Photovoltaic station cleaning robot based on machine vision and cleaning method thereof

Abstract

This application relates to the field of photovoltaic power generation technology and discloses a photovoltaic station cleaning robot and its cleaning method based on machine vision. The robot includes: a vision module for acquiring RGB images of photovoltaic panels; a photovoltaic station data interface port for obtaining the electrical parameters of the photovoltaic panel when the cleaning robot enters the corresponding area of ​​the photovoltaic panel; a visual analysis model for generating a visual dirt probability map based on the RGB images; an electrical dirt analysis model for obtaining the dirt density distribution based on the electrical parameters; and a cleaning module for establishing a cleaning strategy based on the visual dirt probability map and the dirt density distribution. This invention integrates machine vision with the electrical parameters of the photovoltaic panel itself for dirt detection. The vision branch provides spatial distribution information of dirt, while the electrical branch provides a global calibration of the overall panel power attenuation. These two branches mutually verify and complement each other, providing the cleaning robot with accurate path planning and cleaning priority guidance.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation technology, and in particular to a machine vision-based photovoltaic station cleaning robot and its cleaning method. Background Technology

[0002] As a clean and renewable energy source, photovoltaic (PV) power generation plays an increasingly important role in the global energy transition. However, PV modules are typically installed outdoors, and their surfaces are prone to the accumulation of pollutants such as dust, bird droppings, and pollen, forming a dust layer. This dust layer significantly reduces the light transmittance of the panel, leading to a decrease in PV power generation efficiency. In addition to power loss, pollutants such as bird droppings can also generate hot spots on the panel surface, causing permanent damage to the performance of the affected areas. Therefore, it is necessary to conduct timely and accurate detection of the dirt status on the surface of PV modules and implement precise cleaning based on the detection results to ensure the power generation efficiency of PV power plants and extend the service life of the modules.

[0003] Existing methods for detecting dirt on photovoltaic panels acquire images of the panel surface using industrial cameras or visible light cameras mounted on drones, and then employ deep learning algorithms to detect and classify dust areas, bird droppings, physical damage, etc., within the images. However, the accuracy of purely visual detection methods is heavily dependent on image quality, and the accuracy drops significantly under conditions such as overexposed light, shadow occlusion, and hazy weather. Furthermore, the accuracy of deep learning algorithms is highly dependent on the richness and quality of the labeled data. Existing methods also assess the level of dirt on photovoltaic panels using electrical parameters; however, this method cannot accurately locate heavily soiled areas, making it difficult to guide cleaning robots for precise, targeted cleaning. Therefore, one of the problems this invention aims to solve is how to provide cleaning robots with accurate path planning and cleaning priority guidance by integrating machine vision with the electrical parameters of the photovoltaic panels. Summary of the Invention

[0004] To provide precise path planning and cleaning priority guidance for cleaning robots by integrating machine vision with the electrical parameters of photovoltaic panels, this application provides a machine vision-based photovoltaic station cleaning robot and its cleaning method.

[0005] Firstly, this application provides a machine vision-based photovoltaic station cleaning robot, employing the following technical solution:

[0006] A machine vision-based photovoltaic power station cleaning robot includes:

[0007] The vision module is used to acquire RGB images of the photovoltaic panel;

[0008] The photovoltaic station data interface port is used to obtain the electrical parameters of the photovoltaic panel when the cleaning robot of the photovoltaic station enters the area corresponding to the photovoltaic panel;

[0009] A visual analysis model for generating visual dirt probability maps from RGB images;

[0010] An electrical fouling analysis model is used to obtain the fouling density distribution based on electrical parameters;

[0011] The cleaning module is used to establish cleaning strategies based on the visual dirt probability map and dirt density distribution.

[0012] Optionally, the process of running the visual analysis model includes:

[0013] Normalize the pixel values ​​of the RGB image to the [0,1] range;

[0014] The dark channel of each pixel is calculated based on the normalized pixel value, the global atmospheric light is estimated, and the transmittance of each pixel is calculated based on the global atmospheric light and the dark channel.

