An intelligent detection method and system for photovoltaic module damage based on UAV images

Abstract

The present application provides a method and system for intelligently detecting photovoltaic module damage based on drone imagery, relating to the field of drone image technology. The method comprises: a drone collects multi-angle, high-altitude images of a photovoltaic module array along a predetermined path and records geographic location information; using this information, image stitching and overlap correction are performed to form a high-resolution panoramic image; combining a pre-built three-dimensional topological structure model, the photovoltaic modules in the panoramic image are identified and marked to generate a module distribution map; by analyzing the actual spectral response characteristics of each module and comparing them with standard data in a materials science database, suspected damaged areas are screened; adaptive threshold algorithms and morphological operations are applied to these areas to generate feature maps; and by integrating real-time environmental factors and historical performance changes, target features that affect photovoltaic module performance degradation are extracted, damage status is predicted, and ultimately a detection report is generated. The present application improves the efficiency of photovoltaic module detection.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of drone image technology, and in particular to a method and system for intelligent detection of photovoltaic module damage based on drone images. Background Art

[0002] The widespread adoption of solar photovoltaic systems has placed higher demands on the maintenance and monitoring of large-scale PV arrays. To ensure the efficient and reliable operation of these systems, regular inspections of PV panels are necessary to detect potential issues such as cracks, dirt, and aging. Traditional manual inspection methods are inefficient and costly, but the development of drone technology has made it possible to achieve rapid, efficient, and cost-effective PV panel monitoring. Drones can automatically capture multi-angle aerial images along a predetermined flight path and record geographic location information, providing a precise data foundation for subsequent image analysis.

[0003] Currently, the main methods for monitoring the condition of photovoltaic modules include manual on-site inspections, fixed camera monitoring, and drone-based image acquisition. Drone-based methods can already capture and generate panoramic images, identify the location of photovoltaic modules using 3D models, and locate problem areas by comparing the characteristics of intact and damaged modules through spectral analysis. Furthermore, some solutions are beginning to incorporate machine learning algorithms for automated feature extraction and damage prediction.

[0004] Although existing solutions have achieved automated monitoring of PV module status to a certain extent, several shortcomings still exist. First, the traditional image stitching and correction process is prone to introducing errors, affecting the accuracy of the final diagnostic results. Second, existing spectral response characteristic analysis often lacks comprehensive consideration of real-time environmental factors and historical performance data, resulting in inaccurate damage prediction. Finally, although some solutions have begun to adopt machine learning methods, most have not yet fully utilized advanced technologies such as deep learning, transfer learning, and random forests, which limits the accuracy of feature extraction and the generalization ability of prediction models. Therefore, developing a solution that integrates advanced image processing technology and artificial intelligence algorithms is crucial to improving the efficiency and accuracy of PV module monitoring. Summary of the Invention

[0005] The embodiments of the present application provide a method and system for intelligent detection of photovoltaic module damage based on drone images, which is used to solve the problem of low efficiency of photovoltaic module detection in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a method for intelligently detecting photovoltaic module damage based on drone images, comprising:

[0007] Collect multi-angle high-altitude images of the photovoltaic component array collected by the drone on a predetermined flight path, and record geographic location information corresponding to the multi-angle high-altitude images;

[0008] Based on the geographic location information, the multi-angle high-altitude images are stitched and overlapped areas are corrected to obtain a high-resolution panoramic image;

[0009] Using a pre-built three-dimensional topological structure model of photovoltaic modules, identifying and locating photovoltaic modules in the high-resolution panoramic image, and marking the photovoltaic modules to generate a module distribution map;

[0010] Based on the component distribution map, the actual spectral response characteristics corresponding to each photovoltaic component are analyzed. The reflection characteristic data of each photovoltaic component in a good state and the spectral response characteristics under different damage modes are obtained from a pre-built material science database. These data are used as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas.

[0011] Processing the suspected damage area using an adaptive threshold algorithm and morphological operations to generate a suspected damage area;

[0012] Combined with the suspected damaged area, the real-time environmental factors of the photovoltaic module and the historical performance change data, the target features are extracted, and the influence relationship between the target features and the performance degradation of the photovoltaic module is identified. Based on the target features and the influence relationship, the damage status of the photovoltaic module is predicted to generate a detection report.

[0013] Optionally, the suspected damage area is processed using an adaptive threshold algorithm and a morphological operation to generate a suspected damage area, including:

[0014] Processing the suspected damaged area according to the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain an optimal threshold corresponding to each pixel;

[0015] Based on the optimal threshold, pixels in the suspected damage area are divided into two categories, foreground and background, to obtain classified pixels; based on the classified pixels, a suspected damage area map is generated; the suspected damage area map is analyzed to identify suspected damage plaques and generate suspected damage plaque data;

[0016] Determining a target structural element based on the suspected damage plaque data, and performing dilation, erosion, and morphological operations on the suspected damage region map based on the target structural element to generate an optimized suspected damage region map to remove noise from the suspected damage region;

[0017] Performing frequency domain filtering on the optimized suspected damage area map to obtain a preliminary suspected damage feature map;

[0018] The preliminary image of suspected damage features is processed to generate a suspected damage area.

[0019] Optionally, the suspected damage area is processed according to the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain the optimal threshold corresponding to each pixel, including:

[0020] performing a smoothing process on a local neighborhood defined by each pixel point in the high-resolution panoramic image to obtain a local neighborhood image;

[0021] Analyzing the grayscale value distribution of all pixels in the local neighborhood image to establish a local grayscale distribution model, and applying a histogram equalization technique to process the local neighborhood image to obtain a local image;

[0022] Based on the local image and the local grayscale distribution model, constructing a composite objective function based on a structural similarity index and edge regularization, evaluating the segmentation effect of the suspected damage area under different thresholds based on the composite objective function to obtain a segmentation evaluation result, and optimizing the composite objective function by combining the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function;

[0023] An initial candidate threshold is defined, and the initial candidate threshold is optimized in combination with the optimization objective function to obtain an optimized threshold, and the optimized threshold is used as the optimal threshold of the local grayscale distribution characteristic.

[0024] Optionally, based on the local image and the local grayscale distribution model, a composite objective function based on a structural similarity index and edge regularization is constructed; based on the composite objective function, the segmentation effect of the suspected damage area under different thresholds is evaluated to obtain a segmentation evaluation result; and the composite objective function is optimized in combination with the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function, including:

[0025] Based on the local image and the local grayscale distribution model, the segmentation effect of the suspected damage area is evaluated, and edge detection is performed on the local image to obtain edge information. Based on the edge information and an edge regularization term, a composite objective function is constructed;

[0026] Based on the composite objective function, the suspected damage area is segmented to obtain a segmentation result, and the segmentation effect of the suspected damage area is evaluated using the segmentation result to generate a segmentation evaluation result set;

[0027] The segmentation evaluation result set is divided into multiple subsets. According to the subsets, a part of the subsets is used as a test set, and the remaining part of the subsets is used as a training set. The training set and the test set are used to perform multiple iterative evaluations on the segmentation effect of the suspected damage area to obtain an optimized objective function.

