A Deep Learning-Based System and Method for Dust Identification and Cleaning Assistance in Photovoltaic Panels

Abstract

This invention discloses a deep learning-based photovoltaic panel dust accumulation identification and cleaning assistance system and method, belonging to the field of photovoltaic panel dust accumulation identification technology. Specifically, it includes: acquiring surface images of the photovoltaic panel; preprocessing the acquired surface images, including noise reduction, image enhancement, and image segmentation operations; constructing a photovoltaic panel dust accumulation identification model; training the photovoltaic panel dust accumulation identification model; identifying dust accumulation on the photovoltaic panel; quantifying and classifying the dust accumulation level; obtaining the degree of dust accumulation on the photovoltaic panel; segmenting the dust accumulation area; extracting the dust accumulation area from the photovoltaic panel; dynamically planning a cleaning path based on the distribution of the dust accumulation area; and cleaning the photovoltaic panel. Through the identification and deep restoration of obstructions, it can effectively identify dust accumulation under obstructions, better handle complex environmental factors, and improve the accuracy and reliability of dust accumulation analysis.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic panel dust accumulation identification technology, specifically a photovoltaic panel dust accumulation identification and cleaning assistance system and method based on deep learning. Background Technology

[0002] Photovoltaic panels (also known as solar panels), as an important component of clean energy, are widely used in residential, industrial, and commercial sectors to convert solar energy into electricity. However, during long-term use, photovoltaic panels are susceptible to environmental factors, especially dust, rain, and wind. These factors can cause dust, sand, leaves, and other pollutants to accumulate on the surface of the panels. This contamination reduces the amount of sunlight received by the panels, thereby decreasing their energy efficiency and potentially causing permanent damage.

[0003] Traditional cleaning methods typically rely on regular manual inspections and cleaning, which is time-consuming and labor-intensive, may result in incomplete cleaning, and cannot respond promptly to changes in the dirt on the photovoltaic panel surface.

[0004] For example, patent application CN116030409A discloses a method for identifying the dust accumulation status of photovoltaic panels based on adaptive image segmentation, including the following steps: S1, image preprocessing; S2, photovoltaic string segmentation; S3, dust segmentation; S4, image block division; S5, feature extraction; S6, dust status identification. This invention considers that both the acquired and segmented images are color images. Therefore, to improve the algorithm's processing speed, the segmented images are processed into blocks. This not only improves the processing speed but also increases the number of image texture features. The LBP algorithm is then used to extract dust features, and finally, a softmax classifier is used for status identification. Identifying the dust accumulation status of photovoltaic panels can prevent hot spots and breakage, reduce photovoltaic panel losses, and increase service life, thereby maximizing the utilization rate of photovoltaic panels, reducing maintenance costs, and providing an assessment basis for the subsequent health status of the photovoltaic array.

[0005] For example, patent application CN118628791A discloses a method for identifying the degree of dust accumulation on photovoltaic panels, specifically including: step 1, image acquisition; step 2, image data preparation and preprocessing; step 3, constructing an improved DenseNet169 network model for identifying the degree of dust accumulation; step 4, training dataset; after training, the result of identifying the degree of dust accumulation on photovoltaic panels is obtained. This invention replaces the original DenseNet169 model with an improved DenseBlock module. In addition, during the training process, the Ranger optimization algorithm is used to update the model parameters, and transfer learning is introduced. The improved DenseNet169 network model improves the model accuracy while reducing the number of model parameters and floating-point computation, which is a significant improvement compared with the existing dust accumulation identification model and is more practical.

[0006] The shortcomings of the above-mentioned patents are: they are relatively crude in handling complex environments and have a large error in identifying dust accumulation under obstructions. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a deep learning-based photovoltaic panel dust accumulation identification and cleaning assistance system and method. This system can accurately identify the dust accumulation on the surface of photovoltaic panels and provide cleaning assistance suggestions based on the identification results, thereby optimizing the cleaning process of photovoltaic panels and improving the efficiency and economy of photovoltaic power generation systems.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning includes: an image acquisition module, an image preprocessing module, a dust accumulation degree analysis module, a dust accumulation area extraction module, and an intelligent cleaning module;

[0010] The image acquisition module uses a drone or ground-based camera equipment to acquire surface images of the photovoltaic panel;

[0011] The image preprocessing module is used to preprocess the acquired surface image of the photovoltaic panel, including noise reduction, image enhancement and image segmentation operations;

[0012] The dust accumulation analysis module is used to construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level.

