A control method for photovoltaic cleaning robot differential execution with closed-loop logic

Abstract

This invention discloses a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, relating to the field of photovoltaic cleaning. The method includes: acquiring a surface image of a photovoltaic panel in a test area; inputting the photovoltaic panel surface image into a pre-trained YOLO target detection model to identify the location and type of contaminants on the photovoltaic panel surface and obtaining target detection results; dividing the photovoltaic panel surface image into multiple grids and determining the coverage rate of contaminants within a single grid based on the target detection results; classifying the contamination level of contaminants into multiple levels based on the coverage rate, and the photovoltaic cleaning robot performing corresponding cleaning operations for contaminants of different contamination levels; acquiring a cleaned photovoltaic surface image and again using the pre-trained YOLO target detection model to determine if residual contaminants exist; if so, the photovoltaic cleaning robot remains in the current area and re-executes the contaminant identification and cleaning operations; otherwise, the cleaning of the current area is completed.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic cleaning, and in particular to a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic. Background Technology

[0003] Currently, some technologies have attempted to automate the cleaning of photovoltaic panels. For example, Chinese patent application CN116094448A discloses a ground-walking photovoltaic panel cleaning and inspection robot system. This system can inspect the panel surface after completing the cleaning process to determine whether it needs to be cleaned again. However, its working mode is relatively passive: its vision system is only used for post-cleaning effect evaluation and cannot achieve precise operation based on the actual soiling condition. This cleaning mode, lacking recognition and classification capabilities, often leads to over-cleaning of slightly polluted areas and under-cleaning of heavily polluted areas, resulting in low overall cleaning efficiency. Summary of the Invention

[0004] This invention provides a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, to solve the aforementioned problems in the prior art, namely, how to effectively improve the cleaning efficiency of photovoltaic panel surfaces. The method includes: A photovoltaic cleaning robot was used to obtain images of the surface of the photovoltaic panel in the area to be tested before cleaning. Based on the constructed pollutant object library, diverse training samples are generated. The constructed YOLO target detection model is trained using these diverse training samples. The surface image of the photovoltaic panel to be tested before cleaning is input into the pre-trained YOLO target detection model to identify the location and type of pollutants on the photovoltaic panel surface and obtain target detection results. The surface image of the photovoltaic panel to be tested before cleaning is divided into multiple grids. Based on the target detection results, the proportion of pollutant pixels in each grid to the total number of pixels in the grid is determined to obtain the coverage rate corresponding to the type of pollutant in a single grid. Based on coverage, the degree of pollution of pollutants is divided into multiple pollution levels, and the photovoltaic cleaning robot performs corresponding cleaning operations for pollutants of different pollution levels on the surface of photovoltaic panels. After acquiring an image of the cleaned photovoltaic surface, the pre-trained YOLO object detection model is used again to determine whether there are any residual pollutants. If so, the photovoltaic cleaning robot stays in the current area, re-executes pollutant identification, and performs corresponding cleaning operations for pollutants of different pollution levels; otherwise, the cleaning of the current area is completed.

[0005] Optionally, the generation of diverse training samples based on the constructed pollutant object library specifically includes: Construct an object library containing specific pollutant samples; select pollutant samples at a preset ratio from the object library, and perform random scaling and rotation operations on each pollutant sample to determine the transformed pollutant sample; The transformed pollutant sample is pasted to a random position in the current training image, and then synthesized with the background image using a weighted fusion technique to generate a synthesized image and corresponding annotation information. The synthesized image and corresponding annotation information are used as diversified training samples.

[0006] Optionally, the composite image and the background image are synthesized using a weighted fusion technique to generate a composite image and corresponding annotation information; wherein obtaining the composite image specifically includes: roi= ; Where roi represents the generated synthetic image, src1 represents the background image, and src2 represents the transformed contaminant sample image. and These represent the contribution of the background image and the contribution of the pollutant sample image, respectively. This is a correction value; The annotation information includes the coordinates of the new bounding box at the location where the contaminant sample is pasted. The coordinates of the new bounding box are obtained using the following formula: ; Where, paste_x is the pixel coordinate of the left boundary of the paste location of the contaminant sample; paste_w is the pixel width of the contaminant sample; paste_y is the pixel coordinate of the upper boundary of the paste location of the contaminant sample; paste_h is the pixel height of the contaminant sample; image.shape[1] is the total width of the background image; image.shape[0] is the total height of the background image; new_x_center and new_y_center are the x and y coordinates of the new relative position of the center point of the contaminant sample, respectively; new_width and new_height are the new relative width and relative height of the contaminant sample.

