Multi-angle cleaning brush head device of photovoltaic module cleaning robot

Abstract

The invention provides a multi-angle cleaning brush head device of a photovoltaic module cleaning robot, and belongs to the technical field of photovoltaic module cleaning, and the multi-angle cleaning brush head device comprises a brush head main body device, a multi-angle adjusting mechanism device, a self-cleaning system device and an environment sensing device; the brush head main body device comprises a cleaning unit array and nano composite material bristles, the multi-angle adjusting mechanism device comprises a three-axis joint system, the self-cleaning system device comprises a high-pressure water pump, a micro atomizing nozzle and a centrifugal fan, and the environment sensing device comprises various sensors. In the control aspect, intelligentization is achieved through ten steps of processes including system initialization, preparation before cleaning, posture pre-adjustment, intelligent cleaning operation, real-time monitoring, self-cleaning execution, parameter optimization, fault early warning and self-recovery, quality evaluation and self-learning optimization, a CleanFormer deep learning model is adopted, self-adaptive cleaning aiming at different environment conditions and dirt characteristics is achieved, and the intelligent cleaning system has the advantages of being simple in structure, convenient to operate and high in practicability. The efficiency is improved; and the photovoltaic panel is protected.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic module cleaning, and more particularly, relates to a multi-angle cleaning brush head device for a photovoltaic module cleaning robot. Background Art

[0002] Photovoltaic power generation, as an important renewable energy technology, its power generation efficiency is closely related to the cleanliness of the module surface. Traditional photovoltaic module cleaning technologies mainly include manual cleaning, fixed automatic cleaning systems, and simple robot cleaning systems. In actual application scenarios, these traditional technologies usually adopt an operation mode with a fixed brush head angle, a single brushing force, and a unified water spraying amount, which are suitable for standard photovoltaic arrays with flat surfaces and uniform dirt distribution.

[0003] However, traditional technologies face various defects: First, the fixed brush head angle cannot adapt to photovoltaic modules with different installation inclinations; Second, a single brushing force is likely to cause damage to the photovoltaic panel when dealing with stubborn dirt, while excessive energy is consumed when cleaning slightly dirty areas; Third, the unified water spraying amount cannot be optimized according to the dirt degree in different areas, resulting in waste of water resources; Finally, the lack of self-cleaning function makes it difficult to ensure the cleanliness of the brush head itself, reducing the continuous cleaning efficiency.

[0004] Therefore, it can be seen that the existing technology is difficult to solve the technical problem that the brush head system of a photovoltaic module cleaning robot cannot adaptively adjust cleaning parameters according to environmental conditions and dirt characteristics, which not only affects the cleaning efficiency and quality, but also may increase the risk of photovoltaic module loss and cause unnecessary waste of energy and water resources. Summary of the Invention

[0005] In view of this, the present invention provides a multi-angle cleaning brush head device for a photovoltaic module cleaning robot, which can solve the technical problem that the brush head system of a photovoltaic module cleaning robot in the prior art cannot adaptively adjust cleaning parameters according to environmental conditions and dirt characteristics.

[0006] The present invention is implemented as follows:

[0007] The present invention provides a multi-angle cleaning brush head device for a photovoltaic module cleaning robot, which includes a control chip, a brush head main body device, a multi-angle adjustment mechanism device, a self-cleaning system device, an environment perception device, a communication device, a power management device, a fault diagnosis device, and a human-machine interaction device. The control chip is electrically connected to other devices respectively. An automatic cleaning control module is provided in the control chip. The automatic cleaning control module is used to execute steps such as system initialization, preparation before cleaning, pre-adjustment of the brush head posture, intelligent cleaning operation, real-time monitoring of the cleaning effect, execution of the brush head self-cleaning, optimization of adaptive cleaning parameters, fault warning and self-recovery, evaluation of the cleaning quality, and system self-learning optimization. By introducing a multi-parameter cleaning resistance prediction function for cleaning resistance evaluation and optimization, and a pre-trained CleanFormer deep learning model for parameter optimization and self-learning process optimization, intelligent and efficient cleaning is achieved and damage to the photovoltaic panel is prevented.

[0008] Among them, the execution steps of the automatic cleaning control module include: system initialization, receiving a cleaning task instruction and judging whether the environmental conditions meet the requirements of the cleaning operation; preparation before cleaning, calculating the optimal cleaning water temperature and presetting the cleaning intensity parameters; pre-adjustment of the brush head posture, adjusting the initial posture of the brush head; intelligent cleaning operation, performing cleaning and dynamically adjusting the pressure of the cleaning unit; real-time monitoring of the cleaning effect, drawing a heat map of the cleaning resistance distribution; execution of the brush head self-cleaning, triggering the brush head self-cleaning program; optimization of adaptive cleaning parameters, dynamically adjusting the cleaning parameters; fault warning and self-recovery, detecting abnormalities and performing fault handling; evaluation of the cleaning quality, calculating the cleaning efficiency and quality score; system self-learning optimization, optimizing the cleaning strategy model.

[0009] Further, the pre-adjustment of the brush head posture is specifically that, according to the actual installation angle of the photovoltaic module and the cleaning path planning, the initial posture of the brush head is adjusted through the multi-angle adjustment mechanism device. Specifically, through the coordinated movement of the horizontal rotation joint, the vertical pitching joint, and the torsion joint, the brush head main body device is kept at the best contact angle with the surface of the photovoltaic panel. At the same time, the joint angle data is fed back in real time through the angle sensor array to ensure accurate posture adjustment.

[0010] Further, the brush head main body device includes a cleaning unit array, an independent drive motor array, nano-composite material bristles, a dirt collection channel, and a current sensing circuit. The nano-composite material bristles are composed of a high-strength carbon fiber skeleton, a high-elastic polyurethane matrix, and a nano-Ag + ion-modified microporous surface layer.

[0011] Further, the intelligent cleaning operation is specifically as follows: start the independent drive motor array in the brush head main body device to drive the nano-composite material bristles to rotate for cleaning. At the same time, calculate the cleaning resistance according to the change of the motor operating current collected by the current sensing circuit, and combine the joint load torque data monitored by the torque sensor array to dynamically adjust the pressure applied by the cleaning unit on the photovoltaic panel. If the cleaning resistance suddenly increases, the system automatically reduces the pressure of the corresponding cleaning unit and slows down the rotation speed of the bristles to prevent damage to the photovoltaic panel.

[0012] Further, the real-time monitoring of the cleaning effect is specifically as follows: according to the motor operating current data collected by the current sensing circuit, draw a heat map of the cleaning resistance distribution to identify the areas with uneven dirt distribution, and increase the cleaning times for the areas with more dirt. At the same time, analyze the deformation data of the bristles to analyze the contact situation between the bristles and the panel surface to ensure that the cleaning coverage rate reaches 100%.

[0013] Further, the self-cleaning of the brush head is specifically as follows: when the data collected by the current sensing circuit indicates that the cleaning resistance continues to increase or the turbidity sensor monitors that the turbidity of the water after separation exceeds the threshold, the system automatically triggers the self-cleaning program of the brush head, including stopping the cleaning operation and moving the brush head main body device to a safe position, starting the high-pressure water pump and the centrifugal fan to separate the sewage and dirt, collecting the dirt in the sewage collection box, and recycling the clean water back to the water tank.

[0014] Further, the optimization of the adaptive cleaning parameters is specifically as follows: comprehensively analyze the environmental parameters collected by the environmental perception device, the cleaning resistance data collected by the current sensing circuit, and the water quality cleanliness data monitored by the turbidity sensor, establish a multi-parameter correlation model, and dynamically adjust the cleaning water temperature, water injection intensity, cleaning time arrangement, cleaning unit pressure, and bristle rotation speed.

[0015] Further, the intelligent cleaning operation stage is specifically as follows: introduce a multi-parameter cleaning resistance prediction function for cleaning resistance evaluation and optimization. The multi-parameter cleaning resistance prediction function is used to predict the cleaning resistance distribution before the start of the intelligent cleaning operation. The inputs include the surface temperature of the photovoltaic panel, environmental humidity, dirt adhesion time, dust particle size distribution data, and the contact angle data between the bristles and the panel. The outputs are the predicted cleaning resistance values and their spatial distribution heat map.

[0016] Furthermore, the CleanFormer deep learning model is a multi-modal deep learning network based on the Transformer architecture, which includes four main components: an environmental perception encoder, a cleaning parameter encoder, a temporal information encoder, and a parameter optimization decoder. The environmental perception encoder is responsible for processing multi-dimensional environmental data from infrared thermal imagers, humidity sensors, light sensors, wind speed sensors, and dust concentration sensors, and mapping it to a unified feature space; the sparsity parameter in the adaptive sparse attention mechanism of the CleanFormer deep learning model is determined by three key parameters: the surface temperature of the photovoltaic panel, the environmental humidity, and the dirt distribution uniformity.

[0017] Compared with the prior art, the beneficial effects of the multi-angle cleaning brush head device of a photovoltaic module cleaning robot provided by the present invention are as follows: The present invention provides a multi-angle self-cleaning brush head system for a photovoltaic module cleaning robot. Through the collaborative work of a multi-angle adjustment mechanism device, a self-cleaning system device, an environmental perception device, and an intelligent control system, the adaptive adjustment of cleaning parameters is realized. This system constructs a multi-parameter correlation model based on the CleanFormer deep learning model, and dynamically optimizes the brush head posture, brushing force, water spray volume, and cleaning path according to the real-time collected environmental data and dirt characteristics; compared with the traditional technology, the present invention solves the above-mentioned defects: First, the multi-angle adjustment mechanism realizes the best contact angle between the brush head and the photovoltaic panel; Second, the intelligent control based on current sensing and torque sensing realizes the precise adjustment of the brushing force, avoiding damage to the photovoltaic panel; Third, the micro atomizing nozzle array combined with the intelligent control system realizes the on-demand regulation of the water volume and water droplet diameter, saving water resources; Finally, the self-cleaning system device ensures that the brush head continuously maintains a highly efficient cleaning state; Therefore, the present invention solves the technical problem that the brush head system of a photovoltaic module cleaning robot cannot adaptively adjust cleaning parameters according to environmental conditions and dirt characteristics, not only improving the cleaning efficiency and quality, but also reducing the risk of damage to photovoltaic modules, and realizing the optimized utilization of energy and water resources. Brief Description of the Drawings

[0018] Figure 1 It is a schematic diagram of the composition of a multi-angle cleaning brush head device of a photovoltaic module cleaning robot;

[0019] Figure 2 It is a schematic cross-sectional view of the brush head main body device of the present invention;

[0020] Figure 3 It is a schematic diagram of the cleaning unit structure of the present invention;

[0021] Figure 4 It is a flow chart of the present invention;

[0022] Figure 5 It is an example diagram of the multi-angle adjustment mechanism device. Detailed implementation manners

[0023] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] As Figure 1 shown, it is a schematic diagram of the composition of a multi-angle cleaning brush head device of a photovoltaic module cleaning robot provided by the present invention. This device includes a control chip, a brush head main body device, a multi-angle adjustment mechanism device, a self-cleaning system device, an environment perception device, a communication device, a power management device, a fault diagnosis device, and a human-computer interaction device. The control chip is electrically connected to the brush head main body device, the multi-angle adjustment mechanism device, the self-cleaning system device, the environment perception device, the communication device, the power management device, the fault diagnosis device, and the human-computer interaction device respectively. An automatic cleaning control module is provided in the control chip. The brush head main body device is used to perform the cleaning operation on the surface of the photovoltaic panel and collect the cleaning resistance data and the bristles deformation data. The multi-angle adjustment mechanism device is used to adjust the brush head posture and collect the joint angle data and the joint load torque data. The self-cleaning system device is used to clean the bristles, recover the dirt, and collect the water pressure data, the suction data, and the dirt concentration data. The environment perception device is used to collect the surface temperature of the photovoltaic panel, the ambient humidity, the light intensity, the wind speed, and the dust concentration data. The communication device is used to realize the data exchange with the main control system of the cleaning robot and the remote monitoring center. The power management device is used to monitor the system power supply voltage and the current consumption of each executing component. The fault diagnosis device is used to collect the system vibration frequency, abnormal sound, and the motor temperature data. The human-computer interaction device is used to display the system working state and receive operation instructions. The sampling frequency of the cleaning unit in the brush head main body device is 20 times per second. The sampling frequency of the multi-angle adjustment mechanism device is 10 times per second. The sampling frequency of the self-cleaning system device is 5 times per second. The sampling frequency of the environment perception device is 1 time per minute.

