Method and system for real-time online monitoring of photovoltaic module status

Abstract

The application discloses a photovoltaic module state real-time online monitoring method and system, relates to the technical field of real-time monitoring, and collects infrared images of suspected hot spot areas through a unmanned aerial vehicle, combines real-time irradiance data, preliminarily investigates static state of collected multi-modal image information, and thus quickly filters out false hot spots caused by significant physical obstruction or stains. For the highly suspected fault area screened out, a dynamic characteristic analysis mechanism is introduced, the unmanned aerial vehicle is driven to intelligently hover and continuously observe the target, infrared thermal image sequences of the area in a specific time window are actively collected, temperature time sequence characteristics capable of representing thermodynamic response characteristics of the area are extracted from the infrared thermal image sequences, and finally, classification and discrimination are carried out based on the temperature time sequence characteristics, so that true and false hot spots can be accurately distinguished. In this way, fine diagnosis of the health state of the photovoltaic module can be realized, and the recognition accuracy and intelligent level of the diagnosis model are significantly improved.

Description

Technical Field

[0001] This application relates to the field of real-time monitoring technology, and more specifically, to a method and system for real-time online monitoring of the status of photovoltaic modules. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, photovoltaic (PV) power generation, as a crucial component of renewable energy, is experiencing unprecedented growth in installed capacity. The stable and efficient operation of large-scale PV power plants is key to ensuring power supply and investment returns. However, during long-term outdoor operation, PV modules inevitably experience various malfunctions due to manufacturing defects, material aging, environmental corrosion, and other factors, with the "hot spot effect" being particularly prominent. Hot spots not only significantly reduce the power generation efficiency of the modules but can also cause permanent damage or even fires, posing a serious threat to the safety of the entire power plant. Therefore, establishing a system capable of real-time, accurate, and efficient online monitoring of the operating status of PV modules is of paramount practical significance and economic value for ensuring the safe and stable operation of PV power plants, improving operation and maintenance efficiency, and reducing the cost per kilowatt-hour.

[0003] In existing technologies, the inspection of photovoltaic (PV) modules commonly employs drones equipped with infrared thermal imagers. These drones capture large-scale aerial photographs of the PV array, obtaining infrared thermal images. Image processing algorithms then identify areas with temperatures exceeding normal thresholds, marking them as suspected hotspot fault points. Upon detecting such suspected faults, the system typically generates a report containing the locations of these anomalies for maintenance personnel to conduct subsequent on-site investigations. However, this monitoring method has significant limitations. Its diagnostic logic is essentially based on a single, static temperature snapshot. This approach cannot deeply analyze the dynamic physical processes of abnormal temperature rises, making it difficult to effectively distinguish the root cause of the fault. In large-scale PV power plants, many temperature anomalies are not caused by internal module defects (i.e., true hotspots), but rather by temporary temperature increases caused by external obstructions such as leaves, bird droppings, and dust (i.e., pseudo-hotspots). Because existing technologies lack the ability to differentiate between these two phenomena, they often categorize them collectively as faults, resulting in a high false alarm rate in diagnostic results. This not only forces the operations and maintenance team to spend a lot of money to verify pseudo-hot spots that do not actually require maintenance, but also fails to provide sufficiently refined data support to drive the operations and maintenance system to make accurate and efficient decisions, thus making it difficult to fundamentally improve the accuracy of fault diagnosis and the utilization efficiency of operations and maintenance resources.

[0004] Therefore, there is an urgent need for an optimized method and system for real-time online monitoring of photovoltaic module status. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a method for real-time online monitoring of photovoltaic module status is provided, comprising: obtaining a list of suspected hot spot regions collected by a drone based on real-time irradiance data; performing multimodal static screening on the list of suspected hot spot regions to obtain a filtered list of suspected hot spot regions and a static pseudo hot spot log; extracting a first filtered suspected hot spot region from the filtered list of suspected hot spot regions; acquiring an infrared image time series of the first filtered suspected hot spot region through intelligent hovering of the drone; extracting temperature time series features from the infrared image time series of the first filtered suspected hot spot region; and performing classification judgment based on the temperature time series features to obtain a classification result, wherein the classification result includes a category and a confidence level, and the category includes true hot spots and pseudo hot spots.

[0007] According to another aspect of this application, a real-time online monitoring system for photovoltaic module status is provided, comprising: a preliminary screening module for suspected hot spot areas, used to obtain a list of suspected hot spot areas collected by a drone based on real-time irradiance data; a multimodal static screening module, used to perform multimodal static screening on the list of suspected hot spot areas to obtain a filtered list of suspected hot spot areas and a static pseudo hot spot log; a filtered suspected area extraction module, used to extract a first filtered suspected hot spot area from the filtered list of suspected hot spot areas; an infrared image time series acquisition module, used to acquire an infrared image time series of the first filtered suspected hot spot area through intelligent hovering of the drone; a temperature time series feature extraction module, used to extract temperature time series features from the infrared image time series of the first filtered suspected hot spot area; and a true / false hot spot classification and judgment module, used to perform classification and judgment based on the temperature time series features to obtain a classification result, the classification result including a category and a confidence level, the category including true hot spots and pseudo hot spots.

[0008] Compared with existing technologies, this application provides a real-time online monitoring method and system for photovoltaic modules. It uses a drone to collect infrared images of suspected hotspot areas and combines them with real-time irradiance data to perform a preliminary static screening of the collected multimodal image information, thereby quickly filtering out false hotspots caused by significant physical obstructions or dirt. For highly suspected fault areas, a dynamic feature analysis mechanism is introduced. By driving the drone to intelligently hover and continuously observe the target, it actively collects infrared thermal image sequences of the area within a specific time window and extracts temperature time-series features that characterize its thermodynamic response. Finally, classification is performed based on these temperature time-series features to accurately distinguish between true and false hotspots. This enables refined diagnosis of the health status of photovoltaic modules, significantly improving the identification accuracy and intelligence level of the diagnostic model. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a method for real-time online monitoring of photovoltaic module status according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram of a real-time online monitoring method for the status of photovoltaic modules according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of sub-step S1 of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of sub-step S13 of the real-time online monitoring method for the status of photovoltaic modules according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of sub-step S2 of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of sub-step S5 of the real-time online monitoring method for the status of photovoltaic modules according to an embodiment of this application.

[0016] Figure 7 This is a block diagram of a photovoltaic module status real-time online monitoring system according to an embodiment of this application. Detailed Implementation

[0017] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0018] To address the problems mentioned above in the background technology, this application proposes a method for real-time online monitoring of the status of photovoltaic modules. Figure 1 This is a flowchart of a method for real-time online monitoring of photovoltaic module status according to an embodiment of this application. Figure 2 This is a data flow diagram of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application. Figure 1 and Figure 2 As shown, the real-time online monitoring method for photovoltaic module status includes the following steps: S1, obtaining a list of suspected hot spot areas collected by a drone based on real-time irradiance data; S2, performing multimodal static screening on the list of suspected hot spot areas to obtain a filtered list of suspected hot spot areas and a static pseudo hot spot log; S3, extracting a first filtered suspected hot spot area from the filtered list of suspected hot spot areas; S4, collecting the infrared image time series of the first filtered suspected hot spot area through intelligent hovering of the drone; S5, extracting temperature time series features from the infrared image time series of the first filtered suspected hot spot area; S6, performing classification judgment based on the temperature time series features to obtain a classification result, the classification result including category and confidence level, the category including true hot spots and pseudo hot spots.

