Photovoltaic panel cleaning method and device, photovoltaic cleaning robot

Abstract

This invention discloses a method, apparatus, and robot for cleaning photovoltaic panels. The method, executed by a robot control center within the photovoltaic panel cleaning robot, includes: whenever a stain detection time is determined, judging whether standard cleaning conditions are met based on a first image acquired by a vision sensor; when the standard cleaning conditions are met, determining a first cleaning path based on the robot's initial position acquired by a position sensor; and dynamically updating the photovoltaic panel cleaning parameters during cleaning based on a second image acquired periodically by a vision sensor, real-time rainfall data acquired by a rain gauge, and real-time position information acquired by a position sensor. This technical solution combines multi-sensor data for photovoltaic panel cleaning decisions, exhibiting high system robustness and the ability to dynamically adjust cleaning parameters, thereby improving cleaning efficiency and avoiding energy waste.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a method, apparatus, and robot for cleaning photovoltaic panels. Background Technology

[0002] When photovoltaic panels are used in outdoor environments, their surfaces are prone to accumulating pollutants such as dust, bird droppings, and fallen leaves. These pollutants can reflect, absorb, and block light, which may lead to reduced light absorption, decreased light transmittance, and reduced performance of the photovoltaic panels, thereby affecting power generation efficiency.

[0003] With the development of artificial intelligence and automation technology, photovoltaic panel cleaning is gradually moving towards intelligentization, and the application of photovoltaic panel cleaning robots is becoming more and more widespread.

[0004] Traditional photovoltaic cleaning robots rely heavily on preset programs or simple sensors, making them unable to dynamically adapt to complex environments. The data from each sensor is processed independently, resulting in insufficient system robustness and a lack of real-time optimization of cleaning parameters, leading to energy waste or incomplete cleaning. Summary of the Invention

[0005] This invention provides a photovoltaic panel cleaning method, device, and photovoltaic cleaning robot, which can make photovoltaic panel cleaning decisions by combining multi-sensor data. The system has high robustness and can dynamically adjust cleaning parameters to improve cleaning efficiency and avoid energy waste.

[0006] According to one aspect of the present invention, a photovoltaic panel cleaning method is provided, executed by a robot control center in a photovoltaic panel cleaning robot, comprising: Whenever the stain detection time is determined, the system uses the first image captured by the vision sensor to determine whether the standard cleaning conditions are met. When the standard cleaning conditions are met, the first cleaning path is determined based on the robot's initial position collected by the position sensor; During the cleaning process of the photovoltaic panels, the cleaning parameters of the photovoltaic panels are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time location information collected by the position sensor.

[0007] According to another aspect of the present invention, a photovoltaic panel cleaning device is provided, which is executed by a robot control center in a photovoltaic panel cleaning robot, comprising: The standard cleaning condition judgment module is used to determine whether the standard cleaning conditions are met based on the first image collected by the vision sensor whenever the stain detection time is determined. The first cleaning path determination module is used to determine the first cleaning path based on the robot's initial position collected by the position sensor when it is determined that the standard cleaning conditions are met. The cleaning parameter update module is used to dynamically update the photovoltaic panel cleaning parameters during the cleaning process based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time location information collected by the position sensor.

[0008] According to another aspect of the present invention, a photovoltaic cleaning robot is provided, comprising a vision sensor, a position sensor, a rain sensor, a robot control center, and a robot actuator; wherein, The vision sensor is used to capture images of the photovoltaic panel and send the images to the robot control center; The position sensor is used to determine the position of the photovoltaic cleaning robot in the photovoltaic panel array map and send the position information to the robot control center; The rain sensor is used to acquire rainfall data and send the rainfall data to the robot control center; The robot control center is used to execute the photovoltaic panel cleaning method according to any embodiment of the present invention. The robot actuator is used to perform photovoltaic panel cleaning operations according to the cleaning path and cleaning parameters sent by the robot control center.

[0009] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; And, a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the photovoltaic panel cleaning method according to any embodiment of the present invention.

[0010] The technical solution of this invention determines whether standard cleaning conditions are met based on the first image collected by the vision sensor each time a stain detection time is determined. When the standard cleaning conditions are met, a first cleaning path is determined based on the robot's initial position collected by the position sensor. During the cleaning of the photovoltaic panels, the cleaning parameters are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time position information collected by the position sensor. This method combines multi-sensor data to make photovoltaic panel cleaning decisions, fully considers the importance of rainfall for photovoltaic panel cleaning, has high system robustness, and can dynamically adjust cleaning parameters to adapt to complex environments, improve cleaning efficiency, and avoid energy waste.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a photovoltaic panel cleaning method according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of another photovoltaic panel cleaning method provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of a photovoltaic panel cleaning device according to Embodiment 3 of the present invention; Figure 4 This is a structural schematic diagram of a photovoltaic cleaning robot according to Embodiment 4 of the present invention; Figure 5 This is a schematic diagram of the operation of a photovoltaic panel cleaning robot provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device that implements the photovoltaic panel cleaning method of this invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. Example

[0016] Figure 1 This is a flowchart of a photovoltaic panel cleaning method provided in Embodiment 1 of the present invention. This embodiment is applicable to situations where the contamination status of photovoltaic panels is detected and automatically cleaned. The method can be executed by a photovoltaic panel cleaning device, which can be implemented in hardware and / or software and configured in a robot control center with data processing and image processing capabilities. Figure 1 As shown, the method includes: S110. Whenever the stain detection time is determined, the standard cleaning conditions are determined based on the first image collected by the vision sensor.

[0017] Optionally, the stain detection time can be determined comprehensively based on factors such as environmental factors and the rate of past pollution accumulation in the area where the photovoltaic panel is located. For example, for areas with more dust, the time interval between each detection is shorter, while for areas with more rain, the time interval between each detection is longer. By pre-setting a time interval that matches the area, stain detection can be performed on a regular schedule.

[0018] Optionally, a target number of days can be preset, and the next stain inspection time can be determined based on the last stain inspection time and the target number of days.

[0019] Optionally, the first image may refer to the image of the photovoltaic panel surface collected by the vision sensor at the time of stain detection. The vision sensor may include a camera and an infrared sensor. The image content collected by the vision sensor includes the image of the photovoltaic panel surface in the area surrounding the photovoltaic panel cleaning robot.