[0025] As Figure 2 、 Figure 3 shown, the brush head main body device includes a cleaning unit array, an independent drive motor array, nano-composite material bristles, a dirt collection channel, and a current sensing circuit. The cleaning unit array is composed of multiple detachable and independently working cleaning units. Each cleaning unit is connected to the adjacent cleaning unit through a snap connection structure. Each drive motor in the independent drive motor array is fixed inside the corresponding cleaning unit and is used to drive the nano-composite material bristles to rotate. The nano-composite material bristles are composed of a high-strength carbon fiber skeleton, a high-elastic polyurethane matrix, and a nano-Ag + ion-modified microporous surface layer. The dirt collection channel runs through the inside of the brush head main body and is connected to the self-cleaning system device. The current sensing circuit is arranged in the power supply circuit of each drive motor to monitor the motor running current.

[0026] As Figure 5 shown, the multi-angle adjustment mechanism device includes a horizontal rotation joint, a vertical pitch joint, a torsion joint, a rotation motor, an electric push rod, a harmonic reducer, an angle sensor array, and a torque sensor array. The horizontal rotation joint is driven by the rotation motor to achieve the rotational movement of the brush head main body device in the horizontal direction. The vertical pitch joint is controlled by the electric push rod to achieve the pitching movement of the brush head main body device in the vertical direction. The torsion joint is driven by the harmonic reducer to achieve the torsional movement of the brush head main body device around its own central axis. The angle sensor array is respectively installed at the three joints to measure the joint angles in real time. The torque sensor array is respectively installed on the output shafts of the driving devices of the three joints to monitor the joint load conditions.

[0027] The self-cleaning system device includes a high-pressure water pump, a water tank, a micro atomizing nozzle array, a centrifugal fan, a dirt separator, a cyclone separation chamber, a filter screen assembly, a dirt collection box, a pressure sensor, a flow sensor, and a turbidity sensor. The high-pressure water pump is connected to the water tank and supplies water to the micro atomizing nozzle array. The micro atomizing nozzle array is arranged on the surface of the brush head main body device and is used to spray atomized water droplets onto the nano-composite material bristles. The centrifugal fan is connected to the dirt collection channel and is used to generate suction to suck sewage and dirt into the dirt separator. The dirt separator includes the cyclone separation chamber and the filter screen assembly and is used to separate sewage and dirt. The separated dirt is collected in the dirt collection box, and the clean water flows back to the water tank for recycling. The pressure sensor is installed at the outlet of the high-pressure water pump to monitor the water pressure. The flow sensor is installed at the outlet of the centrifugal fan to monitor the air flow. The turbidity sensor is installed at the outlet of the dirt separator to monitor the water quality cleanliness.

[0028] The environmental perception device includes an infrared thermal imager, a humidity sensor, a light sensor, a wind speed sensor, and a dust concentration sensor. The infrared thermal imager is used for non-contact measurement of the surface temperature distribution of the photovoltaic panel. The humidity sensor is used for measuring the relative environmental humidity. The light sensor is used for measuring the environmental light intensity. The wind speed sensor is used for measuring the environmental wind speed. The dust concentration sensor is used for measuring the concentration of dust particles in the ambient air. The communication device includes an industrial Ethernet module, a wireless communication module, and a data encryption processor. The industrial Ethernet module is connected to the main control system of the cleaning robot through a network cable. The wireless communication module uses 5G communication technology to establish a connection with the remote monitoring center. The data encryption processor encrypts the transmitted data to ensure communication security. The power management device includes a voltage monitoring circuit, a current monitoring circuit, an overcurrent protection circuit, and a power distribution circuit. The voltage monitoring circuit monitors the system power supply voltage in real time. The current monitoring circuit monitors the current consumption of each executing component in real time. The overcurrent protection circuit cuts off the power supply of the corresponding circuit when abnormal current is detected. The power distribution circuit reasonably distributes the power supply according to the working states of each component. The fault diagnosis device includes a vibration sensor, a sound sensor, and a temperature sensor array. The vibration sensor is installed on the brush head main body device to monitor the system vibration frequency. The sound sensor is used for collecting the system operation sound. The temperature sensor array is distributed on the surfaces of each motor to monitor the motor temperature. The human-computer interaction device includes a display screen, operation buttons, and an indicator light array. The display screen is used for displaying the system working state and parameters. The operation buttons are used for manually inputting operation instructions. The indicator light array is used for visually displaying the system working mode and fault warnings.

[0029] As Figure 4 shown, the automatic cleaning control module is used to execute the following steps:

[0030] S01. In the system initialization stage, receive the cleaning task instruction through the communication device, read the installation parameters of the photovoltaic module and the cleaning path planning data, and at the same time collect the current environmental parameters through the environmental perception device, and judge whether the environmental conditions meet the cleaning operation requirements. If they meet, continue to execute; otherwise, issue an environmental discomfort warning and wait for the conditions to improve or for manual intervention.

[0031] S02. In the pre-cleaning preparation stage, calculate the optimal cleaning water temperature based on the surface temperature data of the photovoltaic panel collected by the environmental perception device, and adjust the water temperature in the water tank through the self-cleaning system device; at the same time, evaluate the dirt severity according to the data collected by the dust concentration sensor, and preset the cleaning intensity parameters, including the brush rotation speed, the cleaning unit pressure, and the water spraying intensity.

[0032] S03. Pre-adjustment stage of the brush head posture. According to the actual installation angle of the photovoltaic module and the cleaning path planning, the initial posture of the brush head is adjusted through the multi-angle adjustment mechanism device. Specifically, through the coordinated movement of the horizontal rotation joint, vertical pitch joint and torsion joint, the brush head main body device is kept at the best contact angle with the surface of the photovoltaic panel. At the same time, the joint angle data is fed back in real time through the angle sensor array to ensure accurate posture adjustment;

[0033] S04. Intelligent cleaning operation stage. Start the independent drive motor array in the brush head main body device to drive the nano-composite material bristles to rotate for cleaning. At the same time, calculate the cleaning resistance according to the change of the motor running current collected by the current sensing circuit, and combine the joint load torque data monitored by the torque sensor array to dynamically adjust the pressure applied by the cleaning unit on the photovoltaic panel; if the cleaning resistance suddenly increases, the system automatically reduces the pressure of the corresponding cleaning unit and slows down the rotation speed of the bristles to prevent damage to the photovoltaic panel;

[0034] S05. Real-time monitoring stage of cleaning effect. According to the motor running current data collected by the current sensing circuit, draw a heat map of the cleaning resistance distribution to identify the areas with uneven dirt distribution, and increase the cleaning times in the areas with more dirt; at the same time, analyze the bristle deformation data to analyze the contact situation between the bristles and the panel surface to ensure that the cleaning coverage rate reaches 100%; <0,

[0035] S06. Self-cleaning execution stage of the brush head. When the data collected by the current sensing circuit shows that the cleaning resistance continues to increase or the turbidity sensor monitors that the turbidity of the separated water exceeds the threshold, the system automatically triggers the self-cleaning program of the brush head. The specific process includes: first, stop the cleaning operation and move the brush head main body device to a safe position at the edge of the photovoltaic panel, then start the high-pressure water pump, spray fine water mist on the nano-composite material bristles through the micro atomizing nozzle array for rinsing, and at the same time start the centrifugal fan to generate suction, suck the rinsed sewage and dirt into the dirt separator through the dirt collection channel, use centrifugal force to preliminarily separate water and dirt in the cyclone separation chamber, and then carry out fine filtration through the filter screen assembly. Finally, collect the separated dirt in the dirt collection box, and the clean water flows back to the water tank for recycling;

[0036] S07. Adaptive cleaning parameter optimization stage: comprehensively analyze the environmental parameters collected by the environmental perception device, the cleaning resistance data collected by the current sensing circuit, and the water quality cleanliness data monitored by the turbidity sensor, establish a multi-parameter correlation model, and dynamically adjust the following cleaning parameters: adjust the optimal cleaning water temperature according to the surface temperature of the photovoltaic panel, the higher the temperature, the lower the water temperature to avoid thermal shock; adjust the water injection intensity according to the environmental humidity, increase the water spray volume when the humidity is low to prevent dust; adjust the cleaning time arrangement according to the light intensity, avoid the noon period when the light is strong to prevent water droplets from concentrating and damaging the panel; adjust the cleaning unit pressure according to the wind speed, increase the pressure when the wind speed is large to prevent the brush head from shaking; adjust the bristle rotation speed according to the dust concentration, reduce the rotation speed when there is a lot of dust to reduce secondary pollution;

[0037] S08. Fault warning and self-recovery stage: based on the vibration frequency, abnormal sound, and motor temperature data collected by the fault diagnosis device, combined with historical operation data, construct a fault prediction model. When abnormal vibration frequency, abnormal sound, or motor temperature exceeding the threshold is detected, the system performs the following processing: if the cleaning unit is blocked, automatically increase the water pressure of the corresponding micro-atomizing nozzle and increase the suction of the centrifugal fan; if the joint is stuck, automatically perform the forward and reverse reciprocating movement of the joint to release the stuck; if the motor is overheated, automatically reduce the motor load and increase the cooling time; if the sewage collection box is nearly full, the system issues a reminder to replace the sewage collection box and marks the current cleaning position, and automatically resumes cleaning after replacement; if the fault cannot be automatically recovered, send the fault code and detailed parameters to the remote monitoring center through the communication device, and at the same time the system enters the safe mode to stop the cleaning task;

[0038] S09. Cleaning quality evaluation stage: after the cleaning operation is completed, the system calculates the cleaning efficiency and cleaning quality score by analyzing the changes in cleaning resistance, motor current consumption, and dirt collection volume before and after cleaning, and at the same time combines the data on the improvement of the power generation efficiency of the photovoltaic panel to evaluate the cleaning economic benefits;

[0039] S10. System self-learning optimization stage: store the environmental parameters, cleaning parameters, and cleaning quality evaluation data of this cleaning task in the database, analyze the historical cleaning data based on the deep learning algorithm, extract the optimal cleaning parameter combination under different environmental conditions, and continuously optimize the cleaning strategy model to guide the parameter setting of the next cleaning task.