[0019] In the aforementioned real-time online monitoring method for photovoltaic module status, step S1 involves obtaining a list of suspected hot spot areas collected by a drone based on real-time irradiance data. It should be understood that the temperature field distribution of photovoltaic modules is closely related to real-time irradiance. Under different irradiance conditions, the temperature baseline of normal modules and the temperature difference characteristics between hot spots and normal areas show significant differences. Relying solely on image data collected by drones or fixed detection parameters cannot adapt to dynamically changing lighting environments, making it difficult to accurately distinguish between actual hot spots and non-fault-related temperature anomalies caused by light fluctuations. Therefore, this application integrates real-time irradiance data with image information collected by drones to construct a detection logic adapted to the current lighting conditions. This logic filters out areas whose temperature characteristics meet the preliminary criteria for hot spot identification, forming a structured list of suspected hot spot areas, providing clear targeted analysis objects for subsequent refined diagnosis. In this way, under complex lighting environments, interference from non-fault-related temperature fluctuations caused by changes in light intensity can be effectively eliminated, accurately locking down areas with abnormal temperature characteristics and ensuring that subsequent processing objects have clear hot spot orientation.

[0020] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S1 of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application. Figure 3 As shown, step S1 includes: S11, controlling the UAV to perform wide-area flight inspection according to a preset route to obtain infrared images and visible light images of the photovoltaic module; S12, calculating an adaptive temperature difference threshold based on the real-time irradiance data; S13, performing suspected hot spot ROI segmentation and generation on the infrared images of the photovoltaic module based on the adaptive temperature difference threshold to obtain a list of suspected hot spot regions.

[0021] Specifically, in step S11, the UAV is controlled to perform a wide-area flight inspection along a preset route to obtain infrared and visible light images of the photovoltaic modules. It should be understood that photovoltaic power plants typically have large areas and densely packed modules. The flight characteristics of the UAV can meet the needs of large-scale inspection. Simultaneously, infrared images reflect the temperature distribution of the modules, and visible light images provide surface condition information. The combination of these two provides multi-dimensional data support for hot spot analysis. Specifically, through preset route planning, the UAV is ensured to efficiently traverse the photovoltaic array along a predetermined path, simultaneously acquiring infrared and visible light images to form a multimodal data set containing temperature and appearance characteristics. This meets the needs of large-area, comprehensive preliminary inspection data, achieving a comprehensive scan of the entire photovoltaic power plant. The acquired multimodal images completely record the temperature distribution and surface condition of the modules, providing a comprehensive data foundation for the subsequent preliminary identification of suspected hot spots, improving the coverage completeness and data richness of the inspection.

[0022] Specifically, in one possible embodiment, step S11 is implemented as follows: First, based on the distribution range, arrangement, and location of surrounding obstacles of the photovoltaic power station's module array, a preset flight path covering the entire module area is planned. The flight path avoids supports, cables, and ground protrusions between modules, ensuring that the drone's flight trajectory can completely scan each photovoltaic module. Then, a drone equipped with a high-definition infrared thermal imager, a visible light camera, and a high-precision positioning module is selected. The infrared thermal imager is used to capture the surface temperature distribution of the modules, the visible light camera is used to record the surface appearance features of the modules, and the positioning module is used to obtain real-time location information. After taking off from the designated take-off and landing point of the power station, the drone autonomously cruises along the preset flight path, maintaining a stable altitude and speed during flight to ensure the clarity and consistency of image acquisition. When the drone reaches the preset image acquisition point in the flight path, a synchronous acquisition command is triggered, and the infrared thermal imager and the visible light camera simultaneously capture images. For each frame of infrared image acquired, a corresponding frame of visible light image is generated synchronously. The two images are appended with the same time stamp and positioning information to ensure spatial and temporal correspondence.

[0023] Specifically, step S12 involves calculating an adaptive temperature difference threshold based on the real-time irradiance data. In a specific example of this application, step S12 includes: inputting the real-time irradiance data into a preset irradiance-temperature difference benchmark model to obtain a benchmark temperature difference value as the adaptive temperature difference threshold. Specifically, this application uses a preset irradiance-temperature difference benchmark model to convert the real-time collected irradiance data into a benchmark temperature difference value adapted to the current illumination conditions as the adaptive temperature difference threshold. This enables real-time matching of the temperature field characteristics corresponding to the current irradiance, accurately defining the critical temperature difference state between normal components and hot spots under the current environment, effectively eliminating misjudgments or omissions of fixed thresholds in scenarios with fluctuating illumination, ensuring the consistency of the temperature difference judgment standard with environmental conditions, providing a precise and dynamically adjustable judgment standard for subsequent hot spot identification, and improving the adaptability and accuracy of hot spot detection in complex illumination environments.

[0024] Specifically, in one possible embodiment, step S12 is implemented as follows: First, a standard photovoltaic module array is selected as the experimental object, including fault-free normal modules and artificially set typical hot spot modules. Under different weather conditions, gradient changes are formed by monitoring different natural light intensities to ensure that the module temperature reaches a steady state under each light condition. The average temperature of normal modules and the highest temperature of hot spot modules are collected simultaneously, the temperature difference is calculated, and samples are continuously accumulated. The collected raw data is preprocessed to remove invalid data during extreme weather and equipment malfunctions. Intervals are divided according to the light intensity range, and the average reference value of the temperature difference in each interval is calculated. A curve fitting algorithm is used to obtain a continuous mapping relationship between light intensity and reference temperature difference. Independent module arrays are selected to verify the model deviation, and adjustments and optimizations are made until the error meets the requirements. The final mapping relationship is solidified into an irradiance-temperature difference reference model and stored in the system's onboard processing unit. In actual inspections, the irradiance sensor on the drone monitors the ambient light intensity in real time, converts it into real-time irradiance data, and transmits it to the onboard processing unit. The processing unit performs noise reduction on the data to eliminate instantaneous fluctuations, and then inputs the real-time irradiance data into a preset irradiance-temperature difference benchmark model. The model outputs the corresponding benchmark temperature difference value according to the built-in mapping relationship. This benchmark temperature difference value is the adaptive temperature difference threshold under the current environment, which is used to determine the abnormal temperature areas in the infrared image of the photovoltaic module.

[0025] Specifically, in step S13, based on the adaptive temperature difference threshold, the infrared image of the photovoltaic module is segmented and generated into a list of suspected hot spot regions (ROIs) to obtain the list of suspected hot spot regions. It should be understood that infrared images contain a large amount of background and normal component areas; directly analyzing the entire image would be inefficient and easily interfered with by irrelevant information. Therefore, this application extracts regions with abnormal temperatures through threshold segmentation to obtain a structured list of suspected hot spots, clarifying the specific objects for subsequent processing, improving the targeting and processing efficiency of subsequent static screening, and ensuring the structured and accurate nature of the initial screening results.