[0015] The clean endmember and dust endmember are determined based on the pixel values ​​of the RGB image. The pixel value of each pixel is compared with the clean endmember and dust endmember to obtain the dust coverage of each pixel.

[0016] A visual dirt probability map is generated by using transmittance and dust coverage.

[0017] Optionally, the process of estimating global atmospheric light using the visual analysis model includes:

[0018] Take the top 0.1% of the brightest pixels in the dark channel, map these top 0.1% of pixels to the RGB image, and take the average brightness of these pixels in the three RGB channels. Calculate the average of the brightness of the three RGB channels to obtain the global atmospheric light.

[0019] Optionally, the process by which the visual analysis model calculates the transmittance of each pixel includes:

[0020] The initial transmittance of each pixel is calculated based on the dark channel and global atmospheric light.

[0021] By setting the filter radius and regularization parameters, and using the grayscale image of the RGB image as a guide image, the initial transmittance is guided by filtering to obtain the transmittance of each pixel.

[0022] Optionally, the process by which the visual analysis model determines the clean end-user and the dust end-user includes:

[0023] Take the average RGB value of the m% of pixels with the highest brightness in the RGB image as the clean endmember;

[0024] Take the average RGB value of the m% of pixels with the lowest brightness in the RGB image as the dust endmember;

[0025] Calculate the Euclidean distance between each pixel and the clean endmember. and the Euclidean distance between the dust end-unit and the clean end-unit ,pass and The ratio is used to characterize the dust coverage of each pixel.

[0026] Optionally, the process of generating a visual dirt probability map by the visual analysis model includes:

[0027] Transmittance is converted into occlusion rate, and the occlusion rate is weighted and fused with dust coverage to obtain an initial score for each pixel;

[0028] The initial scores are normalized, calibrated, and isolated noise is removed to obtain a visual dirt probability map.

[0029] Optionally, the process of running the electrical contamination analysis model includes:

[0030] Based on the electrical parameters, a dirt electrical feature vector is defined, and the average dirt scalar is calculated according to the dirt electrical feature vector and its corresponding weights.

[0031] The prior weight function is determined based on the location of the pixel, and the prior weight function is determined based on the height of the photovoltaic panel surface relative to the bottom edge and the distance from the frame.

[0032] By mapping the relationship between the average dirt scalar and the prior weight function, the dirt density distribution can be obtained.

[0033] Optionally, the dirty electrical characteristic vector includes power attenuation, short-circuit current attenuation, fill factor attenuation, pollution loss index, and open-circuit voltage attenuation.

[0034] Optionally, the cleaning module operation process includes:

[0035] A negative logarithmic transformation is used to convert the visual dirt probability map into a visual density map; the visual density map has the same dimensions and the same magnitude as the dirt density distribution.

[0036] The confidence weight of each pixel is obtained based on the local gradient of the visual dirt probability map;

[0037] The visual density map and the dirt density distribution are weighted and summed based on confidence weights to obtain the final dirt density map.

[0038] Cleaning areas and cleaning routes are generated based on the final dirt density map, and these areas and routes are used as cleaning strategies.

[0039] Secondly, this application provides a cleaning method for a photovoltaic station cleaning robot based on machine vision, employing the following technical solution:

[0040] A cleaning method for a photovoltaic power station cleaning robot based on machine vision, comprising:

[0041] The system collects environmental information and RGB images of the photovoltaic panels using a vision module.

[0042] The electrical parameters of the photovoltaic panel are obtained through the data interface port of the photovoltaic station when the cleaning robot of the photovoltaic station enters the area corresponding to the photovoltaic panel.

[0043] A visual dirt probability map is generated based on the RGB image using a visual analysis model;

[0044] The dirt density distribution is obtained based on electrical parameters using an electrical dirt analysis model;

[0045] The cleaning module establishes cleaning strategies based on visual dirt probability maps and dirt density distribution.

