A method and system for sensing the pollution state of a marine photovoltaic panel

Abstract

This invention belongs to the field of photovoltaic detection technology. To address the problem of inaccurate detection of existing photovoltaic panels, it proposes a method and system for sensing the pollution status of marine photovoltaic panels. The method involves inputting high-altitude images of marine photovoltaic panels into a coarse-grained pollution detection model to obtain the pollution probability, pollution coverage ratio, and pollution level of the photovoltaic panels. Photovoltaic panels with severe pollution levels are then selected to generate a list of suspected high-pollution targets. Next, close-range RGB images and depth information of the high-pollution targets are acquired and stitched together. The stitched data is then input into a fine-grained pollution detection model to obtain a fine mask of the polluted area, pollution type, and pixel-level thickness map. Based on the pollution type, fine coverage ratio, and thickness level, a cleaning strategy is generated to achieve cleaning. This invention enables rapid pollution detection over a wide area and provides high-precision pollution type and thickness information, comprehensively improving pollution monitoring capabilities.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic detection technology, and in particular relates to a method and system for sensing the pollution status of marine photovoltaic panels. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the increasing global demand for renewable energy, offshore photovoltaic (PV) power generation, as an important form of green energy, has received widespread attention. During the construction and operation of offshore PV power plants, surface contamination of the photovoltaic panels has become a key factor affecting PV power generation efficiency. Due to the unique marine environment, factors such as seawater evaporation and high air humidity make it easy for pollutants such as sea salt, bird droppings, oil, and algae to adhere to the surface of the PV panels, leading to a decrease in PV conversion efficiency and even affecting their long-term stability.

[0004] Existing pollution detection technologies mainly rely on manual inspections, traditional image recognition techniques, and drone-based inspection systems. While some methods have achieved automated monitoring, they still suffer from several drawbacks: Current detection methods often only provide a rough assessment of whether photovoltaic panels are contaminated, lacking a refined evaluation of the type and degree of contamination. This coarse-grained detection is particularly inadequate for large-scale offshore photovoltaic power plants, failing to meet the demands of precise cleaning. Furthermore, the thickness of contaminants significantly influences the choice of cleaning method; current technologies cannot accurately assess contamination thickness, resulting in cleaning strategies that cannot precisely match the actual contamination situation. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for sensing the pollution status of marine photovoltaic panels, which can detect pollution rapidly over a wide area and provide high-precision information on pollution type and thickness, thus comprehensively improving pollution monitoring capabilities.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for sensing the pollution status of marine photovoltaic panels, comprising:

[0008] The collected aerial images of marine photovoltaic panels are input into a coarse-grained pollution detection model to obtain the pollution probability, pollution coverage ratio, and pollution level of the photovoltaic panels. A list of suspected high-pollution targets is generated by screening photovoltaic panels with severe pollution levels.

[0009] Based on the list of suspected high-pollution targets, the inspection route is planned, and the drone is controlled to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel. The close-range RGB images and depth information are stitched together and input into the fine-grained pollution detection model to obtain a fine mask of the polluted area, pollution type and pixel-level thickness map.

[0010] Fit the reference plane of the photovoltaic panel, calculate the depth difference of the contaminated area and map it to the physical thickness, statistically analyze the average thickness and thickness level of the module, generate a cleaning strategy based on the contamination type, fine coverage ratio and thickness level, and control the drone to perform the cleaning task according to the cleaning strategy.

[0011] Secondly, the present invention provides a marine photovoltaic panel pollution status sensing system, comprising:

[0012] The coarse-grained detection module is configured to: input the collected aerial images of marine photovoltaic panels into the pollution coarse-grained detection model to obtain the pollution probability, pollution coverage ratio and pollution level of the photovoltaic panels, and filter photovoltaic panels with severe pollution levels to generate a list of suspected high-pollution targets;

[0013] The fine-grained detection module is configured to: plan the inspection path based on the list of suspected high-pollution targets, control the drone to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel, stitch the close-range RGB images and depth information together, and input them into the fine-grained pollution detection model to obtain a fine mask of the pollution area, pollution type and pixel-level thickness map.