[0026] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S13 of the real-time online monitoring method for the status of photovoltaic modules according to an embodiment of this application. Figure 4 As shown, step S13 includes: S131, performing histogram analysis on the infrared image of the photovoltaic module to obtain the background temperature; S132, performing pixel-level screening on the infrared image of the photovoltaic module based on the background temperature and the adaptive temperature difference threshold to obtain foreground pixels; S133, performing connected component analysis on all foreground pixels to obtain the list of suspected hot spot regions.

[0027] More specifically, step S131 involves performing histogram analysis on the infrared image of the photovoltaic module to obtain the background temperature. It should be understood that directly using simple statistical quantities is easily affected by a small number of abnormally high-temperature pixels, and cannot accurately reflect the temperature baseline of a normal module. Histogram analysis is needed to capture the main distribution characteristics of pixel temperatures to determine the temperature level representing a normal state. Specifically, this application analyzes the histogram distribution of temperature values ​​of all pixels in the infrared image of the photovoltaic module, extracting the temperature value with the highest proportion as the background temperature, such as the mode or median. This background temperature can truly reflect the temperature level of a healthy module, effectively avoiding interference from abnormally high-temperature pixels on the normal temperature baseline. This characterizes the general temperature state of healthy photovoltaic modules in the current image, ensuring a stable and accurate reference for subsequent temperature anomaly screening, and providing a reliable benchmark for subsequent anomaly temperature determination.

[0028] Specifically, in one possible embodiment, step S131 is implemented as follows: First, the infrared image of the photovoltaic module is preliminarily processed to remove pixels in non-photovoltaic module areas caused by the shooting angle at the image edges, retaining only valid pixels belonging to the photovoltaic module area. Next, the temperature values ​​of the valid pixels are extracted, and the number of pixels corresponding to different temperature values ​​is counted to generate a temperature distribution histogram, where the horizontal axis represents the temperature value and the vertical axis represents the number of pixels with that temperature value. Analyzing the temperature distribution histogram identifies the temperature range with the largest number of pixels. The core temperature value of this temperature range is the background temperature. This temperature value represents the general temperature level of a fault-free photovoltaic module under current environmental conditions, effectively avoiding interference from a small number of abnormally high-temperature pixels on the normal temperature benchmark, and providing a stable and reliable reference standard for subsequent judgment of whether pixel temperatures are abnormal.

[0029] More specifically, in step S132, based on the background temperature and the adaptive temperature difference threshold, pixel-level screening is performed on the infrared image of the photovoltaic module to obtain foreground pixels. It should be understood that hot spots in photovoltaic modules typically manifest as abnormally high temperatures in localized areas. These abnormal areas can be accurately located through pixel-by-pixel temperature comparison. Specifically, this application combines the background temperature and the adaptive temperature difference threshold, traverses each pixel in the infrared image, quantifies and defines the critical value of abnormal temperatures, and filters out all pixels with temperatures exceeding this critical value as foreground pixels. This marks potential hot spot region pixels, clearly defining the pixel range of suspected hot spots, providing a precise pixel-level basis for subsequent region aggregation, ensuring that no possible tiny hot spot pixels are missed, and achieving accurate screening of abnormal pixels.

[0030] Specifically, in one possible embodiment, step S132 is implemented as follows: First, each pixel in the infrared image of the photovoltaic module is examined one by one, and the temperature value corresponding to each pixel is extracted. This value is compared with the background temperature, and the difference between the two is calculated. If the difference between the temperature value of a certain pixel and the background temperature is greater than the adaptive temperature difference threshold, it indicates that the temperature at the location of the pixel is outside the normal range, and it is marked as a foreground pixel, and the specific coordinates of the pixel in the image are recorded; if the difference is less than or equal to the adaptive temperature difference threshold, the pixel is determined to be a normal background pixel and is not marked.

[0031] More specifically, step S133 involves performing connected component analysis on all foreground pixels to obtain the list of suspected hotspot regions. It should be understood that since the selected foreground pixels may be scattered, and single or isolated pixels do not possess the physical meaning of actual hotspots, spatial correlation analysis can aggregate adjacent foreground pixels into continuous regions to reflect the spatial morphology of actual hotspots, facilitating subsequent targeted processing of complete hotspot regions. Specifically, this application performs spatial connectivity detection on all foreground pixels, merging spatially adjacent foreground pixels into independent connected regions, assigning a unique identifier to each region, and recording its location, extent, and other information. This transforms scattered abnormal pixels into suspected hotspot regions with clear spatial boundaries and integrity, giving each region identifiable and analyzable independent unit attributes. This provides specific and complete analysis objects for subsequent multimodal static screening and dynamic monitoring, improving the targeting and efficiency of subsequent processing.

[0032] Specifically, in one possible embodiment, step S133 is implemented as follows: First, all foreground pixels obtained through pixel-level screening are aggregated. These pixels appear as scattered temperature anomalies in the infrared image, each corresponding to a local high-temperature location on the component surface. Next, a spatial connectivity analysis algorithm is used to process these foreground pixels, determining the positional correlation between pixels. If two foreground pixels are adjacent in the image, including horizontal, vertical, and diagonal adjacency, they are grouped into the same connected group. This correlation determination and merging operation is continuously performed on all foreground pixels, gradually aggregating adjacent foreground pixels into multiple independent continuous regions. For each formed continuous region, its boundary range in the infrared image is determined, the start and end coordinates of the region are recorded, and a unique identification identifier is assigned to each region to distinguish different suspected hot spot regions. Finally, the range, identifier, and other information of these regions are organized into a structured list, thus obtaining a list of suspected hot spot regions. Each region serves as an independent analysis unit, providing a specific and clear object for subsequent multimodal static investigation.

[0033] In the aforementioned real-time online monitoring method for photovoltaic modules, step S2 involves performing a multimodal static screening of the suspected hot spot area list to obtain a filtered list of suspected hot spot areas and a static pseudo-hot spot log. It should be understood that in the initially screened list of suspected hot spot areas, some temperature anomalies may be caused by external physical obstructions rather than internal module defects. If these pseudo-hot spots are not eliminated in advance, they will lead to invalid objects being processed in subsequent dynamic monitoring processes, increasing system load and reducing diagnostic accuracy. Visible light images can intuitively present information about surface obstructions, and combined with the temperature characteristics of infrared images, they can effectively identify pseudo-hot spots. Based on this, this application effectively filters pseudo-hot spots caused by external physical obstructions through multimodal static screening, making the objects in the filtered list of suspected hot spot areas more focused on temperature anomalies that may be caused by internal module defects. This reduces the invalid workload of subsequent dynamic monitoring. Simultaneously, the static pseudo-hot spot log provides traceable pseudo-hot spot information for operation and maintenance, thus improving the overall accuracy and efficiency of the initial screening process in the hot spot diagnosis process.

[0034] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S2 of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application. Figure 5 As shown, step S2 includes: S21, extracting a first suspected hot spot region from the suspected hot spot region list; S22, inputting a visible light image of the photovoltaic module of the first suspected hot spot region into a pre-trained object detection model to obtain a detection result; S23, in response to the detection result indicating the presence of an obstruction, adding the first suspected hot spot region to the static pseudo hot spot log; S24, in response to the detection result indicating the absence of an obstruction, adding the first suspected hot spot region to the filtered suspected hot spot region list.

