No further operation is required. S4611. Update the currently displayed image index, display the next image, obtain the currently locked row index, check if there is any data that can be deleted, if so, delete the current row data, update the currently displayed image index to prevent the index from exceeding the range, if the last row has been deleted, move the index to the last row, if all rows have been deleted, reset it to 1, and refresh the display. S4612. Obtain the project name, check if the project name has been entered, check if the table data is empty, and set the saved file name to the project name. S4613. Create a Word document, initialize the progress bar, process each group of images, update the progress bar, obtain the names of the infrared and visible light images, read the images, the labeled infrared images, insert target annotations, combine the three images into one, and save the combined image as a temporary file. S4614. Insert the combined image into Word, add latitude and longitude information (below the image), center the latitude and longitude information, add a newline and insert a blank line, delete temporary files, close the progress bar, move the cursor to the beginning of the document, insert "defect details", move the cursor to the beginning of the current text, generate the report and format it, and prompt the user that the save was successful.

[0037] On the other hand, the present invention also provides the following technical solutions: A photovoltaic module defect intelligent detection and location system includes: The data acquisition end uses a drone equipped with a dual-mode imaging module; In the data transmission and storage layer, the drone uploads the collected image data to the cloud server in real time through the 5G / 4G communication module, or collects the data offline using the onboard storage card and then uploads it in a unified manner. The cloud server is configured with object storage service, which stores the data according to the site, time, and image type (visible light / infrared). The intelligent recognition and processing layer deploys the YOLOX target detection model on a cloud server, provides a RESTful API interface, receives image data, calls the model to perform inference, returns defect detection results, and deploys a MySQL database to record detection history, manual review results, and report generation records. At the application layer, users access the system via web or desktop software to perform functions such as data import, one-click recognition, manual verification, report generation, and download.

[0038] The data acquisition terminal includes: Visible light camera, used to capture images of the photovoltaic panel surface, with a resolution of no less than 20 megapixels and support for 2x optical zoom; An infrared thermal imaging camera is used to acquire images of the temperature distribution of photovoltaic panels, with a resolution of no less than 640×512 pixels, a temperature measurement range of -20℃ to 150℃, and an accuracy of ±2℃. A high-precision RTK-GPS module is used to record the latitude and longitude coordinates of each image, with positioning accuracy at the centimeter level; The airborne edge computing unit enables real-time image preprocessing and transmission, based on NVIDIA Jetson series modules.

[0039] Compared with the prior art, the beneficial effects that this invention can achieve are: 1. This invention constructs an "image-annotation" alignment mechanism at the data level, extracts high-dimensional semantic features of visible light and infrared images respectively through the YOLOX decoupled detection head at the feature level, and comprehensively judges the defect type and location through a confidence fusion strategy at the decision level.

[0040] 2. YOLOX's Anchor-Free mechanism, combined with a multi-scale feature pyramid structure, can effectively capture the features of small targets; at the same time, it enhances the simulation of defect manifestations at different scales through Mosaic data, thereby improving the model's response to small targets.

[0041] 3. This invention introduces enhancement strategies such as HSV color dithering, random translation, and random scaling during the training phase, and combines UAV terrain-following flight and orthophoto acquisition during the deployment phase to ensure the consistency of input image quality and improve the robustness of the model in complex environments.

[0042] 4. In order to solve the problems of low efficiency, poor accuracy and high cost of defect inspection of photovoltaic power station components, this invention proposes an intelligent inspection and positioning method based on visible light and infrared thermal imaging image data, and develops a software system that realizes the functions of data import, one-click intelligent identification, manual defect verification, accurate positioning and report generation.

[0043] This invention trains a neural network model to identify various defects in photovoltaic (PV) panels, including shading, wiring problems, inherent defects (such as microcracks / damage), and missing parts. Compared to traditional image processing methods, deep learning algorithms can better capture complex features in images, thus improving the accuracy of defect identification. Simultaneously, it integrates visible light images and infrared thermal imaging data for a more comprehensive analysis of the PV panel's condition. Visible light images provide surface information, while infrared thermal imaging provides thermal distribution information; combining the two allows for more accurate identification of defect types and locations. It not only identifies PV panel defects but also precisely locates them and provides a defect identification report. The report includes information such as defect type, defect location, and identification accuracy, facilitating defect repair and management by maintenance personnel. The system employs a graphical user interface, making operation simple, intuitive, and user-friendly. It offers one-click data import, one-click identification, manual verification, and report viewing functions for convenient user operation. Attached Figure Description

[0044] Figure 1 This is the PR curve of the model of the present invention.