[0013] The dust accumulation area extraction module is used to segment the dust accumulation area after obtaining the degree of dust accumulation on the photovoltaic panel, and extract the dust accumulation area in the photovoltaic panel.

[0014] The intelligent cleaning module dynamically plans a cleaning path based on the distribution of dust accumulation areas to clean the photovoltaic panels.

[0015] Specifically, the dust accumulation analysis module includes: a model training unit and a dust accumulation recognition unit;

[0016] The model training unit is used to construct and train a photovoltaic panel dust accumulation identification model.

[0017] The dust accumulation identification unit uses a trained photovoltaic panel dust accumulation identification model to identify dust accumulation on the surface of the photovoltaic panel and classify the identified dust accumulation on the photovoltaic panel surface.

[0018] Specifically, the construction and training of the photovoltaic panel dust accumulation identification model includes:

[0019] Obtain a public image dataset of dust accumulation on photovoltaic panels, label the images of dust accumulation on photovoltaic panels in the dataset with category labels, and divide the dataset into training and testing sets in a 7:3 ratio.

[0020] A photovoltaic panel dust accumulation recognition model is constructed, including: a convolutional layer for extracting local features from the photovoltaic panel dust accumulation image; a pooling layer for downsampling to reduce data dimensionality and computational cost; a fully connected layer for mapping the photovoltaic panel dust accumulation image features to the final category output; and an activation function for outputting the final classification result.

[0021] The loss function for the photovoltaic panel dust accumulation identification model is defined as follows:

[0022]

[0023] Where Γ represents the loss function of the photovoltaic panel dust accumulation identification model, C represents the number of photovoltaic panel dust accumulation categories, and yc i yc represents the true label of a photovoltaic panel dust accumulation image in the public image dataset. i ' represents the probability predicted by the photovoltaic panel dust accumulation identification model; i represents the index of the number of photovoltaic panel dust accumulation categories;

[0024] The training set is input into the photovoltaic panel dust accumulation recognition model in batches. For each batch, forward propagation is performed to calculate the predicted value, and the loss is calculated through the loss function. Through iterative training, the loss value converges or reaches the preset stopping condition, and the trained photovoltaic panel dust accumulation recognition model is obtained.

[0025] Specifically, the category labels include: no dust accumulation, light dust accumulation, moderate dust accumulation, and heavy dust accumulation.

[0026] Specifically, the process of using a trained photovoltaic panel dust accumulation recognition model to identify dust accumulation on the surface of the photovoltaic panel and classifying the identified dust accumulation includes:

[0027] The pre-processed surface image of the photovoltaic panel is converted to the HSV color space and then normalized.

[0028] A deep learning model is used to detect obstructions in the surface image of a photovoltaic panel, resulting in a set D of obstructions in the surface image, where D = {(zb1,c1,p1),(zb2,c2,p2),...,(zb...}. m ,c m ,p m )}, where zb m c represents the coordinates of the m-th obstruction in the surface image of the photovoltaic panel. m This represents the category of the m-th obstruction in the surface image of the photovoltaic panel, usually represented by a category label, p. m This represents the confidence level of the m-th obstruction in the surface image of the photovoltaic panel, i.e., the probability of it belonging to that category;

[0029] A regression model is used to predict the bounding box position of an object in the surface image of a photovoltaic panel. Combined with the surface image feature map of the photovoltaic panel extracted during detection by a deep learning model, the offset (Δx, Δy, Δw, Δh) of each occlusion candidate box is output, where Δx represents the offset of the horizontal coordinate of the occlusion candidate box, Δy represents the offset of the vertical coordinate of the occlusion candidate box, Δw represents the offset of the width of the occlusion candidate box, and Δh represents the offset of the height of the occlusion candidate box. The predicted bounding boxes are then adjusted using these offsets. The specific formula for adjusting the predicted bounding boxes is as follows:

[0030]

[0031] in, x represents the predicted bounding box location, b represents the true bounding box location (i.e., the actual location of the occluded object), and x represents the actual location of the bounding box. center ,y center w and h represent the predicted center x-coordinate, center y-coordinate, width, and height of the occluded object, respectively. w * ,h * This indicates the actual x-coordinate of the center, y-coordinate of the center, width, and height of the obstructing object.