[0028] Optionally, based on the component distribution map, analyzing the actual spectral response characteristics corresponding to each photovoltaic component, obtaining the reflection characteristic data of each photovoltaic component in an intact state and the spectral response characteristics under different damage modes from a pre-built material science database as standard spectral response characteristics, comparing the actual spectral response characteristics with the standard spectral response characteristics, and screening out suspected damaged areas, including:

[0029] Based on the component distribution diagram, analyzing the actual spectral response characteristics corresponding to each photovoltaic component to obtain characteristic information;

[0030] Based on the characteristic information, the high-resolution panoramic image is partitioned to obtain photovoltaic module areas, and the boundaries of the photovoltaic module areas are optimized to obtain segmentation results, and the segmentation results are analyzed to generate a photovoltaic module area map;

[0031] Based on the photovoltaic module area map, a deep belief network is constructed, and hyperparameters of the deep belief network are adjusted to obtain an optimized deep belief network. The optimized deep belief network is used to analyze and process reflection characteristic data of the photovoltaic module area map to generate reflection characteristic data corresponding to the material of each photovoltaic module in the photovoltaic module area;

[0032] Extracting key spectral features from spectral data in the reflectance characteristic data, and performing abnormal classification processing on the key spectral features to obtain classified key spectral features, using the classified key spectral features to determine spectral response characteristics to be optimized, and dynamically adjusting the spectral response characteristics to be optimized based on actual ambient lighting conditions to obtain standard spectral response characteristics;

[0033] The actual spectral response characteristic is compared with the standard spectral response characteristic to generate spectral analysis parameters. Based on the spectral analysis parameters, a spectral analysis is performed on each photovoltaic module area to screen out suspected damaged areas.

[0034] Optionally, based on the photovoltaic module area map, a deep belief network is constructed, and hyperparameters of the deep belief network are adjusted to obtain an optimized deep belief network. The optimized deep belief network is used to analyze and process reflection characteristic data of the photovoltaic module area map to generate reflection characteristic data corresponding to the materials of each photovoltaic module in the photovoltaic module area, including:

[0035] Constructing a deep belief network based on the data of the photovoltaic module area map;

[0036] In combination with a Gaussian process surrogate model and an acquisition function, a search process is performed on the hyperparameter space of the deep belief network to obtain an optimal hyperparameter combination, a performance evaluation process is performed on the optimal hyperparameter combination to obtain an evaluation result, and an optimal hyperparameter configuration is selected based on the evaluation result to optimize the deep belief network and generate an optimized deep belief network;

[0037] Based on the optimized deep belief network, the reflection characteristic data of the photovoltaic module area map is analyzed and processed to generate reflection behavior data. According to the standard reflection behavior data of the existing standard samples, the reflection behavior data is corrected and verified to obtain the reflection characteristic data corresponding to the materials of each photovoltaic module in the photovoltaic module area.

[0038] Optionally, the using a pre-built three-dimensional topological structure model of photovoltaic components to identify and locate photovoltaic components in the high-resolution panoramic image and mark the photovoltaic components to generate a component distribution map includes:

[0039] Using a pre-built three-dimensional topological structure model of a photovoltaic module, automatically identifying and locating the photovoltaic module in the high-resolution panoramic image to generate an identification result;

[0040] Marking the identification results to establish a topological connection relationship between the photovoltaic modules and generate a photovoltaic module distribution network;

[0041] Based on the photovoltaic component distribution network, a component distribution map is generated, wherein the component distribution map records the positions and mutual relationships of all photovoltaic components.

[0042] In a second aspect, an embodiment of the present application provides an intelligent photovoltaic module damage detection system based on drone images, comprising:

[0043] A collection module is used to collect multi-angle high-altitude images of the photovoltaic module array collected by the drone on a predetermined flight path, and record the geographical location information corresponding to the multi-angle high-altitude images;

[0044] A correction module, configured to stitch and correct overlapping areas of the multi-angle high-altitude images based on the geographic location information to obtain a high-resolution panoramic image;

[0045] an identification module, configured to identify and locate photovoltaic modules in the high-resolution panoramic image using a pre-built three-dimensional topological structure model of photovoltaic modules, and mark the photovoltaic modules to generate a module distribution map;

[0046] an analysis module for analyzing the actual spectral response characteristics corresponding to each photovoltaic module based on the module distribution map, obtaining the reflectance characteristic data of each photovoltaic module in an intact state and the spectral response characteristics under different damage modes from a pre-built material science database as standard spectral response characteristics, comparing the actual spectral response characteristics with the standard spectral response characteristics, and screening out suspected damaged areas;

[0047] A generation module is used to extract target features by combining the suspected damaged area, the real-time environmental factors of the photovoltaic module and the historical performance change data, and to identify the influence relationship between the target features and the performance degradation of the photovoltaic module. Based on the target features and the influence relationship, the damage status of the photovoltaic module is predicted to generate a detection report.

[0048] In a third aspect, an embodiment of the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the methods described in the first aspect for intelligent detection of photovoltaic component damage based on drone images.

[0049] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon, which, when executed by a processor, implements an intelligent photovoltaic component damage detection method based on drone images as described in any one of the first aspects.

[0050] In an embodiment of the present application, multi-angle high-altitude images of a photovoltaic module array collected by a drone along a predetermined flight path are collected, and geographic location information corresponding to the multi-angle high-altitude images is recorded. Based on the geographic location information, the multi-angle high-altitude images are stitched and overlap correction is performed on the multi-angle high-altitude images to obtain a high-resolution panoramic image. A pre-constructed three-dimensional topological structure model of the photovoltaic module is used to identify and locate the photovoltaic modules in the high-resolution panoramic image, and the photovoltaic modules are marked to generate a module distribution map. Based on the module distribution map, the actual spectral response characteristics corresponding to each photovoltaic module are analyzed. The reflection characteristic data of each photovoltaic module in an intact state and the spectral response characteristics under different damage modes are obtained from a pre-constructed material science database as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas. The suspected damaged areas are processed using an adaptive threshold algorithm and morphological operations to generate suspected damaged areas. Target features are extracted by combining the suspected damaged areas, real-time environmental factors of the photovoltaic modules, and historical performance change data. The influence relationship between the target features and the performance degradation of the photovoltaic modules is identified. Based on the target features and the influence relationship, the damage state of the photovoltaic modules is predicted to generate a test report.

[0051] The technical solution of this application has the following beneficial effects:

[0052] This application improves PV module inspection efficiency and reduces labor costs. Accurate damage location and prediction facilitate timely maintenance, extending the lifespan of PV systems. Automated analysis reduces human error and improves the reliability of diagnostic results. The combination of real-time monitoring and historical data analysis provides strong support for optimizing PV system operations.