[0040] In the step S04, a multi-parameter cleaning resistance prediction function is introduced to optimize the cleaning resistance evaluation. The multi-parameter cleaning resistance prediction function is used to predict the cleaning resistance distribution before the start of the intelligent cleaning operation. The inputs include the surface temperature of the photovoltaic panel collected by the infrared thermal imager, the ambient humidity collected by the humidity sensor, the dirt adhesion time calculated by the automatic cleaning control module, the dust particle size distribution data collected by the dust concentration sensor, and the brush and panel contact angle data collected by the angle sensor array. The output is the predicted cleaning resistance value and its spatial distribution heat map. The predicted cleaning resistance value is used by the automatic cleaning control module to initially set the pressure of each cleaning unit and the power of the drive motor. The spatial distribution heat map is used to determine the cleaning path priority and the key cleaning area.

[0041] In the steps S07 and S10, a pre-trained CleanFormer deep learning model is introduced for parameter optimization and self-learning process optimization. The sparsity parameter in the adaptive sparse attention mechanism of the CleanFormer deep learning model is determined by three key parameters: the surface temperature of the photovoltaic panel collected by the infrared thermal imager, the environmental humidity collected by the humidity sensor, and the fouling distribution uniformity calculated by the current sensing circuit. The output result of the CleanFormer deep learning model is used by the automatic cleaning control module to dynamically adjust the brush rotation speed, cleaning unit pressure, water injection intensity, water temperature, and joint movement speed. The specific structure of the CleanFormer deep learning model is a multi-modal deep learning network based on the Transformer architecture, which includes four main components: an environmental perception encoder, a cleaning parameter encoder, a temporal information encoder, and a parameter optimization decoder. The environmental perception encoder is responsible for processing multi-dimensional environmental data from the infrared thermal imager, the humidity sensor, the light sensor, the wind speed sensor, and the dust concentration sensor, and mapping it to a unified feature space. The cleaning parameter encoder is responsible for encoding the current cleaning parameter settings, including the brush rotation speed, cleaning unit pressure, water injection intensity, water temperature, and joint movement speed. The temporal information encoder captures the dynamic change features during the cleaning process and establishes a dependency relationship in the time dimension. The parameter optimization decoder generates an optimal cleaning parameter adjustment scheme based on the fused feature representation. The steps for establishing the training data set during the pre-training process of the CleanFormer deep learning model specifically include four main links: real-scene data collection, virtual-scene data generation, data annotation and preprocessing, and data augmentation. The real-scene data collection records environmental parameters, cleaning parameters, and cleaning effect evaluation data by deploying a data collection system at photovoltaic power stations in different geographical locations, different climate conditions, and different seasons, forming an original data set. The virtual-scene data generation uses a physics engine to construct a photovoltaic cleaning simulation environment, simulating the cleaning process under various extreme weather conditions and complex fouling situations, and generating boundary data that is difficult to obtain in the actual environment. The data annotation and preprocessing link annotates the corresponding cleaning efficiency and energy consumption for each set of environmental parameters and cleaning parameters, filters out abnormal data, and interpolates and supplements missing data. The data augmentation link expands the data set scale by adding Gaussian noise, random occlusion, parameter micro-perturbation, etc., to improve the generalization ability of the model.The steps for pre-training the CleanFormer deep learning model specifically include four key aspects: parameter initialization, phased training, multi-task learning, and model distillation. In the parameter initialization stage, the weights pre-trained on a large-scale industrial control dataset are used as the starting point to accelerate model convergence. In the phased training, warm-up training is first carried out on virtual scenario data to enable the model to initially master basic physical laws and parameter relationships. Then, fine-tuning is performed on real-scenario data, and finally, joint training is carried out on a mixed dataset to comprehensively improve the model performance. In multi-task learning, two objectives of cleaning efficiency and energy consumption are optimized simultaneously, and a dynamic weight adjustment mechanism is used to balance the trade-off between the two, minimizing energy consumption while achieving efficient cleaning. Model distillation transfers the knowledge of a large and complex model to a lightweight model to meet the deployment requirements of embedded systems.

[0042] The nano-composite material bristles in the brush head main body device have the characteristic of intelligent variable stiffness. When the ambient temperature is lower than 5°C, the material stiffness automatically increases by 10% to adapt to the low-temperature environment. When the ambient temperature is higher than 40°C, the material stiffness automatically decreases by 15% to prevent scratching the photovoltaic panel. The micro atomizing nozzle array uses ultrasonic atomization technology to atomize water into tiny particles with a particle size between 50 microns and 150 microns. The size of the atomized particles is automatically adjusted according to the degree of dirt adhesion. Larger water droplets are sprayed to increase the impact force when the dirt adhesion force is strong, and smaller water droplets are sprayed to reduce water resource consumption when the dirt adhesion force is weak. The centrifugal fan adopts intelligent variable frequency control technology and automatically adjusts the rotational speed within the range of 800 revolutions per minute to 3000 revolutions per minute according to the air flow monitored by the flow sensor and the water quality cleanliness monitored by the turbidity sensor. When the dirt concentration in the air flow is detected to increase, the fan speed automatically increases, and vice versa, optimizing energy consumption while ensuring the dust suction effect. The joint movement speed and acceleration of the multi-angle adjustment mechanism device are automatically adjusted according to the urgency of the cleaning task. In the normal cleaning mode, the maximum rotational speed of the horizontal rotation joint is 45 degrees per second, the maximum rotational speed of the vertical pitch joint is 30 degrees per second, and the maximum rotational speed of the torsion joint is 60 degrees per second. In the emergency cleaning mode, the maximum rotational speed of each joint increases by 50% to meet the emergency cleaning requirements before sudden severe weather. The control chip uses an industrial-grade 64-bit quad-core processor with a main frequency of 3 gigahertz, an built-in artificial intelligence acceleration unit, has edge computing capabilities, and adopts a distributed computing architecture to complete complex cleaning strategy optimization calculations locally, reducing the dependence on the remote server and improving the system response speed.

[0043] The following will describe the specific implementation manners of the above steps in detail.

[0044] The specific implementation of step S01 is to ensure the normal startup and environmental adaptability assessment of the brush head system of the multi-angle self-cleaning photovoltaic module cleaning robot through the system initialization process. First, the communication device receives cleaning task instructions from the main control system of the cleaning robot, including information such as the position coordinates, area size, installation angle, and cleaning priority of the photovoltaic module. Next, the system reads the pre-stored installation parameters of the photovoltaic module, including panel type, material thickness, surface coating characteristics, etc., and loads the cleaning path planning data. Subsequently, the environmental perception device starts to comprehensively collect the current environmental parameters. The infrared thermal imager measures that the surface temperature of the photovoltaic panel must be in the range of -10°C to 60°C, the humidity sensor measures that the environmental relative humidity needs to be less than 95%, the wind speed sensor measures that the environmental wind speed should be less than 12 meters per second, and the dust concentration sensor measures that the dust concentration in the air must be less than 150 μg / m³. The system comprehensively evaluates the deviation degree between the environmental parameters and the preset threshold through the fuzzy logic control algorithm, calculates the environmental suitability score, and determines that the environmental conditions meet the cleaning operation requirements when the score is higher than 75 points, and continues to execute the subsequent steps; if the score is between 50 points and 75 points, the system issues a mild environmental discomfort warning but continues to execute the cleaning task, and appropriately adjusts the cleaning parameters at the same time; if the score is lower than 50 points, a severe environmental discomfort warning is issued, the cleaning task is suspended, and waiting for the environmental conditions to improve or manual intervention.

[0045] The specific implementation of step S02 is to optimize the cleaning parameter settings based on the environmental perception data in the pre-cleaning preparation stage. The system first analyzes the surface temperature distribution data of the photovoltaic panel collected by the infrared thermal imager, and uses the Fourier heat conduction equation to calculate the heat exchange rate between the panel and water, so as to determine the optimal cleaning water temperature. When the panel temperature is more than 5°C lower than the environmental temperature, the water temperature is set to the environmental temperature plus 2°C; when the panel temperature is higher than the environmental temperature, the water temperature is set to the average surface temperature of the panel minus 3°C to avoid panel damage caused by thermal shock. The electric heating device and water temperature cooling system built in the water tank accurately adjust the water temperature according to the calculation results, and the control accuracy is ±0.5°C. At the same time, the system constructs a dirt distribution model based on the data collected by the dust concentration sensor. When the dust concentration is higher than 100 μg / m³, a high-intensity cleaning mode is adopted, and the brush rotation speed is set to 2800 revolutions per minute; when the dust concentration is between 50 μg / m³ and 100 μg / m³, a medium-intensity cleaning mode is selected, and the brush rotation speed is set to 2200 revolutions per minute; when the dust concentration is lower than 50 μg / m³, a low-intensity cleaning mode is adopted, and the brush rotation speed is set to 1600 revolutions per minute. The cleaning unit pressure and water injection intensity are also set in gradients accordingly, forming a complete set of cleaning intensity parameter combinations to provide a basic configuration for the subsequent cleaning operation.

[0046] The specific implementation of step S03 is to achieve the best fitting state between the brush head and the surface of the photovoltaic panel during the pre-adjustment stage of the brush head posture. The system first analyzes the received installation parameters of the photovoltaic module, extracts the data of the panel inclination angle, azimuth angle, and installation height, and calculates the theoretical optimal contact angle by combining the three-dimensional coordinate transformation algorithm. Subsequently, the multi-angle adjustment mechanism device realizes the precise positioning of the brush head by coordinating the control of three degrees-of-freedom joints. The horizontal rotation joint realizes the rotational positioning within the range of 0 degrees to 360 degrees with a minimum step of 0.1 degree under the drive of a stepper motor. The vertical pitch joint realizes the angle adjustment within the range of -45 degrees to 90 degrees through an electric push rod, and the telescopic precision of the push rod reaches 0.5 mm. The torsion joint realizes the torsional movement within the range of ±180 degrees through a harmonic reducer with a reduction ratio of 100:1, ensuring high-precision and low-backlash angle control. During the adjustment process, the angle sensor array monitors the actual angle values of each joint in real time at a frequency of 100 Hz, compares them with the theoretical angle values, and triggers the PID controller to perform error compensation when the deviation exceeds 0.5 degrees, so that each joint quickly reaches the target position and remains stable through closed-loop feedback control. At the same time, the system monitors the joint load conditions through the torque sensor array to ensure that the torque does not exceed 80% of the rated value during the adjustment process, preventing the mechanical structure from being overloaded. After the posture adjustment is completed, the system executes a touch-type contact force confirmation program to ensure uniform contact between the brush head and the panel surface, and the contact pressure is controlled within the range of 5 N to 15 N, laying a foundation for efficient and non-destructive cleaning operations.