[0007] Optionally, the step of selecting pollutant samples at a preset ratio from the object library and performing random scaling and rotation operations on each pollutant sample specifically includes: Select pollutant samples at a preset ratio from the object library, and determine the pixel coordinates of the pollutant samples using the following formula: Where (x1, y1) are the pixel coordinates of the upper left corner of the contaminant sample bounding box, (x2, y2) are the pixel coordinates of the lower right corner of the contaminant sample bounding box, x_center and y_center are the relative positions of the center point of the contaminant sample, width and height are the relative width and relative height of the contaminant sample, and w and h are the actual width and actual height of the image, respectively. Based on the pixel coordinates of the pollutant samples, each pollutant sample is randomly scaled and rotated.

[0008] Optionally, the step of dividing the photovoltaic panel surface image of the area to be tested before cleaning into multiple grids, and combining the target detection results to determine the proportion of the number of pollutant grids to the total number of grids, in order to obtain the pollutant coverage, specifically includes: The pre-trained convolutional neural network (CNN) divides the photovoltaic panel surface image into 64 grids in 8 rows and 8 columns. Combined with the target detection results, each grid in the image is classified to determine the number of grids containing pollutants. The single-cell coverage coefficient is determined based on the proportion of pollutant cells to the total number of cells. After assigning values, the coverage C is obtained using the following formula based on the assigned grid: ; Where i and j are the single-cell coverage coefficients, respectively. The row and column numbers.

[0009] Optionally, the classification of pollutant pollution levels into multiple levels based on coverage includes: When the coverage rate C is less than or equal to 20%, it is considered light pollution; when the coverage rate C is between 20% and 60%, it is considered moderate pollution; and when the coverage rate C is greater than 60%, it is considered heavy pollution.

[0010] Optionally, when the number of cleaning operations is greater than or equal to 3, the current photovoltaic panel surface image is uploaded to the cloud database for subsequent strategy optimization.

[0011] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention provides a control method for differential execution of photovoltaic cleaning robots with closed-loop logic. This method generates a massive amount of diverse training samples through a constructed pollutant object library, improving the generalization ability and recognition accuracy of the YOLO object detection model under complex lighting and varying pollution patterns. Furthermore, based on the acquired coverage of pollutants within a single cell, the pollution level of pollutants is divided into multiple levels. The photovoltaic cleaning robot performs corresponding cleaning operations for pollutants of different pollution levels on the photovoltaic panel surface, ensuring the targeting and effectiveness of the cleaning strategy, thereby effectively improving the cleaning efficiency of the photovoltaic panel surface. Finally, after the cleaning operation is completed, this invention again uses a visual sensing module to acquire images of the photovoltaic panel surface and uses the YOLO object detection model to analyze the images to determine whether there are residual pollutants. If there are no residual pollutants, the cleaning is confirmed to be complete; if residual pollutants still exist, the cleaning operation is re-executed, forming a closed-loop control mechanism that ensures the integrity and effectiveness of the cleaning operation and effectively improves the cleaning efficiency of pollutants. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0013] Figure 1 A flowchart illustrating a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, provided as an embodiment of the present invention; Figure 2 This is a flowchart of the "online copy-paste data enhancement" module provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the implementation of the differential adjustment cleaning strategy provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0015] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart of a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, provided by an embodiment of the present invention. Figure 1 As shown in the figure, this embodiment illustrates a control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, including: S1: Obtain an image of the photovoltaic panel surface in the area to be tested before cleaning using a photovoltaic cleaning robot.

[0017] For example, contaminants in the obtained image of the photovoltaic panel surface of the area to be tested before cleaning may include, for example, “sand,” “dust,” or “bird droppings.”