[0046] In summary, this application includes at least one of the following beneficial technical effects:

[0047] This invention integrates machine vision with the electrical parameters of the photovoltaic panel itself for dirt detection, overcoming the inherent defect of pure vision detection methods where the accuracy of recognition drops significantly in complex environments such as overexposed light, shadow occlusion, and smog. The vision branch provides spatial distribution information of dirt, while the electrical branch provides global calibration of the power attenuation of the entire panel. The two branches verify and complement each other, thereby providing a precise path planning basis and cleaning priority guidance for the cleaning robot. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the module of the photovoltaic station cleaning robot in an embodiment of the present invention.

[0049] Figure 2 This is a flowchart of the cleaning method of the photovoltaic station cleaning robot in an embodiment of the present invention. Detailed Implementation

[0050] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0051] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0052] The following description, with reference to the accompanying drawings, illustrates a machine vision-based photovoltaic station cleaning robot and its cleaning method according to embodiments of the present invention.

[0053] Figure 1 This is a schematic diagram of a photovoltaic power station cleaning robot according to an embodiment of the present invention. Figure 1 As shown in the embodiments of this application, a machine vision-based photovoltaic station cleaning robot includes:

[0054] The vision module is used to acquire RGB images of the photovoltaic panel. The vision module is implemented through hardware such as a visible light camera and corresponding supplementary lights.

[0055] The photovoltaic station data interface port is used to obtain the electrical parameters of the photovoltaic panel when the photovoltaic station cleaning robot enters the area corresponding to the photovoltaic panel; the photovoltaic station cleaning robot maintains data communication with the photovoltaic station, and each photovoltaic panel is set with a corresponding detection area. When the position coordinates of the photovoltaic station cleaning robot enter the area, the photovoltaic station sends the electrical parameters of the photovoltaic panel corresponding to the area to the photovoltaic station cleaning robot.

[0056] A visual analysis model for generating visual dirt probability maps from RGB images;

[0057] An electrical fouling analysis model is used to obtain the fouling density distribution based on electrical parameters;

[0058] The cleaning module is used to establish a cleaning strategy based on the visual dirt probability map and dirt density distribution; and to control the photovoltaic station cleaning robot to clean the photovoltaic panels according to the cleaning strategy.

[0059] It should be noted that the scope of protection of this invention is the process by which a photovoltaic station cleaning robot obtains a cleaning strategy through machine vision and the electrical parameters of the photovoltaic panels. The cleaning structure and hardware layout of the photovoltaic station cleaning robot can be implemented based on existing solutions, but any cleaning structure and hardware layout that uses the process of obtaining a cleaning strategy using this invention is within the scope of protection of this invention.

[0060] Figure 2 This is a flowchart of the cleaning method of the photovoltaic station cleaning robot according to an embodiment of the present invention. Figure 2 As shown in the embodiments of this application, a cleaning method for a photovoltaic station cleaning robot based on machine vision includes:

[0061] S1. Collect environmental information and RGB images of photovoltaic panels through the vision module;

[0062] S2. Through the photovoltaic station data interface port, when the cleaning robot of the photovoltaic station enters the area corresponding to the photovoltaic panel, obtain the electrical parameters of the photovoltaic panel;

[0063] S3. Generate a visual dirt probability map based on the RGB image using a visual analysis model;

[0064] S4. Obtain the dirt density distribution based on electrical parameters using an electrical dirt analysis model;

[0065] S5. The cleaning module establishes a cleaning strategy based on the visual dirt probability map and dirt density distribution.

[0066] This application embodiment detects dirt by integrating machine vision with the electrical parameters of the photovoltaic panel itself. This overcomes the inherent defect of pure vision detection methods, which have a significantly reduced accuracy in complex environments such as overexposed light, shadow occlusion, and smog. The vision branch provides spatial distribution information of dirt, while the electrical branch provides global calibration of the power attenuation of the entire panel. The two branches verify and complement each other, thereby providing the cleaning robot with accurate path planning and cleaning priority guidance.