[0047] The specific implementation of step S04 is to achieve efficient and precise cleaning execution and dynamic adjustment during the intelligent cleaning operation phase. The system first starts the independent drive motor array in the brush head main body device according to the cleaning intensity parameters set in step S02. Each drive motor individually controls a cleaning unit, and a vector control algorithm is used to precisely adjust the speed and torque. A soft start strategy is adopted during startup to avoid impact. During the cleaning process, the current sensing circuit collects the operating current of each drive motor at a frequency of 20 Hz, and calculates the real-time cleaning resistance through the linear relationship model between the motor current and the cleaning resistance. The conversion coefficient between the cleaning resistance and the motor current is calibrated according to the parameters of different motor models, and the typical value is 0.25 N / A. The torque sensor array synchronously monitors the joint load torque data, and the contact force distribution between the brush head and the panel is calculated by inverse calculation using the Jacobian matrix of the robotic arm. When the deviation between the detected value and the predicted value of the cleaning resistance exceeds 20%, the system performs adaptive adjustment: if the cleaning resistance in a local area suddenly increases and exceeds the threshold (usually set as 2 times the reference resistance), the intelligent control system immediately reduces the pressure of the corresponding cleaning unit, and the pressure reduction range is 30% to 50%. At the same time, the bristle rotation speed is reduced to 70% of the original value to prevent scratching damage to the photovoltaic panel; if the cleaning resistance is lower than expected, the pressure of the cleaning unit is appropriately increased to improve the cleaning efficiency, and the increase range is controlled within 20%. The whole process realizes smooth transition through the fuzzy adaptive control algorithm, ensures the best matching of the cleaning pressure and the dirt adhesion force, and protects the surface of the photovoltaic panel while achieving efficient cleaning.

[0048] The specific implementation of step S05 is to establish a visualization mechanism for cleaning resistance distribution and an evaluation mechanism for bristle contact status during the real-time monitoring of cleaning effect. Based on the motor operating current data collected by the current sensing circuit, the system extracts the spectral characteristics of the current signal through Fourier transform to identify the characteristic frequencies of different types of dirt. When fluctuations of 12 Hz to 18 Hz appear in the motor current, it usually indicates the presence of dust-like dirt; when fluctuations of 5 Hz to 8 Hz appear, it may be organic dirt such as leaves or bird droppings. The system calculates the cleaning resistance using the current data and generates a heat map of cleaning resistance distribution with a resolution of 5 cm × 5 cm using the bilinear interpolation algorithm. Different colors in the heat map represent different levels of cleaning resistance. The red area indicates that the cleaning resistance is greater than 1.5 N, the yellow area indicates that the cleaning resistance is between 0.8 N and 1.5 N, and the green area indicates that the cleaning resistance is less than 0.8 N. Through the analysis of the heat map, the system identifies areas with uneven dirt distribution and automatically increases the cleaning times for the red areas in the heat map. The increased times are proportional to the cleaning resistance value of the area, usually increasing by 1 to 3 times on the original basis. At the same time, the system estimates the bristle deformation state by analyzing the current change pattern of the bristles. When the amplitude of the current fluctuation is less than 50% of the reference value, it indicates that the bristles are not in sufficient contact with the panel; when the current fluctuation shows high-frequency small-amplitude oscillations, it indicates that the bristles are overly bent. Based on these analysis results, the system dynamically adjusts the position and pressure of the cleaning unit to ensure that the cleaning coverage rate reaches 100% without any missed areas.

[0049] The specific implementation of step S06 is to achieve timely dirt removal and water recycling during the brush head self-cleaning phase. The system continuously monitors data collected by the current sensing circuit to assess the changing trend of the cleaning resistance. When the average cleaning resistance over 10 consecutive samples increases by more than 30% compared to the initial value, or when the turbidity sensor detects that the turbidity of the separated water exceeds 200 NTU (Nephelometric Turbidity Units), the brush head self-cleaning process is automatically triggered. First, the control system uses a multi-angle adjustment mechanism to smoothly move the brush head assembly to a pre-set safe position at the edge of the photovoltaic panel. A fifth-order polynomial trajectory planning is used to ensure a smooth transition. Once in the safe position, the high-pressure water pump activates and gradually increases the outlet pressure to 0.8 MPa. A micro-atomizing nozzle array sprays a fine mist onto the nanocomposite bristles, distributed in a 120-degree fan pattern to ensure full coverage of the bristle surface. Simultaneously, the centrifugal fan activates and reaches a set speed of 2400 rpm, generating a negative pressure of 8 kPa. This draws the flushed wastewater and dirt into the dirt separator through the dirt collection channel. Within the cyclonic separation chamber, the mixture rotates at a tangential speed of 30 meters per second. Centrifugal force causes denser dirt particles to move toward the outer wall and sink. After initial separation, the dirt particles enter the dirt collection box. The water flows upward and passes through a 50-micron pore size filter assembly for secondary filtration. The filtered water is then returned to the water tank for recycling. The entire self-cleaning process lasts 20 to 30 seconds. The system controls the self-cleaning intensity and duration through data feedback from pressure sensors, flow sensors, and turbidity sensors. When the turbidity sensor reading is below 50 NTU and remains stable for 5 consecutive seconds, self-cleaning is considered complete and the system returns to cleaning mode.

[0050] The specific implementation of step S07 is to establish a multi-parameter correlation model and achieve dynamic parameter adjustment during the adaptive cleaning parameter optimization phase. The system first constructs a correlation model between environmental parameters and cleaning effect based on a Bayesian network. The input nodes include the surface temperature of the photovoltaic panel collected by an infrared thermal imager (effective range: -10°C to 60°C), the relative environmental humidity collected by a humidity sensor (effective range: 10% to 95%), the environmental light intensity collected by a light sensor (effective range: 0 to 120,000 lux), the environmental wind speed collected by a wind speed sensor (effective range: 0 to 20 m / s), and the dust concentration collected by a dust concentration sensor (effective range: 0 to 500 μg / m³); the intermediate layer nodes represent the combined effects of environmental factors; the output node is the adjustment amount of the cleaning parameters. The system dynamically adjusts the cleaning parameters according to the inference results of the Bayesian network: when the surface temperature of the photovoltaic panel is more than 10°C higher than the environmental temperature, the water temperature is reduced to the panel temperature minus 5°C to avoid thermal shock; when the environmental humidity is lower than 30%, the water injection intensity is increased to 140% of the standard value to suppress dust flying; when the light intensity is higher than 80,000 lux, the system automatically avoids cleaning during the period from 10:00 to 14:00 to prevent damage to the panel caused by the water droplet concentrating effect; when the wind speed is greater than 8 m / s, the pressure of the cleaning unit is increased to 125% of the standard value to overcome the jitter caused by the wind; when the dust concentration is higher than 100 μg / m³, the bristle rotation speed is reduced to 80% of the standard value to reduce secondary pollution. The system continuously optimizes the parameter adjustment strategy through a reinforcement learning algorithm to minimize energy consumption while maximizing the cleaning effect. The adjustment range changes dynamically according to the feedback of the actual cleaning effect and is usually controlled within ±50% of the benchmark value.

[0051] The specific implementation of step S08 is to realize system health status monitoring and intelligent fault handling during the fault warning and self-recovery stage. The system establishes a multi-level fault detection and handling mechanism based on the data collected by the fault diagnosis device. First, the vibration sensor monitors the system vibration signal at a sampling frequency of 100 Hz, extracts the vibration characteristics through wavelet packet analysis, and compares them with the vibration characteristic library in the normal working state. When the energy of a certain frequency band exceeds 150% of the normal value, it is determined as abnormal vibration; the sound sensor collects the system operation sound and extracts the acoustic characteristics through Mel Frequency Cepstral Coefficients (MFCC), and compares them with the normal working sound pattern. When the acoustic similarity is lower than 0.6, it is determined as abnormal sound; the temperature sensor array monitors the surface temperature of each motor. When the temperature exceeds 70°C or the temperature rise rate is greater than 2°C / minute, it is determined as abnormal temperature. The system inputs these abnormal signals into a fault classifier based on support vector machines to identify the specific fault type and automatically execute the corresponding processing strategy: for the cleaning unit blockage fault (characterized by a sudden increase in the drive motor current exceeding 150% of the rated value and accompanied by low-frequency vibration), the system increases the water pressure of the corresponding micro-atomizing nozzle to 1.2 MPa and increases the suction of the centrifugal fan to 10 kPa, and tries to dredge for 10 seconds; for the joint jamming fault (characterized by the stagnation of the joint angle sensor data and a sudden increase in the torque sensor data), the system controls the joint to reciprocate 3 times, with an amplitude of ±5 degrees each time, to relieve the jamming; for the motor overheating fault (characterized by the motor surface temperature exceeding 75°C), the system reduces the motor load to 60% of the rated value and increases the cooling interval time to 2 minutes; for the near-full dirt collection box fault (characterized by the detection of a gas flow decrease by the flow sensor exceeding 40%), the system issues a replacement reminder and records the current cleaning position coordinates, and automatically resumes cleaning after replacement. If the fault persists after taking the above measures, the system will send a detailed fault report including the fault code, abnormal parameter data, fault occurrence time, and system operation log to the remote monitoring center through the communication device. At the same time, the system enters the safe mode, shuts down the power supply of high-risk components, and moves the brush head to a safe position.

[0052] The specific implementation of step S09 is to establish a multi-dimensional cleaning effect evaluation system during the cleaning quality assessment phase. After the cleaning operation is completed, the system first calculates the change rate of the cleaning resistance before and after cleaning. The cleaning area is divided into 10 cm × 10 cm units, and the average cleaning resistance before and after cleaning of each unit is compared. Areas where the cleaning resistance reduction rate exceeds 80% are judged as excellent cleaning, areas where the reduction rate is between 60% and 80% are judged as good cleaning, and areas where the reduction rate is less than 60% are judged as average cleaning. The system calculates the percentage of the excellent cleaning area, which is used as the cleaning coverage rate index. At the same time, the time integral value of the driving motor current is analyzed, and the energy consumption per unit area of cleaning is calculated. Under normal circumstances, the cleaning energy consumption of each square meter of photovoltaic panel should be less than 0.05 kWh. The system also evaluates the dirt recovery rate by analyzing the ratio of the dirt collection amount in the dirt collection box to the theoretically estimated dirt amount, and this ratio should usually be greater than 85%. Based on the above indicators, the system uses the weighted average method to calculate the comprehensive cleaning quality score. The weight distribution is that the cleaning coverage rate accounts for 50%, the energy consumption index accounts for 30%, and the dirt recovery rate accounts for 20%. The score range is from 0 to 100 points. A cleaning quality score greater than 85 points is judged as excellent, a score between 70 and 85 points is judged as good, and a score less than 70 points is judged as average. The system also obtains the data of the improvement in the power generation efficiency of the photovoltaic panel before and after cleaning through data interaction with the photovoltaic power station monitoring system, compares the improvement amplitude of the power generation efficiency with the cleaning cost (including water resource consumption, power consumption, and equipment loss), calculates the cleaning return on investment, and provides a decision-making basis for the optimization of the next cleaning task.

[0053] The specific implementation of step S10 is to establish a historical data analysis and parameter adaptive optimization mechanism during the system self-learning and optimization stage. The system stores all the parameter data of the current cleaning task in the local database, including environmental parameters (temperature, humidity, light, wind speed, dust concentration), cleaning parameters (water temperature, brush rotation speed, unit pressure, water volume, joint speed), and the cleaning quality assessment results. The data is stored in a time series format, with a sampling period of once per minute for environmental parameters and once per second for cleaning parameters. The system builds a deep neural network model based on the stored historical data. The network structure adopts a hybrid architecture combining long short-term memory network and fully connected layers. The input layer nodes correspond to environmental parameters and cleaning parameters. The hidden layer contains 3 LSTM layers and 2 fully connected layers, with the number of nodes in each layer being 128, 256, 128, 64, and 32 respectively. The output layer corresponds to the cleaning quality assessment index. The training process uses the stochastic gradient descent algorithm, with a learning rate set to 0.001, a batch size of 32, and a training cycle of 1000 rounds. The trained model is used to analyze the impact of different combinations of environmental conditions and cleaning parameters on the cleaning effect, and to find the optimal cleaning parameter combination under specific environmental conditions through the gradient search algorithm. The system integrates the optimal parameter combination with the preset cleaning strategy to generate a parameter lookup table for different environmental conditions, where each environmental parameter combination corresponds to a set of optimal cleaning parameters.