[0014] The cleaning module is configured to: fit the reference plane of the photovoltaic panel, calculate the depth difference of the contaminated area and map it to the physical thickness, statistically analyze the average thickness and thickness level of the module, generate a cleaning strategy based on the contamination type, fine coverage ratio and thickness level, and control the drone to perform the cleaning task according to the cleaning strategy.

[0015] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0017] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0018] The above one or more technical solutions have the following beneficial effects:

[0019] In this invention, coarse-grained detection yields the probability of contamination, the proportion of contamination coverage, and the level of contamination in photovoltaic panels, while fine-grained detection provides a detailed mask of the contaminated area, the type of contamination, and a pixel-level thickness map. This approach enables rapid contamination detection over a wide area and provides high-precision information on the type and thickness of contamination, comprehensively enhancing contamination monitoring capabilities. By combining high-altitude and close-range images with multimodal data fusion, contaminated photovoltaic panels are accurately identified, and the degree of contamination is assessed based on the specific type and thickness of the contamination. Compared to traditional manual inspections or single-image detection methods, this approach offers higher accuracy and efficiency. Furthermore, by utilizing depth maps to estimate the contamination thickness, the accumulation layer of pollutants on the photovoltaic panels is accurately assessed, providing a reliable basis for developing cleaning strategies.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a schematic diagram of a drone detecting dust on a photovoltaic panel in an embodiment of the present invention;

[0023] Figure 2 This is a flowchart of the coarse-grained dust detection process for drones in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of a drone cleaning dust from a photovoltaic panel in an embodiment of the present invention;

[0025] Figure 4 This is a flowchart illustrating the process of cleaning dust using a drone in an embodiment of the present invention;

[0026] Figure 5 This is a flowchart illustrating the overall process of drone detection and cleaning in the implementation of this invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0029] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] Example 1

[0031] Existing pollution detection technologies mainly focus on manual inspection, traditional image recognition technology, and drone-based inspection systems. Manual inspection relies on regular inspections and manual cleaning by hand, which suffers from high labor costs, low efficiency, and the risk of oversights. Regular cleaning, on the other hand, is energy-intensive and lacks targeted cleaning capabilities, as not all photovoltaic panels require large-scale cleaning due to varying levels of contamination. Some studies have attempted to use traditional image processing methods for pollution detection, but these methods often rely on manual feature design, resulting in poor performance and difficulty adapting to the changing marine environment. In recent years, drone-based inspection systems have been widely used, enabling efficient regular inspections of offshore photovoltaic power plants. However, existing drone inspection systems rely on traditional visual recognition technology and can only perform coarse-grained pollution detection, failing to effectively identify pollutant types and their specific coverage areas, thus hindering the precise formulation of cleaning strategies. The inability of existing technologies to effectively achieve accurate detection of photovoltaic panel contamination, estimate contamination thickness, and formulate cleaning strategies leaves significant room for optimization in the pollution control and operation and maintenance management of offshore photovoltaic power plants. Therefore, this embodiment proposes a method for sensing the pollution status of offshore photovoltaic panels, including:

[0032] The collected aerial images of marine photovoltaic panels are input into a coarse-grained pollution detection model to obtain the pollution probability, pollution coverage ratio, and pollution level of the photovoltaic panels. A list of suspected high-pollution targets is generated by screening photovoltaic panels with severe pollution levels.

[0033] Based on the list of suspected high-pollution targets, the inspection route is planned, and the drone is controlled to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel. The close-range RGB images and depth information are stitched together and input into the fine-grained pollution detection model to obtain a fine mask of the polluted area, pollution type and pixel-level thickness map.

[0034] Fit the reference plane of the photovoltaic panel, calculate the depth difference of the contaminated area and map it to the physical thickness, statistically analyze the average thickness and thickness level of the module, generate a cleaning strategy based on the contamination type, fine coverage ratio and thickness level, and control the drone to perform the cleaning task according to the cleaning strategy.

[0035] This embodiment utilizes coarse-grained contamination detection based on high-altitude images, combined with photovoltaic panel detection and contamination coverage ratio prediction, to quickly identify severely contaminated photovoltaic panels. Fine-grained contamination detection based on close-range images and depth information accurately estimates the type and thickness of contaminants, providing a precise basis for generating cleaning strategies. Combining the contamination detection results with the assessment of photovoltaic panel contamination type and thickness, a cleaning task is intelligently generated, including cleaning mode, cleaning path, and cleaning priority. This embodiment's solution enables both rapid, wide-area contamination detection and provides high-precision information on contamination type and thickness, comprehensively enhancing contamination monitoring capabilities.