[0032] Specifically, the step of using a trained photovoltaic panel dust accumulation recognition model to identify dust accumulation on the surface of the photovoltaic panel and classifying the identified dust accumulation also includes:

[0033] The obtained shading object prediction bounding boxes are removed from the surface image of the photovoltaic panel, and the blank areas after shading object removal are repaired using the mean fill method, resulting in the initial repaired image I of the shading object prediction bounding boxes. init (x,y) optimizes the boundary of the predicted bounding box for occlusions. The specific formula is as follows:

[0034] L boun =v (x',y')∈θ |▽I init (x',y')-▽I orig (x',y')|;

[0035] Among them, L boun This represents the boundary difference between the initial restored image of the occlusion prediction bounding box and the surface image of the photovoltaic panel, i.e., the difference in the boundary region of the occlusion prediction bounding box. (x', y') represents the pixel coordinates of the boundary region of the occlusion prediction bounding box, θ represents the boundary of the occlusion prediction bounding box, i.e., the boundary contour formed by all pixels of the restored region, and |·| represents the absolute value function. init (x',y') represents the gradient of the initial inpainted image at pixel (x',y') of the predicted bounding box of the occluded object. orig (x',y') represents the gradient of the surface image of the photovoltaic panel at pixel (x',y');

[0036] The surface image of the repaired photovoltaic panel is input into the trained photovoltaic panel dust accumulation recognition model to obtain the degree of dust accumulation on the photovoltaic panel.

[0037] Specifically, the step of segmenting the dust accumulation area and extracting the dust accumulation area from the photovoltaic panel includes:

[0038] A local adaptive threshold is set, and region segmentation is performed based on the brightness differences in the surface image of the repaired photovoltaic panel. The specific formula for the local adaptive threshold is as follows:

[0039] T(x,y)=μ(x,y)-α×σ(x,y);

[0040] Where T(x,y) represents the local adaptive threshold, i.e. the adaptive threshold of pixel (x,y), μ(x,y) represents the local average brightness of the neighborhood around pixel (x,y), σ(x,y) represents the standard deviation of the neighborhood around pixel (x,y), α represents the adjustment factor; and (x,y) represents the pixel in the surface image of the repaired photovoltaic panel.

[0041] Based on the local adaptive threshold T(x,y), the surface image of the repaired photovoltaic panel is divided into a dust accumulation area and a non-dust accumulation area. The pixel value corresponding to the dust accumulation area is greater than T(x,y), while the pixel value corresponding to the non-dust accumulation area is less than T(x,y).

[0042] Specifically, the preprocessing of the acquired photovoltaic panel surface image includes denoising, image enhancement, and image segmentation operations. Denoising uses algorithms such as Gaussian filtering to remove noise from the photovoltaic panel surface image. Image enhancement is used to adjust the brightness or contrast of the photovoltaic panel surface image. Image segmentation is used to segment the photovoltaic panel area and distinguish the background from the photovoltaic panel area.

[0043] Specifically, the intelligent cleaning module includes: an intelligent planning unit and a cleaning unit;

[0044] The intelligent planning unit dynamically plans the cleaning path based on the distribution of dust accumulation areas;

[0045] The cleaning unit cleans the surface of the photovoltaic panel according to a planned path.

[0046] A deep learning-based method for identifying and assisting in cleaning dust accumulation on photovoltaic panels, implemented based on a deep learning-based photovoltaic panel dust accumulation identification and cleaning assistance system, specifically includes the following steps:

[0047] Step 1: Acquire surface images of the photovoltaic panel;

[0048] Step 2: Preprocess the acquired surface image of the photovoltaic panel, including noise reduction, image enhancement, and image segmentation operations;

[0049] Step 3: Construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level;

[0050] Step 4: After obtaining the degree of dust accumulation on the photovoltaic panel, the dust accumulation area is segmented and extracted from the photovoltaic panel;

[0051] Step 5: Based on the distribution of dust accumulation areas, dynamically plan the cleaning path and clean the photovoltaic panels.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] 1. This invention proposes a photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning. Through the training of a deep learning model, it is possible to quantitatively assess the degree of dust accumulation on photovoltaic panels. This assessment can not only identify the location of the dust accumulation area, but also determine the severity of the dust accumulation.