[0053] Furthermore, the present embodiment optimizes the processing of suspected damaged areas. By combining adaptive thresholding algorithms based on Bayesian optimization, connected component analysis, adaptive selection of structural elements, and morphological operations, this method achieves precise segmentation and feature extraction of suspected damaged areas in high-resolution panoramic images of photovoltaic modules. Furthermore, the application of fast Fourier transforms and U-Net deep learning models effectively improves the quality of suspected damaged areas and enhances the reliability of subsequent analysis.

[0054] The above methods improve the accuracy and efficiency of identifying suspected damaged areas during photovoltaic module monitoring. First, the adaptive threshold algorithm using Bayesian optimization can find the optimal threshold for each pixel based on the local grayscale distribution characteristics, thereby more accurately separating the foreground and background and reducing misjudgments. Second, connected component analysis helps identify independent suspected damaged patches and generate more detailed suspected damaged patch data, while adaptive structural element selection ensures the effectiveness of morphological operations (such as dilation and erosion), better removing noise without compromising true damage information. Finally, frequency domain filtering using fast Fourier transform can eliminate unnecessary high-frequency noise, while the U-Net network model can automatically and intelligently extract the most representative suspected damage features from the preliminary image, ultimately generating high-quality suspected damaged areas. This not only improves detection accuracy but also provides a solid foundation for subsequent damage assessment.

[0055] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0057] Figure 1 A flowchart of a method for intelligently detecting photovoltaic module damage based on drone images provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the structure of an intelligent photovoltaic module damage detection system based on drone images provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0062] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0063] Figure 1 A flowchart of a method for intelligently detecting photovoltaic module damage based on drone images is provided for the embodiment of the present application. Figure 1 As shown, the method includes:

[0064] Step 101: collecting multi-angle high-altitude images of a photovoltaic module array captured by a drone on a predetermined flight path, and recording geographic location information corresponding to the multi-angle high-altitude images;

[0065] In this step, geographic location information refers to location data obtained through positioning technologies such as GPS, including latitude, longitude, and altitude. This data accurately describes the ground location of each image captured by the drone. Multi-angle aerial imagery refers to images of the PV array acquired from different perspectives, providing rich visual information.

[0066] In practice, the drone flew along a pre-set path, capturing images of the PV array at varying heights and angles. With each shot, the drone recorded its precise geographic location, which not only helped ensure geographic consistency between images but also provided the necessary spatial coordinate information for subsequent data processing.

[0067] For example, during a maintenance inspection of a large solar power plant, operators programmed one or more drones to autonomously fly in a specific pattern, covering the entire plant area. The drones' cameras captured images of the photovoltaic panels from multiple angles during flight, recording the exact location of each image. This data was then transmitted to a ground station for further processing.

[0068] Step 102: Based on the geographic location information, the multi-angle high-altitude images are stitched and overlapped areas are corrected to obtain a high-resolution panoramic image;

[0069] In this step, stitching is the process of merging multiple images into a larger image; overlap correction is to eliminate inconsistencies between adjacent images caused by changes in perspective or illumination to ensure the quality of the stitched image.

[0070] In practice, image stitching software, combined with geolocation information, stitches together images from different angles of the same PV array into a complete, high-resolution panorama. The software automatically identifies and adjusts overlapping areas between images to achieve a smooth transition while preserving image detail.

[0071] For example, continuing with the previous example, after the drone completes its image capture mission, engineers at the ground station use specialized image processing software to stitch all uploaded images together into a detailed panoramic image based on their geographic location. This panoramic image clearly illustrates the entire power plant's photovoltaic panel layout, laying a solid foundation for further analysis.

[0072] Step 103: using a pre-built three-dimensional topological structure model of photovoltaic modules, identifying and locating photovoltaic modules in the high-resolution panoramic image, and marking the photovoltaic modules to generate a module distribution map;

[0073] In this step, the three-dimensional topological structure model is a mathematical model that represents the spatial relationship and connection method of objects; the deep learning algorithm is a type of artificial intelligence method that can automatically extract features from large amounts of data; and the graph theory algorithm is used to solve problems in the network structure, such as the shortest path between nodes.

[0074] In practice, based on a trained deep learning model, the system automatically scans panoramic images, locates and annotates all photovoltaic modules. Then, using graph theory algorithms, the system labels each module and its connection method in a distribution diagram, establishing logical relationships between the modules.

[0075] For example, based on the aforementioned panoramic image, engineers ran a deep learning model based on the U-Net architecture to identify each PV module and labeled them using graph theory algorithms. The result is a detailed component distribution map that intuitively shows the specific location of all PV panels in the power plant and their interconnections.

[0076] Step 104: Based on the component distribution map, the actual spectral response characteristics corresponding to each photovoltaic component are analyzed. The reflectance characteristic data of each photovoltaic component in a good state and the spectral response characteristics under different damage modes are obtained from a pre-built material science database as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas.

[0077] In this step, the spectral response characteristics refer to the degree to which the material reacts to light of different wavelengths; the reflective characteristic data in the intact state is the ideal spectral response of the material when it is undamaged; and the spectral response characteristics in the damage mode refer to the spectral performance of the material when it is damaged in different forms.

[0078] In practice, the actual spectral response characteristics of each photovoltaic module are first obtained. Then, a materials science database is consulted to obtain the spectral response standards of the corresponding module under intact conditions and various possible damage conditions. Finally, the two are compared to identify any deviations and thus identify potential problem areas.

[0079] For example, engineers conducted spectral analysis on each photovoltaic module based on the module distribution map. By referencing standard spectral response characteristics in the materials science database, they successfully identified several areas with abnormal spectral response. These areas were marked as suspected damage areas, providing key targets for further in-depth analysis.

[0080] Step 105: Processing the suspected damage area using an adaptive threshold algorithm and morphological operations to generate a suspected damage area;

[0081] In this step, the adaptive threshold algorithm is a method for dynamically adjusting the image segmentation threshold, which can optimize the segmentation effect according to the local characteristics of the image; morphological operations are a series of techniques used to change the shape of the image, such as dilation and erosion, which help to clean up image noise and enhance target features.

[0082] In actual operation, the suspected damage area is segmented through the Bayesian optimized adaptive threshold algorithm, and then unnecessary noise is removed by morphological operations such as dilation and erosion, and finally the optimized suspected damage area is generated, providing key evidence for damage assessment.

[0083] For example, engineers further analyzed previously marked suspected damage areas and applied advanced image processing techniques, including adaptive threshold segmentation and morphological operations, to generate clearer suspected damage areas. This made it easier for technicians to identify where actual damage might exist.

[0084] Step 106: Extract target features based on the suspected damaged area, the real-time environmental factors of the photovoltaic module, and historical performance change data, and identify the influence relationship between the target features and the performance degradation of the photovoltaic module. Based on the target features and the influence relationship, predict the damage status of the photovoltaic module to generate a test report.