[0018] S2: Based on the constructed pollutant object library, diverse training samples are generated. The constructed YOLO target detection model is trained using these diverse training samples. The surface image of the photovoltaic panel to be tested before cleaning is input into the pre-trained YOLO target detection model to identify the location and type of pollutants on the photovoltaic panel surface and obtain the target detection results. The surface image of the photovoltaic panel to be tested before cleaning is divided into multiple grids. Based on the target detection results, the proportion of pollutant pixels in each grid to the total number of pixels in the grid is determined to obtain the coverage rate corresponding to the type of pollutant in a single grid.

[0019] For example, this invention uses a pre-trained YOLO target detection model to identify the location and type of contaminants on the photovoltaic panel surface based on real-time images of the photovoltaic panel surface acquired before cleaning operations. This model is an advanced convolutional neural network that processes the entire photovoltaic panel surface image as a single input. After convolution operations and feature extraction within the model, it can directly output the category and location information of all targets of interest in the image end-to-end.

[0020] For example, the YOLO object detection model can be the YOLOv8 model. As an advanced convolutional neural network, YOLOv8 can process the entire photovoltaic panel surface image as a single input. After multi-layer convolution operations and feature extraction within the model, it can directly output the category and location information of all targets of interest in the image end-to-end.

[0021] The "online copy-paste data augmentation" module proposed in this invention generates diverse training samples to train the recognition model. This results in a final model with significantly higher robustness to recognition under complex lighting conditions and varying pollution patterns (especially bird droppings) compared to conventionally trained models. The model's data processing, especially during the training phase, can be found in [reference needed]. Figure 2 (The flowchart for copying and pasting the data augmentation module online) will be explained in detail, and its core steps include:

[0022] Offline Preparation: First, based on the annotation information in the training dataset (e.g., YOLO format class_id, x_center, y_center, width, height), one or more object libraries containing specific pollutant samples (e.g., "bird droppings") are constructed offline. This process involves calculating the absolute pixel coordinates (x1, y1, x2, y2) of each bounding box, "copying" and extracting the corresponding pollutant image patches from the original training images, and using them as material for subsequent synthesis.

[0023] Online Training Phase: During model training and data loading in batches, the following online data augmentation process is triggered with a preset probability (e.g., p>=0.75): (1) Sample transformation: Randomly select a pollutant sample from the object library and apply a series of random geometric and optical transformations to it, such as random scaling, rotation, etc.

[0024] (2) Dynamic synthesis: The transformed pollutant sample is pasted into a random position on a training image in the current batch using a weighted fusion technique (e.g., using the cv2.addWeighted function). The weighted fusion technique (roi= It can simulate a semi-transparent overlay effect to generate more realistic composite images.

[0025] (3) Synchronous label generation: While completing the image synthesis, the new normalized label information (new_x_center, new_y_center, new_width, new_height) of the newly added pollutant in the synthesized image is calculated in real time and dynamically according to the pasting position, sample size and background image size, and then added to the label list of the current training image.

[0026] Finally, the image data and updated label data after the above online enhancement processing are fed into the YOLOv8 architecture model for training. By dynamically generating massive amounts of diverse training samples during the training process, the model is able to learn the characteristics of pollutants under various backgrounds, lighting conditions, and poses, thereby achieving excellent generalization ability and recognition accuracy.

[0027] The model outputs detection results in the standard YOLO format. Each detection result includes: location coordinates, which are the precise coordinates that define the bounding box of the pollutant; and category labels, which identify the preset category to which the pollutant within the bounding box belongs, such as "sand", "dust" or "bird droppings".

[0028] Specifically, the model outputs one or more detection results, each containing: location coordinates, defining the precise coordinates of the contaminant's bounding box, specifying its exact location on the photovoltaic panel; and a category label, identifying which preset category the contaminant within the bounding box belongs to, such as "sand," "dust," or "bird droppings." This identifies the type of contaminant and its spatial location on the photovoltaic panel.

[0029] For example, diverse training samples can be generated based on the constructed pollutant object library, specifically including: Construct an object library containing specific pollutant samples; select pollutant samples at a preset ratio from the object library, and perform random scaling and rotation operations on each pollutant sample to determine the transformed pollutant sample; The transformed pollutant sample is pasted to a random position in the current training image, and then synthesized with the background image using a weighted fusion technique to generate a synthesized image and corresponding annotation information. The synthesized image and corresponding annotation information are used as diversified training samples.