[0035] Specifically, in step S21, a first suspected hot spot region is extracted from the list of suspected hot spot regions. In other words, a region is extracted from the list of suspected hot spot regions as the current processing unit. This first suspected hot spot region establishes a specific analysis object for subsequent targeted detection, enabling the multimodal static screening to proceed in an orderly manner. This clarifies the specific region that needs in-depth analysis, providing a clear target for subsequent visible light image detection, ensuring the screening process proceeds sequentially without overlooking any potential objects.

[0036] Specifically, in one possible embodiment, step S21 is implemented as follows: First, a list of suspected hot spot regions is arranged according to their distribution location or generation order in the photovoltaic array. Then, a region extraction program is started, locating the first entry in the list according to the storage order of the regions. The data reading interface is called to extract the region information corresponding to the entry, including the coordinates of the upper left and lower right corners of the region in the infrared image, the associated visible light image number, and the region's unique identifier. Simultaneously, the extraction status of the region is marked in the list to prevent repeated processing in subsequent operations, ensuring that the extracted region information is complete and unique, thereby obtaining the first suspected hot spot region.

[0037] Specifically, in step S22, the visible light image of the photovoltaic module in the first suspected hot spot area is input into a pre-trained object detection model to obtain the detection result. It should be understood that the visible light image can intuitively present the physical state of the module surface, such as the presence of external obstructions like bird droppings or leaves. The pre-trained object detection model has the ability to accurately identify such obstructions and can quickly determine whether the temperature anomaly is caused by external physical factors. Therefore, this application utilizes an object detection model to perform feature analysis on the visible light image of the photovoltaic module in the first suspected hot spot area, identifies the presence of obstructions, and obtains a clear detection result indicating whether obstructions exist in the area. This provides a reliable basis for subsequent differentiation between false hot spots and potential true hot spots, improving the scientific accuracy of false hot spot identification.

[0038] Specifically, in one possible embodiment, step S22 is implemented as follows: First, a large number of visible light image samples of photovoltaic modules are collected, covering different lighting conditions, weather conditions, and module types. The samples must include images with obstructions (such as leaves, bird droppings, dust accumulation, branches, etc.) and normal module images without obstructions, ensuring that the types of obstructions are consistent with common interference objects in photovoltaic power plants. The collected images are labeled, and professional labeling tools are used to define the location boundaries of obstructions in the images and label the obstruction categories, forming a dataset with labeled information. The dataset is preprocessed, and images are cropped to a uniform size. The number of samples is expanded by random rotation, brightness adjustment, contrast changes, etc., to avoid model overfitting. Then, the processed dataset is divided into training and validation sets according to the proportion for model training and performance evaluation. Then, a pre-trained model based on a convolutional neural network is selected as the basic architecture. This model has been initially trained on a large-scale image dataset and has strong image feature extraction capabilities. The pre-trained model's weights are loaded, and some shallow network layers are frozen to retain general image feature extraction capabilities. Only the deep network layers related to occlusion detection are adjusted. The training set is input into the model, and an appropriate loss function is set. Considering both the accuracy of occlusion bounding box localization and the accuracy of category judgment, a gradient descent optimizer is used to update the network parameters. During training, the loss value changes on the training set are monitored in real time. The model performance is evaluated using a validation set at fixed intervals, and the accuracy and false negative rate of occlusion detection are calculated. Finally, the training parameters are adjusted based on the validation set evaluation results. If the model has a low accuracy in recognizing a certain type of occlusion, the proportion of that type of sample in the training set is increased; if overfitting occurs, the data augmentation strategy is adjusted or a regularization term is added. After multiple rounds of iterative training, when the detection performance on the validation set stabilizes and reaches the preset standard, training is stopped, and the final model parameters are saved, thus completing the training of the object detection model.

[0039] Then, based on the association information of the first suspected hot spot region, the corresponding visible light image of the photovoltaic module is retrieved from the stored visible light image data. This image clearly shows the surface details of the region, including whether there are foreign objects, stains, or occlusion marks on the module's glass surface. This visible light image is input into a pre-trained object detection model. The model first preprocesses the image, such as cropping it to match the first suspected hot spot region and adjusting the color channels to optimize feature extraction. Subsequently, the model extracts features such as edges and textures from the image through multi-layer convolution operations. Combining this with the feature patterns of occlusions learned during training, the model performs pixel-by-pixel scanning and feature comparison of the regions in the image. By calculating the feature matching degree, the model determines whether there is a preset type of occlusion in the region, and finally outputs a clear detection result, namely "occlusion exists" or "occlusion does not exist".

[0040] In particular, in another possible preferred embodiment, step S22 includes: generating a photovoltaic module temperature image based on the pixel temperature of the photovoltaic module infrared image of the first suspected hot spot region; extracting image semantic features from the visible light image of the photovoltaic module and the photovoltaic module temperature image to obtain a first feature map and a second feature map; performing multimodal feature fusion on the first feature map and the second feature map to obtain image semantic coding features; and obtaining the detection result based on the image semantic coding features.

[0041] Here, when the object detection model obtains the detection results, in addition to the visible light image of the photovoltaic module in the first suspected hot spot area, it can also refer to a photovoltaic module temperature image generated based on the pixel temperature of each pixel in the infrared image of the photovoltaic module in the first suspected hot spot area, thereby achieving information complementarity and ambiguity resolution. It should be understood that RGB information provides rich texture, shape, and color features, which can be used to identify the type of object (bird droppings, leaves), while infrared information provides the thermal radiation distribution of the object's surface, which can be used to identify the object's thermodynamic state (abnormally high temperature points, evaporative cooling zones). By combining the visible light image and the temperature image, for example, when a suspicious spot is seen in the RGB image, the thermal features of the corresponding location in the infrared image can be queried. If the thermal features show a sustained high temperature, it may be an internal defect of the module; while if the thermal features show a normal temperature or low temperature, it is more likely an external stain.

[0042] Therefore, semantic feature extraction is first performed on the visible light image and the temperature image of the photovoltaic module to obtain a first feature map and a second feature map. Specifically, the visible light image of the photovoltaic module is input into a convolutional neural network model. Shallow convolutional layers extract basic visual features such as edges, colors, and textures, for example, scratches, stain edges, and outlines of obstructions on the module's glass surface. As the network depth increases, deeper convolutional layers further aggregate these basic features to extract more abstract semantic information, such as the overall shape of obstructions and the integrity of the module's frame, ultimately generating the first feature map. For the temperature image of the photovoltaic module, a convolutional neural network model is also used. Shallow convolutional layers capture basic thermodynamic features such as local temperature gradient changes and the spatial location of temperature extreme points, while deeper networks further integrate this information to mine abstract semantics such as the aggregation patterns of high-temperature areas and the shape features of temperature anomaly areas, ultimately generating the second feature map. In this way, appearance semantic features such as texture and shape of the photovoltaic module surface are extracted from the visible light image of the photovoltaic module, such as the outline of obstructions and the shape of stains. At the same time, thermodynamic semantic features such as temperature distribution gradient and extreme value regions, such as the spatial clustering of high-temperature points, are extracted from the temperature image of the photovoltaic module. This yields a first feature map and a second feature map that can characterize the essential attributes of the two types of images. For example, the irregular edge features of bird droppings in the visible light image of the photovoltaic module are enhanced in the first feature map, while the high-temperature core region features of hot spots in the temperature image of the photovoltaic module are highlighted in the second feature map, providing basic features rich in discriminative information for subsequent feature fusion.