[0085] In this step, transfer learning refers to applying knowledge learned in one field to another related field; convolutional neural network is an artificial intelligence model that is good at processing image data; random forest is an integrated learning algorithm that performs classification or regression prediction by combining the results of multiple decision trees.

[0086] In practice, all available information is integrated to model and predict the status of PV panels using machine learning algorithms. Transfer learning is used to improve the model's generalization capabilities, and a random forest algorithm is used for classification and prediction, outputting prediction results and recommended measures.

[0087] For example, engineers ultimately used an enhanced convolutional neural network and random forest algorithm to predict the health of PV panels by integrating suspected damaged areas, current environmental conditions, and historical performance data. Based on these predictions, they compiled a detailed inspection report, identifying specific panels requiring maintenance and providing appropriate repair recommendations.

[0088] The six steps described above effectively improve the accuracy and efficiency of PV module damage detection. From the precise acquisition of multi-angle, high-altitude images, to efficient analysis enabled by deep learning and image processing technologies, to damage prediction based on machine learning, each step enhances the intelligence of PV power plant operation and maintenance management. This solution not only promptly detects and locates potential problems but also predicts future damage trends through scientific analysis, significantly ensuring the safe and stable operation of PV power plants.

[0089] In order to resolve the noise interference that may exist in the suspected damage area and improve the accuracy of the damaged plaque identification, in some embodiments, the suspected damage area is processed using an adaptive threshold algorithm and morphological operations in step 105 to generate the suspected damage area, including:

[0090] According to the local grayscale distribution characteristics of each pixel point in the high-resolution panoramic image, the suspected damage area is processed to obtain the optimal threshold corresponding to each pixel point; based on the optimal threshold point, the pixels in the suspected damage area are divided into two categories, foreground and background, to obtain classified pixels; based on the classified pixels, a suspected damage area map is generated, and the suspected damage area map is analyzed to identify suspected damage plaques and generate suspected damage plaque data; based on the suspected damage plaque data, a target structural element is determined; based on the target structural element, dilation, erosion and morphological operations are performed on the suspected damage area map to generate an optimized suspected damage area map to remove noise in the suspected damage area; the optimized suspected damage area map is subjected to frequency domain filtering processing to obtain a preliminary suspected damage feature map; the preliminary suspected damage feature map is processed to generate a suspected damage area.

[0091] In this embodiment, the local grayscale distribution characteristics refer to the statistical properties of pixel values within a certain range around each pixel, such as average brightness or contrast. The adaptive threshold algorithm based on Bayesian optimization is a method for dynamically adjusting the segmentation threshold. It can automatically adjust the threshold according to the image content, making the segmentation result more consistent with the actual scene. The division of foreground and background categories is part of the image binarization process, which is used to distinguish important information from other irrelevant parts in the image. Connected component analysis is an image processing method used to find and mark groups of pixels that are connected to each other in the image, that is, connected regions, which helps to identify independent damage patches. The structural element adaptive selection algorithm is used to determine the shape and size of the template used in morphological operations to adapt to damage patches of different shapes and sizes. Morphological operations can enhance specific shape features in the image while reducing noise. Fast Fourier transform is a technology that converts spatial domain signals to frequency domain, which can be used to filter out unnecessary frequency components, thereby cleaning the image. U-Net is a convolutional neural network architecture that is particularly suitable for image segmentation tasks and can accurately locate and describe objects in the image.

[0092] In an embodiment of the present application, first, for the suspected damage area, by analyzing the local grayscale distribution characteristics of each pixel, a Bayesian optimized adaptive threshold algorithm is used to calculate the most appropriate segmentation threshold for each pixel. Next, according to these thresholds, the pixels in the image are divided into two categories: foreground and background, and a suspected damage area map is constructed based on this. Afterwards, the suspected damage plaques are identified through the connected component analysis algorithm, and their position, shape and other information are recorded. In order to remove the noise in the image, the most suitable structural elements are selected for morphological operations to make the image clearer. Then, the image is frequency-domain filtered using fast Fourier transform technology to eliminate high-frequency noise and obtain a preliminary map of suspected damage features. Finally, the preliminary map is further processed using the U-Net network model to generate a more refined and accurate suspected damage area.

[0093] Here is a specific example:

[0094] During routine maintenance inspections at a large solar power plant, technicians discovered potential problems with some photovoltaic modules. Following the aforementioned method, a drone was first used to capture a high-resolution panoramic image of the power plant. Subsequently, in the laboratory, engineers used an adaptive thresholding algorithm based on Bayesian optimization to identify suspected damage areas. This algorithm determined the optimal threshold for each pixel and used this threshold to classify the pixels in the image into foreground and background. Connected component analysis was then used to identify multiple suspected damage patches and record their data. To ensure the accuracy of the identified suspected damage areas, engineers applied morphological operations to remove noise and performed a fast Fourier transform to eliminate frequency-domain noise. Finally, engineers used a U-Net network model to perform a detailed analysis of the processed image to generate a final map of suspected damage areas. This feature map not only helped technicians accurately locate all suspected damage areas but also provided valuable guidance for subsequent repairs.

[0095] In order to further improve the accuracy of detecting suspected damage areas and reduce the impact of image noise on segmentation results, in some embodiments, step 105 processes the suspected damage areas based on the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain the optimal threshold corresponding to each pixel, and further includes:

[0096] The local neighborhood defined by each pixel in the high-resolution panoramic image is smoothed to obtain a local neighborhood image; the grayscale value distribution of all pixels in the local neighborhood image is analyzed to establish a local grayscale distribution model, and the local neighborhood image is processed by applying a histogram equalization technique to obtain a local image; based on the local image and the local grayscale distribution model, a composite objective function based on a structural similarity index and edge regularization is constructed; based on the composite objective function, the segmentation effect of the suspected damaged area under different thresholds is evaluated to obtain a segmentation evaluation result; the composite objective function is optimized in combination with the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function; an initial candidate threshold is defined, and the initial candidate threshold is optimized in combination with the optimization objective function to obtain an optimized threshold, and the optimized threshold is used as the optimal threshold of the local grayscale distribution characteristic. Optionally, based on the local image and the local grayscale distribution model, a composite objective function based on a structural similarity index and edge regularization is constructed; based on the composite objective function, the segmentation effect of the suspected damage area under different thresholds is evaluated to obtain a segmentation evaluation result; the composite objective function is optimized in combination with the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function, including: based on the local image and the local grayscale distribution model, the segmentation effect of the suspected damage area is evaluated, and edge detection is performed on the local image to obtain edge information; based on the edge information and combined with an edge regularization term, a composite objective function is constructed; based on the composite objective function, the suspected damage area is segmented to obtain a segmentation result; using the segmentation result, the segmentation effect of the suspected damage area is evaluated to generate a segmentation evaluation result set; the segmentation evaluation result set is divided into multiple subsets; according to the subsets, a part of the subsets is used as a test set, and the remaining part of the subsets is used as a training set; using the training set and the test set, the segmentation effect of the suspected damage area is iteratively evaluated multiple times to obtain an optimized objective function.