[0054] 2. This invention proposes a photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning. Through the identification, processing and deep repair of obstructions, it can effectively identify dust accumulation under obstructions and better handle complex environmental factors (such as light, shadow, and type of dirt), thereby improving the accuracy and reliability of dust accumulation analysis.

[0055] 3. This invention proposes a photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning, which automatically completes dust accumulation detection and cleaning plan generation without manual intervention, greatly reducing labor costs and operational risks. At the same time, due to the personalized design of the cleaning plan, it can also reduce unnecessary resource waste and achieve more efficient and economical photovoltaic panel maintenance. Attached Figure Description

[0056] Figure 1 This invention provides an architecture diagram of a photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning.

[0057] Figure 2 The flowchart of a deep learning-based photovoltaic panel dust accumulation identification and cleaning assistance method provided by the present invention is shown. Detailed Implementation

[0058] To facilitate understanding of the technical means, creative features, objectives, and effects of this invention, it should be noted in the description that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "number one," "number two," and "number three" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The invention will be further elaborated below with reference to specific embodiments.

[0059] Example 1

[0060] Please see Figure 1 The present invention provides an embodiment of a photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning, comprising: an image acquisition module, an image preprocessing module, a dust accumulation degree analysis module, a dust accumulation area extraction module, and an intelligent cleaning module;

[0061] The image acquisition module uses a drone or ground-based camera equipment to periodically capture and photograph the surface images of the photovoltaic panel;

[0062] The image acquisition frequency and angle should be set according to the layout of the photovoltaic panel and the actual application scenario. The images should include the surface condition of the photovoltaic panel under different time periods and environmental conditions. The images should have sufficient resolution and clarity to ensure that the details of the dust can be captured.

[0063] The image preprocessing module is used to preprocess the acquired surface image of the photovoltaic panel, including operations such as noise reduction, image enhancement and image segmentation, in order to improve the accuracy of model recognition.

[0064] The denoising process uses algorithms such as Gaussian filtering to remove noise from the image and improve image quality; the image enhancement process adjusts the brightness or contrast of the image to enhance the model's adaptability to changes in lighting; the image segmentation process segments the photovoltaic panel area to distinguish the background from the photovoltaic panel area.

[0065] Image enhancement also includes: rotation, randomly rotating the image by a certain angle to simulate different shooting angles; and flipping, randomly flipping the image horizontally or vertically to increase image diversity.

[0066] The dust accumulation analysis module is used to construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level.

[0067] The dust accumulation degree analysis module includes: a model training unit and a dust accumulation recognition unit;

[0068] The model training unit is used to construct and train a photovoltaic panel dust accumulation identification model.

[0069] The dust accumulation identification unit uses a trained photovoltaic panel dust accumulation identification model to identify dust accumulation on the surface of the photovoltaic panel and classify the identified dust accumulation on the photovoltaic panel surface.

[0070] The specific steps for constructing and training a photovoltaic panel dust accumulation identification model include:

[0071] Step S1: Obtain the public image dataset of photovoltaic panel dust accumulation, label the photovoltaic panel dust accumulation images in the public image dataset with category labels, and divide the public image dataset of photovoltaic panel dust accumulation into training set and test set in a 7:3 ratio;

[0072] Step S2: Construct a photovoltaic panel dust accumulation recognition model, including: a convolutional layer for extracting local features from the photovoltaic panel dust accumulation image; a pooling layer for downsampling to reduce data dimensionality and computational cost; a fully connected layer for mapping the photovoltaic panel dust accumulation image features to the final category output; and an activation function for outputting the final classification result.

[0073] Step S3: Define the loss function for the photovoltaic panel dust accumulation identification model. The specific formula is as follows:

[0074]

[0075] Where Γ represents the loss function of the photovoltaic panel dust accumulation identification model, C represents the number of photovoltaic panel dust accumulation categories, and yci yc represents the true label of a photovoltaic panel dust accumulation image in the public image dataset. i ' represents the probability predicted by the photovoltaic panel dust accumulation identification model; i represents the index of the number of photovoltaic panel dust accumulation categories;

[0076] Step S4: Input the training set into the photovoltaic panel dust accumulation recognition model in batches. For each batch, perform forward propagation to calculate the predicted value and calculate the loss through the loss function. Iterate through training until the loss value converges or the preset stopping condition is reached to obtain the trained photovoltaic panel dust accumulation recognition model.