[0020] The determination of whether standard cleaning conditions are met based on the first image acquired by the vision sensor may include: The first image is processed by a convolutional neural network in a deep learning algorithm model to obtain the type and level of dirt in the first image. Based on the type and level of dirt, determine whether the standard cleaning conditions are met; If the cleaning conditions are not met, record the stain detection time and do not perform the cleaning operation.

[0021] Optionally, the convolutional neural network is a pre-trained classification model. After the image is input into the convolutional neural network, the type and level of dirt in each region of the image can be obtained. The types of dirt include, but are not limited to, dust, fallen leaves, mud, bird droppings, algae, etc. Among them, dirt such as dust, fallen leaves, and mud can be classified as the first type of dirt, which can refer to dirt types that are relatively easy to clean. Dirt such as bird droppings and algae can be classified as the second type of dirt, which can refer to dirt types that are not easy to clean. The dirt level can be a pre-set level, for example, it can be pre-set to level 1-5.

[0022] It is understandable that standard cleaning refers to cleaning the photovoltaic panels in the entire area. When the conditions for standard cleaning are met in the first image, it means that the photovoltaic panels in the entire area need to be cleaned according to standard. The purpose of standard cleaning is to clean the entire area as a whole to remove dirt caused by weather, environment and other reasons.

[0023] Optionally, the standard cleaning conditions can be: when the proportion of areas belonging to the first type of dirt in the first image reaches a preset first proportion, and the proportion of areas reaching the first level of dirt reaches a second proportion, the standard cleaning conditions can be determined to be met. If the dirt belongs to the first type of dirt, the photovoltaic panel can be cleaned by conventional cleaning methods. The first level of dirt can be a specified level of dirt, such as level 3-5. When the level of dirt is low, no cleaning is performed. When the level of dirt reaches the first level of dirt, it has a certain impact on the photovoltaic panel. At this time, when the conditions related to the first type of dirt and the first level of dirt are met simultaneously in the first image, the standard cleaning conditions are determined to be met.

[0024] Optionally, determining whether the standard cleaning conditions are met based on the type and level of dirt may further include: scoring the first image based on the level and type of dirt in the first image using a deep learning algorithm model; when the score reaches a preset target score, it is determined that the standard cleaning conditions are met.

[0025] Optionally, when cleaning conditions are not met, the stain detection time is recorded, and the next stain detection time is determined based on the recorded stain detection time.

[0026] S120. When it is determined that the standard cleaning conditions are met, the first cleaning path is determined based on the robot's initial position collected by the position sensor.

[0027] Optionally, the photovoltaic panel array map is a pre-built map that can be generated by collecting data from position sensors and combining it with SLAM (Simultaneous Localization and Mapping) technology before the photovoltaic panel cleaning robot is used. The robot control center can mark abnormal locations on the photovoltaic panel array map based on the images collected by the robot during each cleaning process and the cleaning situation.

[0028] Optionally, one specific implementation of the step of determining the first cleaning path based on the robot's initial position collected by the position sensor is as follows: After confirming that the standard cleaning conditions are met, the initial position coordinates (x0, y0, θ0) of the photovoltaic panel cleaning robot in the pre-established photovoltaic panel array map are first obtained through position sensors (such as data that integrates IMU, wheel odometry and UWB signals), where (x0, y0) are planar coordinates and θ0 is the heading angle.

[0029] After determining the initial location, the system needs to plan a traversal path that covers the preset standard cleaning range. The preset standard cleaning range can be the entire photovoltaic panel array in the robot's area of ​​responsibility by default, or it can be dynamically adjusted based on historical data, weather forecasts, or the dirt distribution identified in the first image.

[0030] This method designs a hybrid path planning strategy to balance efficiency and robustness: Path library pre-stored and recall mode: The system maintains a path library that stores pre-calculated optimal traversal paths for several common fixed starting locations (such as charging pile locations and array entrance locations). These paths are calculated offline using a complete photovoltaic array map and global optimization algorithms (such as genetic algorithms and reinforcement learning) to select the best paths that cover the standard cleaning range and have the shortest total travel distance. When the system determines that the error between the current initial position coordinates and a fixed starting position in the path library is within an acceptable threshold (such as 10 cm), it directly recalls the corresponding pre-stored path as the first cleaning path. This mode offers fast response and ensures global optimality of the path.

[0031] Real-time dynamic planning mode: If the current initial position is not near any pre-stored fixed starting position (e.g., the robot stops in the middle of the array for some reason), the system starts real-time path planning. This mode takes the initial position coordinates as the starting point and, based on the latest photovoltaic array map (including static obstacles and dynamically marked temporary obstacles), uses an improved path search algorithm to calculate online a path that can traverse all passable areas within the preset cleaning range and has an approximately optimal total path length, which is then used as the first cleaning path.

[0032] This method employs a hybrid strategy, enabling the rapid invocation of high-quality pre-stored paths under most normal circumstances (starting from a fixed point), thereby improving response efficiency and planning quality. In abnormal or special situations, it can switch to real-time planning mode, ensuring the system's flexibility and adaptability, allowing the robot to start working efficiently from any passable location.

[0033] The core of the step of calculating and generating the shortest traversal path covering the preset standard cleaning range in real time lies in an improved coverage path planning algorithm. The specific implementation is as follows: First, environmental modeling is performed: the pre-defined standard cleaning area is discretized into a grid on a digitized photovoltaic panel array map. The continuous two-dimensional plane is divided into uniform square grid cells. Based on obstacle information in the map (such as photovoltaic panel edges, supports, and marked damaged areas) and impassable areas (such as steep slopes and puddles), each grid cell is marked as either a "passable area" or a "non-passable area".

[0034] Next, an improved A-Star algorithm is used for global path search to generate a closed or open path that starts from the starting point, traverses all passable grid areas, and finally returns to the starting point or reaches the destination, while minimizing the total travel distance. The cost function of the standard A-Star algorithm is F(n) = G(n) + H(n), where G(n) is the actual cost from the starting point to the current node n, and H(n) is the estimated cost from the current node n to the target point (such as Manhattan distance or Euclidean distance). This algorithm excels at finding the shortest path between two points, but it is not optimized for the "coverage" problem.