[0045] Figure 2 This is the overall system architecture and data flow diagram of the present invention.

[0046] Figure 3 This is the software interface of the present invention.

[0047] Figure 4 This is the test report for the present invention. Detailed Implementation

[0048] The intelligent detection and location system and method for photovoltaic module defects provided by the present invention will be further described with reference to the accompanying drawings.

[0049] This embodiment was practically verified in a 50MW mountain photovoltaic power station. A total of 12,000 visible light and infrared images were collected, and 8,500 defect samples were labeled, covering typical defect types such as shading, hidden cracks, hot spots, and loose connections. After model training, the average accuracy on the test set reached 92.3%, meeting the requirements for real-time inspection. Compared with existing methods based on YOLOv3 or traditional image segmentation, this invention improves defect identification accuracy by approximately 12%, improves positioning accuracy by approximately 8%, and reduces manual review workload by more than 70%, significantly improving the operation and maintenance efficiency of photovoltaic power stations.

[0050] The detailed definitions of the model evaluation metrics are shown in Table 1:

[0051] The experimental verification data are shown in Table 2:

[0052] The AP values ​​for various defects are shown in Table 3:

[0053] The system and method described in this embodiment have significant advantages in the field of photovoltaic panel defect detection. They can effectively improve the operation and maintenance efficiency of photovoltaic power generation systems, reduce operation and maintenance costs, and provide technical support for the healthy development of the photovoltaic power generation industry. The specific system and method are as follows: 1. UAV flight and image acquisition parameters; To ensure image quality and subsequent recognition accuracy, this invention has established standardized flight and acquisition procedures, as shown in Table 4:

[0054] 2. Data acquisition and model training; Start the drone, set the infrared image to 1x zoom and the zoomed image to 2x zoom, set the camera to orthogonal acquisition, plan the drone's flight path, and collect image data of defects in different photovoltaic power station components.

[0055] After obtaining the data, manual classification and labeling are performed. Based on the collected infrared hot spot data and visible light image data, software is used to manually label the defect categories of the problem images, generating a dataset with defect label boxes.

[0056] Build a YOLOX recognition model, load the dataset, set network parameters, perform data augmentation, train the recognition model, evaluate the model using the test set, and make optimization adjustments.

[0057] Model training and optimization process: During the model training phase, this invention further optimizes the adaptability of YOLOX in photovoltaic module defect detection tasks: Input image preprocessing: The visible light images and infrared thermal images acquired by the UAV are uniformly scaled to 640×640 resolution to maintain the original aspect ratio and avoid image distortion affecting feature extraction.

[0058] Feature extraction network: CSPNet (Cross Stage Partial Network) is used as the backbone network to enhance feature reuse capability while ensuring computational efficiency and improve sensitivity to defects in small targets (such as hidden cracks and occluded edges).

[0059] Loss function optimization: Multi-task joint optimization is adopted to ensure that the model achieves a balance between defect classification and localization accuracy. The specific form of the loss function is as described above.

[0060] Training strategy: A cosine annealing learning rate decay strategy was adopted, with an initial learning rate of 0.0001, a momentum parameter of 0.937, a weight decay coefficient of 0.0005, and 100 training epochs. Model weights were saved every 10 epochs during training, and the model with the highest mean average precision (mAP) on the validation set was selected as the final deployment model.

[0061] 3. Defect detection and location process; Drones are used to inspect specific photovoltaic power stations that need to be inspected. For mountainous power stations, in order to ensure that orthophotos can be collected, they need to be set to the terrain-following flight mode, and at the same time, visible light and infrared thermal imaging data are acquired.

[0062] Images are uploaded to a cloud server via image transmission, and the server then calls the intelligent recognition system to identify defects in the collected data.