[0077] After training the photovoltaic panel dust accumulation recognition model, it is necessary to evaluate the model using a test set to monitor overfitting. Cross-validation can be used to further improve the model's robustness. To enhance the model's recognition performance, transfer learning strategies can be employed. This involves fine-tuning pre-trained networks (such as ResNet and VGG) and utilizing feature representations obtained from training these networks on large-scale datasets (such as ImageNet) to improve model performance even with limited sample learning.

[0078] The category labels in step S1 include: no dust accumulation, light dust accumulation, moderate dust accumulation, and heavy dust accumulation.

[0079] The process of using a trained photovoltaic panel dust accumulation recognition model to identify dust accumulation on the surface of photovoltaic panels and classifying the identified dust accumulation includes the following steps:

[0080] Step A1: Convert the preprocessed surface image of the photovoltaic panel to the HSV color space and perform normalization processing;

[0081] Step A2: Use a deep learning model to detect obstructions in the surface image of the photovoltaic panel, obtaining a set D of obstructions in the surface image of the photovoltaic panel, D = {(zb1,c1,p1),(zb2,c2,p2),...,(zb...} m ,c m ,p m )}, where zb m c represents the coordinates of the m-th obstruction in the surface image of the photovoltaic panel. m This represents the category of the m-th obstruction in the surface image of the photovoltaic panel, usually represented by a category label, p. m This represents the confidence level of the m-th obstruction in the surface image of the photovoltaic panel, i.e., the probability of it belonging to that category;

[0082] In this embodiment, the deep learning model is an object detection model, that is, a model established by an object detection algorithm. During the detection process, it is necessary to extract the feature map of the surface image of the photovoltaic panel for use in subsequent steps.

[0083] Step A3: Predict the bounding box position of the object in the surface image of the photovoltaic panel using a regression model. Combine this with the surface image feature map of the photovoltaic panel extracted by the deep learning model during detection, and output the offset (Δx, Δy, Δw, Δh) of each occlusion candidate box. Here, Δx represents the offset of the horizontal coordinate of the occlusion candidate box, Δy represents the offset of the vertical coordinate of the occlusion candidate box, Δw represents the offset of the width of the occlusion candidate box, and Δh represents the offset of the height of the occlusion candidate box. The predicted bounding boxes are then adjusted using these offsets. The specific formula for adjusting the predicted bounding boxes is as follows:

[0084]

[0085] in, x represents the predicted bounding box location, b represents the true bounding box location (i.e., the actual location of the occluded object), and x represents the actual location of the bounding box. center ,y center w and h represent the predicted center x-coordinate, center y-coordinate, width, and height of the occluded object, respectively. w * ,h * Indicates the actual x-coordinate of the center, y-coordinate of the center, width, and height of the obstruction;

[0086] Step A4: Remove the obtained shading object prediction bounding box from the surface image of the photovoltaic panel, and use the mean fill method to repair the blank area after shading object removal, to obtain the initial repaired image I of the shading object prediction bounding box. init (x,y) optimizes the boundary of the predicted bounding box for occlusions. The specific formula is as follows:

[0087] L boun =∑ (x',y')∈θ |▽I init (x',y')-▽I orig (x',y')|;

[0088] Among them, L boun This represents the boundary difference between the initial restored image of the occlusion prediction bounding box and the surface image of the photovoltaic panel, i.e., the difference in the boundary region of the occlusion prediction bounding box. (x', y') represents the pixel coordinates of the boundary region of the occlusion prediction bounding box, and θ represents the boundary of the occlusion prediction bounding box, i.e., the boundary contour formed by all pixels of the restored region. Typically, the restored region is identified by a mask image M, and the boundary is the boundary line from the restored region to the non-restored region in the mask image. |·| represents the absolute value function, ▽I init (x',y') represents the gradient of the initial inpainted image at pixel (x',y') of the predicted bounding box of the occluded object. orig(x',y') represents the gradient of the surface image of the photovoltaic panel at pixel (x',y');

[0089] In this embodiment, ▽I init (x', y') represents the intensity change at that location, typically reflecting important features such as texture and edges in the image. orig (x',y') represents the variations in texture and edges in the surface image of the photovoltaic panel, typically used to preserve image detail. boun The repair effect is measured by calculating the gradient difference between the repair area and the surface image of the photovoltaic panel at the boundary location. This is used to minimize the difference between the gradient of the repair area and the gradient of the surface image of the photovoltaic panel, making the transition of the repair area smoother and more natural.