[0043] Then, multimodal feature fusion is performed on the first feature map and the second feature map. Specifically, the first feature map and the second feature map are first multiplied by matrix to obtain the third feature map. The third feature map The feature matrix of each channel is used to represent the degree of interaction between features in the RGB feature map and features in the infrared feature map. This is used to capture the strength of the interaction between visible light semantic features and temperature semantic features at different spatial locations, quantify the matching degree of the two types of features, and obtain a third feature map that can reflect the collaborative relationship between cross-modal features. For example, in the pseudo hot spot area caused by foliage shading, the interaction value between visible light features and temperature features is high, while in the true hot spot area, the interaction value between the two is low, clarifying the association pattern between features.

[0044] Then, in the third feature map As the basis for interactive mapping, the third feature map is multiplied by the first feature map and the second feature map respectively to obtain the first interactive feature map and the second interactive feature map, that is: ;in, The first feature map represents the first feature map. Feature matrices of each channel The second feature map represents the first... Feature matrices of each channel The third feature map represents the first... Feature matrices of each channel Represents matrix multiplication. The first interactive feature map represents the first... Feature matrices of each channel The second interactive feature map represents the first... The feature matrix of each channel. That is, the original features are weighted and modulated by the strength of the interaction correlation, which strengthens the regional features that are strongly correlated with the other features. For example, in the first interaction feature map, the occlusion features that are strongly correlated with the high temperature region are enhanced, and in the second interaction feature map, the temperature features that are strongly correlated with the occlusion features are highlighted, thus improving the cross-modal discrimination ability of the features.

[0045] The positional similarity between the first interaction feature map and the second interaction feature map is calculated to obtain a similarity feature map. In other words, similarity feature map This is used to measure the similarity between RGB interaction features and temperature interaction features, with high similarity values ​​indicating a strong correlation between RGB interaction features (such as irregular dark spots) and infrared interaction features (such as gentle hot spots). Furthermore, by using similarity feature maps... by Probabilistic similarity feature maps are obtained by activation using probabilistic functions. Furthermore, it represents a pixel correlation metric based on the spatial distribution of pixel semantics, ensuring that each position value in the probabilistic similarity feature map is within the [0,1] interval. High values ​​correspond to high certainty of feature association, while low values ​​correspond to low certainty. For example, in the pseudo hot spot area formed by leaves covering the surface of a photovoltaic module, the probabilistic similarity value of this area approaches 1, clearly demonstrating the strong certainty of the association between the appearance features of the covering object and the local high temperature features. In contrast, in the true hot spot area caused by defects inside the module, the probabilistic similarity value approaches 0, intuitively reflecting the low certainty of the association between the absence of appearance abnormalities and the continuous high temperature features.

[0046] Then, based on the probabilistic similarity feature map, the first interaction feature map, and the second interaction feature map, the posterior first feature map and the posterior second feature map are generated, including: performing element-wise multiplication of the first interaction feature map with the reciprocal of the probabilistic similarity feature map to obtain the posterior first feature map; and performing element-wise multiplication of the second interaction feature map with the reciprocal of the probabilistic similarity feature map to obtain the posterior second feature map, i.e.: ;in, Indicates to Calculate the reciprocal of each element. Indicates to Calculate the reciprocal of each element. This represents the first interaction feature map. This represents the second interactive feature map. Represents element-wise multiplication. Represents the probabilistic similarity feature map. This represents the posterior first feature map. This represents the posterior second feature map.

[0047] In other words, given the semantic features of an image, the posterior probability distribution of the visual feature-thermodynamic feature association state based on interactive association is obtained in the image semantic space. For example, it can be simply understood as how much probability there is of presenting another association feature at the corresponding semantic representation position of an image based on a certain visual feature or thermodynamic feature.

[0048] Finally, the weighted sum of the first posterior feature map and the second posterior feature map is performed to obtain the image semantic encoding features, and the detection result is obtained based on the image semantic encoding features. Specifically, the image feature extraction capability of convolutional neural networks can be utilized. First, the image semantic encoding features are compressed through convolutional layers to reduce the computational load, while strengthening key features related to occlusion. Then, global average pooling is used to aggregate the spatial global information of the feature map, transforming the high-dimensional feature map into a low-dimensional vector that centrally reflects the core features of whether or not an occlusion exists. Subsequently, this vector is processed through multiple fully connected networks for feature mapping, gradually extracting key information related to the presence or absence of an occlusion. Finally, the output of the fully connected network is processed through the Softmax activation function to obtain the probability distributions of two classes: the presence of an occlusion and the absence of an occlusion. The class with the higher probability value is the detection result, and this probability value serves as the confidence level, reflecting the reliability of the result. This allows for the effective differentiation between true hotspots (no features in RGB, high heat in infrared) caused by internal defects and false hotspots (features in RGB, varying infrared temperatures) caused by external obstructions, thereby significantly reducing the number of invalid ROIs entering subsequent steps. It also makes the detection of obstructions with different thermal properties, such as bird droppings, fallen leaves, stagnant water, and tape, more accurate and robust.

[0049] Specifically, in steps S23 and S24, in response to the detection result indicating the presence of an obstruction, the first suspected hotspot area is added to the static pseudo-hotspot log; in response to the detection result indicating the absence of an obstruction, the first suspected hotspot area is added to the filtered suspected hotspot area list. Specifically, this application, based on the obstruction detection result, clearly categorizes suspected hotspot areas: pseudo-hotspots caused by external obstruction are included in the static pseudo-hotspot log for system recording and traceability; high-risk areas without obstruction are included in the filtered list as objects for subsequent dynamic feature analysis. This effectively separates pseudo-hotspots from high-risk suspected areas, ensuring resources are concentrated on potential fault points that truly require in-depth diagnosis. The static pseudo-hotspot log provides clear records of external interference for maintenance, facilitating subsequent targeted cleanup; the filtered suspected hotspot area list focuses on potential internal defect areas, making subsequent dynamic monitoring more targeted, reducing ineffective process consumption, and overall improving the accuracy of hotspot diagnosis and resource utilization efficiency.

[0050] In the aforementioned real-time online monitoring method for photovoltaic module status, step S3 involves extracting a first filtered suspected hot spot region from the filtered list of suspected hot spot regions. It should be understood that the list of filtered suspected hot spot regions contains multiple potentially high-risk regions that have undergone preliminary screening. Processing each region individually ensures that each filtered suspected hot spot region receives targeted dynamic feature analysis, thereby avoiding information contamination or insufficient analysis due to batch processing and guaranteeing the systematic nature and depth of subsequent monitoring. Specifically, this application extracts one region from the list of filtered suspected hot spot regions as the specific object of current dynamic monitoring, namely the first filtered suspected hot spot region. This clarifies the specific region that needs in-depth monitoring, providing a clear direction for subsequent precise data collection by the drone, ensuring that the dynamic monitoring process proceeds sequentially without overlooking any potential internal defect areas.