[0035] To make the A-Star algorithm applicable to scenes covering all grid cells, this invention improves its cost function by introducing an exploration incentive term. The improved cost function is as follows: F(n) = G(n) + H(n) + λ * T(n) in: G(n): Same as the standard A-Star, representing the actual cost (such as path length) incurred in moving from the starting point to node n.

[0036] H(n): Same as the standard A-Star, representing the estimated cost from node n to the traditional target point. In the coverage problem, this target point can be set as the center point of the unvisited grid furthest from n, or a virtual global endpoint.

[0037] T(n): Exploration incentive. Its core idea is to evaluate the average distance (or shortest distance) from the current node n to all unvisited accessible grates. A smaller T(n) value indicates that it seems more convenient to access the remaining unswept areas from n. When calculating T(n), a set of unvisited grates can be maintained, and the average distance (e.g., Manhattan distance) from n to all grates in that set can be calculated in real-time.

[0038] λ: A balance coefficient, a weighting parameter greater than 0. λ is used to adjust the weight of the incentive to explore unvisited areas relative to the proximity to the final goal. By adjusting λ, you can control whether the path is more biased towards covering the surrounding area as quickly as possible or moving more directly in a certain direction.

[0039] During algorithm execution, whenever the robot moves to a new grid, that grid is marked as visited and removed from the unvisited set calculated by T(n). The algorithm continues to run until all passable grids within the preset cleaning range are marked as visited. The resulting path sequence is the first cleaning path sought.

[0040] This improved algorithm, by introducing an exploration incentive term T(n), successfully transforms the traditional point-to-point optimal path search algorithm into a path planning algorithm suitable for area coverage traversal. It can autonomously generate a near-optimal coverage path in complex photovoltaic array environments with obstacles, ensuring no areas are missed during cleaning, while effectively controlling the total travel distance and improving the overall efficiency of the cleaning operation.

[0041] S130. During the cleaning process of the photovoltaic panel, the cleaning parameters of the photovoltaic panel are dynamically updated based on the second image collected by the vision sensor at regular intervals, the real-time rainfall data collected by the rain sensor, and the real-time position information collected by the position sensor.

[0042] Optionally, after each sensor acquires images, the acquired data is preprocessed. For example, after the vision sensor acquires an image, image denoising is performed; after the position sensor acquires position information, the robot's current position is calibrated; and after the rain sensor acquires rainfall data, the rainfall range to which the rainfall data belongs is determined. After preprocessing the data, the preprocessed data can be used in the photovoltaic panel cleaning method described above. The first image, the second image, the real-time rainfall data, and the real-time position information can all be preprocessed data.

[0043] Optionally, the second image may refer to the image of the photovoltaic panel surface collected by the vision sensor at regular intervals during the photovoltaic panel cleaning process. Since the collected first image can only show the dirt situation in some areas around the cleaning robot, if cleaning is carried out with fixed cleaning parameters, it may lead to energy waste or incomplete cleaning of some photovoltaic panels. Therefore, the dirt situation and cleaning status of photovoltaic panels in other areas can be detected in real time during the cleaning process to dynamically adjust the photovoltaic panel cleaning parameters.

[0044] Optional parameters for cleaning photovoltaic panels include, but are not limited to, the wheel speed of each wheel of the photovoltaic panel cleaning robot, the rotation speed of the roller brush, and the downward pressure of the roller brush.

[0045] Understandably, this invention integrates rainfall data, location data, and images to improve energy efficiency and robot robustness, thereby obtaining optimal cleaning parameters. Existing technologies often only consider location data and image analysis results, neglecting the impact of rainfall on photovoltaic panel cleaning. During rainfall, rainwater itself can wash away some of the photovoltaic panels, with different rainfall amounts resulting in different cleaning effects. Less rainfall may not completely clean the panels, while heavier rainfall may wash away most of the dirt. By acquiring real-time rainfall data, the robot's water spray volume and brush head pressure can be adjusted. Furthermore, continuous rainfall can make the photovoltaic panel surface slippery. If the robot continues to clean at its usual speed and intensity, it may slip on the panels, affecting cleaning effectiveness and potentially damaging both the robot and the panels. By acquiring rainfall data, the robot can reduce its cleaning speed and adjust its cleaning intensity based on rainfall conditions, protecting the equipment and avoiding energy waste and increased maintenance costs due to equipment failure, thus improving energy efficiency in the long run.

[0046] Optionally, the photovoltaic panel cleaning parameters can be dynamically updated based on the second image acquired periodically by the vision sensor, the real-time rainfall data acquired by the rain sensor, and the real-time location information acquired by the position sensor. This can include: Based on the second images collected during the photovoltaic panel cleaning process and the real-time location information at the time of collection of each second image, abnormal locations are marked in the pre-established photovoltaic panel array map. The convolutional neural network in the deep learning algorithm model is used to process each second image to determine and mark the abnormal information at each abnormal location in the photovoltaic panel array map. The abnormal information includes whether the abnormality type is dirt or damage. When the abnormality type is dirt, the abnormal information also includes the dirt type and dirt level; when the abnormality type is damage, the abnormal information also includes the damage level. To achieve accurate abnormality identification and map marking, this embodiment provides a specific implementation method for processing each second image using the convolutional neural network in the deep learning algorithm model to determine and mark the abnormal information at each abnormal location in the photovoltaic panel array map, including the following steps: Step S1, Image Preprocessing and Geolocation: First, the second images acquired periodically by the vision sensor are preprocessed. Preprocessing aims to improve image quality, facilitating subsequent neural network processing. Simultaneously, the robot control center receives real-time position information from position sensors (such as UWB and laser SLAM modules) and uses a coordinate transformation matrix to accurately map the pixel coordinates in the images to the world coordinate system of a pre-constructed two-dimensional or three-dimensional photovoltaic panel array map. This step ensures that each frame of the image corresponds to a specific area on the map, laying the spatial foundation for subsequent anomaly location marking.

[0047] Step S2, Multi-task Neural Network Analysis and Preliminary Segmentation: The preprocessed and geolocated second image is input into a pre-trained multi-task convolutional neural network model. The core component of this model is a semantic segmentation network based on an encoder-decoder structure (such as U-Net or DeepLabV3+). The encoder (such as ResNet or VGG) is responsible for extracting deep features of the image, while the decoder is responsible for upsampling these features and restoring them to the original image resolution for pixel-level classification. This segmentation network is trained to output a segmentation mask of the same size as the input image, where each pixel is classified as background (such as a clean solar panel or sky), a dirty area, or a suspected damaged area. This step achieves coarse-grained localization and preliminary classification of abnormal regions in the image.