[0067] In this embodiment, the process of running the visual analysis model includes:

[0068] The pixel values ​​of the RGB image are normalized to the [0,1] range. In this embodiment, the RGB image is an 8-bit image, so the normalization process is achieved by dividing the pixel value by 255. The dark channel of each pixel is calculated based on the normalized pixel value. In this embodiment, the local window is set to 15×15 (pixels). The acquisition process is as follows: traverse each pixel in the image, compare the values ​​of the R, G, and B channels of the pixel itself, keep only the minimum value of the three, discard the other two larger values, and generate a temporary single-channel grayscale image. For the single-channel grayscale image, the local window is used to perform a pixel-by-pixel sliding scan. Whenever the window stops at a certain position, the grayscale values ​​of all pixels (225) covered by the window are extracted, and the minimum value is taken out. This minimum value is used to replace the final value of the pixel in the center of the window. After traversing the entire image, the dark channel of each pixel is obtained.

[0069] Next, global atmospheric light is estimated by selecting the top 0.1% of brightest pixels in the dark channel. These pixels are then mapped to the RGB image, and the average brightness of these pixels across the three RGB channels is calculated. This average yields the global atmospheric light. Since there are always points in a local neighborhood where a particular color channel is exceptionally dark, the resulting value will be extremely low, approaching zero. However, when the photovoltaic panel surface is covered with dust, the grayish-white scattered light from the dust brightens the original dark areas, making it impossible to find sufficiently dark points within the local window. Therefore, the normalized dark channel value at that location will significantly increase. Thus, by estimating global atmospheric light, the brightest and purest pixels in the RGB image can be identified, representing the intensity and color of ambient light.

[0070] The transmittance of each pixel is calculated based on global atmospheric light and the dark channel; firstly, the initial transmittance of each pixel is calculated based on the dark channel and global atmospheric light. Where x=(u,v) are pixel coordinates. To ensure fixed parameters, in this embodiment... Set to 0.95, where A represents global atmospheric light. The dark channel is defined by pixel coordinate x. Then, the filter radius and regularization parameter are set. In empirical data, the filter radius ranges from 40 to 60; in this embodiment, it is set to 45. The regularization parameter is set to 0.03. A grayscale image of the RGB image is used as a guide image to perform guided filtering on the initial transmittance, obtaining the transmittance of each pixel. The specific process of guided filtering is not detailed here.

[0071] Next, clean and dust endmembers are determined based on the pixel values ​​of the RGB image. The average RGB value of the 5% brightest pixels in the RGB image is taken as the clean endmember; the average RGB value of the 5% brightest pixels in the RGB image is taken as the dust endmember. The pixel value of each pixel is compared with the clean and dust endmembers, and the Euclidean distance between each pixel and the clean endmember is calculated. and the Euclidean distance between the dust end-unit and the clean end-unit , obtain and The ratio will and The ratio is used as the dust coverage of each pixel.

[0072] A visual dirt probability map is generated by using transmittance and dust coverage. First, the transmittance is... Converted to occlusion rate ; occlusion rate With dust coverage Weighted fusion is performed to obtain an initial score for each pixel; in this embodiment, the weighting coefficient for occlusion rate is 0.58, and the weighting coefficient for dust coverage is... The weighting coefficient is 0.42; it was set after continuous optimization based on actual measurement data.

[0073] The initial score is normalized, calibrated, and isolated noise is removed. First, the minimum value of the initial score of the RGB image is obtained. and maximum value Obtain the normalized initial score The normalized initial score is calibrated to obtain the initial dirt probability. k is the steepness factor, and its optimal value range in empirical data is 8~12. In this embodiment, it is set to 12. The mean of the initial scores is normalized; finally, the initial dirt probability is filtered by median using a 3×3 filtering window to obtain the visual dirt probability map.