[0051] Specifically, in one possible embodiment, step S3 is implemented as follows: First, a list of suspected hot spot regions after filtering, generated through multimodal static screening, is obtained. This list is stored in a structured form, containing the unique identifier of each region, its location boundaries in the infrared and visible light images, associated acquisition time information, and verification markers after excluding obstructions, and is arranged according to the distribution order of the regions in the photovoltaic array. The region extraction process is initiated, locating the first entry in the list according to the storage order of the regions. The complete information corresponding to this entry is retrieved through a data interface, including the precise coordinate range of the region in the infrared image, the corresponding photovoltaic module number, associated previous screening records, and real-time irradiance data. Simultaneously, the extraction status of this region is marked as "to be monitored" in the list to avoid repeated extraction in subsequent operations, ensuring that the extracted region information is complete and unique, thereby obtaining the first filtered suspected hot spot regions.

[0052] In the aforementioned real-time online monitoring method for photovoltaic module status, step S4 involves acquiring a time series of infrared images of the first filtered suspected hot spot area using intelligent hovering of a drone. It should be understood that intelligent hovering of the drone ensures stable residence in the target area, guaranteeing spatial consistency of the acquired image sequence and avoiding temperature data deviations caused by positional changes. Specifically, this application uses a drone to continuously acquire infrared images within a specific time window by stably hovering in the target area. The length of this specific time window is set to a duration sufficient to observe the dynamic temperature response, for example, 30 to 90 seconds. The acquisition frequency is set to a level capable of capturing subtle temperature fluctuations, for example, 1 Hz to 5 Hz. This yields an infrared image time series containing information on temperature changes over time, completely recording the temperature change trajectory in the time dimension. This provides reliable data support for subsequent extraction of temperature fluctuation and trend characteristics, making it possible to distinguish between true and false hot spots based on dynamic features.

[0053] Specifically, in one possible embodiment, step S4 is implemented as follows: First, the location coordinates of the first filtered suspected hot spot area are acquired. This information includes the specific location of the area within the photovoltaic array and its corresponding geographical coordinates. The UAV receives the coordinate command for this area from the ground control system, adjusts its flight trajectory using its onboard positioning module, and moves towards the target area. During this process, real-time position calibration ensures that the flight path accurately points to the area. When the UAV reaches a preset altitude directly above the target area, it activates intelligent hovering mode, using its body sensors to perceive airflow changes in real time and adjust propeller speed to maintain stability, ensuring the lens remains focused on the first filtered suspected hot spot area. Subsequently, the UAV activates the continuous acquisition function of the infrared thermal imager according to a preset time window, continuously capturing infrared images of the area at fixed intervals within a set time period. Each frame of the image is appended with the same timestamp and location information to ensure the temporal and spatial continuity of the image sequence. During the acquisition process, the image data is transmitted to the ground processing system in real time and stored in the acquisition order, forming an infrared image time series containing the temperature changes of the area over time, providing continuous dynamic data support for subsequent extraction of temperature time series features.

[0054] In the aforementioned real-time online monitoring method for photovoltaic modules, step S5 involves extracting temperature time-series features from the infrared image time series of the first filtered suspected hot spot region. It should be understood that the essential difference between genuine and false hot spots lies in the dynamic pattern of temperature changes. Feature extraction can transform this pattern into quantifiable indicators to support subsequent classification decisions. Specifically, key features characterizing temperature change patterns are extracted from continuous infrared image sequences, transforming the original image data into structured feature information. This yields temperature time-series features that accurately depict the dynamic temperature change pattern, effectively capturing the essential differences in thermodynamic response between genuine and false hot spots. This lays a reliable quantitative foundation for subsequent classification judgment, enabling the classification process to be based on dynamic features rather than static information, thus improving the scientific rigor of the distinction.

[0055] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S5 of the real-time online monitoring method for photovoltaic module status according to an embodiment of this application. Figure 6 As shown, step S5 includes: S51, extracting the highest temperature of each infrared image from the infrared image time series to obtain a highest temperature vector composed of the highest temperatures of each infrared image; S52, performing fluctuation feature calculation and trend feature calculation on the highest temperature vector to obtain a temperature time series feature vector composed of fluctuation feature and trend feature as the temperature time series feature.

[0056] Specifically, step S51 involves extracting the highest temperature from each infrared image in the infrared image time series to obtain a highest temperature vector composed of the highest temperatures from each infrared image. It should be understood that the highest temperature in the hotspot region can centrally reflect the abnormal temperature characteristics of that region and is a key indicator reflecting the core state of the hotspot. Other temperature values ​​may be affected by the background or non-core areas, making it difficult to accurately characterize the dynamic changes of the hotspot. Therefore, this application extracts the highest temperature values ​​from each infrared image and combines them in chronological order to obtain a concise highest temperature vector that centrally reflects the dynamics of the hotspot core temperature. This eliminates interference from non-critical temperature information, transforms the two-dimensional image sequence into a one-dimensional temperature sequence, simplifies the data dimension while retaining the trajectory of the hotspot core temperature change over time, and provides an accurate and efficient data foundation for subsequent calculations of fluctuation and trend characteristics.

[0057] Specifically, in one possible embodiment, step S51 is implemented as follows: First, an infrared image time series of the first filtered suspected hot spot area is acquired through intelligent hovering of the UAV. This series is arranged in the order of acquisition time and contains multiple frames of infrared images continuously captured within a specific time window. Each frame clearly shows the temperature distribution details of the area. An image processing interface is called to process the infrared images in the sequence one by one in chronological order. For a single frame of infrared image, the pixel range belonging to the first filtered suspected hot spot area is first located, and interference from non-target areas at the edges is eliminated. Then, the temperature values ​​of all pixels within this range are extracted, and the maximum value is selected as the highest temperature of that frame. The highest temperature extracted from each frame is recorded sequentially according to its acquisition order in the time series, forming an ordered temperature value sequence, i.e., the highest temperature vector, which provides basic data for subsequent feature calculations.

[0058] In particular, in another possible preferred embodiment, extracting the highest temperature of each infrared image from the infrared image time series includes: calculating the feature mean and feature variance of the posterior second feature map; performing Gaussian parameterization reshaping on the pixel temperature values ​​of each infrared image based on the feature mean and feature variance to obtain the reshaped pixel temperature values; and extracting the maximum value from the reshaped pixel temperature values ​​as the highest temperature.

[0059] Specifically, for extracting the highest temperature of each infrared image from the infrared image time series, the feature mean and feature variance of the posterior second feature map are calculated and used as a gate threshold. Based on the feature mean and feature variance, the pixel temperature values ​​of each infrared image are Gaussian parameterized renormalized to obtain the renormalized pixel temperature value, for example, denoted as... Perform Gaussian parametric renormalization, i.e.: ;in, and Let represent the mean and variance of the posterior second feature map, respectively. This represents the pixel temperature value of each infrared image. Represents the natural constant. This represents the re-adjusted pixel temperature value.

[0060] In other words, by utilizing the visual-temperature interaction semantic information contained in the posterior second feature map, a Gaussian constraint model is constructed through its statistical characteristics to reorganize the distribution of the original pixel temperature values, so that the quantitative representation of the temperature value forms an association constraint with the semantic features of the image, thereby filtering out non-essential high-temperature interference caused by false hot spots and accurately focusing on temperature features related to the internal state of the component.