[0063] First, import the cloud dataset by entering information such as the project name, detection time, and weather conditions. Second, enter the intelligent recognition module and use the pre-trained YOLOX model to perform batch intelligent recognition on the imported data and extract the category labels of the problem components; Third, manual review is conducted. The results of the intelligent recognition are checked manually to see if they are correct. If they are correct, click "Confirm" to proceed to the next image. If they are incorrect, click "Negative" to delete the data. Fourth, after the manual review is completed, the report generation module will be entered. The manually confirmed data will form a defect database. The system will extract information such as the defect type, detection time, and latitude and longitude of each image, and generate a Word version of the detection report, which will be uploaded to the cloud server for on-site maintenance personnel to download and perform targeted repairs according to the defect category.

[0064] The present invention has been described with reference to preferred embodiments; however, those skilled in the art will understand that various changes in form and detail may be made within the scope of the claims.

Claims

1. A method for intelligent detection and location of defects in photovoltaic modules, characterized in that... Includes the following steps: S1. Collect image data: Use a drone equipped with a dual-mode infrared camera to conduct aerial inspections of photovoltaic power stations and collect zoom and infrared image data of photovoltaic panels. S2. Classify the image data obtained in step S1, and classify the infrared data according to the different types of the field station. S3. Classify and label the data obtained in step S2 and generate a dataset. Use image labeling software to classify and label the hotspots according to their types, including at least occlusion, short circuit, no load, damage, missing, and water accumulation. After classification and labeling, divide the data into training set, test set, and validation set in a ratio of 80:10:

10. S4. Train the recognition network and generate a network model using the dataset obtained in step S3. The YOLOX model is used for training. S5. Based on the network model in step S4, establish an identification and positioning system to realize data import, one-click intelligent identification, manual defect verification and report generation functions.

2. The intelligent detection and location method for photovoltaic module defects according to claim 1, characterized in that... Step S3 further includes reading all the original images and labeled data of the training set, validation set, and test set respectively, associating the image data and labeled data, checking whether the label boxes are correct, that is, the label boxes cannot exceed the image edges, the width and height must be positive values, and displaying the labeling effect in the training set.

3. The intelligent detection and location method for photovoltaic module defects according to claim 1, characterized in that... Step S4 includes: S41. Construct the YOLOX model; S42. Adjust the image size, perform data augmentation, and display the data augmentation effect in the visualization window; S43. Training parameter settings: Select adam as the optimization function, set the initial learning rate to 0.0001, set the minibatchsize to 16, and the maximum number of iterations to 100. S44. Train the YOLOX detector, evaluate the model using the test set, start detecting targets, evaluate the difference between the detected targets and the real targets, and calculate precision, recall, and average accuracy (AP). Plot the PR curve and calculate the average precision. S45. Use the YOLOX detector to detect targets, load the trained network model, read the image to be detected, perform target detection, and display the detection results; S46. Design a batch target detection system based on MATLAB appdesigner.

4. The intelligent detection and location method for photovoltaic module defects according to claim 3, characterized in that... Step S41 includes: S411. The Anchor-Free method is adopted to transform target detection into a direct regression of the target's center point and its width and height. S412. Decouple the classification task from the regression task and process them separately through two independent convolutional branches; S413. A dynamic label allocation strategy is adopted, which dynamically allocates the optimal positive sample to each target by calculating the cost matrix between the predicted box and the ground truth box. S414. The total loss function consists of classification loss, regression loss, and confidence loss.

5. The intelligent detection and location method for photovoltaic module defects according to claim 3, characterized in that... Step S42 includes color dithering enhancement, random translation, and random scaling in the HSV space; The color dithering enhancement in the HSV space is achieved by adding random perturbations to the hue (H), saturation (S), and brightness (V) channels, respectively. The random translation performs a horizontal or vertical translation on the image with probability p, and the translation amounts Δx and Δy follow a uniform distribution. The random scaling is based on probability. Scaling the image, scaling factor And adjust the coordinates of the annotation box accordingly.