[0036] The following is combined Figure 1 The following is a detailed description of the pollution status sensing method for marine photovoltaic panels proposed in this embodiment:

[0037] Step 1: Take photos with a drone.

[0038] The inspection drone is equipped with a gimbal camera. After takeoff, it rises to an altitude of 80-120 meters and covers the entire offshore photovoltaic power station area according to a pre-planned route.

[0039] The drone acquires overhead RGB images at fixed intervals, taking one image with a resolution of 3840×2160 every 2 seconds, while simultaneously recording information such as GPS coordinates, heading, and pitch angle at the time of capture. All images are transmitted to the ground control terminal in real time or in batches via wireless link, and stored according to timestamps and location information for subsequent offline or online inference.

[0040] Step 2, image preprocessing.

[0041] Distortion correction is performed on each aerial image, using camera calibration parameters to remove barrel or pincushion distortion. Adaptive histogram equalization (CLAHE) is employed to enhance image brightness and improve the contrast between the photovoltaic panel and the background. Significant reflective areas on the sea surface are suppressed by using simple thresholding and morphological operations to assign large areas of high-brightness reflective regions to low-weight regions, reducing false detection interference in subsequent networks.

[0042] Step 3, design of coarse-grained detection network.

[0043] like Figure 2 As shown, in this embodiment, the coarse-grained detection network adopts a single-stage detector with a YOLO-like structure and has been extended for photovoltaic scenarios. The model is denoted as YOLO-PV-Coarse.

[0044] The basic structure of YOLO-PV-Coarse consists of: a backbone network using a convolutional neural network to extract multi-scale features; a neck network using a feature pyramid structure to fuse features from different scales; and a head network outputting component-related information for each scale feature.

[0045] To address the challenge of highly irregular surface contamination morphology and low contrast with background texture on photovoltaic panels, traditional detection heads using standard convolution or AKConv algorithms are prone to "sampling point drift" when extracting contamination features, leading to the introduction of significant background noise. To resolve this, our network innovatively introduces arbitrary-morphological convolution with spatial consistency constraints (SC-AKConv) as the core feature extraction operator in the detection head. This operator constrains the sampling offset through total variational regularization, providing high signal-to-noise ratio feature support for subsequent high-dimensional joint sensing.

[0046] Based on the high-quality features extracted by SC-AKConv, unlike traditional YOLO which only outputs "category information and location information", YOLO-PV-Coarse outputs the following high-dimensional information for each candidate box:

[0047]

[0048] in, This represents the relative coordinates of the detection box center. and This indicates the width and height of the frame. This represents the probability of photovoltaic panel category. This indicates the probability that the component is significantly contaminated. This represents a rough estimate of the pollution coverage ratio. . This indicates the pollution level classification result (0=clean, 1=mild, 2=moderate, 3=severe).

[0049] To achieve accurate regression and classification of the aforementioned multidimensional physical quantities, the size of three additional detection heads was increased from the original detection head. The original detection head size was... , , In this case, the 4 in parentheses represents The 1 in parentheses represents the target confidence level, the probability of including the target; The number in parentheses represents the category confidence level; the number of candidate boxes generated is indicated by the number of boxes outside the parentheses.

[0050] The improved detection head size is expanded to:

[0051] , , The 4 in parentheses represents The 1 in parentheses represents the target confidence level; The parentheses indicate the category confidence (there is only one category here); the last 3 outside the parentheses indicates the number of candidate boxes generated; the newly added 3 inside the parentheses indicates the independent channels for the probability of contamination, the proportion of contamination coverage, and the classification results of the contamination level.

[0052] The YOLO-PV-Coarse detection model employs a coupling head based on SC-AKConv enhancement. It not only stabilizes the underlying spatial topological features using improved operators but also decouples the tasks along the output channel dimension of the detection head, essentially constructing three collaborative logical sensing branches: a contamination probability branch for predicting the probability of photovoltaic panel contamination. A scalar regression branch is used to predict the percentage of pollution coverage. A multi-class branch is used to predict pollution level classification. This collaborative design, which decouples operator-level spatial constraints from logic-level tasks, significantly improves the perception accuracy of complex and irregular contamination.