[0061] For example, if an infrared image contains both a slight, genuine hotspot representing an early fault and a higher-temperature pseudo-hotspot (such as heat-absorbing dark bird droppings), the original maximum temperature might incorrectly pinpoint the bird droppings, and its temperature timeline might fluctuate drastically due to moisture evaporation, causing the component containing the real fault to be misclassified as a pseudo-hotspot, resulting in a missed detection. However, by associating the temperature with visually characteristic states through a posterior interactive association representation... The temperature distribution can be constrained by using its distribution as a priori, and the maximum value is extracted from the reorganized pixel temperature values ​​as the highest temperature. This allows the time series of temperatures to reflect not only thermodynamic characteristics but also the correlation with visual features, thus achieving sequence distribution reorganization based on semantic information. This can be understood as follows: during thermal anomaly diagnosis, known visual features have been correlated with thermodynamic features, thereby obtaining the hottest temperature associated with the visual state, which is more accurate in a physical sense.

[0062] Specifically, in step S52, fluctuation characteristics and trend characteristics are calculated on the highest temperature vector to obtain a temperature time-series feature vector composed of fluctuation characteristics and trend characteristics as the temperature time-series feature. In a specific example of this application, the fluctuation characteristic is the temperature standard deviation, and the trend characteristic is the slope of the temperature change rate. Specifically, by calculating the fluctuation characteristics of the highest temperature vector, such as indicators reflecting temperature stability, and the trend characteristics, such as indicators reflecting the overall direction of temperature change, the temperature sequence is transformed into a temperature time-series feature vector that can directly distinguish between true and false hot spots. This accurately captures the essential differences between true and false hot spots in dynamic changes, provides input parameters with clear physical meaning for the classification model, significantly improves the subsequent classification model's ability to identify the two types of hot spots, and makes the classification results more reliable and physically based.

[0063] Specifically, in one possible embodiment, step S52 is implemented as follows: First, a maximum temperature vector composed of the highest temperatures from each infrared image is obtained. This vector records the core temperature changes of the target area in chronological order. Fluctuation characteristics are calculated for this vector. By statistically analyzing the dispersion of all temperature values ​​in the vector, an index reflecting temperature stability, namely the temperature standard deviation, is calculated. This value reflects the fluctuation range of the target area temperature during the monitoring period. Then, trend characteristics are calculated. A coordinate system is established with time as the horizontal axis and the highest temperature at the corresponding moment as the vertical axis. The trend line of temperature change in this coordinate system is solved using a linear fitting method to obtain the slope of the trend line, namely the slope of the temperature change rate. This value reflects the overall upward or downward trend of the target area temperature. The calculated temperature standard deviation and the temperature change rate slope are combined in a fixed order to form an ordered array containing two features, namely the temperature time-series feature vector.

[0064] In the aforementioned real-time online monitoring method for photovoltaic modules, step S6 involves classifying the temperature time-series features to obtain a classification result. The classification result includes a category and a confidence level. The category includes true hot spots and false hot spots. It should be understood that true hot spots caused by internal defects primarily derive their heat from the continuous Joule heating effect. Under stable irradiation, their temperature exhibits low fluctuations and a relatively stable or slowly rising trend. False hot spots caused by external factors, on the other hand, show temperature changes strongly correlated with the dynamic process of external disturbances, typically exhibiting drastic, irregular fluctuations or rapid temperature rises and falls. In a specific example of this application, step S6 includes: inputting the temperature time-series feature vector into a trained support vector machine to obtain the classification result. Specifically, the trained support vector machine is used to perform pattern matching on the temperature time-series feature vector containing fluctuation and trend characteristics. Based on the distribution patterns of true and false hot spot features learned by the model, the classification result for that region is output, effectively distinguishing between true and false hot spots. This transforms abstract feature data into clear and reliable diagnostic conclusions, providing a standardized basis for operation and maintenance decisions. The classification results clearly define the fault attributes, while the confidence level provides a quantitative reference for prioritizing maintenance. This avoids ineffective maintenance costs caused by false hot spots and ensures that genuine hot spots can be identified and dealt with in a timely manner, significantly improving the accuracy of hot spot diagnosis and the utilization efficiency of maintenance resources. This lays a solid foundation for the refined management of the health status of photovoltaic modules.

[0065] Specifically, in one possible embodiment, step S6 is implemented as follows: First, a large number of historical inspection cases accumulated during system operation are selected as the training data source for the support vector machine. These cases have all been manually verified on-site and clearly marked as true hot spots or false hot spots. True hot spot cases originate from temperature anomalies caused by defects such as internal microcracks in photovoltaic modules, junction box failures, and cell aging. False hot spot cases are caused by temperature increases due to external obstructions such as leaves and bird droppings or temporary stains. The corresponding infrared image time series are extracted from these cases, and the highest temperature vector is obtained according to the previously determined method. Then, the fluctuation characteristics and trend characteristics are calculated to form feature vectors for model training. The extracted feature vectors are preprocessed, and the influence of differences in the dimensions between different features is eliminated through standardization operations to ensure that the feature values ​​are within a suitable numerical range and that the influence weight of each feature is balanced during model training. The processed dataset is divided into a training set and a validation set, with most of the data used for model parameter learning and a small portion used to verify the training effect. A support vector machine (SVM) model was constructed using a radial basis function kernel. Cross-validation was employed to optimize the model within a pre-defined parameter range, adjusting the regularization parameter and kernel coefficients. The optimal parameter combination was determined based on the classification performance on the validation set. During training, the model learned the distribution patterns of features from both classes of samples, continuously optimizing the classification hyperplane to maximize the gap between genuine and pseudo-hot spot samples. This enabled the model to better capture the essential differences in temperature time-series characteristics between the two types of hot spots. Training was stopped when the classification results on the validation set remained stable for several consecutive rounds and could accurately distinguish the vast majority of genuine and pseudo-hot spot cases. The resulting model can then be used to classify newly acquired temperature time-series feature vectors, outputting the corresponding category and confidence level, thus supporting the accurate identification of hot spot faults in photovoltaic modules.

[0066] Then, the trained Support Vector Machine (SVM) model is invoked. This model, trained on a large amount of labeled sample data of both true and false hot spots, is capable of identifying the differences in temperature time-series characteristics between the two types of hot spots. The temperature time-series feature vectors are input into the model, which analyzes the features within the vectors and transforms them into a high-dimensional space using a kernel function. It then searches for the optimal classification hyperplane and determines the category based on the relative position of the vector to the hyperplane. Simultaneously, the model calculates the reliability of this determination, i.e., the confidence score, reflecting the certainty of the classification. Finally, the model outputs a clear classification result, including whether the target area belongs to a true or false hot spot, and the corresponding confidence score. This result provides direct evidence for distinguishing between true and false hot spots, supporting subsequent operational decisions.

[0067] In summary, the real-time online monitoring method for photovoltaic module status based on the embodiments of this application is explained. It uses a drone to collect infrared images of suspected hot spot areas and combines them with real-time irradiance data to perform a preliminary static screening of the collected multimodal image information, thereby quickly filtering out false hot spots caused by significant physical obstructions or dirt. For highly suspected fault areas, a dynamic feature analysis mechanism is introduced. By driving the drone to intelligently hover and continuously observe the target, it actively collects infrared thermal image sequences of the area within a specific time window and extracts temperature time-series features that characterize its thermodynamic response. Finally, classification is performed based on the temperature time-series features to accurately distinguish between true and false hot spots. This enables refined diagnosis of the health status of photovoltaic modules, significantly improving the identification accuracy and intelligence level of the diagnostic model.