[0097] In this embodiment, the weighted average filter is a linear filtering method that performs smoothing by applying different weights to each pixel in the image and its surrounding pixels, effectively reducing random noise while protecting important edge features. The local grayscale distribution model is a statistical method used to describe the distribution of grayscale values of pixels in a specific area, which helps to understand the image content. Histogram equalization is a common image enhancement technique that aims to expand the grayscale range of an image and increase visual contrast. The structural similarity index is an indicator used to measure the similarity between two images, which takes into account brightness, contrast, and structure. The edge regularization term is a constraint on the image edge, ensuring that edge features can be accurately maintained during the optimization process. The composite objective function combines multiple evaluation criteria into a single objective to facilitate the optimization process. The segmentation evaluation result set is a record of the segmentation effect under a series of different thresholds, which is used to guide subsequent optimization work. The cross-validation mechanism is a method for evaluating model performance by dividing the dataset into several subsets and repeatedly training and testing the model to find the optimal parameter configuration.

[0098] In an embodiment of the present application, a weighted average filter is first applied to each pixel in a high-resolution panoramic image to smooth the image while protecting edge details and generate a local neighborhood image. Then, by analyzing the grayscale value distribution of pixels in these local neighborhood images, a local grayscale distribution model is established. This step is crucial for subsequent threshold calculation. Subsequently, histogram equalization technology is applied to improve image quality, making the image clearer. Next, based on the improved local image and the previously established local grayscale distribution model, a composite objective function that combines structural similarity index and edge regularization is constructed to evaluate the segmentation effect under different threshold settings. In order to ensure the robustness and generalization ability of the segmentation effect, a cross-validation mechanism is introduced to iteratively optimize the composite objective function until a satisfactory result is obtained. Finally, through the Bayesian optimization technique, under the guidance of the composite objective function, the initially set threshold is fine-tuned to finally find the optimal threshold that best suits the local grayscale distribution characteristics of the current image.

[0099] Here is a specific example:

[0100] In an automobile manufacturing plant, engineers regularly inspect car body surfaces for defects such as scratches and dents. An automated inspection system equipped with a high-resolution camera rapidly captures panoramic images of the car bodies on the production line. However, due to varying lighting conditions and shooting angles, the raw images can contain significant noise, posing a challenge for automated inspection. Following this approach, the engineers first applied a weighted average filter to the captured panoramic images, effectively reducing image noise while preserving the clarity of key features such as scratches and dents. They then established a local grayscale distribution model and enhanced image contrast through histogram equalization, making potential damage more visible. Next, they constructed a composite objective function based on a structural similarity index and edge regularization to evaluate segmentation performance at different thresholds. To ensure the optimal threshold, they employed a cross-validation mechanism to iteratively optimize the composite objective function, ultimately finding the most appropriate threshold. This process not only improved the accuracy of the inspection system but also significantly shortened inspection time, boosting overall production line efficiency.

[0101] To address the problem that potential hidden damage to photovoltaic modules is difficult to detect through conventional means, in some embodiments, the actual spectral response characteristics corresponding to each photovoltaic module are analyzed based on the module distribution map in step 104, and the reflection characteristic data of each photovoltaic module in a good state and the spectral response characteristics under different damage modes are obtained from a pre-built material science database as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas, including:

[0102] Based on the component distribution map, the actual spectral response characteristics corresponding to each photovoltaic component are analyzed to obtain characteristic information; based on the characteristic information, the high-resolution panoramic image is partitioned to obtain photovoltaic component areas, and the boundaries of the photovoltaic component areas are optimized to obtain segmentation results, and the segmentation results are analyzed to generate a photovoltaic component area map; based on the photovoltaic component area map, a deep belief network is constructed, and the hyperparameters of the deep belief network are adjusted to obtain an optimized deep belief network, and the optimized deep belief network is used to analyze the reflection characteristic data of the photovoltaic component area map to generate reflection characteristic data corresponding to the materials of each photovoltaic component in the photovoltaic component area; key spectral features are extracted from spectral data in the reflection characteristic data, and the key spectral features are classified to obtain classified key spectral features, and the classified key spectral features are used to determine the spectral response characteristics to be optimized, and the spectral response characteristics to be optimized are dynamically adjusted based on actual ambient lighting conditions to obtain standard spectral response characteristics; the actual spectral response characteristics are compared with the standard spectral response characteristics to generate spectral analysis parameters, and spectral analysis is performed on each photovoltaic component area based on the spectral analysis parameters to screen out suspected damaged areas. Optionally, based on the photovoltaic component area map, a deep belief network is constructed, and the hyperparameters of the deep belief network are adjusted to obtain an optimized deep belief network, and the optimized deep belief network is used to analyze and process the reflection characteristic data of the photovoltaic component area map to generate reflection characteristic data corresponding to the materials of each photovoltaic component in the photovoltaic component area, including: constructing a deep belief network based on the data of the photovoltaic component area map; searching and processing the hyperparameter space of the deep belief network in combination with a Gaussian process proxy model and an acquisition function to obtain an optimal hyperparameter combination, performing performance evaluation processing on the optimal hyperparameter combination to obtain an evaluation result, and selecting an optimal hyperparameter configuration based on the evaluation result to optimize the deep belief network and generate an optimized deep belief network; based on the optimized deep belief network, analyzing and processing the reflection characteristic data of the photovoltaic component area map to generate reflection behavior data, and correcting and verifying the reflection behavior data according to the standard reflection behavior data of existing standard samples to obtain reflection characteristic data corresponding to the materials of each photovoltaic component in the photovoltaic component area.

[0103] In this embodiment, spectral response refers to the efficiency of a photovoltaic module's response to light of different wavelengths, a key metric for evaluating its performance. The Materials Science Database is a pre-built repository containing spectral response data for various materials in good condition and under various damage patterns. This data serves as a standard reference for comparing measured spectral response characteristics of photovoltaic modules to identify potential issues. Superpixel segmentation is an image segmentation method that decomposes an image into several small regions with similar properties, known as superpixels, thereby simplifying image processing tasks. The maximum flow minimum cut algorithm is a graph theory algorithm used to find optimal segmentation boundaries in an image that maximizes the difference between adjacent regions. Morphological operations and connected component analysis are techniques used to clean and enhance the segmentation results, ensuring that the boundaries of each photovoltaic module are clear and continuous. A deep belief network is a deep neural network architecture that can learn complex nonlinear mapping relationships, while Bayesian optimization is used to automatically tune network hyperparameters to improve prediction accuracy. Principal component analysis is a statistical method that reduces data dimensionality while retaining the most important information. Support vector machines are supervised learning models that excel at recognizing small samples, nonlinear patterns, and high-dimensional patterns. The principle of maximizing margin is the core concept of SVM, which aims to find a hyperplane that best separates data points of different categories. Genetic algorithm is a random search algorithm that simulates natural selection and genetic mechanisms to find the global optimal solution.