[0053] The core architecture of the SC-AKConv operator mainly consists of two branches working together: a branch for performing feature extraction, 3D feature sampling, and convolution, and a branch for dynamically guiding the receptive field morphology through offset prediction and constraint. Its core computational flow is as follows:

[0054] Offset Feature Map Generation: Let the input feature map be... A lightweight offset prediction convolutional layer inside SC-AKConv first processes... Generate a dynamic offset field, denoted as .in, , indicating that for the entire graph For each pixel location, the network adaptively predicts... Each sampling point is and Relative coordinate offset in two directions; Batch size; Input the number of channels; W represents the height; W represents the width. This represents the number of channels in the offset feature map.

[0055] Spatial consistency constraints are applied: this is the core difference between SC-AKConv and other methods. The generated offset field... They are simultaneously fed into the sampling branch and the spatial consistency regularization computation unit, which defines the training phase. .when When attempting to generate non-smooth, topologically torn local coordinate jumps under background noise, the regularization unit will produce a strong gradient penalty, thereby forcing the generation of non-smooth, topologically torn local coordinate jumps during training. Generate a smooth and continuous deformation field.

[0056] Adaptive Sampling and Convolution: The sampling branch utilizes an offset field optimized by spatial constraints. The bilinear interpolation algorithm is used to obtain the input from the input. Feature sampling is performed on the deformed position. Let the optimized sampling coordinates be... , Using the initial grid coordinates, the sampled features are then multiplied by the standard convolution kernel parameters to obtain the high signal-to-noise ratio feature output after SC-AKConv processing.

[0057] Training Labels and Loss Function: For each photovoltaic panel in the training data, the bounding box coordinates are labeled, and the true coverage ratio is obtained through pixel-level mask statistics. Assume the total number of pixels within a single photovoltaic panel area is... The number of pixels marked as contaminated is The actual coverage ratio Recorded as:

[0058]

[0059] Will Divide into multiple intervals and generate level labels. As a monitoring signal for pollution level branches.

[0060] The loss function consists of the following parts: bounding box regression and classification loss for "whether it is a photovoltaic panel" (the classification loss is the same as the regular YOLO). Coverage ratio regression loss is used... or , recorded as The pollution level classification loss is expressed as cross-entropy, denoted as... Spatial consistency regularization loss, denoted as .

[0061] Coverage ratio regression loss Represented as:

[0062]

[0063] in, The true label represents the coverage percentage. The predicted coverage percentage.

[0064] Pollution level classification loss Represented as:

[0065]

[0066] in, This represents the probability of the true class c.

[0067] Spatial consistency regularization loss Represented as:

[0068]

[0069] in, Indicates in batch ,aisle Spatial coordinates The predicted value of deformation offset at the location. For batch size, This represents the number of channels in the offset feature map. and These represent the height and width of the current feature map, respectively. This is a normalization coefficient used to eliminate numerical magnitude differences caused by different feature map scales and batch sizes.

[0070] To drive SC-AKConv to learn according to the expected topological smoothness and jointly optimize the upper-layer multidimensional sensing branch, this study completely reconstructs the objective function of the YOLO detection network. Total loss function. The definition is as follows:

[0071]

[0072] in, For the bounding box regression and object classification loss in the conventional detection branch, and These correspond to the coverage ratio loss and pollution level prediction loss of the newly added multidimensional coupling head branch output, respectively. , , To balance the hyperparameter weights of the various subtasks, the following is introduced: This model not only achieves collaboration among multiple tasks at the upper level (detection, regression, classification), but also completes an end-to-end joint optimization closed loop from the underlying operator space structure constraints to macroscopic multidimensional perception.

[0073] Step 4, Reasoning Process and Coarse-grained Effects.

[0074] The preprocessed aerial image is input into the YOLO-PV-Coarse model to obtain a series of photovoltaic panel candidate frames and their corresponding pollution probabilities, coverage ratios, and levels. Non-maximum suppression is applied to the candidate frames to filter out frames with large overlaps, retaining high-confidence component detection results.