[0054] The detailed structure of the CleanFormer deep learning model adopts a multi-layer Transformer encoder-decoder architecture, specifically including 1 input layer, 6 encoder layers, 6 decoder layers, and 1 output layer. The input layer maps different types of input data (environmental data, cleaning parameters, time series data) to a unified 512-dimensional feature space through independent embedding layers and adds position encoding to preserve sequence information. The environmental perception encoder contains 3 multi-head self-attention layers and 3 feed-forward network layers. Each multi-head self-attention layer has 8 attention heads, and the number of attention heads is the same as the number of cleaning units, realizing targeted attention to the states of different cleaning units. The cleaning parameter encoder adopts a similar structure but adjusts the attention mechanism to correlation attention between parameters to capture the mutual influence between different cleaning parameters. The time series information encoder adopts a causal attention mechanism, and the attention mask ensures that only historical data is focused on at the current time step. The parameter optimization decoder fuses the outputs of the three encoders, establishes a mapping relationship between the environmental state, cleaning parameters, and time series changes through the cross-attention mechanism, and finally outputs the optimal parameter adjustment scheme. To better adjust the cleaning parameters, the model also introduces an adaptive sparse attention mechanism. The sparsity parameter is determined by the surface temperature of the photovoltaic panel, environmental humidity, and the uniformity of dirt distribution. The calculation formula is sparsity = 0.5×(1 + dirt distribution uniformity index) - 0.2×|normalized temperature value| - 0.1×|normalized humidity value|. Optionally, the sparsity range is controlled between 0.1 and 0.9. The model uses residual connections with a Dropout rate of 0.1 and layer normalization techniques. The activation function is GELU, the optimizer is AdamW, the learning rate is 0.0001, and the weight decay is 0.01. The detailed steps for establishing the training dataset are as follows: First, deploy data acquisition systems in photovoltaic power stations in 10 different climate regions around the world. Select 10 photovoltaic panels with different orientations and inclinations in each power station and continuously collect environmental data and cleaning data for 1 year. The data sampling frequency is once every 10 minutes for environmental data and 10 times per second for cleaning process data. The total amount of real-scene data is approximately 2.5 million records, including cleaning parameter and effect data for different dirt types (dust, pollen, resin, bird droppings, industrial soot) under different environmental conditions. Virtual scene data is generated by simulating the cleaning process under extreme environmental conditions using a physical engine, including a low-temperature environment of -20°C, a high-temperature environment of 55°C, an extremely dry environment with a relative humidity of 5%, an extremely wet environment with a relative humidity of 98%, and a strong wind environment with a wind speed of 18 meters per second. A total of approximately 500,000 boundary condition data are generated. During the data annotation process, for each combination of environmental parameters and cleaning parameters, the corresponding cleaning efficiency (percentage reduction in dirt before and after cleaning) and energy consumption (energy consumption per square meter for cleaning) are annotated, and the annotation accuracies are 0.1% and 0.001 kWh respectively.Anomaly data filtering adopts the 3-sigma principle to eliminate data points that exceed 3 times the standard deviation of the normal distribution; missing data supplementation uses the K-nearest neighbor algorithm to perform interpolation based on the weighted average of the nearest 5 similar data points. In the data augmentation step, the original data is perturbed by adding Gaussian noise with a mean of 0 and a standard deviation of 10% of the original data; at the same time, the random occlusion technique is adopted to randomly set 20% of the data points as missing to train the model to adapt to incomplete data; also, through the parameter micro-perturbation method, a random perturbation of ±5% is added to the clean parameters to expand the data volume and improve the model robustness. The final training dataset contains approximately 5 million groups of samples, which are divided into a training set, a validation set, and a test set according to the ratio of 7:2:1.

[0055] The following details the mathematical models or calculation processes involved in the present invention.

[0056] In step S01, the system uses the fuzzy logic control algorithm to comprehensively evaluate the deviation degree between the environmental parameters and the preset thresholds, and calculates the environmental suitability score. The specific formula for calculating the environmental suitability score is as follows:

[0057]

[0058] In the formula, S e is the environmental suitability score, with a value range of 0 to 100; n is the number of environmental parameters, and in this system, n = 4; w i is the weight coefficient of the i-th environmental parameter, satisfying μ i (x i ) is the membership function value corresponding to the i-th environmental parameter x i , with a value range of 0 to 100; x i is the measured value of the environmental parameter. Among them, the parameter acquisition method is: the environmental parameter x1 is the surface temperature of the photovoltaic panel, which is measured by an infrared thermal imager; x2 is the environmental relative humidity, which is measured by a humidity sensor; x3 is the environmental wind speed, which is measured by a wind speed sensor; x4 is the dust concentration in the air, which is measured by a dust concentration sensor. The default values of the weight coefficients w i are w1 = 0.3, w2 = 0.2, w3 = 0.3, and w4 = 0.2. The membership function μ i (x i ) adopts the trapezoidal membership function, which is defined according to the suitable range of each environmental parameter. For example, the membership function of the temperature parameter x1 is:

[0059]

[0060] The fuzzy logic control algorithm realizes the comprehensive evaluation of complex environmental conditions by converting multiple environmental parameters into a unified scoring standard. The membership function adopts a trapezoidal form, which can reflect the smooth transition of environmental parameters from unsuitable to completely suitable. The introduction of weight coefficients reflects the differences in the impacts of different environmental parameters on the cleaning operation. The entire evaluation process considers the synergistic effects of multiple environmental factors, providing a numerical basis for system decision-making.

[0061] In step S02, the system calculates the optimal cleaning water temperature based on the surface temperature of the photovoltaic panel. The specific calculation formula is as follows:

[0062]

[0063] In the formula, T w is the cleaning water temperature, with the unit of °C; T e is the environmental temperature, with the unit of °C; T p is the average surface temperature of the photovoltaic panel, with the unit of °C.

[0064] Meanwhile, the system determines the cleaning intensity parameter according to the dust concentration. The setting formula for the bristle rotation speed is:

[0065]

[0066] In the formula, ω b is the bristle rotation speed, with the unit of revolutions per minute; D d is the dust concentration, with the unit of μg / cubic meter.

[0067] The pressure P c of the cleaning unit and the bristle rotation speed ω b have the following relationship:

[0068] P c = k p ·ω b / 2800 + P min ;

[0069] In the formula, k p is the pressure coefficient, with a value of 5 N; P min is the minimum pressure value, with a value of 5 N.

[0070] The water injection intensity F w and the dust concentration D d have the following relationship:

[0071] F w = F base ·(1 + 0.5·D d / 100);

[0072] In the formula, F baseis the benchmark water injection intensity, with a value of 0.1 liters per minute per square meter. The parameter acquisition method is: T e and T p are measured by an infrared thermal imager; D d is measured by a dust concentration sensor. This set of calculation formulas is designed based on the principle of heat conduction and cleaning experience, taking into account the influence of temperature difference on the thermal stress of materials and the influence of dirt concentration on the cleaning intensity requirements. The water temperature calculation adopts a piecewise function form to avoid panel damage caused by thermal shock for different cleaning requirements under different temperature conditions. The setting of the cleaning intensity parameter uses threshold judgment and linear proportional relationship, which not only simplifies the calculation process but also meets the cleaning requirements under different dirt levels, providing a reliable parameter initialization scheme for the system.

[0073] In step S03, the system calculates the theoretical optimal contact angle through a three-dimensional coordinate transformation algorithm, which involves the operation of the coordinate transformation matrix. The specific representation of the coordinate transformation matrix is as follows:

[0074] T = R z (α)·R x (β)·R y (γ);

[0075] In the formula, T is the comprehensive coordinate transformation matrix; R z (α), R x (β), R y (γ) are the basic rotation matrices for rotating α angle around the z-axis, β angle around the x-axis, and γ angle around the y-axis respectively; α is the azimuth angle of the photovoltaic panel, β is the tilt angle of the photovoltaic panel, and γ is the torsion angle of the photovoltaic panel.

[0076] Among them, the basic rotation matrix is defined as:

[0077]

[0078] The calculation formula for the theoretical optimal contact angle vector θ best is:

[0079] θ best = T·n0;

[0080] In the formula, n0 is the normal vector [0, 0, 1] in the standard coordinate system T .

[0081] The error vector e between the actual brush head angle and the theoretical optimal angle is calculated as:

[0082] e = θ actual - θ best ;

[0083] In the formula, θ actualThe actual brush head angle vector measured by the angle sensor array, including horizontal rotation angle, vertical pitch angle, and torsion angle.

[0084] During joint control, the output calculation formula of the PID controller is:

[0085]

[0086] Where u(t) is the control output; e(t) is the current error; K p , K i , K d They are proportional, integral and differential coefficients respectively, and the typical value is K p =2.0, K i =0.1, K d =0.5. The parameter acquisition method is: α, β, γ are obtained from the cleaning task instruction; θ actual Obtained from measurements by an angle sensor array. This coordinate transformation calculation process is based on Euler angle rotation theory, achieving angular transformation in three-dimensional space through matrix multiplication. The rotation order of first around the z-axis, then around the x-axis, and finally around the y-axis is chosen to directly correspond to the photovoltaic panel installation parameters (azimuth, inclination, and torsion angle), facilitating engineering implementation. The PID controller achieves precise angle adjustment and stable control through proportional, integral, and differential control, and is a key step in achieving high-precision attitude control.

[0087] In step S04, the system calculates the cleaning resistance based on the motor current. The calculation formula is as follows:

[0088] F r =k c ·(I m -I0)+F0+ΔF T ;

[0089] Where, F r is the cleaning resistance, in Newton; k c Is the conversion coefficient between current and force, with a typical value of 0.25 Newtons / ampere; I m is the actual operating current of the motor, measured by the current sensing circuit, in amperes; I0 is the no-load current of the motor, in amperes; F0 is the basic resistance constant, in Newtons; ΔF T is the temperature compensation term, in Newtons.

[0090] Temperature compensation term ΔF T The calculation formula is:

[0091] ΔF T =α T (T-T0);

[0092] Where, αT is the temperature compensation coefficient, with a typical value of -0.01 N / °C; T is the current ambient temperature in °C; T0 is the standard calibration temperature, with a value of 25 °C.

[0093] The adaptive adjustment formula for the cleaning unit pressure is:

[0094]

[0095] In the formula, P adj is the adjusted cleaning unit pressure; P cur is the current cleaning unit pressure; F base is the reference cleaning resistance, usually the mean of the initial cleaning resistances. The parameter acquisition method is: I m is measured by the current sensing circuit; I0 and F0 are obtained through initial calibration; T is measured by the temperature sensor. This cleaning resistance calculation formula is based on the motor load theory, establishing a linear relationship between the motor current increment and the mechanical load force. The introduction of the temperature compensation term takes into account the influence of the ambient temperature on the motor performance and the characteristics of the brush material, improving the accuracy of the force value calculation. The adaptive pressure adjustment adopts a piecewise function form, realizing intelligent regulation based on resistance feedback, which not only protects the photovoltaic panel from damage caused by excessive pressure but also ensures that the cleaning efficiency is not significantly affected by the resistance change.