[0048] Step S3, Fine-grained anomaly classification and attribute determination: Connectivity analysis is performed on the segmentation mask output from step S2 to identify each independent dirt region and suspected damage region block. For each independent region, richer visual features are extracted from its corresponding original second image region, such as color histograms, texture features (LBP, HOG), shape descriptors, etc., or depth features extracted directly from the intermediate layers of the neural network. Subsequently: For soiled areas: their feature vectors are input into a soiling classification sub-model (which can be a fully connected layer classifier). This sub-model determines the specific type of soiling, such as dust, bird droppings, oil stains, fallen leaves, etc. Simultaneously, it combines the area of ​​the region (calculated by converting pixel count to map resolution), coverage (the proportion of the region's area to the area of ​​its corresponding photovoltaic panel unit), and thickness information reflected by texture features to comprehensively determine its soiling level.

[0049] For suspected damaged areas: their features are input into a damage determination and classification sub-model. This model first performs binary classification to determine whether it is genuine damage (such as cracks, fragments, hot spots, color anomalies caused by PID attenuation) rather than interference such as reflections or shadows. If it is determined to be genuine damage, its damage type is further determined, and its damage level is evaluated based on features such as damage area, shape complexity, and edge sharpness, combined with preset rules or another regression / classification model.

[0050] Step S4, Anomaly Information Integration and Map Update: The final judgment results of each independent area obtained in Step S3 are structurally encapsulated to form a complete anomaly information record. This record includes at least: a unique identifier, anomaly type (dirt / damage), specific subtype, level, and the precise geographic coordinates of its circumscribed polygon outline in the photovoltaic array map. The robot control center associates these anomaly information records with the corresponding real-time location information (timestamp, coordinates) and updates them to the globally shared photovoltaic array map data structure. Each anomaly location in the map is visually marked with a specific graphic symbol (such as polygons of different colors) and attribute labels. This provides accurate and structured input data for subsequently determining the contamination level of each area and dynamically updating the photovoltaic panel cleaning parameters.

[0051] This method, through a two-stage process of "preliminary segmentation + refined judgment," combined with a multi-task neural network and attribute sub-models, significantly improves the accuracy of identifying dirt and damage on photovoltaic panel surfaces and enhances classification precision. By strictly binding with real-time location information, it achieves precise mapping from image anomalies to map coordinates, enabling continuous tracking and management of abnormal states and providing core data support for intelligent, adaptive cleaning decisions.

[0052] Based on the anomaly information marked in the photovoltaic panel array map, the pollution level of each area is determined, and the photovoltaic panel cleaning parameters are dynamically updated according to the pollution level of each area and real-time rainfall data through a deep learning algorithm model.

[0053] The photovoltaic panel cleaning parameters are dynamically updated based on the pollution level of each area and real-time rainfall data, including: By using a multimodal fusion model in a deep learning algorithm, the pollution level and real-time rainfall data of each region are fused, and the photovoltaic panel cleaning parameters of each region are determined based on the fusion results. The fusion result is G(L,R)=w1·G L + w2·G R L represents the pollution level of the target area, R represents the real-time rainfall, and G represents the pollution level of the target area. L G represents the pollution level characteristics of the target area. R The fusion features of real-time rainfall data are represented by w1 and w2, which are the weights of the pollution level fusion features and the real-time rainfall data fusion features of the target area, respectively.

[0054] Optionally, the deep learning algorithm model may include a multimodal fusion model. This model can further fuse pollution levels and real-time rainfall data from different areas. Based on the fusion results, the photovoltaic panel cleaning parameters are determined. The fusion result G(L,R) can be obtained using the following formula: G(L,R)=w1·G L + w2·G R Where L represents the pollution level of the target area, which can refer to any area defined in the photovoltaic array map; R represents the real-time rainfall characteristic; and G represents the real-time rainfall characteristic. L G represents the pollution level characteristics of the target area. R The fusion feature of real-time rainfall data is represented by w1 and w2, which are the weights of the pollution level fusion feature and the real-time rainfall data fusion feature, respectively, and w1+w2=1.

[0055] Furthermore, G L =gL·Γ L (L)+(1-g L )·Cattn(L||R); where, g L g is a gating signal representing the pollution level characteristics of the target area. L =Sigmoid(W L ;[L;R]), W L The learnable parameter matrix is ​​obtained through pre-training. [L;R] represents the first feature concatenation result, which is obtained by concatenating the pollution level feature of the target area with the real-time rainfall feature. Sigmoid is the Sigmoid function, and Γ... L (L) represents the intramodal nonlinear transformation result of the pollution level characteristics of the target area. Using a multilayer perceptron and layer normalization, the intramodal nonlinear transformation of the pollution level characteristics of the target area can be performed to obtain Γ. L(L), Cattn is the cross-attention mechanism, and Cattn(L||R) is the result of the cross-attention feature interaction between the pollution degree feature of the target area and the real-time rainfall feature. It is obtained by concatenating the pollution degree feature of the target area and the real-time rainfall feature and then interacting the features through the cross-attention mechanism.

[0056] Similarly, G R =g R ·Γ R (R)+(1-g R )·Cattn(R||L); where, g R g is a gated signal for real-time rainfall characteristics. R =Sigmoid(W R ;[R;L]), W R The learnable parameter matrix is ​​obtained through pre-training. [R;L] represents the second feature concatenation result, which is obtained by concatenating real-time rainfall features with the pollution level features of the target area. Sigmoid is the Sigmoid function, and Γ... R (R) represents the intramodal nonlinear transformation result of the real-time rainfall characteristics. A multilayer perceptron and layer normalization can be used to perform an intramodal nonlinear transformation on the real-time rainfall characteristics to obtain Γ. R (R), Cattn is the cross-attention mechanism, and Cattn(R||L) is the result of the cross-attention feature interaction between the real-time rainfall feature and the pollution level feature of the target area. It is obtained by concatenating the real-time rainfall feature and the pollution level feature of the target area and then interacting the features through the cross-attention mechanism.