[0090] The basic principle of boundary optimization is to guide the filling of the restoration area through gradient difference, making the transition between the restoration area and the surrounding area smoother and more natural. In image restoration, the boundary between the restoration area and the unoccluded area in the original image is a crucial visual element. If the boundary between the restoration area and the surrounding area is too abrupt, it may lead to an unnatural restoration result and obvious artifacts. The gradient reflects the rate of change in an image, i.e., the degree of change in image brightness or color. In natural images, boundaries usually correspond to important structural information, such as object edges and texture variations. If the boundary of the restoration area differs significantly in gradient from the boundary of the original image, the restored part will appear very abrupt. Therefore, by minimizing the gradient difference, the boundary of the restoration area can be made smoother, making the restored image more consistent with the original image.

[0091] For example, suppose there is an area of ​​obscured leaves in the surface image of a photovoltaic panel, and this area is filled in by a restoration algorithm. During the restoration process, the transition between the restored area and the surrounding photovoltaic panel background is crucial. By minimizing gradient differences, the restored leaf area will not produce obvious jagged edges or unnatural texture transitions at the boundary with the background, but will instead blend seamlessly with the background, giving a continuous visual effect.

[0092] Step A5: Input the repaired surface image of the photovoltaic panel into the trained photovoltaic panel dust accumulation recognition model to obtain the degree of dust accumulation on the photovoltaic panel.

[0093] The dust accumulation area extraction module is used to segment the dust accumulation area after obtaining the degree of dust accumulation on the photovoltaic panel, and extract the dust accumulation area in the photovoltaic panel.

[0094] The specific steps for segmenting and extracting the dust accumulation area from the photovoltaic panel include:

[0095] Step B1: Set a local adaptive threshold and perform region segmentation based on the brightness differences in the surface image of the repaired photovoltaic panel. The specific formula for the local adaptive threshold is:

[0096] T(x,y)=μ(x,y)-α×σ(x,y);

[0097] Where T(x,y) represents the local adaptive threshold, i.e., the adaptive threshold of pixel (x,y), μ(x,y) represents the local average brightness of the neighborhood around pixel (x,y), σ(x,y) represents the standard deviation of the neighborhood around pixel (x,y), α represents the adjustment factor used to control the sensitivity of the threshold, which determines the distance between the threshold and the local average brightness μ(x,y) under a given standard deviation σ(x,y). Adjusting α can balance the segmentation accuracy between the gray area and the non-gray area; (x,y) represents the pixel in the surface image of the repaired photovoltaic panel.

[0098] For example, if α is large, the threshold T(x,y) will be further away from the average brightness μ(x,y), which may result in the segmentation of too many dust accumulation areas; conversely, if α is small, some dust accumulation areas may be missed.

[0099] The surrounding neighborhood is typically a window (such as a 3×3 or 5×5 pixel area), within which the average brightness is calculated. By calculating the local mean, the system can adapt to brightness variations under different lighting conditions in the image, reducing the errors that may be introduced by the global thresholding method.

[0100] The standard deviation of the neighborhood around a pixel (x,y) describes the dispersion of brightness in that area. The larger the standard deviation, the greater the brightness variation in that area, and vice versa. The standard deviation can capture texture variations in an image and help distinguish between gray areas (usually with small brightness variations) and non-gray areas (with large brightness variations).

[0101] The brightness here refers to the grayscale value of the restored photovoltaic panel surface image;

[0102] Step B2: Based on the local adaptive threshold T(x,y), the surface image of the repaired photovoltaic panel is divided into a dust accumulation area and a non-dust accumulation area. The pixel value corresponding to the dust accumulation area is greater than T(x,y), while the pixel value corresponding to the non-dust accumulation area is less than T(x,y).

[0103] The intelligent cleaning module dynamically plans a cleaning path based on the distribution of dust accumulation areas to clean the photovoltaic panels.

[0104] The intelligent cleaning module includes: an intelligent planning unit and a cleaning unit;

[0105] The intelligent planning unit dynamically plans the cleaning path based on the distribution of dust accumulation areas;

[0106] The cleaning unit cleans the surface of the photovoltaic panel according to a planned path.