[0074] In this embodiment, transmittance reflects the overall degree of light attenuation, which is sensitive to large areas of thin dust, but not sharp enough for the boundaries of local patches; dust coverage is based on the distance ratio between pixels and endmembers, reflecting the proportion of the area occupied by dust in the pixel, which can accurately capture local non-uniform pollution; this application fuses the two, so that the information of large-area dust and local patches complement each other, effectively improving the model's ability to identify different types of dirt (area-like dust accumulation vs. point-like bird droppings); in addition, the clean endmember and dust endmember are dynamically extracted directly from the currently acquired image, without the need to pre-calibrate the factory color of the photovoltaic panel, which can be applied to photovoltaic modules of different manufacturers and colors, and can be directly deployed in distributed photovoltaic and large-scale ground power station scenarios; in addition, the calibration process can apply gain amplification to medium and high-level dirty areas and suppress and compress slight background noise, thereby significantly widening the numerical gap between clean and dirty areas in the final probability map, making the detection results more sensitive and specific, and reducing the false judgment rate.

[0075] In this embodiment, the process of running the electrical contamination analysis model includes:

[0076] Based on the definition of a fouling electrical characteristic vector using electrical parameters, the fouling electrical characteristic vector in this embodiment includes power attenuation, short-circuit current attenuation, fill factor attenuation, pollution loss index, and open-circuit voltage attenuation; wherein, the electrical parameters include real-time output voltage, real-time output current, short-circuit current, open-circuit voltage, and output power at the maximum power point; power attenuation P(t) represents the real-time power, which is obtained based on the real-time output voltage and real-time output current. The theoretical maximum power is calculated based on the current irradiance and temperature. This is obtained by collecting the irradiance and temperature at the current time point and substituting them into the photovoltaic panel power model, which is set based on test data; short-circuit current attenuation. The short-circuit current is a limiting parameter of photovoltaic panels; it is not a real-time parameter but is measured periodically. This represents the short-circuit current from the most recent measurement at time t. The short-circuit current is under clean conditions; the open-circuit voltage decays. Open-circuit voltage is also a limiting parameter for photovoltaic panels. This represents the open-circuit voltage from the most recent measurement at time t. Open-circuit voltage under clean conditions; fill factor decay ,in, , Output power at the maximum power point; This is the baseline value for the fill factor under clean conditions, which can be set according to the type of photovoltaic panel. For conventional monocrystalline / polycrystalline silicon photovoltaic panels, The value is between 0.70 and 0.85. Furthermore, by retaining a reference photovoltaic panel in the photovoltaic power station that is regularly cleaned by robots, the current... As Pollution loss index ,in, ,in, This refers to the short-circuit current under standard test conditions. ; This is the temperature coefficient of short-circuit current, which is set based on the photovoltaic panel data. For real-time temperature, ; The global horizontal irradiance measured by the irradiance meter is used. The average scalar value of dirt is calculated based on the dirt electrical feature vector and its corresponding weights. The weights are set according to the sensitivity of the dirt electrical feature vector to dust. In this embodiment, the weights corresponding to power attenuation, short-circuit current attenuation, fill factor attenuation, pollution loss index and open-circuit voltage attenuation are 0.30, 0.30, 0.20, 0.15 and 0.05, respectively. Since the values ​​of power attenuation, short-circuit current attenuation, fill factor attenuation, pollution loss index and open-circuit voltage attenuation are all within [0, 1], the sum of the corresponding weights is 1. Therefore, the range of the obtained average scalar value of dirt is [0, 1].