[0068] Figure 7 This is a block diagram of a real-time online monitoring system for the status of photovoltaic modules according to an embodiment of this application. Figure 7 As shown, the photovoltaic module real-time online monitoring system 100 according to an embodiment of this application includes: a suspected hot spot area screening module 110, used to obtain a list of suspected hot spot areas collected by a drone based on real-time irradiance data; a multimodal static screening module 120, used to perform multimodal static screening on the list of suspected hot spot areas to obtain a filtered list of suspected hot spot areas and a static pseudo hot spot log; a filtered suspected area extraction module 130, used to extract a first filtered suspected hot spot area from the filtered suspected hot spot area list; an infrared image time series acquisition module 140, used to acquire the infrared image time series of the first filtered suspected hot spot area through intelligent hovering of the drone; a temperature time series feature extraction module 150, used to extract temperature time series features from the infrared image time series of the first filtered suspected hot spot area; and a true and false hot spot classification judgment module 160, used to perform classification judgment based on the temperature time series features to obtain a classification result, the classification result including category and confidence level, the category including true hot spots and pseudo hot spots.

[0069] As described above, the real-time online monitoring system 100 for photovoltaic module status according to the embodiments of this application can be implemented in various wireless terminals, such as servers with real-time online monitoring algorithms for photovoltaic module status. In one possible implementation, the real-time online monitoring system 100 for photovoltaic module status according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the real-time online monitoring system 100 for photovoltaic module status can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the real-time online monitoring system 100 for photovoltaic module status can also be one of many hardware modules of the wireless terminal.

[0070] Alternatively, in another example, the real-time online monitoring system 100 for photovoltaic module status and the wireless terminal can also be separate devices, and the real-time online monitoring system 100 for photovoltaic module status can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0071] Those skilled in the art will understand that the specific operations of each step in the above-described real-time online monitoring system for photovoltaic module status have been referenced above. Figures 1 to 6 The method for real-time online monitoring of photovoltaic module status has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for real-time online monitoring of the status of photovoltaic modules, characterized in that, include: Based on real-time irradiance data, a list of suspected hot spot areas collected by drones is obtained; A multimodal static screening is performed on the list of suspected hot spot areas to obtain a filtered list of suspected hot spot areas and a static pseudo hot spot log; a first filtered suspected hot spot area is extracted from the filtered list of suspected hot spot areas; an infrared image time series of the first filtered suspected hot spot area is collected by intelligent hovering of a UAV, the infrared image time series consisting of multiple infrared images within a predetermined time period; temperature time series features are extracted from the infrared image time series of the first filtered suspected hot spot area. Classification is performed based on the temperature time-series characteristics to obtain classification results, which include categories and confidence levels. Categories include true hot spots and false hot spots. The process of performing multimodal static screening on the list of suspected hot spot regions to obtain a filtered list of suspected hot spot regions and a static pseudo hot spot log includes: extracting a first suspected hot spot region from the list of suspected hot spot regions; inputting a visible light image of the photovoltaic module of the first suspected hot spot region into a pre-trained object detection model to obtain a detection result; in response to the detection result indicating the presence of an obstruction, adding the first suspected hot spot region to the static pseudo hot spot log; and in response to the detection result indicating the absence of an obstruction, adding the first suspected hot spot region to the filtered list of suspected hot spot regions.

2. The method for real-time online monitoring of photovoltaic module status according to claim 1, characterized in that, Based on real-time irradiance data, a list of suspected hot spot regions collected by a drone is obtained, including: controlling the drone to perform wide-area flight inspection according to a preset route to obtain infrared images and visible light images of the photovoltaic module; calculating an adaptive temperature difference threshold based on the real-time irradiance data; and performing suspected hot spot ROI segmentation and generation on the infrared images of the photovoltaic module based on the adaptive temperature difference threshold to obtain the list of suspected hot spot regions.

3. The method for real-time online monitoring of photovoltaic module status according to claim 2, characterized in that, Calculating an adaptive temperature difference threshold based on the real-time irradiance data includes: inputting the real-time irradiance data into a preset irradiance-temperature difference benchmark model to obtain a benchmark temperature difference value as the adaptive temperature difference threshold.

4. The method for real-time online monitoring of photovoltaic module status according to claim 2, characterized in that, Based on the adaptive temperature difference threshold, the infrared image of the photovoltaic module is segmented and generated for suspected hot spot regions (ROIs) to obtain a list of suspected hot spot regions. This includes: performing histogram analysis on the infrared image of the photovoltaic module to obtain the background temperature; performing pixel-level screening on the infrared image of the photovoltaic module based on the background temperature and the adaptive temperature difference threshold to obtain foreground pixels; and performing connected component analysis on all foreground pixels to obtain the list of suspected hot spot regions.

5. The method for real-time online monitoring of photovoltaic module status according to claim 1, characterized in that, Extracting temperature time-series features from the infrared image time series of the first filtered suspected hot spot region includes: extracting the highest temperature of each infrared image from the infrared image time series to obtain a highest temperature vector composed of the highest temperatures of each infrared image; performing fluctuation feature calculation and trend feature calculation on the highest temperature vector to obtain a temperature time-series feature vector composed of fluctuation feature and trend feature as the temperature time-series feature.

6. The method for real-time online monitoring of photovoltaic module status according to claim 5, characterized in that, The volatility characteristic is the temperature standard deviation, and the trend characteristic is the slope of the temperature change rate.

7. The method for real-time online monitoring of photovoltaic module status according to claim 5, characterized in that, The classification judgment based on the temperature time series features to obtain the classification result includes: inputting the temperature time series feature vector into the trained support vector machine to obtain the classification result.

8. A real-time online monitoring system for the status of photovoltaic modules, characterized in that, include: The preliminary screening module for suspected hot spot areas is used to obtain a list of suspected hot spot areas collected by UAVs based on real-time irradiance data. A multimodal static screening module is used to perform multimodal static screening on the list of suspected hot spot areas to obtain a filtered list of suspected hot spot areas and a static pseudo hot spot log; a filtered suspected area extraction module is used to extract a first filtered suspected hot spot area from the list of filtered suspected hot spot areas; an infrared image time series acquisition module is used to acquire the infrared image time series of the first filtered suspected hot spot area through intelligent hovering of a drone, wherein the infrared image time series consists of multiple infrared images within a predetermined time period; and a temperature time series feature extraction module is used to extract temperature time series features from the infrared image time series of the first filtered suspected hot spot area. The module for classifying and judging true and false hot spots is used to perform classification and judgment based on the temperature time-series characteristics to obtain classification results. The classification results include categories and confidence levels. Categories include true hot spots and false hot spots. The process of performing multimodal static screening on the list of suspected hot spot regions to obtain a filtered list of suspected hot spot regions and a static pseudo hot spot log includes: extracting a first suspected hot spot region from the list of suspected hot spot regions; inputting a visible light image of the photovoltaic module of the first suspected hot spot region into a pre-trained object detection model to obtain a detection result; in response to the detection result indicating the presence of an obstruction, adding the first suspected hot spot region to the static pseudo hot spot log; and in response to the detection result indicating the absence of an obstruction, adding the first suspected hot spot region to the filtered list of suspected hot spot regions.

Images