[0075] For each detected photovoltaic panel: record its location in the image and its corresponding GPS coordinates; record the pollution coverage ratio estimated by the YOLO-PV-Coarse model. and level .

[0076] Based on the set thresholds, such as coverage ratio or level The photovoltaic panels are categorized into a "suspected high-pollution target list" for subsequent fine-grained detection and cleaning. Across the entire site, the pollution level of each photovoltaic panel is mapped to a color code, generating a coarse-grained pollution heat map that visually displays areas of concentrated pollution.

[0077] Based on the "list of suspected high-pollution photovoltaic panels" obtained above, this embodiment focuses on the design of a fine-grained pollution detection network and the generation and execution process of a cleaning strategy.

[0078] Step 5: Target selection and close-range acquisition task planning.

[0079] From the results obtained in step 4, photovoltaic panel IDs with moderate or severe pollution levels are selected to form a target set for fine-grained detection. These target photovoltaic panels are then clustered based on their geographical coordinates, grouping spatially close panels into the same task cluster to reduce drone flight time.

[0080] Each photovoltaic panel is equipped with 1 to 2 hovering points, located approximately 3 to 5 meters above the center of the panel; the camera's downward angle (45° to 60°) and flight path sequence are set to ensure that the photovoltaic panel occupies a large proportion of the image field of view and that reflections are controllable.

[0081] Step 6: Acquisition and preprocessing of close-up images and geometric information.

[0082] like Figure 3 As shown, the close-range drone is equipped with a high-resolution camera and can optionally be equipped with a depth sensor (such as a single-line laser or a binocular camera). The drone flies to the mission cluster area and hovers near each hovering point in sequence.

[0083] Collect one or more high-resolution RGB images at each target photovoltaic panel, denoted as: The depth sensor synchronously acquires a depth map or disparity map aligned with it, denoted as... .

[0084] The collected data is preprocessed: image sizes are scaled to the network input size and normalized. Depth maps are aligned with RGB images in pixel coordinates, and outliers are cropped and interpolated.

[0085] Step 7, fine-grained detection network design (subdivision type and thickness).

[0086] like Figure 4 As shown, in this embodiment, the fine-grained detection network is denoted as PV-FineNet, which is used to output fine contamination region mask, contamination type and contamination thickness on close-up images.

[0087] In this embodiment, the fine-grained detection network is a shared encoder + dual decoder structure; the encoder is a Unet encoder, and the decoder is a Unet decoder.

[0088] The RGB image and the depth map are concatenated channel by channel to obtain a multi-channel input tensor:

[0089]

[0090] The encoder uses a convolutional backbone network to extract multi-scale features.

[0091] The decoder section is configured with two output branches:

[0092] 1) Branching: Outputting a fine-grained binary mask for the contaminated region. and multi-type probability ,in Binary values ​​are obtained through thresholding. Multiple probability types are used to distinguish between sea salt crystals, bird droppings, oil stains, biological deposits, etc.

[0093] 2) Thickness branch: Outputs pixel-level thickness map .

[0094] By binary mask Statistical analysis of pollution coverage ratio:

[0095]

[0096] in, This represents the total number of pixels within the image area of ​​the photovoltaic panel.

[0097] Step 8: With a depth map available, estimate the contamination thickness by combining geometric constraints.

[0098] 1) Fitting the reference plane.

[0099] Using the results of the segmentation, the photovoltaic panel area is divided into clean and contaminated areas:

[0100]

[0101] On the depth data corresponding to the clean region, the least squares method is used to fit the planar model:

[0102]

[0103] in, To set pixel coordinates 3D coordinates transformed from camera intrinsic and extrinsic parameters Let be the plane parameters to be estimated.

[0104] The reference plane obtained by fitting can be represented as:

[0105]

[0106] 2) Height residual and thickness diagram.

[0107] For pixels within the contaminated area, the depth D acquired by the depth sensor is calculated. With reference plane The difference:

[0108]

[0109] To reduce the effects of noise and indentation, negative values ​​can be truncated:

[0110]

[0111] Mapping the height residual to physical thickness yields a pixel-level thickness map:

[0112]

[0113] in, , These are the proportionality coefficients and biases obtained through calibration experiments.

[0114] At the component level, the thickness of the contaminated area is statistically analyzed, and the average thickness is:

[0115]

[0116] Through the The sorting yielded the 90th percentile thickness. .