[0077] The prior weight function is determined based on the pixel's location (u, v), and this prior weight function is determined based on the height of the photovoltaic panel surface relative to the bottom edge and its distance from the frame; specifically, based on... Calculate and obtain the initial weight function Where H is the height of the photovoltaic panel (v=0 is the bottom edge, v=H is the top edge), and W is the width of the photovoltaic panel (u=0 and u=W are the left and right borders, respectively). The vertical stacking gradient has a value ranging from 0.3 to 0.8, as shown in this embodiment. Based on empirical data, it was set to 0.45; This is the edge enhancement coefficient, which ranges from 0.1 to 0.3. In this embodiment... Based on empirical data, it was set to 0.16; therefore, Reflects gravitational settling; the index is highest at the bottom (v=0) and lowest at the top (v=H); Reflecting the wind-blocking effect of the frame, dust accumulates more at the edges than in the center; through right Perform homogenization to obtain the prior weight function. .

[0078] By mapping the relationship between the average dirt scalar G and the prior weighting function, the dirt density distribution can be obtained. It should be noted that the dirt density distribution must meet the following requirements. , for The maximum value in the entire graph, if it does not satisfy... Then truncate G and let .

[0079] By introducing a prior weighting function based on gravity settling and the wind-blocking effect of the frame during the process of obtaining the dirt density distribution, the generated dirt density distribution can still approximate the actual spatial pattern of dust accumulation even without visual guidance. This avoids systematic bias caused by blind assumptions and provides a global dirt benchmark independent of vision for the entire cleaning control closed loop. It complements the visual detection branch, which provides spatial details through vision and amplitude benchmarks through electrical detection, thus enabling more accurate and reliable acquisition of cleaning decisions.

[0080] In this embodiment, the cleaning module operation process includes:

[0081] Using negative logarithmic transformation to transform the visual dirt probability map Convert to a visual density map; ensure the visual density map has the same dimensions and amplitudes as the dirt density distribution, first calculate the initial visual density map. , ; Calculate the amplitude scaling factor based on the initial visual density map Obtain visual density map .

[0082] The confidence weight of each pixel is obtained based on the local gradient g(x) of the visual dirt probability map. Local gradient , The gradient magnitude is the average of the entire image; when the gradient magnitude of a pixel is higher than the average value of the entire image, that pixel is located at a dirty edge or in a region with rich texture; using Using a threshold allows the weights to be adaptively distributed across the entire graph, avoiding the problem of fixed thresholds failing under different lighting conditions.

[0083] The final dirt density map is obtained by weighted summation of the visual density map and the dirt density distribution based on confidence weights. .

[0084] Finally, cleaning areas and routes are generated based on the final dirt density map, and these areas and routes are used as the cleaning strategy. This process first compares the final dirt density map with the corresponding threshold to determine the areas that need to be cleaned. The corresponding threshold is set based on measured data. Then, the cleaning routes are determined based on the cleaning areas. This process can be implemented using existing path planning algorithms, such as prioritizing based on dirt density combined with the nearest neighbor method or the 2-Opt optimization method to generate the optimal access order, or using the A* algorithm to plan the shortest transfer path between areas. The specific implementation methods are not detailed here. Through the above process, cleaning resources can be concentrated on the most heavily soiled areas, significantly reducing operation and maintenance costs while ensuring cleaning effectiveness.

[0085] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A machine vision-based photovoltaic station cleaning robot, characterized in that, include: The vision module is used to acquire RGB images of the photovoltaic panel; The photovoltaic station data interface port is used to obtain the electrical parameters of the photovoltaic panel when the cleaning robot of the photovoltaic station enters the area corresponding to the photovoltaic panel; A visual analysis model for generating visual dirt probability maps from RGB images; An electrical fouling analysis model is used to obtain the fouling density distribution based on electrical parameters; The cleaning module is used to establish cleaning strategies based on visual dirt probability maps and dirt density distribution.

2. The photovoltaic station cleaning robot based on machine vision according to claim 1, characterized in that, The process of running the visual analysis model includes: Normalize the pixel values ​​of the RGB image to the [0,1] range; The dark channel of each pixel is calculated based on the normalized pixel value, the global atmospheric light is estimated, and the transmittance of each pixel is calculated based on the global atmospheric light and the dark channel. The clean endmember and dust endmember are determined based on the pixel values ​​of the RGB image. The pixel value of each pixel is compared with the clean endmember and dust endmember to obtain the dust coverage of each pixel. A visual dirt probability map is generated by using transmittance and dust coverage.