6. The intelligent detection and location method for photovoltaic module defects according to claim 3, characterized in that... Step S46 includes: S4601. Add an attribute to the class definition to store the index of the currently displayed image, initialize it to 1, convert the degree, minute, and second array to a string in degree, minute, and second format, and set the selected folder as the working path; S4602. Prompt the user to select multiple image files, create a progress bar, ensure that the file name is a cell array, initialize a counter, iterate through each selected image file, update the progress bar, and read the image metadata; S4603. Obtain longitude, latitude, altitude, and modification date information from metadata; obtain the modification date; convert longitude and latitude into strings in degree, minute, and second format; add image information and date to the table data; update the number of successfully imported images; and close the progress bar. S4604, Update the text boxes for the total number of images imported successfully and those that failed to import, display the data in a table, load the YOLOX object detector, obtain the total number of images, initialize the total time, create and display the progress, iterate through each image, and update the progress bar; S4605. Obtain the image name, read the image and start timing, use the YOLOX model for object detection, set the confidence threshold, stop timing and accumulate the detection time. If the detection result is not empty, process the detection result, convert the detection label into a string array, and iterate through each detection result. S4606. Obtain the longitude, latitude, and sampling time of the image, add the image information to the table data, ensuring that it is a character and numeric type, close the progress bar dialog box, calculate the average frame rate and display it; S4607. Obtain the names of the infrared and visible light images, read and display the infrared image, read and display the visible light image, perform target detection, add a label to each detected target, and display a prompt on the image indicating that no target was detected; S4608. Obtain and display relevant defect information, the index of the currently displayed image, the name of the infrared image, construct the filename of the visible light image, read the infrared image and refresh the display; S4609. Read the visible light image and refresh the display to show the defect text information and detection results; S4610. Check if the last image has been reached. A pop-up prompt box asks whether to re-verify. Perform the corresponding operation according to the user's selection, re-verify, return to the first image, and end. No further operation is required. S4611. Update the currently displayed image index, display the next image, obtain the currently locked row index, check if there is any data that can be deleted, if so, delete the current row data, update the currently displayed image index to prevent the index from exceeding the range, if the last row has been deleted, move the index to the last row, if all rows have been deleted, reset it to 1, and refresh the display. S4612. Obtain the project name, check if the project name has been entered, check if the table data is empty, and set the saved file name to the project name. S4613. Create a Word document, initialize the progress bar, process each group of images, update the progress bar, obtain the names of the infrared and visible light images, read the images, the labeled infrared images, insert target annotations, combine the three images into one, and save the combined image as a temporary file. S4614. Insert the combined image into Word, add latitude and longitude information (below the image), center the latitude and longitude information, add a newline and insert a blank line, delete temporary files, close the progress bar, move the cursor to the beginning of the document, insert "defect details", move the cursor to the beginning of the current text, generate the report and format it, and prompt the user that the save was successful.

7. A photovoltaic module defect intelligent detection and location system, characterized in that... include: The data acquisition end uses a drone equipped with a dual-mode imaging module; In the data transmission and storage layer, the drone uploads the collected image data to the cloud server in real time through the 5G / 4G communication module, or collects the data offline using the onboard storage card and then uploads it in a unified manner. The cloud server is configured with object storage service, which stores the data according to the site, time, and image type (visible light / infrared). The intelligent recognition and processing layer deploys the YOLOX target detection model on a cloud server, provides a RESTful API interface, receives image data, calls the model to perform inference, returns defect detection results, and deploys a MySQL database to record detection history, manual review results, and report generation records. At the application layer, users access the system via web or desktop software to perform functions such as data import, one-click recognition, manual verification, report generation, and download.

8. The intelligent detection and location system for photovoltaic module defects according to claim 7, characterized in that... The data acquisition terminal includes: Visible light camera, used to capture images of the photovoltaic panel surface, with a resolution of no less than 20 megapixels and support for 2x optical zoom; An infrared thermal imaging camera is used to acquire images of the temperature distribution of photovoltaic panels, with a resolution of no less than 640×512 pixels, a temperature measurement range of -20℃ to 150℃, and an accuracy of ±2℃. A high-precision RTK-GPS module is used to record the latitude and longitude coordinates of each image, with positioning accuracy at the centimeter level; The airborne edge computing unit enables real-time image preprocessing and transmission, based on NVIDIA Jetson series modules.

Images