[0117] according to or With preset threshold The relationship maps the thickness to thickness levels 0 through 3:

[0118]

[0119] Thickness map can be output directly. And obtained through the above statistics It can also directly output discrete thickness level predictions.

[0120] 3) When training PV-FineNet, a multi-task joint loss is used:

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127] Where C represents the number of pollutant types, and H and W represent the height and width of the image. In the image The probability of class c at location c in the image location At this point, is the actual category class c (1 if yes, 0 otherwise); Represents pixel-level thickness map; This indicates that PV-FineNet outputs a pixel-level thickness map; Indicates the predicted component-level thickness level; The splitting loss can be composed of binary cross-entropy and Dice loss; This represents the L1 or L2 regression loss for pixel-level thickness maps; This represents the cross-entropy loss at the component-level thickness level. This is the loss weight.

[0128] Step 9: Cleaning strategy generation.

[0129] For each photovoltaic panel, the comprehensive fine-grained detection results yielded: pollution type and fine-grained pollution coverage ratio. and thickness grade Simple rule mapping functions can be constructed. Map "type + thickness" to cleaning mode parameters:

[0130]

[0131] For sea salt crystals, when the thickness level is 0–1, use medium-pressure water spray followed by a short rinse. When the thickness level is 2–3, use high-pressure spray combined with short-stroke soft brush scrubbing. For bird droppings, pre-wet with low pressure first, then remove with medium-pressure spray or soft brush. For oil stains, spray a small amount of specialized cleaning solution first, wait for a period of time, then rinse with medium pressure. For biological attachments, [details to be added]. Choose a high-pressure spray + scrubbing strategy and appropriately extend the cleaning time.

[0132] The cleaning strategy is automatically converted into drone cleaning task parameters, including the corresponding photovoltaic panel location, cleaning mode (spraying / spraying + brushing / cleaning liquid + spraying, etc.), spraying time, flight speed, nozzle angle, etc.

[0133] Step 10: Cleaning execution and effect verification.

[0134] The cleaning drone executes cleaning actions based on generated task parameters, moving along a pre-set path above the target photovoltaic panel while simultaneously controlling the spraying and scrubbing devices. After cleaning, the drone again acquires close-up images from the same location and repeatedly uses PV-FineNet for fine-grained detection to obtain the post-cleaning contamination coverage ratio. and thickness grade .

[0135] The cleaning effect indicators are defined as follows:

[0136]

[0137] like Greater than the threshold, and thickness level If the level drops below the target level, the cleaning effect is considered to be up to standard; if there is still a significant thick layer of dirt in some areas, a "secondary cleaning" task will be automatically generated for the photovoltaic panel, which can increase the spray intensity or extend the scrubbing time.

[0138] Fine-grained detection results and cleaning parameters before and after cleaning are recorded in a database for subsequent statistical analysis and model retraining. This allows for analysis of the effectiveness differences between different types of contamination and different cleaning modes, enabling gradual optimization of cleaning strategies.

[0139] like Figure 5 As shown, in the coarse-grained stage of pollution detection, this invention rapidly identifies contaminated photovoltaic panels using high-altitude images; in the fine-grained detection stage, it accurately assesses the thickness of pollutant accumulation by combining geometric thickness estimation from depth maps, thus providing precise data for cleaning tasks. By generating intelligent cleaning tasks, this method optimizes the cleaning process, improves the operating efficiency of photovoltaic power plants, reduces energy waste, and lowers maintenance costs.

[0140] Example 2

[0141] The purpose of this embodiment is to provide a pollution status sensing system for marine photovoltaic panels, including:

[0142] The coarse-grained detection module is configured to: input the collected aerial images of marine photovoltaic panels into the pollution coarse-grained detection model to obtain the pollution probability, pollution coverage ratio and pollution level of the photovoltaic panels, and filter photovoltaic panels with severe pollution levels to generate a list of suspected high-pollution targets;

[0143] The fine-grained detection module is configured to: plan the inspection path based on the list of suspected high-pollution targets, control the drone to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel, stitch the close-range RGB images and depth information together, and input them into the fine-grained pollution detection model to obtain a fine mask of the pollution area, pollution type and pixel-level thickness map.