[0107] Example 2

[0108] Please see Figure 2 Another embodiment of the present invention provides: a method for identifying and assisting in cleaning dust accumulation on photovoltaic panels based on deep learning, the specific steps of which include:

[0109] Step 1: Acquire surface images of the photovoltaic panel;

[0110] Step 2: Preprocess the acquired surface image of the photovoltaic panel, including noise reduction, image enhancement, and image segmentation operations;

[0111] Step 3: Construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level;

[0112] Step 4: After obtaining the degree of dust accumulation on the photovoltaic panel, the dust accumulation area is segmented and extracted from the photovoltaic panel;

[0113] Step 5: Based on the distribution of dust accumulation areas, dynamically plan the cleaning path and clean the photovoltaic panels.

[0114] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0115] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning, characterized in that, include: Image acquisition module, image preprocessing module, dust accumulation degree analysis module, dust accumulation area extraction module, and intelligent cleaning module; The image acquisition module uses a drone or ground-based camera equipment to acquire surface images of the photovoltaic panel; The image preprocessing module is used to preprocess the acquired surface image of the photovoltaic panel, including noise reduction, image enhancement and image segmentation operations; The dust accumulation analysis module is used to construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level. The dust accumulation area extraction module is used to segment the dust accumulation area after obtaining the degree of dust accumulation on the photovoltaic panel, and extract the dust accumulation area in the photovoltaic panel. The intelligent cleaning module dynamically plans the cleaning path based on the distribution of dust accumulation areas to clean the photovoltaic panels; The dust accumulation degree analysis module includes: a model training unit and a dust accumulation recognition unit; The model training unit is used to construct and train a photovoltaic panel dust accumulation identification model. The dust accumulation identification unit uses a trained photovoltaic panel dust accumulation identification model to identify dust accumulation on the surface of the photovoltaic panel and classify the identified dust accumulation on the surface of the photovoltaic panel. The process involves using a trained photovoltaic panel dust accumulation recognition model to identify dust accumulation on the surface of photovoltaic panels and classifying the identified dust accumulation, including: The pre-processed surface image of the photovoltaic panel is converted to the HSV color space and then normalized. A deep learning model is used to detect obstructions in the surface image of a photovoltaic panel, resulting in a set D of obstructions in the surface image, where D = {(zb1,c1,p1),(zb2,c2,p2),...,(zb...}. m ,c m ,p m )}, where zb m c represents the coordinates of the m-th obstruction in the surface image of the photovoltaic panel. m p represents the category of the m-th obstruction in the surface image of the photovoltaic panel. m This represents the confidence level of the m-th obstruction in the surface image of the photovoltaic panel; A regression model is used to predict the bounding box position of objects in the surface image of a photovoltaic panel. Combined with the surface image feature map of the photovoltaic panel extracted by the deep learning model during detection, the offset of each occlusion candidate box is output. ,in, This represents the offset of the horizontal coordinate of the candidate bounding box for the occluded object. This represents the offset of the ordinate of the candidate bounding box for the occluded object. This represents the offset of the candidate bounding box width for occluders. This represents the offset of the candidate bounding box height, and the predicted bounding box is adjusted using this offset. The specific formula for adjusting the predicted bounding box is as follows: ， ， in, 'b' represents the predicted bounding box location, and 'b' represents the actual bounding box location, i.e., the actual location of the occluding object. These represent the predicted center x-coordinate, center y-coordinate, width, and height of the occluded object, respectively. Indicates the actual x-coordinate of the center, y-coordinate of the center, width, and height of the obstruction; The obtained shading object prediction bounding boxes are removed from the surface image of the photovoltaic panel, and the blank areas after shading object removal are repaired using the mean fill method, resulting in the initial repaired image I of the shading object prediction bounding boxes. init (x,y) optimizes the boundary of the predicted bounding box for occlusions. The specific formula is as follows: ， Among them, L boun This represents the boundary difference between the initial restored image of the predicted bounding box of the occluder and the surface image of the photovoltaic panel, i.e., the difference in the boundary region of the predicted bounding box of the occluder. This represents the pixel coordinates of the predicted bounding box region of the occluded object. This represents the boundary of the predicted bounding box of the occluded object, that is, the boundary contour formed by all the pixels of the repaired region. This represents the function for finding the absolute value. The initial inpainted image representing the predicted bounding box of the occluded object in pixels gradient at, This indicates the surface image of the photovoltaic panel in pixels. gradient at; The surface image of the repaired photovoltaic panel is input into the trained photovoltaic panel dust accumulation recognition model to obtain the degree of dust accumulation on the photovoltaic panel. The step of segmenting the dust accumulation area and extracting the dust accumulation area from the photovoltaic panel includes: A local adaptive threshold is set, and region segmentation is performed based on the brightness differences in the surface image of the repaired photovoltaic panel. The specific formula for the local adaptive threshold is as follows: ， in, This represents the local adaptive threshold, i.e., the adaptive threshold for pixel (x, y). This represents the local average brightness of the neighborhood surrounding a pixel (x, y). This represents the standard deviation of the neighborhood surrounding a pixel (x, y). (x,y) represents the adjustment factor; (x,y) represents the pixels in the surface image of the repaired photovoltaic panel. Based on local adaptive threshold The surface image of the repaired photovoltaic panel was divided into dust-accumulated areas and non-dust-accumulated areas. The pixel value corresponding to the dust-accumulated area was greater than [value missing]. The non-dust accumulation area is smaller. .