[0096] In step S05, the system uses the bilinear interpolation algorithm to generate the cleaning resistance distribution heat map, and its interpolation calculation formula is specifically expressed as follows:

[0097]

[0098] In the formula, F(x, y) is the cleaning resistance interpolation result at the position (x, y); (x1, y1), (x1, y2), (x2, y1), (x2, y2) are the four nearest cleaning unit center positions surrounding (x, y); F(x1, y1), F(x1, y2), F(x2, y1), F(x2, y2) are the cleaning resistance measurement values at the corresponding positions.

[0099] The decision formula for increasing the cleaning times is:

[0100]

[0101] In the formula, N add(x, y) is the increased cleaning frequency at position (x, y); F(x, y) is the cleaning resistance value at this position, with the unit of Newton. The parameter acquisition method is as follows: F(x1, y1), F(x1, y2), F(x2, y1), and F(x2, y2) are obtained by measuring the motor current through the current sensing circuit and converting it into the cleaning resistance. This bilinear interpolation algorithm calculates the resistance value at any position based on the resistance values of four known points through the method of weighted average. The bilinear interpolation is used instead of the nonlinear interpolation to achieve a smooth transition of the resistance distribution on the two-dimensional plane and avoid the problem of discontinuous interpolation. The cleaning frequency decision adopts a threshold piecewise function, dynamically adjusting the cleaning intensity according to the resistance size, realizing an accurate regional cleaning strategy, and improving the cleaning efficiency and quality.

[0102] In step S06, the system uses the cyclone separation technology to separate dirt and water, and its separation efficiency calculation formula is specifically expressed as follows:

[0103]

[0104] In the formula, η is the separation efficiency, and its value range is from 0 to 1; ρ p is the dirt particle density, with the unit of kg / m³; d p is the dirt particle diameter, with the unit of m; v t is the tangential flow velocity, with the unit of m / s; μ is the dynamic viscosity of water, with the unit of Pa·s; D is the diameter of the cyclone separation chamber, with the unit of m; N is the effective number of rotation circles; Q is the volume flow rate, with the unit of m³ / s.

[0105] The relationship between the tangential flow velocity v t and the pressure drop ΔP is:

[0106]

[0107] In the formula, C d is the flow coefficient, and its typical value is from 0.7 to 0.8; ΔP is the pressure difference between the inlet and outlet of the cyclone separator, with the unit of Pa; ρ w is the density of water, with the unit of kg / m³.

[0108] The judgment formula for the self-cleaning trigger condition is:

[0109]

[0110] In the formula, S clean is the self-cleaning trigger signal, 1 indicates triggering, and 0 indicates non-triggering; is the average cleaning resistance of 10 consecutive samples; F init is the initial cleaning resistance value; T w is the water turbidity measured by the turbidity sensor, with the unit of NTU. The parameter acquisition method is: ρp Determined according to the physical properties of different types of dirt, the value for dust - like dirt is approximately 2,600 kg / m³; d p Indirectly estimated by a turbidity sensor, the typical range is 10 to 100 microns; v t Calculated by a formula after measuring the pressure drop with a pressure sensor; μ is obtained by looking up a table according to the water temperature; D is a design parameter with a value of 0.1 m; N is an empirical parameter with a value of 5 to 10; Q is directly measured by a flow sensor; Obtained by measuring the motor current with a current - sensing circuit and converting it into cleaning resistance; T w Obtained by measuring with a turbidity sensor. The cyclone separation efficiency formula is based on the principles of particle kinematics and fluid mechanics theory, expressing the separation process of dirt particles under the action of centrifugal force. The exponential - function form in the formula reflects the cumulative effect of the separation process, and the square term of the particle diameter reflects the significant influence of particle size on the separation efficiency, which is consistent with the actual physical process. The self - cleaning trigger condition uses a simple threshold judgment method, considering both the resistance change during the cleaning process and the water quality condition, providing a clear start basis for the self - cleaning process.

[0111] In step S07, the system uses a Bayesian network to establish an association model between environmental parameters and cleaning effects. In the Bayesian network, the calculation formula for the conditional probability distribution between nodes is specifically expressed as follows:

[0112]

[0113] In the formula, C i is the adjustment amount of the i - th cleaning parameter; Pa(C i ) is the set of parent nodes affecting C i , that is, the relevant environmental parameters; μ i (Pa(C i )) is the conditional mean function of C i given the values of the parent nodes; σ i is the conditional standard deviation.

[0114] The conditional mean function μ i (Pa(C i )) is represented by a linear regression model:

[0115]

[0116] In the formula, β i0 is the intercept term; β ij is the linear effect coefficient of the j - th environmental parameter E j ; β ijk is the interaction effect coefficient between environmental parameters E j and E k ; m is the number of environmental parameters, and in this system, m = 5.

[0117] The dynamic adjustment formula for cleaning parameters based on the Bayesian network is as follows:

[0118]

[0119] In the formula, ΔT w , ΔF w , ΔP c , Δω b are the adjustment amounts of water temperature, water volume injection intensity, cleaning unit pressure, and brush rotation speed respectively; T p is the surface temperature of the photovoltaic panel; T w is the current water temperature; T e is the ambient temperature; F base is the reference water volume injection intensity; H is the ambient relative humidity; P cur is the current cleaning unit pressure; V is the ambient wind speed; ω cur is the current brush rotation speed; D d is the dust concentration. The method for obtaining parameters is as follows: The environmental parameters E1 to E5 are the surface temperature of the photovoltaic panel, ambient relative humidity, ambient light intensity, ambient wind speed, and dust concentration respectively, which are measured by corresponding sensors. The coefficients β i0 , β ij and β ijk are obtained through training with historical data, and the maximum likelihood estimation method is used for training. The standard deviation σ i is estimated through residual analysis, and the typical value range is from 0.05 to 0.2. This Bayesian network model represents the dependence relationship between variables through a probabilistic graph structure, introduces a Gaussian conditional probability distribution to characterize the uncertainty of parameter adjustment amounts, and combines the conditional mean functions of linear and non-linear (interaction) terms, which can capture the complex influence of environmental parameters on cleaning parameters. The specific parameter adjustment adopts a conditional judgment form, and clear adjustment strategies are set for different environmental conditions. This design not only integrates prior knowledge but also retains the adaptive characteristics of the model, providing a reliable parameter optimization mechanism for the system.

[0120] In step S08, the system classifies abnormal signals based on a support vector machine-based fault classifier. The decision function of the support vector machine is specifically expressed as follows:

[0121]

[0122] In the formula, f(x) is the output of the decision function, a positive value represents one class, and a negative value represents the other class; x is the input feature vector, including vibration features, acoustic features, and temperature features; x i is the support vector; y i is the class label corresponding to the support vector; α i is the Lagrange multiplier; K(x i, \(K(x, x)\) is the kernel function; \(b\) is the bias term; \(N\) is the number of support vectors.

[0123] The system adopts a radial basis kernel function:

[0124]

[0125] In the formula, \(\sigma\) is the width parameter of the kernel function, determined by cross-validation, with typical values ranging from 0.5 to 2.0, \(d\) is the dimension of the feature vector; \(x\) i,j and \(x\) i are respectively the \(j\)-th eigenvalue of \(x\) i and \(x\).

[0126] The composition of the feature vector \(x\) is:

[0127] \(x = [V1, V2,..., V\) p , \(A1, A2,..., A\) q , \(T1, T2,..., T\) r T ;

[0128] In the formula, \(V1\) to \(V\) p are vibration features, including the energy ratio of different frequency bands; \(A1\) to \(A\) q are acoustic features, using Mel Frequency Cepstral Coefficients; \(T1\) to \(T\) r are temperature features, including the absolute temperature value and the temperature change rate.

[0129] The triggering condition of the fault handling strategy is:

[0130]

[0131] In the formula, \(S\) fault is the fault handling strategy trigger signal; \(P\) fault is the fault probability, obtained by mapping the output value of the decision function of the support vector machine. The method for obtaining parameters is: the vibration features \(V1\) to \(V\) p are obtained by collecting signals through vibration sensors and performing wavelet packet decomposition; the acoustic features \(A1\) to \(A\) q are obtained by collecting sounds through sound sensors and calculating MFCC coefficients; the temperature features \(T1\) to \(T\) r are directly measured by a temperature sensor array. The support vector \(x\) i , the coefficient \(\alpha\) i ​They are obtained through offline training, and the training samples come from various typical fault scenarios. This support vector machine classifier uses the kernel trick to map complex fault features into a high-dimensional space to achieve non-linear classification. The radial basis kernel function is selected because it has good generalization ability for abnormal data and is suitable for solving complex pattern recognition problems such as equipment fault diagnosis. The decision function adopts a weighted sum form, and each support vector contributes its influence through the kernel function, ultimately achieving accurate identification and classification of different types of faults, providing a reliable decision basis for subsequent fault handling strategies.

[0132] In step S09, the system calculates the comprehensive cleaning quality score, and its calculation formula is specifically expressed as follows:

[0133] Q = w c ·C + w e ·E + w r ·R;

[0134] In the formula, Q is the comprehensive cleaning quality score, and its value range is from 0 to 100; C is the cleaning coverage rate index, and its value range is from 0 to 100; E is the energy consumption index, and its value range is from 0 to 100; R is the dirt recovery rate, and its value range is from 0 to 100; w c 、w e 、w r are the weight coefficients of the corresponding indicators respectively, and satisfy w c + w e + w r = 1.

[0135] The calculation formula for the cleaning coverage rate index C is:

[0136]

[0137] In the formula, A good is the area of the excellent cleaning area, in square meters; A fair is the area of the good cleaning area, in square meters; A total is the total cleaning area, in square meters.

[0138] The calculation formula for the energy consumption index E is:

[0139]

[0140] In the formula, P actual is the actual cleaning energy consumption per unit area, in kilowatt-hours per square meter; P standard is the standard cleaning energy consumption per unit area, with a value of 0.05 kilowatt-hours per square meter.

[0141] The calculation formula for the dirt recovery rate R is:

[0142]

[0143] Wherein, M collected is the mass of the dirt actually collected in the dirt collection box, in grams; M estimated is the theoretical dirt mass estimated based on the dirt density and the cleaning area, in grams. The parameter acquisition method is as follows: A good and A fair are determined by analyzing the change rate of the resistance before and after cleaning; P actual is calculated by analyzing the time integral of the motor current; M collected is obtained by weighing the dirt collection box; M estimated is calculated by predicting the dirt density and the cleaning area. The default weight coefficients are w c = 0.5, w e = 0.3, w r = 0.2. This comprehensive scoring formula adopts the weighted average method to integrate the evaluation indexes of three dimensions, namely the cleaning coverage rate, the energy consumption index and the dirt recovery rate, into a single quality score. The weight setting reflects the relative importance of different indexes, giving priority to the cleaning effect while taking into account the energy efficiency and environmental protection requirements. The formula design is simple and intuitive, facilitating practical engineering applications and providing a quantitative basis for the objective evaluation of the cleaning quality.

[0144] The calculation formula of the multi-parameter cleaning resistance prediction function is specifically expressed as follows:

[0145]

[0146] Wherein, F pred (x, y) is the predicted cleaning resistance value at the position (x, y), in Newtons; T is the surface temperature of the photovoltaic panel, in °C; H is the ambient relative humidity, in %; t a is the dirt adhesion time, in hours; D p is the average diameter of the dust particles, in microns; θ is the contact angle between the brush and the panel, in degrees; a ijklm is the polynomial coefficient; n is the highest degree of the polynomial, with a value of 3; ε is the error term.