[0104] In the embodiment of the present application, the actual spectral response characteristics of each photovoltaic module are first analyzed based on the component distribution map to obtain characteristic information. Subsequently, the superpixel segmentation algorithm and the maximum flow minimum cut algorithm are applied to optimize the boundaries and form a photovoltaic module area map. Next, a deep belief network is constructed and optimized to analyze the reflectance characteristic data of each photovoltaic module material. Then, the reflectance characteristic data is processed by principal component analysis dimensionality reduction technology and support vector machine classifier to identify abnormal conditions, determine the spectral response characteristics to be optimized, and adjust these characteristics according to the actual lighting conditions. Finally, a genetic algorithm is used to compare the actual spectral response characteristics with the standard characteristics to generate spectral analysis parameters, which are used as a basis to screen out suspected damaged areas.

[0105] Here is a specific example:

[0106] At a solar power plant, technicians regularly inspect photovoltaic (PV) panels for hidden damage that could reduce power generation efficiency but remain invisible to the naked eye. Following the aforementioned method, a drone was used to capture high-resolution panoramic images of the power plant. The actual spectral response characteristics of each PV panel were analyzed based on the panel distribution map. Next, engineers used a superpixel segmentation algorithm to partition the image and optimized the boundaries using a maximum flow minimum cut algorithm, resulting in a precise map of the PV panel regions. They then constructed a deep belief network and used Bayesian optimization to fine-tune its hyperparameters to more accurately analyze the reflectance characteristics of each PV panel material. For further analysis, engineers used principal component analysis to extract key spectral features and a support vector machine classifier to classify anomalies. Finally, a genetic algorithm was used to compare the actual spectral response characteristics with the standard, successfully identifying PV panel areas with potential problems and providing clear guidance for subsequent maintenance work.

[0107] To address the accuracy issues of component identification and positioning in photovoltaic power plants and improve detection efficiency and accuracy, in some embodiments, the steps of step 103 using a pre-built 3D topological structure model of photovoltaic components to identify and locate photovoltaic components in the high-resolution panoramic image and mark the photovoltaic components to generate a component distribution map include:

[0108] Using a pre-built three-dimensional topological structure model of photovoltaic components, the photovoltaic components in the high-resolution panoramic image are automatically identified and positioned to generate an identification result; the identification result is marked to establish a topological connection relationship between the photovoltaic components to generate a photovoltaic component distribution network; based on the photovoltaic component distribution network, a component distribution map is generated, wherein the component distribution map records the positions and mutual relationships of all photovoltaic components.

[0109] In this embodiment, the three-dimensional topological structure model of the photovoltaic module is a detailed digital model that includes the physical size, shape, and relative position of each component in the photovoltaic array in space. This model is not only used to assist the deep learning algorithm in more accurately identifying and locating the photovoltaic modules in the image, but also serves as the basis for subsequent analysis. Deep learning algorithms are a type of machine learning technology, specifically artificial neural networks that contain multiple layers of nonlinear transformations, such as convolutional neural networks. They are good at automatically extracting features from large amounts of data and are particularly effective for image recognition tasks. Graph theory algorithms are a branch of mathematics that involve the study of graph structures. They are used here to mark and define the connection methods between photovoltaic modules and help understand the topological relationship between modules. The component distribution map is a visualization tool generated according to the above process. It clearly marks the location of each photovoltaic module and the connection between them, which is crucial for the management and maintenance of photovoltaic power stations.

[0110] In an embodiment of the present application, a 3D topological structure model of photovoltaic modules and a deep learning algorithm are first used to process high-resolution panoramic images acquired by drones or other imaging devices to achieve automatic identification and precise positioning of photovoltaic modules. Then, the identified components are connected through a graph theory algorithm to build a network that reflects the actual connection relationship between the components. Finally, a component distribution map is generated based on this network. This map not only shows the specific location of the components, but also reveals how they are connected together, which is very useful for monitoring and troubleshooting photovoltaic power stations.

[0111] Here is a specific example:

[0112] To ensure efficient operation, technicians in a large ground-based photovoltaic power plant regularly inspect the status of photovoltaic modules. They first constructed a three-dimensional topological model containing detailed information about all photovoltaic modules within the plant. Then, during a routine inspection, operators flew a drone over the entire plant, capturing a series of high-resolution panoramic images. These images were fed into a deep learning-based system, which automatically identified and precisely located each photovoltaic module. Next, a graph theory algorithm was used to label these modules and establish topological connections between them, forming a complete photovoltaic module distribution network. Finally, a detailed module distribution map was generated based on this network. This map not only facilitates daily management but also allows for the rapid identification of potential problem areas, guiding maintenance teams. This approach enables technicians to more quickly and accurately complete photovoltaic power plant inspections, thereby improving overall operational efficiency.

[0113] This application considers the limitations of traditional threshold segmentation methods when processing photovoltaic module images, such as sensitivity to illumination variations and noise, and an inability to adapt to complex local grayscale distribution characteristics. To improve the accuracy and robustness of identifying suspected damaged areas, a method that can adaptively adjust the threshold to account for image variations under varying environmental conditions is required. Therefore, a new alternative solution is proposed, which includes:

[0114] The suspected damaged area is processed according to the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain the optimal threshold corresponding to each pixel, including:

[0115] The local neighborhood defined by each pixel in the high-resolution panoramic image is smoothed, and a local neighborhood image is obtained by weighted averaging. The pixel value of the smoothed image at position (x, y) is obtained by calculating the weighted average of all pixels in the neighborhood, where the weight is determined by spatial distance and grayscale similarity. The spatial distance weight means that the closer the distance between the pixel and the center point, the greater the weight; the grayscale similarity weight means that the closer the grayscale value of the pixel and the center point, the greater the weight. These two weights are calculated using a Gaussian function and are controlled by the spatial distance standard deviation and the grayscale similarity standard deviation, respectively.

[0116] The smoothed image reflects the grayscale value at position (x, y) after weighted average filtering. The local neighborhood N(x, y) is defined by the area centered at position (x, y), typically a rectangular window or a circular area. The spatial distance standard deviation controls the rate at which the spatial distance weight decays, while the grayscale similarity standard deviation controls the rate at which the grayscale similarity weight decays. These two parameters need to be determined empirically or experimentally.

[0117] Next, the grayscale value distribution of all pixels in the local neighborhood image is analyzed to establish a local grayscale distribution model. The local neighborhood image is processed using histogram equalization technology to obtain a local image. Histogram equalization adjusts the grayscale distribution of the image to make the grayscale distribution more uniform, thereby enhancing the contrast of the image. The equalized grayscale value is obtained by calculating the probability weighted average of the original grayscale, and performing normalization and mapping. Normalization ensures that local images of different sizes can be compared fairly, and mapping ensures that the equalized grayscale value can make full use of the entire grayscale range.