2. The photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning as described in claim 1, characterized in that, The construction and training of the photovoltaic panel dust accumulation identification model includes: Obtain a public image dataset of dust accumulation on photovoltaic panels, label the images of dust accumulation on photovoltaic panels in the dataset with category labels, and divide the dataset into training and testing sets in a 7:3 ratio. A photovoltaic panel dust accumulation recognition model is constructed, including: a convolutional layer for extracting local features from the photovoltaic panel dust accumulation image; a pooling layer for downsampling to reduce data dimensionality and computational cost; a fully connected layer for mapping the photovoltaic panel dust accumulation image features to the final category output; and an activation function for outputting the final classification result. The loss function for the photovoltaic panel dust accumulation identification model is defined as follows: ， in, Let yc represent the loss function of the photovoltaic panel dust accumulation identification model, where C represents the number of photovoltaic panel dust accumulation categories. i This represents the true label of the photovoltaic panel dust accumulation images in the public image dataset. This represents the probability predicted by the photovoltaic panel dust accumulation identification model; i represents the index of the number of photovoltaic panel dust accumulation categories. The training set is input into the photovoltaic panel dust accumulation recognition model in batches. For each batch, forward propagation is performed to calculate the predicted value, and the loss is calculated through the loss function. Through iterative training, the loss value converges or reaches the preset stopping condition, and the trained photovoltaic panel dust accumulation recognition model is obtained.

3. The photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning as described in claim 2, characterized in that, The category labels include: no dust accumulation, light dust accumulation, moderate dust accumulation, and heavy dust accumulation.

4. The photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning as described in claim 3, characterized in that, The preprocessing of the acquired surface image of the photovoltaic panel includes denoising, image enhancement and image segmentation operations. For denoising, a Gaussian filtering algorithm is used to remove noise from the surface image of the photovoltaic panel. Image enhancement is used to adjust the brightness or contrast of the surface image of the photovoltaic panel; image segmentation is used to segment the photovoltaic panel area and distinguish it from the background.

5. The photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning as described in claim 4, characterized in that, The intelligent cleaning module includes: an intelligent planning unit and a cleaning unit; The intelligent planning unit dynamically plans the cleaning path based on the distribution of dust accumulation areas; The cleaning unit cleans the surface of the photovoltaic panel according to a planned path.

6. A method for identifying and assisting in cleaning dust accumulation on photovoltaic panels based on deep learning, implemented based on the photovoltaic panel dust accumulation identification and cleaning assistance system based on deep learning as described in any one of claims 1-5, characterized in that, Specifically, the following steps are included: Step 1: Acquire surface images of the photovoltaic panel; Step 2: Preprocess the acquired surface image of the photovoltaic panel, including noise reduction, image enhancement, and image segmentation operations; Step 3: Construct a photovoltaic panel dust accumulation identification model, train the photovoltaic panel dust accumulation identification model, identify the dust accumulation on the photovoltaic panel, and quantify and classify the dust accumulation level; Step 4: After obtaining the degree of dust accumulation on the photovoltaic panel, the dust accumulation area is segmented and extracted from the photovoltaic panel; Step 5: Based on the distribution of dust accumulation areas, dynamically plan the cleaning path and clean the photovoltaic panels.

Images