[0030] For example, such as Figure 2 As shown, firstly, based on the annotation information in the training dataset, an object library containing samples of specific pollutants is constructed offline. The pollutants are "copied" from the images, and the pixel coordinates of the top-left corner (x1, y1) and bottom-right corner (x2, y2) of the object's bounding box are calculated. ; Where (x1, y1) are the pixel coordinates of the upper left corner of the contaminant sample bounding box, (x2, y2) are the pixel coordinates of the lower right corner of the contaminant sample bounding box, x_center and y_center are the relative positions of the center point of the contaminant sample, width and height are the relative width and relative height of the contaminant sample, and w and h are the actual width and actual height of the image, respectively. Then, during the data loading phase of model training, enhancement is triggered intermittently: a sample is randomly selected from the object library and subjected to random scaling, rotation, and other transformations; further, it is pasted into a random position on the current training image, and then synthesized with the background image using a weighted fusion technique, wherein the weighted fusion technique is as follows: roi= ; The above formula is applied to each corresponding pixel in the transformed pollutant sample image and the background image, where roi represents the region of interest in the background image that will be covered, i.e., the generated composite image, src1 represents the background image, src2 represents the transformed pollutant sample image, and the weight is 1 ( ) and weight 2 ( ) represent the contributions of the background image and the pollutant sample image, respectively, and satisfy the following conditions: If the grid texture of the underlying photovoltaic panel needs to be preserved, the background image weight is 1 ( The value should be greater than 0.5; if the occlusion effect of the pollutant itself needs to be emphasized, the weight of the pollutant sample image should be 2( Values ​​greater than 0.5; correction value The positive and negative values ​​represent the increase or decrease in the overall brightness of the final output image. Positive values ​​simulate pollutants under strong light / reflection; negative values ​​simulate pollutants under weak light / shadow. Taking bird droppings as an example, the following is defined in the module: ; Where, roc represents src1 as the current background image; guano represents src2 as the bird droppings image; if the weight is 1 ( ) and weight 2 ( Both are 0.5, meaning that in the final blending result, the contribution of the current background image (roc) and the contribution of the bird droppings image (guano) are both 50%; the correction value ( This can be used to brighten or darken the blended image as a whole. If you do not need to adjust the brightness, set it to 0. Finally, new annotation information is generated synchronously. The new normalized annotation information can be obtained using the following formula: ; Where, paste_x is the pixel coordinate of the left boundary of the paste location of the contaminant sample; paste_w is the pixel width of the contaminant sample; paste_y is the pixel coordinate of the upper boundary of the paste location of the contaminant sample; paste_h is the pixel height of the contaminant sample; image.shape[1] is the total width of the background image; image.shape[0] is the total height of the background image; new_x_center and new_y_center are the x and y coordinates of the new relative position of the center point of the contaminant sample, respectively; new_width and new_height are the new relative width and relative height of the contaminant sample.

[0031] S3: Based on coverage, the pollution level of pollutants is divided into multiple pollution levels, and the photovoltaic cleaning robot performs corresponding cleaning operations for pollutants of different pollution levels on the surface of the photovoltaic panel.

[0032] For example, after obtaining the specific location and label type of pollutants on the photovoltaic panel, pollutants such as sand and dust are classified into three levels: light, moderate, and heavy, based on their overall coverage.

[0033] Specifically, for different types of pollutants identified on the surface of photovoltaic panels, a pixel-level coverage statistical algorithm based on target detection results is used to classify the degree of pollution. The specific method is as follows: For sand and dust pollutants, the acquired image data can be divided into 64 grids of 8 rows and 8 columns. Within each grid, an image recognition algorithm, based on the trained convolutional neural network (CNN), performs semantic segmentation to classify each pixel in the image, thereby identifying and counting the number of pixels containing sand and dust pollutants. The coverage coefficient of each grid cell is then calculated based on the proportion of sand and dust pollutant pixels to the total number of pixels in the grid. Assign the following values ​​to a grid with a pixel percentage of (0%, 10%): Assign a value of 0; for grids with pixel percentages in the (10%, 40%) range, then... Assign a value of 1; for grids with pixel percentages in the range of (40%, 70%), then... Assign a value of 2; for grids with pixel percentages in the (70%, 100%) range, then... Assign the value 3; after assigning values ​​to 64 grid cells, an 8×8 number matrix is ​​generated. Then, sum all the element values ​​in the matrix and divide by 192 to obtain an overall coverage C of 0-1, which is calculated using the following formula: C= ; If the coverage rate C is 0%-20%, the model identifies the pollutants as "light pollution"; if the coverage rate C is 20%-60%, the model identifies the pollutants as "moderate pollution"; and if the coverage rate C is 60%-100%, the model identifies the pollutants as "heavy pollution".