[0147] The polynomial coefficient a ijklm is determined by the least squares method:

[0148] a = (X T X) -1 X T y;

[0149] Wherein, a is the coefficient vector; X is the input variable design matrix; y is the measured cleaning resistance vector. The parameter acquisition method is as follows: T is measured by an infrared thermal imager; H is measured by a humidity sensor; t a is calculated according to the last cleaning time; D pObtained by analyzing the particle size distribution with a dust concentration sensor; θ is obtained by measuring with an angle sensor array. The range of the error term ε is ±0.2 Newtons. The multi-parameter cleaning resistance prediction function adopts a high-order polynomial model, which can capture the non-linear effects and interaction effects of each input parameter on the cleaning resistance. The polynomial form is selected because of its good approximation ability and mathematical processing convenience. The model considers the influence of environmental factors such as temperature and humidity on the dirt adhesion, as well as key physical parameters such as dirt adhesion time, particle size, and contact angle, comprehensively reflecting the mechanical characteristics in the cleaning process and providing a theoretical support for accurately predicting the cleaning resistance distribution.

[0150] Optionally, in the adaptive sparse attention mechanism of the CleanFormer model mentioned in steps S07 and S10, the calculation formula of the sparsity parameter is specifically expressed as follows:

[0151] s = 0.5·(1 + U) - 0.2·|T n | - 0.1·|H n |;

[0152] In the formula, s is the attention sparsity, and its value range is from 0.1 to 0.9; U is the dirt distribution uniformity index, and its value range is from 0 to 1; T n is the normalized temperature value, and its value range is from -1 to 1; H n is the normalized humidity value, and its value range is from -1 to 1.

[0153] The calculation formula of the dirt distribution uniformity index U is:

[0154]

[0155] [[ID=2**]] F is the standard deviation of the cleaning resistance; is the average value of the cleaning resistance.

[0156] The calculation formulas of the normalized temperature value T n and the humidity value H n are:

[0157]

[0158] In the formula, T is the current temperature; T mid is the temperature median value, and its value is 25 °C; T max and T min are the maximum and minimum temperature values respectively, and their values are 60 °C and -10 °C respectively; H is the current humidity; H mid is the humidity median value, and its value is 50%; H max and H min ​They are the maximum and minimum humidity values, with the values being 95% and 5% respectively. The method for obtaining parameters is as follows: U is calculated by analyzing the cleaning resistance data collected by the current sensing circuit; T is obtained by measuring with an infrared thermal imager; H is obtained by measuring with a humidity sensor. The sparsity parameter calculation formula designs a balance mechanism between the dirt distribution uniformity and environmental parameters: when the dirt distribution is uneven (low U value), the sparsity decreases, making the attention mechanism more focused on key areas; when the temperature or humidity deviates from the intermediate value, the absolute value term increases, and the sparsity decreases accordingly, improving the sensitivity of the model to extreme environmental conditions. The formula adopts a linear combination form, simplifying the calculation process and facilitating real-time adjustment. At the same time, the setting of each coefficient reflects the relative importance of different factors for attention allocation.

[0159] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on a multi-sensor collaborative perception, multi-parameter adaptive control, and deep learning-driven intelligent decision-making system. First, the multi-sensor collaborative perception mechanism forms an environmental perception network through an infrared thermal imager, humidity sensor, light sensor, wind speed sensor, and dust concentration sensor to capture key environmental parameters affecting the cleaning effect in real time; at the same time, the current sensing circuit and the torque sensor array constitute a cleaning process perception system to monitor the cleaning resistance and brush head load conditions in real time. This all-round perception network provides a rich and accurate data basis for cleaning parameter optimization; secondly, the multi-parameter adaptive control mechanism constructs a mapping relationship between cleaning parameters and cleaning effects based on the sensing data. According to the principles of physics, the surface temperature of the photovoltaic panel affects the dirt adhesion strength, so the system dynamically adjusts the water temperature to avoid thermal shock; environmental humidity affects the dust suspension characteristics, and accordingly adjusts the water spray volume to prevent dust from flying; light intensity affects the water droplet concentration effect, so the cleaning time period is intelligently arranged; wind speed affects the stability of the brush head, and the brush head pressure is adjusted accordingly; dust concentration affects the risk of secondary pollution, and the brush hair rotation speed is automatically optimized. This mechanism ensures the best match between cleaning parameters and environmental conditions and dirt characteristics; most importantly, the present invention adopts a CleanFormer deep learning model based on the Transformer architecture, which realizes the optimal parameter decision-making under complex non-linear factors through four main components: an environmental perception encoder, a cleaning parameter encoder, a timing information encoder, and a parameter optimization decoder. Through the multi-head adaptive sparse attention mechanism, this model can dynamically adjust the attention allocation according to the key factors under different environmental conditions and find the optimal solution in the high-dimensional parameter space. CleanFormer is trained with a mixed dataset of virtual scenarios and real scenarios, and has strong generalization ability, capable of coping with various complex environments and dirt conditions, and providing a scientific and accurate parameter optimization scheme for the brush head system; the organic combination of the above three layers of technical principles constitutes the theoretical basis for the present invention to solve the technical problem of the adaptive adjustment of cleaning parameters for the brush head system of the photovoltaic module cleaning robot, realizing the closed-loop intelligent control from perception to decision-making and then to execution.

[0160] A specific Embodiment 1 of the present invention is provided below. The specific implementation manners of each step in this Embodiment 1 are described in detail as follows.

[0161] The specific implementation manner of step S01 is the same as that described above and will not be elaborated here. The specific implementation manner of step S02 is to optimize the cleaning parameter settings based on the environmental perception data during the pre-cleaning preparation stage. The system first analyzes the surface temperature distribution data of the photovoltaic panel collected by the infrared thermal imager, and uses the Fourier heat conduction equation to calculate the heat exchange rate between the panel and water, so as to determine the optimal cleaning water temperature. The formula for calculating the optimal cleaning water temperature based on the surface temperature of the photovoltaic panel is: In the formula, T w is the cleaning water temperature, with the unit of °C; T e is the environmental temperature, with the unit of °C; T p is the average surface temperature of the photovoltaic panel, with the unit of °C. The built-in electric heating device and water temperature cooling system in the water tank precisely adjust the water temperature according to the calculation result, and the control accuracy is ±0.5 °C. At the same time, the system constructs a dirt distribution model based on the data collected by the dust concentration sensor and sets the bristle rotation speed. The calculation formula is: In the formula, ω b is the bristle rotation speed, with the unit of revolutions per minute; D d is the dust concentration, with the unit of μg / cubic meter. The relationship between the cleaning unit pressure and the bristle rotation speed is: P c = k p ·ω b / 2800 + P min ; in the formula, k p is the pressure coefficient, and P min is the minimum pressure value. The relationship between the water injection intensity and the dust concentration is: F w = F base ·(1 + 0.5·D d / 100); in the formula, F base is the reference water injection intensity. Through the above parameter settings, the system forms a complete set of cleaning intensity parameter combinations, providing a basic configuration for subsequent cleaning operations.

[0162] The specific implementation manner of step S03 is to achieve the best fitting state between the brush head and the surface of the photovoltaic panel during the pre-adjustment stage of the brush head attitude. The system first analyzes the received installation parameters of the photovoltaic module, extracts the panel inclination angle, azimuth angle, and installation height data, and calculates the theoretical best contact angle in combination with the three-dimensional coordinate transformation algorithm. The calculation formula of the coordinate transformation matrix is: T = R z (α)·R x (β)·R y (γ); in the formula, T is the comprehensive coordinate transformation matrix; R z(α), R x (β), R y (γ) is the basic rotation matrix for rotation angles α around the z-axis, β around the x-axis, and γ around the y-axis; α is the azimuth angle of the photovoltaic panel, β is the tilt angle of the photovoltaic panel, and γ is the torsion angle of the photovoltaic panel. The theoretical optimal contact angle vector is calculated as: θ best =T·n0; where n0 is the normal vector in the standard coordinate system. Subsequently, the multi-angle adjustment mechanism device realizes precise positioning of the brush head by collaboratively controlling the three-degree-of-freedom joints. The horizontal rotation joint is driven by a stepper motor to realize rotation positioning in the range of 0 to 360 degrees with a minimum step unit of 0.1 degrees. The vertical pitch joint realizes angle adjustment in the range of -45 to 90 degrees through an electric push rod, and the push rod telescopic accuracy reaches 0.5 mm. The torsional joint realizes torsional movement in the range of ±180 degrees through a harmonic reducer, and the reduction ratio is 100:1, ensuring high-precision and low-backlash angle control. During the adjustment process, the angle sensor array monitors the actual angle value of each joint in real time at a frequency of 100 Hz and compares it with the theoretical angle value. The error vector is calculated as: e=θ actual -θ best Where, θ actual is the actual brush head angle vector measured by the angle sensor array. When the deviation exceeds 0.5 degrees, the PID controller is triggered to perform error compensation. The controller output is calculated as: Where u(t) is the control output; e(t) is the current error; K p , K i , K d These are the proportional, integral, and differential coefficients, respectively. Closed-loop feedback control enables each joint to quickly reach its target position and maintain stability. The system also monitors joint loads via a torque sensor array, ensuring that torque does not exceed 80% of the rated value during adjustment, preventing overload on the mechanical structure. After posture adjustment is complete, the system performs a light-touch contact force verification procedure to ensure uniform contact between the brush head and the panel surface, with contact pressure controlled within the range of 5 to 15 Newtons, laying the foundation for efficient and non-destructive cleaning.

[0163] The specific implementation method of step S04 is to achieve efficient and accurate cleaning execution and dynamic adjustment during the intelligent cleaning operation stage. The system first starts the independent drive motor array in the brush head main device according to the cleaning intensity parameters set in step S02. Each drive motor controls a cleaning unit separately, and uses a vector control algorithm to accurately adjust the speed and torque. A soft start strategy is used during startup to avoid impact. During the cleaning process, the current sensing circuit collects the operating current of each drive motor at a frequency of 20 Hz, and calculates the real-time cleaning resistance through the relationship model between the motor current and the cleaning resistance. The calculation formula for cleaning resistance is: F r =k c ·(I m-I0)+F0+ΔF T ; where F r is the cleaning resistance, in Newtons; k c is the conversion coefficient of current to force; I m is the actual operating current of the motor; I0 is the no-load current of the motor; F0 is the basic resistance constant; ΔF T is the temperature compensation term, calculated as: ΔF T = α T ·(T - T0); where α T is the temperature compensation coefficient; T is the current ambient temperature; T0 is the standard calibration temperature. The torque sensor array synchronously monitors the joint load torque data, and the contact force distribution between the brush head and the panel is obtained by inverse calculation of the manipulator Jacobian matrix. When the deviation between the detected value and the predicted value of the cleaning resistance exceeds 20%, the system performs adaptive adjustment. The adjustment formula for the cleaning unit pressure is: where P adj is the adjusted cleaning unit pressure; P cur is the current cleaning unit pressure; F base is the reference cleaning resistance. The whole process realizes smooth transition through the fuzzy adaptive control algorithm, ensures the best matching of the cleaning pressure and the dirt adhesion force, and protects the surface of the photovoltaic panel while achieving efficient cleaning. At the same time, the system uses a multi-parameter cleaning resistance prediction function to predict the cleaning resistance value at the position (x, y), and the calculation formula is: where F pred (x, y) is the predicted cleaning resistance value; T is the surface temperature of the photovoltaic panel; H is the ambient relative humidity; t a is the dirt adhesion time; D p is the average diameter of dust particles; θ is the contact angle between the brush bristles and the panel; a ijklm is the polynomial coefficient; n is the highest degree of the polynomial; ε is the error term. The prediction results are used to initially set the pressure of each cleaning unit and the power of the drive motor, and to determine the cleaning path priority and the key cleaning areas.