[0057] Optionally, during the photovoltaic panel cleaning process, based on the collected second images and the real-time location information at the corresponding collection time, the locations of photovoltaic panels with abnormal conditions are marked on a pre-established photovoltaic panel array map. The abnormalities may involve dirt accumulation, photovoltaic panel damage, etc. Marking the abnormal locations helps with subsequent targeted treatment.

[0058] Optionally, the pollution level can refer to the degree index obtained by quantitatively assessing the pollution status of each area based on the abnormal information marked in the photovoltaic panel array map, taking into account factors such as the coverage area, type, and distribution of dirt. For example, it can be classified as heavy pollution, moderate pollution, or light pollution.

[0059] Optionally, the robot actuator performs standard cleaning actions according to the first cleaning path and dynamically updated photovoltaic panel cleaning parameters.

[0060] The technical solution of this invention determines whether standard cleaning conditions are met based on the first image collected by the vision sensor each time a stain detection time is determined. When the standard cleaning conditions are met, a first cleaning path is determined based on the robot's initial position collected by the position sensor. During the cleaning of the photovoltaic panels, the cleaning parameters are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time position information collected by the position sensor. This method combines multi-sensor data to make photovoltaic panel cleaning decisions, fully considers the importance of rainfall for photovoltaic panel cleaning, has high system robustness, and can dynamically adjust cleaning parameters to adapt to complex environments, improve cleaning efficiency, and avoid energy waste. Example

[0061] Figure 2 This is a flowchart of a photovoltaic panel cleaning method provided in Embodiment 2 of the present invention. This embodiment specifically describes the photovoltaic panel cleaning method based on the above embodiments. Figure 2 As shown, the method includes: S210. Whenever the stain detection time is determined, the first image is processed by the convolutional neural network in the deep learning algorithm model to obtain the type and level of dirt in the first image.

[0062] Optionally, a convolutional neural network can be pre-trained. The training method includes: calling a simulation function to generate simulated visual data, which is a 224x224 RGB image, and labeling the dirt type and dirt level corresponding to each visual data to generate a training set; and using the training set to train the convolutional neural network.

[0063] S220. Determine whether the standard cleaning conditions are met based on the type and level of dirt. If yes, proceed to step S230; otherwise, proceed to step S240.

[0064] S230. When it is determined that the standard cleaning conditions are met, the first cleaning path is determined based on the robot's initial position collected by the position sensor; step S250 is executed.

[0065] S240. When it is determined that the cleaning conditions are not met, record the stain detection time and do not perform the cleaning operation.

[0066] S250. During the photovoltaic panel cleaning process, based on the second images collected during the photovoltaic panel cleaning process and the real-time location information at the time of collection of each second image, abnormal locations are marked in the pre-established photovoltaic panel array map.

[0067] S260. The convolutional neural network in the deep learning algorithm model is used to process each second image to determine and mark the abnormal information of each abnormal location in the photovoltaic panel array map.

[0068] The abnormal information includes whether the abnormality type is dirt or damage; when the abnormality type is dirt, the abnormality information also includes the dirt type and dirt level; when the abnormality type is damage, the abnormality information also includes the damage level.

[0069] Optionally, damage warnings can be generated and sent to management and maintenance personnel based on the damage marked in the photovoltaic panel array map.

[0070] S270. Based on the abnormal information marked in the photovoltaic panel array map, determine the pollution level of each area, and dynamically update the photovoltaic panel cleaning parameters according to the pollution level of each area and real-time rainfall data through a deep learning algorithm model.

[0071] S280. After confirming that the photovoltaic panel cleaning is complete, adjust the dirt level in the photovoltaic panel array map, and determine the stubborn dirt area based on the adjusted dirt level.

[0072] Optionally, after completing a standard photovoltaic panel cleaning, the dirt level can be dynamically reduced. For example, the level can be uniformly lowered by 2 based on the original dirt level. However, for stubborn dirt with a high dirt level, the level may be reduced from level 5 to level 3, but it still does not meet the standard of being clean. Therefore, after each cleaning, areas with a dirt level greater than a preset second dirt level can be identified as stubborn dirt areas. The second dirt level is a preset value, such as level 2. If there are still areas with a dirt level greater than the second dirt level in the photovoltaic panel array map after each cleaning, then these areas are identified as stubborn dirt areas.

[0073] S290: Acquire rainfall data monitored periodically by the rain sensor, and when the rainfall reaches the preset first rainfall range, perform a cleaning action on the stubborn dirt area.

[0074] Optionally, when the rainfall reaches the preset first rainfall range, it indicates that the washing force of the rain is strong. At this time, cleaning stubborn dirt areas consumes less energy and has a better cleaning effect. The first rainfall range is a preset range. For example, when the rain sensor detects continuous rainfall >1mm / h and <5mm / h, it is determined that the first rainfall range has been reached.

[0075] Performing cleaning actions on stubborn dirt areas may include: The second cleaning path is determined based on the stubborn dirt areas in the photovoltaic panel array map using a deep learning algorithm model. Using a deep learning algorithm model, the cleaning parameters of the photovoltaic panels are dynamically updated during the cleaning process based on the dirt level of each stubborn dirt area in the photovoltaic panel array map and the real-time rainfall data collected by the rain sensor. After confirming that the stubborn dirt areas have been cleaned, the dirt level of the stubborn dirt areas in the photovoltaic panel array map is adjusted.

[0076] Understandably, when cleaning stubborn dirt areas, the robot typically cleans specific sections of the photovoltaic panels within the area. In this case, a suitable secondary cleaning path can be planned, allowing the robot to clean all stubborn dirt areas within a shorter distance.

[0077] The method may further include: after each solar panel cleaning is completed, obtaining the robot's energy consumption and the solar panel's power, and determining a reward function based on the robot's energy consumption and the solar panel's power; The reward function is used to perform reinforcement learning training on the photovoltaic panel cleaning parameter calculation sub-model in the deep learning algorithm model; wherein, the photovoltaic panel cleaning parameters include the wheel speed of each walking wheel of the photovoltaic panel cleaning robot, the rotation speed of the brush, and the downward pressure of the brush.

[0078] The advantage of this setup is that by further training the deep learning algorithm model using the robot's energy consumption and the photovoltaic panel's power during each cleaning cycle, the cleaning parameters of the photovoltaic panel output for the next cleaning cycle can be optimized, enabling cleaning with lower robot power consumption, thereby saving robot energy while increasing power generation.