[0118] Based on the local image and the local grayscale distribution model, a composite objective function based on a structural similarity index and edge regularization is constructed. Based on the composite objective function, the segmentation effect of the suspected damage area at different thresholds is evaluated to obtain a segmentation evaluation result. The composite objective function is optimized by combining the segmentation evaluation results with a cross-validation mechanism to obtain an optimized objective function.

[0119] An initial candidate threshold is defined, and the initial candidate threshold is optimized in combination with the optimization objective function to obtain an optimized threshold. The optimized threshold is used as the optimal threshold for the local grayscale distribution characteristic. The goal of Bayesian optimization is to maximize information gain, i.e., expected improvement. Information gain is evaluated by calculating the potential improvement of the new threshold relative to the current optimal value and combining it with a conditional probability density function. The conditional probability density function reflects the model's uncertainty and likelihood for the new threshold.

[0120] Based on the optimized parameter set, the maximum entropy principle is used to determine an initial threshold range. Based on this initial threshold range, Bayesian optimization is used to calculate the optimal threshold for the local grayscale distribution characteristics. The maximum entropy principle accounts for the uncertainty of the local grayscale distribution when determining the threshold range. The optimal threshold is obtained by maximizing the entropy value; a larger entropy value indicates a more uniform distribution and a higher information content.

[0121] The following is a specific embodiment:

[0122] During routine maintenance and inspections at a large ground-based photovoltaic power plant, technicians regularly monitor the status of photovoltaic panels to detect potential damage. To address the inefficiency of traditional threshold segmentation methods under complex lighting conditions, the proposed solution was employed. First, technicians used a drone to acquire high-resolution panoramic images of all photovoltaic panels within the power plant. These images were then preprocessed using a weighted average filter to smooth image noise while preserving the integrity of important features, such as edges. Histogram equalization was then applied to the smoothed images to enhance image contrast, facilitating subsequent analysis. Based on this, a composite objective function was constructed, taking into account structural similarity and edge preservation, to evaluate segmentation quality at different threshold settings. This objective function was continuously optimized through cross-validation to identify the optimal parameter configuration for the current image conditions. An initial set of candidate thresholds was then defined, and Bayesian optimization techniques were used to further refine these thresholds to find the optimal threshold that maximized information gain. Finally, the threshold range was redefined based on the maximum entropy principle, and Bayesian optimization was again applied to determine the optimal threshold for each local region. This method not only improves the accuracy of detection of suspected damaged areas, but also reduces the false alarm rate, greatly improving the efficiency and reliability of photovoltaic power station operation and maintenance management.

[0123] Figure 2 The present invention provides a schematic diagram of a photovoltaic module damage intelligent detection system based on drone images, as shown in FIG. Figure 2 As shown, the system includes:

[0124] A collection module 21 is configured to collect multi-angle high-altitude images of the photovoltaic module array captured by the drone on a predetermined flight path, and record geographic location information corresponding to the multi-angle high-altitude images;

[0125] A correction module 22 is configured to stitch and correct overlapping areas of the multi-angle high-altitude images based on the geographic location information to obtain a high-resolution panoramic image;

[0126] an identification module 23 for identifying and locating photovoltaic modules in the high-resolution panoramic image using a pre-built three-dimensional topological structure model of photovoltaic modules, and marking the photovoltaic modules to generate a module distribution map;

[0127] An analysis module 24 is configured to analyze the actual spectral response characteristics corresponding to each photovoltaic module based on the module distribution map, obtain the reflectance characteristic data of each photovoltaic module in an intact state and the spectral response characteristics under different damage modes from a pre-built material science database, use the data as standard spectral response characteristics, compare the actual spectral response characteristics with the standard spectral response characteristics, and screen out suspected damaged areas;

[0128] The generation module 25 is used to extract target features by combining the suspected damaged area, the real-time environmental factors of the photovoltaic module and the historical performance change data, and to identify the influence relationship between the target features and the performance degradation of the photovoltaic module. Based on the target features and the influence relationship, the damage status of the photovoltaic module is predicted to generate a detection report.

[0129] Figure 2 The photovoltaic module damage intelligent detection system based on drone images can be performed Figure 1 The implementation principles and technical effects of the intelligent photovoltaic module damage detection method based on drone imagery described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the intelligent photovoltaic module damage detection system based on drone imagery in the aforementioned embodiment has been described in detail in the relevant embodiments of the method and will not be further elaborated here.

[0130] In one possible design, Figure 2 The photovoltaic module damage intelligent detection system based on drone images of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0131] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0132] The processing component 32 is used for the above Figure 1 The embodiment provides an intelligent photovoltaic module damage detection method based on drone images.

[0133] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0134] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0135] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0136] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0137] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0138] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0139] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment provides an intelligent method for detecting photovoltaic module damage based on drone images.

[0140] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0141] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0142] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent detection method for photovoltaic module damage based on drone images, characterized in that: include: Collect multi-angle high-altitude images of the photovoltaic component array collected by the drone on a predetermined flight path, and record geographic location information corresponding to the multi-angle high-altitude images; Based on the geographic location information, the multi-angle high-altitude images are stitched and overlapped areas are corrected to obtain a high-resolution panoramic image; Using a pre-built three-dimensional topological structure model of photovoltaic modules, identifying and locating photovoltaic modules in the high-resolution panoramic image, and marking the photovoltaic modules to generate a module distribution map; Based on the component distribution map, the actual spectral response characteristics corresponding to each photovoltaic component are analyzed. The reflection characteristic data of each photovoltaic component in a good state and the spectral response characteristics under different damage modes are obtained from a pre-built material science database. These data are used as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas. Combined with the suspected damaged area, the real-time environmental factors of the photovoltaic module and the historical performance change data, the target features are extracted, and the influence relationship between the target features and the performance degradation of the photovoltaic module is identified. Based on the target features and the influence relationship, the damage status of the photovoltaic module is predicted to generate a detection report.

2. The method according to claim 1, characterized in that The suspected damage area is processed using an adaptive threshold algorithm and morphological operations to generate a suspected damage area, including: Processing the suspected damaged area according to the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain an optimal threshold corresponding to each pixel; Based on the optimal threshold, pixels in the suspected damage area are divided into two categories, foreground and background, to obtain classified pixels; based on the classified pixels, a suspected damage area map is generated; the suspected damage area map is analyzed to identify suspected damage plaques and generate suspected damage plaque data; Determining a target structural element based on the suspected damage plaque data, and performing dilation, erosion, and morphological operations on the suspected damage region map based on the target structural element to generate an optimized suspected damage region map to remove noise from the suspected damage region; Performing frequency domain filtering on the optimized suspected damage area map to obtain a preliminary suspected damage feature map; The preliminary image of suspected damage features is processed to generate a suspected damage area.