[0034] For example, after identifying the location and type of pollutants using the YOLO target detection model, this invention combines the grid division results to determine the number of pollutant pixels within each grid and calculates the pollutant coverage rate. When the pollutant coverage rate C within a grid is less than or equal to 20%, that grid is determined to be a lightly polluted area.

[0035] like Figure 3As shown, for sandy pollutants, the system classifies them into three levels: light, moderate, and heavy, based on real-time detection data. To avoid excessive brushing force that could scratch the photovoltaic panel surface, the robot employs a cleaning strategy of constant brushing force but varying the number of brush strokes for each level of sand contamination. For example, "light contamination" is brushed 1-2 times, "moderate contamination" 3-4 times, and "heavy contamination" 5-6 times, thus achieving an optimal balance between cleaning effectiveness and resource consumption.

[0036] For example, when dealing with bird droppings, once the model identifies the presence of such pollutants through feature recognition, it will not classify them into different levels, but will directly trigger a preset cleaning mode to perform the cleaning task. For example, for bird droppings, a cleaning mode that includes soaking, decomposition, and removal can be executed: First, the nozzle is used to evenly soak the polluted area with a constant low-pressure water flow to soften the surface scale; then, surfactants or special cleaning agents are sprayed to chemically decompose the dirt to reduce its adhesion; finally, the cleaning tool is switched to high-power mode to perform a one-time high-intensity cleaning of the decomposed dirt. For sand and soil pollutants with different levels of pollution, the robot adopts a cleaning strategy of constant brushing force and differentiated brushing frequency. For each level of sand and soil pollution, a corresponding number of brushing times is set to carry out differentiated cleaning work. For example, under different levels of pollution, the robot adopts a composite cleaning process of "rinsing first and then brushing": first, the nozzle is used to rinse the surface with a constant water flow to remove residual particles, and then the cleaning force and cleaning power of the cleaning tools are set for each level of dust pollution to carry out differentiated cleaning work.

[0037] S4: Obtain an image of the cleaned photovoltaic surface, and use the pre-trained YOLO object detection model again to determine whether there are residual pollutants. If so, the photovoltaic cleaning robot stays in the current area, re-executes pollutant identification, and performs corresponding cleaning operations for pollutants of different pollution levels; otherwise, the cleaning of the current area is completed.

[0038] For example, after acquiring an image of the photovoltaic panel surface after cleaning, the image can be analyzed again using a pre-trained YOLO object detection model to determine whether there are any residual pollutants. If there are no residual pollutants, the current area is confirmed to be cleaned, and the robot moves to the next area for cleaning. If there are still residual pollutants, the robot stays in the current area for secondary identification, decision-making, and cleaning. If there are special pollutants that cannot be cleaned even after three cleaning processes, the image information of the pollutants is acquired through the visual sensing module and uploaded to the cloud database for subsequent analysis and optimization strategies. The robot then moves to the next area for pollutant identification and cleaning.