[0079] Optionally, the reward function can be expressed by the following formula: R=α·△P-β·E; where R is the reward function, △P is the difference between the photovoltaic panel power after cleaning and the photovoltaic panel power before cleaning, in kW·h, E is the robot energy consumption for this cleaning, and α and β are weighting coefficients. The weighting coefficients are used to adjust the priority of power generation gain and energy consumption cost. For example, in water-scarce areas, the weight of water resource consumption can be appropriately reduced.

[0080] Optionally, the robot's energy consumption can be obtained through the energy management unit in the photovoltaic cleaning robot. The energy management unit can be a solar photovoltaic panel, which can be used to power the photovoltaic cleaning robot and can feed back the electrical energy consumed in each cleaning process to the robot control center.

[0081] Optionally, historical cleaning records, meteorological data, and a database of dirt images can be used as training data to train the deep learning algorithm model at regular intervals. After each cleaning of the photovoltaic panels, a reward function can be used to train the deep learning algorithm model for reinforcement learning.

[0082] The method may further include: during the process of cleaning the photovoltaic panels, when it is determined that the rainfall reaches a preset second rainfall range, stopping the photovoltaic panel cleaning action and starting the drainage mode.

[0083] Optionally, the rain sensor can trigger priority adjustment. When the rainfall reaches the second rainfall range, it can be determined that the rainfall is too heavy and does not meet the safe cleaning conditions. The cleaning action of the photovoltaic panel will be stopped and the drainage mode will be started. For example, when the rain sensor detects continuous rainfall >10mm / h, the cleaning will be paused and the drainage mode will be started.

[0084] Optionally, the method may further include: whenever snowfall is detected, determining that standard cleaning conditions are met, and performing standard cleaning actions.

[0085] The technical solution of this invention determines whether standard cleaning conditions are met based on the first image collected by the vision sensor each time a stain detection time is determined. When the standard cleaning conditions are met, a first cleaning path is determined based on the robot's initial position collected by the position sensor. During the cleaning of the photovoltaic panels, the cleaning parameters are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time position information collected by the position sensor. This method combines multi-sensor data to make photovoltaic panel cleaning decisions, fully considers the importance of rainfall for photovoltaic panel cleaning, has high system robustness, and can dynamically adjust cleaning parameters to adapt to complex environments, improve cleaning efficiency, and avoid energy waste. Example

[0086] Figure 3 This is a schematic diagram of a photovoltaic panel cleaning device provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes: a standard cleaning condition judgment module 310, a first cleaning path determination module 320, and a cleaning parameter update module 330.

[0087] The standard cleaning condition judgment module 310 is used to determine whether the standard cleaning conditions are met based on the first image collected by the vision sensor whenever the stain detection time is determined.

[0088] The first cleaning path determination module 320 is used to determine the first cleaning path based on the robot's initial position collected by the position sensor when it is determined that the standard cleaning conditions are met.

[0089] The cleaning parameter update module 330 is used to dynamically update the photovoltaic panel cleaning parameters during the photovoltaic panel cleaning process based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time location information collected by the position sensor.

[0090] The technical solution of this invention determines whether standard cleaning conditions are met based on the first image collected by the vision sensor each time a stain detection time is determined. When the standard cleaning conditions are met, a first cleaning path is determined based on the robot's initial position collected by the position sensor. During the cleaning of the photovoltaic panels, the cleaning parameters are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time position information collected by the position sensor. This method combines multi-sensor data to make photovoltaic panel cleaning decisions, fully considers the importance of rainfall for photovoltaic panel cleaning, has high system robustness, and can dynamically adjust cleaning parameters to adapt to complex environments, improve cleaning efficiency, and avoid energy waste.

[0091] Based on the above embodiments, the standard cleaning condition judgment module 310 can be specifically used for: The first image is processed by a convolutional neural network in a deep learning algorithm model to obtain the type and level of dirt in the first image. Based on the type and level of dirt, determine whether the standard cleaning conditions are met; If the cleaning conditions are not met, record the stain detection time and do not perform the cleaning operation.

[0092] Based on the above embodiments, the cleaning parameter update module 330 can be specifically used for: Based on the second images collected during the photovoltaic panel cleaning process and the real-time location information at the time of collection of each second image, abnormal locations are marked in the pre-established photovoltaic panel array map. The convolutional neural network in the deep learning algorithm model is used to process each second image to determine and mark the abnormal information of each abnormal location in the photovoltaic panel array map. The abnormal information includes whether the abnormality type is dirt or damage. When the abnormality type is dirt, the abnormality information also includes the dirt type and dirt level. When the abnormality type is damage, the abnormality information also includes the damage level. Based on the anomaly information marked in the photovoltaic panel array map, the pollution level of each area is determined, and the photovoltaic panel cleaning parameters are dynamically updated according to the pollution level of each area and real-time rainfall data through a deep learning algorithm model.

[0093] Based on the above embodiments, the cleaning parameter update module 330 can be further specifically used for: By using a multimodal fusion model in a deep learning algorithm, the pollution level and real-time rainfall data of each region are fused, and the photovoltaic panel cleaning parameters of each region are determined based on the fusion results. The fusion result is G(L,R)=w1·G L + w2·GR L represents the pollution level of the target area, R represents the real-time rainfall, and G represents the pollution level of the target area. L G represents the pollution level characteristics of the target area. R The fusion features of real-time rainfall data are represented by w1 and w2, which are the weights of the pollution level fusion features and the real-time rainfall data fusion features of the target area, respectively.

[0094] Based on the above embodiments, a stubborn dirt treatment module may also be included, comprising: The stubborn dirt area identification unit is used to adjust the dirt level in the photovoltaic panel array map after the photovoltaic panel cleaning is completed, and to identify the stubborn dirt area based on the adjusted dirt level. The stubborn dirt cleaning unit is used to acquire rainfall data monitored by the rain sensor at regular intervals, and when it is determined that the rainfall has reached the preset first rainfall range, it performs the cleaning action on the stubborn dirt area.

[0095] Based on the above embodiments, the stubborn dirt cleaning unit can be specifically used for: The second cleaning path is determined based on the stubborn dirt areas in the photovoltaic panel array map using a deep learning algorithm model. Using a deep learning algorithm model, the cleaning parameters of the photovoltaic panels are dynamically updated during the cleaning process based on the dirt level of each stubborn dirt area in the photovoltaic panel array map and the real-time rainfall data collected by the rain sensor. After confirming that the stubborn dirt areas have been cleaned, the dirt level of the stubborn dirt areas in the photovoltaic panel array map is adjusted.