3. The method according to claim 2, characterized in that The suspected damaged area is processed according to the local grayscale distribution characteristics of each pixel in the high-resolution panoramic image to obtain the optimal threshold corresponding to each pixel, including: performing a smoothing process on a local neighborhood defined by each pixel point in the high-resolution panoramic image to obtain a local neighborhood image; Analyzing the grayscale value distribution of all pixels in the local neighborhood image to establish a local grayscale distribution model, and applying a histogram equalization technique to process the local neighborhood image to obtain a local image; Based on the local image and the local grayscale distribution model, constructing a composite objective function based on a structural similarity index and edge regularization, evaluating the segmentation effect of the suspected damage area under different thresholds based on the composite objective function to obtain a segmentation evaluation result, and optimizing the composite objective function by combining the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function; An initial candidate threshold is defined, and the initial candidate threshold is optimized in combination with the optimization objective function to obtain an optimized threshold, and the optimized threshold is used as the optimal threshold of the local grayscale distribution characteristic.

4. The method according to claim 3, characterized in that The method comprises: constructing a composite objective function based on a structural similarity index and edge regularization based on the local image and the local grayscale distribution model; evaluating the segmentation effect of the suspected damage area under different thresholds based on the composite objective function to obtain a segmentation evaluation result; and optimizing the composite objective function by combining the segmentation evaluation result and a cross-validation mechanism to obtain an optimized objective function, including: Based on the local image and the local grayscale distribution model, the segmentation effect of the suspected damage area is evaluated, and edge detection is performed on the local image to obtain edge information. Based on the edge information and an edge regularization term, a composite objective function is constructed; Based on the composite objective function, the suspected damage area is segmented to obtain a segmentation result, and the segmentation effect of the suspected damage area is evaluated using the segmentation result to generate a segmentation evaluation result set; The segmentation evaluation result set is divided into multiple subsets. According to the subsets, a part of the subsets is used as a test set, and the remaining part of the subsets is used as a training set. The training set and the test set are used to perform multiple iterative evaluations on the segmentation effect of the suspected damage area to obtain an optimized objective function.

5. The method according to claim 1, wherein Based on the component distribution map, the actual spectral response characteristics corresponding to each photovoltaic component are analyzed, and the reflection characteristic data of each photovoltaic component in an intact state and the spectral response characteristics under different damage modes are obtained from a pre-built material science database as standard spectral response characteristics. The actual spectral response characteristics are compared with the standard spectral response characteristics to screen out suspected damaged areas, including: Based on the component distribution diagram, analyzing the actual spectral response characteristics corresponding to each photovoltaic component to obtain characteristic information; Based on the characteristic information, the high-resolution panoramic image is partitioned to obtain photovoltaic module areas, and the boundaries of the photovoltaic module areas are optimized to obtain segmentation results, and the segmentation results are analyzed to generate a photovoltaic module area map; Based on the photovoltaic module area map, a deep belief network is constructed, and hyperparameters of the deep belief network are adjusted to obtain an optimized deep belief network. The optimized deep belief network is used to analyze and process reflection characteristic data of the photovoltaic module area map to generate reflection characteristic data corresponding to the material of each photovoltaic module in the photovoltaic module area; Extracting key spectral features from spectral data in the reflectance characteristic data, and performing abnormal classification processing on the key spectral features to obtain classified key spectral features, using the classified key spectral features to determine spectral response characteristics to be optimized, and dynamically adjusting the spectral response characteristics to be optimized based on actual ambient lighting conditions to obtain standard spectral response characteristics; The actual spectral response characteristic is compared with the standard spectral response characteristic to generate spectral analysis parameters. Based on the spectral analysis parameters, a spectral analysis is performed on each photovoltaic module area to screen out suspected damaged areas.

6. The method according to claim 5, characterized in that The method comprises: constructing a deep belief network based on the photovoltaic module area map, adjusting the hyperparameters of the deep belief network to obtain an optimized deep belief network, and using the optimized deep belief network to perform reflection characteristic data analysis and processing on the photovoltaic module area map to generate reflection characteristic data corresponding to the materials of each photovoltaic module in the photovoltaic module area, including: Constructing a deep belief network based on the data of the photovoltaic module area map; In combination with a Gaussian process surrogate model and an acquisition function, a search process is performed on the hyperparameter space of the deep belief network to obtain an optimal hyperparameter combination, a performance evaluation process is performed on the optimal hyperparameter combination to obtain an evaluation result, and an optimal hyperparameter configuration is selected based on the evaluation result to optimize the deep belief network and generate an optimized deep belief network; Based on the optimized deep belief network, the reflection characteristic data of the photovoltaic module area map is analyzed and processed to generate reflection behavior data. According to the standard reflection behavior data of the existing standard samples, the reflection behavior data is corrected and verified to obtain the reflection characteristic data corresponding to the materials of each photovoltaic module in the photovoltaic module area.

7. The method according to claim 1, characterized in that The method of using a pre-built three-dimensional topological structure model of photovoltaic components to identify and locate photovoltaic components in the high-resolution panoramic image and mark the photovoltaic components to generate a component distribution map includes: Using a pre-built three-dimensional topological structure model of a photovoltaic module, automatically identifying and locating the photovoltaic module in the high-resolution panoramic image to generate an identification result; Marking the identification results to establish a topological connection relationship between the photovoltaic modules and generate a photovoltaic module distribution network; Based on the photovoltaic component distribution network, a component distribution map is generated, wherein the component distribution map records the positions and mutual relationships of all photovoltaic components.

8. An intelligent photovoltaic module damage detection system based on drone images, characterized in that: include: A collection module is used to collect multi-angle high-altitude images of the photovoltaic module array collected by the drone on a predetermined flight path, and record the geographical location information corresponding to the multi-angle high-altitude images; A correction module, configured to stitch and correct overlapping areas of the multi-angle high-altitude images based on the geographic location information to obtain a high-resolution panoramic image; an identification module, configured to identify and locate photovoltaic modules in the high-resolution panoramic image using a pre-built three-dimensional topological structure model of photovoltaic modules, and mark the photovoltaic modules to generate a module distribution map; an analysis module for analyzing the actual spectral response characteristics corresponding to each photovoltaic module based on the module distribution map, obtaining the reflectance characteristic data of each photovoltaic module in an intact state and the spectral response characteristics under different damage modes from a pre-built material science database as standard spectral response characteristics, comparing the actual spectral response characteristics with the standard spectral response characteristics, and screening out suspected damaged areas; A generation module is used to extract target features by combining the suspected damaged area, the real-time environmental factors of the photovoltaic module and the historical performance change data, and to identify the influence relationship between the target features and the performance degradation of the photovoltaic module. Based on the target features and the influence relationship, the damage status of the photovoltaic module is predicted to generate a detection report.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an intelligent detection method for photovoltaic component damage based on drone images as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for intelligent detection of photovoltaic component damage based on drone images as described in any one of claims 1 to 7 is implemented.

Images