3. The photovoltaic station cleaning robot based on machine vision according to claim 2, characterized in that, The process by which the visual analysis model estimates global atmospheric light includes: Take the top 0.1% of the brightest pixels in the dark channel, map these top 0.1% of pixels to the RGB image, and take the average brightness of these pixels in the three RGB channels. Calculate the average of the brightness of the three RGB channels to obtain the global atmospheric light.

4. A photovoltaic station cleaning robot based on machine vision according to claim 2, characterized in that, The process by which the visual analysis model calculates the transmittance of each pixel includes: The initial transmittance of each pixel is calculated based on the dark channel and global atmospheric light. By setting the filter radius and regularization parameters, and using the grayscale image of the RGB image as a guide image, the initial transmittance is guided by filtering to obtain the transmittance of each pixel.

5. A photovoltaic station cleaning robot based on machine vision according to claim 2, characterized in that, The process by which the visual analysis model determines the clean end-user and the dust end-user includes: Take the average RGB value of the m% of pixels with the highest brightness in the RGB image as the clean endmember; Take the average RGB value of the m% of pixels with the lowest brightness in the RGB image as the dust endmember; Calculate the Euclidean distance between each pixel and the clean endmember. and the Euclidean distance between the dust end-unit and the clean end-unit ,pass and The ratio is used to characterize the dust coverage of each pixel.

6. A photovoltaic station cleaning robot based on machine vision according to claim 2, characterized in that, The process by which the visual analysis model generates a visual dirt probability map includes: Transmittance is converted into occlusion rate, and the occlusion rate is weighted and fused with dust coverage to obtain an initial score for each pixel; The initial scores are normalized, calibrated, and isolated noise is removed to obtain a visual dirt probability map.

7. A photovoltaic station cleaning robot based on machine vision according to claim 1, characterized in that, The process of running the electrical contamination analysis model includes: Based on the electrical parameters, a dirt electrical feature vector is defined, and the average dirt scalar is calculated according to the dirt electrical feature vector and its corresponding weights. The prior weight function is determined based on the location of the pixel, and the prior weight function is determined based on the height of the photovoltaic panel surface relative to the bottom edge and the distance from the frame. By mapping the relationship between the average dirt scalar and the prior weight function, the dirt density distribution can be obtained.

8. A photovoltaic station cleaning robot based on machine vision according to claim 7, characterized in that, The dirty electrical characteristic vector includes power attenuation, short-circuit current attenuation, fill factor attenuation, pollution loss index, and open-circuit voltage attenuation.

9. A photovoltaic station cleaning robot based on machine vision according to claim 1, characterized in that, The cleaning module operates as follows: A negative logarithmic transformation is used to convert the visual dirt probability map into a visual density map; the visual density map has the same dimensions and the same magnitude as the dirt density distribution. The confidence weight of each pixel is obtained based on the local gradient of the visual dirt probability map; The visual density map and the dirt density distribution are weighted and summed based on confidence weights to obtain the final dirt density map. Cleaning areas and cleaning routes are generated based on the final dirt density map, and these areas and routes are used as cleaning strategies.

10. A cleaning method for a photovoltaic station cleaning robot based on machine vision, characterized in that, The method employs a machine vision-based photovoltaic station cleaning robot as described in any one of claims 1-9, comprising: The system collects environmental information and RGB images of the photovoltaic panels using a vision module. The electrical parameters of the photovoltaic panel are obtained through the data interface port of the photovoltaic station when the cleaning robot of the photovoltaic station enters the area corresponding to the photovoltaic panel. A visual dirt probability map is generated based on the RGB image using a visual analysis model; The dirt density distribution is obtained based on electrical parameters using an electrical dirt analysis model; The cleaning module establishes cleaning strategies based on visual dirt probability maps and dirt density distribution.

Images