[0096] Based on the above embodiments, a model reinforcement learning module may also be included, for: After each solar panel cleaning is completed, the robot's energy consumption and the solar panel's power are obtained, and a reward function is determined based on the robot's energy consumption and the solar panel's power. The reward function is used to perform reinforcement learning training on the photovoltaic panel cleaning parameter calculation sub-model in the deep learning algorithm model; wherein, the photovoltaic panel cleaning parameters include the wheel speed of each walking wheel of the photovoltaic panel cleaning robot, the rotation speed of the brush, and the downward pressure of the brush.

[0097] Based on the above embodiments, the reward function is expressed by the following formula: R = α·△P - β·E; Where R is the reward function, ΔP is the difference between the photovoltaic panel power after cleaning and the photovoltaic panel power before cleaning, E is the robot energy consumption for this cleaning, and α and β are weighting coefficients, which are used to adjust the priority of power generation gain and energy consumption cost.

[0098] The photovoltaic panel cleaning device provided in this embodiment of the invention can execute the photovoltaic panel cleaning method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method. Example

[0099] Figure 4 This is a structural schematic diagram of a photovoltaic cleaning robot provided in Embodiment 4 of the present invention. Figure 4 As shown, the photovoltaic cleaning robot includes: a vision sensor 410, a position sensor 420, a rain sensor 430, a robot control center 440, and a robot actuator 450.

[0100] The vision sensor 410 is used to capture images of the photovoltaic panel and send the images to the robot control center.

[0101] The position sensor 420 is used to determine the position of the photovoltaic cleaning robot in the photovoltaic panel array map and send the position information to the robot control center.

[0102] Rain sensor 430 is used to acquire rainfall data and send the rainfall data to the robot control center.

[0103] The robot control center 440 is used to execute the photovoltaic panel cleaning method described in any embodiment of the present invention.

[0104] The robot actuator 450 is used to perform photovoltaic panel cleaning operations according to the cleaning path and cleaning parameters sent by the robot control center.

[0105] Optionally, the photovoltaic cleaning robot body adopts a waterproof design, and the visual sensor 410 is installed on the adjustable gimbal of the photovoltaic cleaning robot, which can effectively prevent the shaking during the movement of the photovoltaic cleaning robot and ensure the clarity of the photos taken by the visual sensor 410. The visual sensor 410 may include a camera and an infrared sensor.

[0106] Optionally, the position sensor 420 can be a Global Positioning System (GPS), an Inertial Measurement Unit (IMU), or an encoder, and the position sensor 420 can be configured at the center of the photovoltaic cleaning robot.

[0107] Optionally, the rain sensor 430 is integrated on the top of the photovoltaic cleaning robot and can be used to monitor rainfall and humidity in real time.

[0108] Optionally, the photovoltaic cleaning robot is equipped with an embedded chip, such as the NVIDIA Jetson series chip, which supports TensorRT accelerated inference. The robot control center 440 can perform control functions based on the embedded chip to achieve high-speed data processing.

[0109] Optionally, the robot actuator 450 may include a cleaning mechanism 451, a mobile chassis 452, and an energy management unit 453. The cleaning mechanism 451 may include, but is not limited to, a rotating brush and a vacuuming device. The mobile chassis 452 may be a wheeled chassis or a tracked chassis. The energy management unit 453 may be a solar photovoltaic panel. The energy management unit 453 may be used to power the photovoltaic cleaning robot and may be able to feed back the electrical energy consumed in each cleaning process to the robot control center 440.

[0110] Figure 5 This is a schematic diagram illustrating the operation of an optional photovoltaic panel cleaning robot. Figure 5 As shown, the front end of the photovoltaic panel cleaning robot is equipped with a vision sensor, which can collect images of the photovoltaic panel surface in the direction the robot is moving. The top of the photovoltaic panel cleaning robot is equipped with a rain sensor, which can accurately obtain rainfall data. The photovoltaic panel cleaning robot is also equipped with a position sensor. The control execution module, preprocessing module and deep learning algorithm module are integrated together to form the robot control center of the photovoltaic panel cleaning robot.

[0111] The technical solution of this invention, by configuring a vision sensor, position sensor, rain sensor, robot control center and robot actuator in the photovoltaic cleaning robot, can make photovoltaic panel cleaning decisions by combining multi-sensor data, fully consider the importance of rainfall for photovoltaic panel cleaning, the system has high robustness, can dynamically adjust cleaning parameters, can dynamically adapt to complex environments, improve cleaning efficiency and avoid energy waste. Example

[0112] Figure 6 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0113] like Figure 6As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0114] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0115] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the photovoltaic panel cleaning method as described in any embodiment of the present invention.

[0116] In some embodiments, the photovoltaic panel cleaning method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the photovoltaic panel cleaning method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the photovoltaic panel cleaning method by any other suitable means (e.g., by means of firmware).

[0117] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0118] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for cleaning photovoltaic panels, characterized in that, Performed by the robot control center in the photovoltaic panel cleaning robot, including: Whenever the stain detection time is determined, the system uses the first image captured by the vision sensor to determine whether the standard cleaning conditions are met. When the standard cleaning conditions are met, the first cleaning path is determined based on the robot's initial position collected by the position sensor; During the cleaning process of the photovoltaic panels, the cleaning parameters of the photovoltaic panels are dynamically updated based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time location information collected by the position sensor.

2. The method according to claim 1, characterized in that, Based on the first image captured by the vision sensor, determine whether the standard cleaning conditions are met, including: The first image is processed by a convolutional neural network in a deep learning algorithm model to obtain the type and level of dirt in the first image. Based on the type and level of dirt, determine whether the standard cleaning conditions are met; If the cleaning conditions are not met, record the stain detection time and do not perform the cleaning operation.