[0039] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A control method for differential execution of a photovoltaic cleaning robot with closed-loop logic, characterized in that, include: A photovoltaic cleaning robot was used to obtain images of the photovoltaic panel surface in the area to be tested before cleaning. Based on the constructed pollutant object library, diverse training samples are generated. The constructed YOLO target detection model is trained using these diverse training samples. The surface image of the photovoltaic panel to be tested before cleaning is input into the pre-trained YOLO target detection model to identify the location and type of pollutants on the photovoltaic panel surface and obtain target detection results. The surface image of the photovoltaic panel to be tested before cleaning is divided into multiple grids. Based on the target detection results, the proportion of pollutant pixels in each grid to the total number of pixels in the grid is determined to obtain the coverage rate corresponding to the type of pollutant in a single grid. Based on coverage, the degree of pollution of pollutants is divided into multiple pollution levels, and the photovoltaic cleaning robot performs corresponding cleaning operations for pollutants of different pollution levels on the surface of photovoltaic panels. After acquiring an image of the cleaned photovoltaic surface, the pre-trained YOLO object detection model is used again to determine whether there are any residual pollutants. If so, the photovoltaic cleaning robot stays in the area to be tested, re-executes pollutant identification, and performs corresponding cleaning operations for pollutants of different pollution levels; otherwise, the cleaning of the current area is completed.

2. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 1, characterized in that, The constructed pollutant object library generates diverse training samples, specifically including: Construct an object library containing specific pollutant samples; select pollutant samples at a preset ratio from the object library, and perform random scaling and rotation operations on each pollutant sample to determine the transformed pollutant sample; The transformed pollutant sample is pasted to a random position in the current training image, and then synthesized with the background image using a weighted fusion technique to generate a synthesized image and corresponding annotation information. The synthesized image and corresponding annotation information are used as diversified training samples.

3. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 2, characterized in that, The weighted fusion technique is used to synthesize the image with the background image, generating a synthesized image and corresponding annotation information; The acquisition of the synthesized image specifically includes: roi= ; Where roi represents the generated synthetic image, src1 represents the background image, and src2 represents the transformed contaminant sample image. and These represent the contribution of the background image and the contribution of the pollutant sample image, respectively. This is a correction value; The annotation information includes the coordinates of the new bounding box at the location where the contaminant sample is pasted. The coordinates of the new bounding box are obtained using the following formula: ； Where, paste_x is the pixel coordinate of the left boundary of the paste location of the contaminant sample; paste_w is the pixel width of the contaminant sample; paste_y is the pixel coordinate of the upper boundary of the paste location of the contaminant sample; paste_h is the pixel height of the contaminant sample; image.shape[1] is the total width of the background image; image.shape[0] is the total height of the background image; new_x_center and new_y_center are the x and y coordinates of the new relative position of the center point of the contaminant sample, respectively; new_width and new_height are the new relative width and relative height of the contaminant sample.

4. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 1, characterized in that, The step of selecting pollutant samples at a preset ratio from the object library and randomly scaling and rotating each pollutant sample specifically includes: Select pollutant samples at a preset ratio from the object library, and determine the pixel coordinates of the pollutant samples using the following formula: Where (x1, y1) are the pixel coordinates of the upper left corner of the contaminant sample bounding box, (x2, y2) are the pixel coordinates of the lower right corner of the contaminant sample bounding box, x_center and y_center are the relative positions of the center point of the contaminant sample, width and height are the relative width and relative height of the contaminant sample, and w and h are the actual width and actual height of the image, respectively. Based on the pixel coordinates of the pollutant samples, each pollutant sample is randomly scaled and rotated.

5. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 1, characterized in that, The acquisition of the coverage rate specifically includes: The pre-trained convolutional neural network (CNN) divides the photovoltaic panel surface image into 64 grids in 8 rows and 8 columns. Combined with the target detection results, each grid in the image is classified to determine the number of grids containing pollutants. The single-cell coverage coefficient is determined based on the proportion of pollutant cells to the total number of cells. After assigning values, the coverage C corresponding to the type of pollutant within a single cell is obtained using the following formula: ； Where i and j are the single-cell coverage coefficients, respectively. The row and column numbers.

6. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 5, characterized in that, The method of classifying pollutant pollution levels into multiple levels based on coverage includes: When the coverage rate C is less than or equal to 20%, it is considered light pollution; when the coverage rate C is between 20% and 60%, it is considered moderate pollution; and when the coverage rate C is greater than 60%, it is considered heavy pollution.

7. The control method for differential execution of a photovoltaic cleaning robot with closed-loop logic as described in claim 1, characterized in that, When the number of cleaning operations is greater than or equal to 3, the current image of the photovoltaic panel surface is uploaded to the cloud database for subsequent strategy optimization.

Images