[0144] The cleaning module is configured to: fit the reference plane of the photovoltaic panel, calculate the depth difference of the contaminated area and map it to the physical thickness, statistically analyze the average thickness and thickness level of the module, generate a cleaning strategy based on the contamination type, fine coverage ratio and thickness level, and control the drone to perform the cleaning task according to the cleaning strategy.

[0145] In further embodiments, the following is also provided:

[0146] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0147] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0148] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0149] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0150] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0151] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0152] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0153] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0154] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0155] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for sensing the pollution status of marine photovoltaic panels, characterized in that, include: The collected aerial images of marine photovoltaic panels are input into a coarse-grained pollution detection model to obtain the pollution probability, pollution coverage ratio, and pollution level of the photovoltaic panels. A list of suspected high-pollution targets is generated by screening photovoltaic panels with severe pollution levels. Based on the list of suspected high-pollution targets, the inspection route is planned, and the drone is controlled to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel. The close-range RGB images and depth information are stitched together and input into the fine-grained pollution detection model to obtain a fine mask of the polluted area, pollution type and pixel-level thickness map. A fine binary mask of the contaminated area is used to divide the photovoltaic panel area into clean and contaminated areas; On the depth data corresponding to the clean area, a reference plane is obtained by fitting a plane model using the least squares method; the difference between the actual depth and the reference plane is calculated for the pixels in the contaminated area, and the height residual is mapped to the physical thickness; the component-level average thickness and thickness level are statistically analyzed, and a cleaning strategy is generated based on the contamination type, fine coverage ratio and thickness level, and the drone is controlled to perform the cleaning task according to the cleaning strategy.

2. The method for sensing the pollution status of marine photovoltaic panels as described in claim 1, characterized in that, The coarse-grained pollution detection model uses an improved YOLO network, adding a scalar regression branch and a multi-classification branch to the original YOLO network detection head. The scalar regression branch is used to predict the pollution coverage ratio, and the multi-class branch is used to predict the pollution level.

3. The method for sensing the pollution status of marine photovoltaic panels as described in claim 1, characterized in that, It also includes preprocessing the acquired close-range RGB images and depth information. The preprocessing includes: uniformly scaling and normalizing the size of the close-range RGB images; aligning the depth information with the close-range RGB images in pixel coordinates; and cropping and interpolating outliers.

4. The method for sensing the pollution status of marine photovoltaic panels as described in claim 1, characterized in that, During training, the fine-grained pollution detection model employs a multi-task joint loss function, which includes segmentation loss, pixel-level thickness map regression loss, and component-level thickness level cross-entropy loss.

5. The method for sensing the pollution status of marine photovoltaic panels as described in claim 1, characterized in that, Also includes: After cleaning, close-up images are taken at the same location on the photovoltaic panel. A fine-grained contamination detection model is used to detect the contamination coverage ratio and thickness level after cleaning. The cleaning effect is determined based on the contamination coverage ratio and thickness level after cleaning.

6. A pollution status sensing system for marine photovoltaic panels, characterized in that, include: The coarse-grained detection module is configured to: input the collected aerial images of marine photovoltaic panels into the pollution coarse-grained detection model to obtain the pollution probability, pollution coverage ratio and pollution level of the photovoltaic panels, and filter photovoltaic panels with severe pollution levels to generate a list of suspected high-pollution targets; The fine-grained detection module is configured to: plan the inspection path based on the list of suspected high-pollution targets, control the drone to collect close-range RGB images and depth information at the hovering point of the target photovoltaic panel, stitch the close-range RGB images and depth information together, and input them into the fine-grained pollution detection model to obtain a fine mask of the pollution area, pollution type and pixel-level thickness map. The cleaning module is configured to: divide the photovoltaic panel area into clean and contaminated areas using a fine binary mask of the contaminated area; fit a reference plane to the depth data corresponding to the clean area using the least squares method; calculate the difference between the actual depth and the reference plane for pixels in the contaminated area, and map the height residual to the physical thickness; statistically analyze the component-level average thickness and thickness level, generate a cleaning strategy based on the contamination type, fine coverage ratio, and thickness level, and control the drone to perform the cleaning task according to the cleaning strategy.

7. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

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

9. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-5.

Images