3. The method according to claim 1, characterized in that, Based on the second images periodically acquired by the vision sensor, the real-time rainfall data acquired by the rain sensor, and the real-time location information acquired by the position sensor, the photovoltaic panel cleaning parameters are dynamically updated, including: Based on the second images collected during the photovoltaic panel cleaning process and the real-time location information at the time of collection of each second image, abnormal locations are marked in the pre-established photovoltaic panel array map. The convolutional neural network in the deep learning algorithm model is used to process each second image to determine and mark the abnormal information of each abnormal location in the photovoltaic panel array map. The abnormal information includes whether the abnormality type is dirt or damage. When the abnormality type is dirt, the abnormality information also includes the dirt type and dirt level. When the abnormality type is damage, the abnormality information also includes the damage level. Based on the anomaly information marked in the photovoltaic panel array map, the pollution level of each area is determined, and the photovoltaic panel cleaning parameters are dynamically updated according to the pollution level of each area and real-time rainfall data through a deep learning algorithm model. The photovoltaic panel cleaning parameters are dynamically updated based on the pollution level of each area and real-time rainfall data, including: By using a multimodal fusion model in a deep learning algorithm, the pollution level and real-time rainfall data of each region are fused, and the photovoltaic panel cleaning parameters of each region are determined based on the fusion results. The fusion result is G(L,R)=w1·G L + w2·G R L represents the pollution level of the target area, R represents the real-time rainfall, and G represents the pollution level of the target area. L G represents the pollution level characteristics of the target area. R The fusion features of real-time rainfall data are represented by w1 and w2, which are the weights of the pollution level fusion features and the real-time rainfall data fusion features of the target area, respectively.

4. The method according to claim 3, characterized in that, Also includes: After confirming that the cleaning of the photovoltaic panels is complete, the dirt level in the photovoltaic panel array map is adjusted, and the stubborn dirt area is identified based on the adjusted dirt level. It acquires rainfall data monitored periodically by a rain sensor, and when the rainfall reaches a preset first rainfall range, it performs a cleaning action on stubborn dirt areas.

5. The method according to claim 4, characterized in that, Perform cleaning actions on stubborn dirt areas, including: The second cleaning path is determined based on the stubborn dirt areas in the photovoltaic panel array map using a deep learning algorithm model. Using a deep learning algorithm model, the cleaning parameters of the photovoltaic panels are dynamically updated during the cleaning process based on the dirt level of each stubborn dirt area in the photovoltaic panel array map and the real-time rainfall data collected by the rain sensor. After confirming that the stubborn dirt areas have been cleaned, the dirt level of the stubborn dirt areas in the photovoltaic panel array map is adjusted.

6. The method according to claim 3, characterized in that, The process of processing each second image using a convolutional neural network in a deep learning algorithm model to determine and mark the abnormal information at each abnormal location in the photovoltaic panel array map includes the following steps: S1. Preprocess the second image and locate it to the corresponding coordinate area in the photovoltaic panel array map based on the real-time location information when the second image was acquired; S2. Input the preprocessed second image into a pre-trained multi-task convolutional neural network model for processing; The multi-task convolutional neural network model includes a segmentation network with an encoder-decoder structure, which outputs a pixel-level segmentation mask for the second image. The segmentation mask is used to identify the background, dirt areas, and suspected damaged areas in the image. S3. For each independent region identified in the segmentation mask, extract its visual features, and determine whether its abnormality type belongs to dirt or damage based on the extracted visual features; When something is identified as dirt, the type and level of dirt are determined based on its visual characteristics. When an injury is determined, the level of injury is determined based on its visual characteristics. S4. Associate the anomaly information of each independent region determined in step S3 with its corresponding position coordinates in the second image, and update it in the photovoltaic array map to complete the marking of the anomaly location.

7. The method according to claim 1, characterized in that, The step of determining the first cleaning path based on the robot's initial position collected by the position sensor includes: The initial position coordinates of the robot relative to a pre-established photovoltaic panel array map are obtained through the position sensor; Based on the initial position coordinates, obtain or generate a traversal path that covers the preset standard cleaning range, and use it as the first cleaning path; The step of obtaining or generating a traversal path covering a preset standard cleaning range based on the initial position coordinates includes: Determine whether the initial position coordinates are one of multiple fixed starting positions pre-stored in the path library; If so, then the pre-stored optimal traversal path corresponding to the initial position coordinates is called from the path library as the first cleaning path; If not, then based on the initial position coordinates, the shortest traversal path covering the preset standard cleaning range is calculated in real time and generated as the first cleaning path.

8. The method according to claim 7, characterized in that, The real-time calculation and generation of the shortest traversal path covering the preset standard cleaning range includes: The preset standard cleaning range is discretized into multiple grids in the photovoltaic panel array map, where grids without obstacles are marked as passable areas and grids with obstacles are marked as impassable areas; Starting from the grid where the robot's initial position is located, the improved A-Star algorithm is used to perform a global path search in the grid of the passable area to generate a path that traverses all passable grids and has the shortest total distance, which is used as the first cleaning path. Wherein, the cost function of the improved A-Star algorithm is F(n)=G(n)+H(n)+λ*T(n), where G(n) is the actual cost from the starting point to the current node n, H(n) is the estimated cost from the current node n to the end point, T(n) is the average distance cost from the current node n to the untraversed passable grid, and λ is the balance coefficient.

9. A photovoltaic panel cleaning device, characterized in that, Performed by the robot control center in the photovoltaic panel cleaning robot, including: The standard cleaning condition judgment module is used to determine whether the standard cleaning conditions are met based on the first image collected by the vision sensor whenever the stain detection time is determined. The first cleaning path determination module is used to determine the first cleaning path based on the robot's initial position collected by the position sensor when it is determined that the standard cleaning conditions are met. The cleaning parameter update module is used to dynamically update the photovoltaic panel cleaning parameters during the cleaning process based on the second image collected periodically by the vision sensor, the real-time rainfall data collected by the rain sensor, and the real-time location information collected by the position sensor.

10. A photovoltaic cleaning robot, characterized in that, This includes visual sensors, position sensors, rain sensors, a robot control center, and robot actuators; among which, The vision sensor is used to capture images of the photovoltaic panel and send the images to the robot control center; The position sensor is used to determine the position of the photovoltaic cleaning robot in the photovoltaic panel array map and send the position information to the robot control center; The rain sensor is used to acquire rainfall data and send the rainfall data to the robot control center; The robot control center is used to execute the photovoltaic panel cleaning method according to any one of claims 1-8; The robot actuator is used to perform photovoltaic panel cleaning operations according to the cleaning path and cleaning parameters sent by the robot control center.

Images