[0164] The specific implementation of step S05 is to establish a visualization mechanism for cleaning resistance distribution and an evaluation mechanism for the contact state of the brush bristles during the real-time monitoring of the cleaning effect. The system extracts the spectral characteristics of the current signal through Fourier transform based on the motor running current data collected by the current sensing circuit, and identifies the characteristic frequencies of different types of dirt. When there are fluctuations in the motor current between 12 Hz and 18 Hz, it usually indicates the presence of dust-like dirt; when there are fluctuations between 5 Hz and 8 Hz, it may be organic dirt such as leaves or bird droppings. The system calculates the cleaning resistance using the current data and generates a heat map of the cleaning resistance distribution using the bilinear interpolation algorithm. The interpolation calculation formula is: Where, F(x, y) is the interpolation result of the cleaning resistance at position (x, y); (x1, y1), (x1, y2), (x2, y1), (x2, y2) are the center positions of the four nearest cleaning units surrounding (x, y); F(x1, y1), F(x1, y2), F(x2, y1), F(x2, y2) are the cleaning resistance measurements at the corresponding positions. Different colors in the heat map represent different levels of cleaning resistance. The red area indicates that the cleaning resistance is greater than 1.5 Newtons, the yellow area indicates that the cleaning resistance is between 0.8 Newtons and 1.5 Newtons, and the green area indicates that the cleaning resistance is less than 0.8 Newtons. Through heat map analysis, the system identifies areas with uneven dirt distribution and automatically increases the number of cleanings for the red areas in the heat map. The calculation formula for the increased number of cleanings is: Where N add (x, y) represents the number of additional cleanings at position (x, y). The system also analyzes the bristle current variation pattern to estimate bristle deformation. When the current fluctuation is less than 50% of the baseline value, it indicates insufficient contact between the bristles and the panel; when the current fluctuation exhibits high-frequency, small oscillations, it indicates excessive bending of the bristles. Based on these analysis results, the system dynamically adjusts the cleaning unit position and pressure to ensure 100% cleaning coverage, with no areas missed.

[0165] The specific implementation of step S06 is to achieve timely dirt removal and water resource recycling during the brush head self-cleaning execution phase. The system continuously monitors the data collected by the current sensing circuit to evaluate the cleaning resistance change trend. When the average cleaning resistance of 10 consecutive samples increases by more than 30% compared to the initial value, or the turbidity sensor detects that the turbidity of the separated water exceeds 200 NTU, the brush head self-cleaning program is automatically triggered. The judgment formula for the self-cleaning trigger condition is: Where S clean It is the self-cleaning trigger signal; is the average cleaning resistance of 10 consecutive samplings; F init is the initial cleaning resistance value; T w The turbidity of the water is measured by the turbidity sensor. First, the control system smoothly moves the brush head body to a preset safe position at the edge of the photovoltaic panel through a multi-angle adjustment mechanism. The movement process uses a fifth-order polynomial trajectory planning to ensure a smooth transition. After reaching the safe position, the high-pressure water pump starts and gradually increases the water outlet pressure to 0.8 MPa. Fine water mist is sprayed onto the nanocomposite bristles through an array of micro-atomizing nozzles. The nozzles are distributed at a 120-degree fan angle to ensure that the bristle surface is fully covered and flushed. At the same time, the centrifugal fan starts and reaches the set speed of 2400 revolutions per minute, generating a negative pressure of 8 kPa, and the flushed sewage and dirt are sucked into the dirt separator through the dirt collection channel. The calculation formula for cyclone separation efficiency is: Where η is the separation efficiency; ρp is the density of dirt particles; d p is the diameter of dirt particles; v t is the tangential flow velocity; μ is the dynamic viscosity of water; D is the diameter of the cyclone separation chamber; N is the number of effective rotation cycles; Q is the volume flow rate. The relationship between the tangential flow velocity and the pressure drop is: In the formula, C d is the flow coefficient; ΔP is the pressure difference between the inlet and outlet of the cyclone separator; ρ w is the density of water. Inside the cyclone separation chamber, the mixture rotates at a tangential velocity of 30 meters per second. The centrifugal force causes the dirt particles with a larger density to move towards the outer wall and sink. The preliminarily separated dirt particles enter the dirt collection box; the water flow moves upward and passes through a filter screen assembly with a pore size of 50 microns for secondary filtration. The filtered clean water returns to the water tank for recycling. The entire self-cleaning process lasts for 20 to 30 seconds. The system controls the self-cleaning intensity and duration through the data feedback of the pressure sensor, flow sensor, and turbidity sensor. When the detection value of the turbidity sensor is lower than 50 NTU and remains stable for 5 consecutive seconds, it is determined that the self-cleaning is completed, and the system resumes to the cleaning operation state.

[0166] The specific implementation of step S07 is to establish a multi-parameter correlation model and achieve dynamic parameter adjustment during the adaptive cleaning parameter optimization stage. The system first constructs a correlation model between environmental parameters and cleaning effects based on a Bayesian network. The calculation formula for the conditional probability distribution between nodes in the Bayesian network is: In the formula, C i is the adjustment amount of the i-th cleaning parameter; Pa(C i ) is the set of parent nodes affecting C i ; μ i (Pa(C i )) is the conditional mean function of C i given the values of the parent nodes; σ i is the conditional standard deviation. The conditional mean function is represented by a linear regression model: In the formula, β i0 is the intercept term; β ij is the linear effect coefficient of the j-th environmental parameter E j ; β ijk is the interaction effect coefficient between environmental parameters E j and E k ; m is the number of environmental parameters. The system dynamically adjusts the cleaning parameters according to the inference results of the Bayesian network. The calculation of the water temperature adjustment amount is: The calculation of the water injection intensity adjustment amount is: The calculation of the cleaning unit pressure adjustment amount is: The calculation of the brush rotation speed adjustment amount is: In the formula, ΔTw , ΔF w , ΔP c , Δω b are the adjustment amounts of water temperature, water injection intensity, cleaning unit pressure, and brush rotation speed respectively; T p is the surface temperature of the photovoltaic panel; T w is the current water temperature; T e is the ambient temperature; F base is the reference water injection intensity; H is the ambient relative humidity; P cur is the current cleaning unit pressure; V is the ambient wind speed; ω cur is the current brush rotation speed; D d is the dust concentration. At the same time, the system also introduces the CleanFormer deep learning model for parameter optimization. The calculation formula for the sparsity parameter in the multi-head adaptive sparse attention mechanism of this model is: s = 0.5·(1 + U) - 0.2·|T n | - 0.1·|H n |; where s is the attention sparsity; U is the fouling distribution uniformity index, calculated as: T n is the normalized temperature value; H n is the normalized humidity value. The system continuously optimizes the parameter adjustment strategy through the reinforcement learning algorithm to minimize energy consumption while maximizing the cleaning effect. The adjustment range changes dynamically according to the actual cleaning effect feedback, usually controlled within ±50% of the reference value.

[0167] The specific implementation of step S08 is to realize system health status monitoring and fault intelligent processing during the fault warning and self-recovery stage. The system establishes a multi-level fault detection and processing mechanism based on the data collected by the fault diagnosis device. First, the vibration sensor monitors the system vibration signal at a sampling frequency of 100 Hz, extracts the vibration characteristics through wavelet packet analysis, and compares them with the vibration characteristic library in the normal working state. When the energy in a certain frequency band exceeds 150% of the normal value, it is determined as abnormal vibration; the sound sensor collects the system operation sound and extracts the acoustic characteristics through Mel-frequency cepstral coefficients, and compares them with the normal working sound pattern. When the acoustic similarity is lower than 0.6, it is determined as abnormal sound; the temperature sensor array monitors the surface temperature of each motor. When the temperature exceeds 70°C or the temperature rise rate is greater than 2°C / minute, it is determined as abnormal temperature. The system inputs these abnormal signals into the fault classifier based on the support vector machine. The decision function of the fault classifier is: where f(x) is the output of the decision function; x is the input feature vector; x i is the support vector; y i is the class label corresponding to the support vector; α i is the Lagrange multiplier; K(xi , x) is the kernel function; b is the bias term; N is the number of support vectors. The system adopts a radial basis kernel function:

[0168] where σ is the kernel function width parameter, determined by cross-validation, with typical values ranging from 0.5 to 2.0, d is the dimension of the feature vector; x i,j and x i are respectively the j-th eigenvalue of x i and x. The trigger condition for the fault handling strategy is: where S fault is the fault handling strategy trigger signal; P fault is the fault probability. The system identifies specific fault types and automatically executes corresponding handling strategies: for the clogging fault of the cleaning unit, the system raises the water pressure of the corresponding micro atomizing nozzle to 1.2 MPa and increases the suction of the centrifugal fan to 10 kPa, and attempts to dredge for 10 seconds; for the joint jamming fault, the system controls the joint to reciprocate 3 times, with an amplitude of ±5 degrees each time, to relieve the jamming; for the motor overheating fault, the system reduces the motor load to 60% of the rated value and increases the cooling interval time to 2 minutes; for the near-full sewage collection box fault, the system issues a replacement reminder and records the current cleaning position coordinates, and automatically resumes cleaning after replacement. If the fault persists after taking the above measures, the system will send a detailed fault report containing the fault code, abnormal parameter data, fault occurrence time, and system operation log to the remote monitoring center through the communication device. At the same time, the system enters the safe mode, shuts down the power supply of high-risk components, and moves the brush head to a safe position.

[0169] The specific implementation manners of steps S09 - S10 are the same as those described above and will not be elaborated here.

[0170] The brush head system of the multi-angle self-cleaning photovoltaic module cleaning robot in this embodiment adopts a modular design concept in terms of hardware implementation. Each functional module can work independently and cooperate with each other to ensure the overall stability and reliability of the system. The core control chip uses an industrial-grade 64-bit quad-core processor with a main frequency of 3 GHz, built-in artificial intelligence acceleration unit, and has powerful edge computing capabilities. Through the distributed computing architecture, complex cleaning strategy optimization calculations are completed locally, significantly reducing the dependence on the remote server and greatly improving the system response speed. The control chip is tightly connected to each functional module through a high-speed bus to achieve efficient data transmission and processing, ensuring the real-time and accuracy of the system operation.

[0171] The brush head main body device, as the key execution unit of the system, adopts the design of a cleaning unit array structure, which is composed of multiple detachable and independently operating cleaning units. Each cleaning unit is connected to the adjacent cleaning unit through a precisely designed snap - type connection structure, ensuring both the rigidity of the overall structure and facilitating maintenance and replacement. An independent drive motor is equipped inside each cleaning unit to directly drive the rotation of the nano - composite material bristles to perform the cleaning task. The nano - composite material bristles are one of the core innovation points of this system. They adopt a three - layer composite structure: a high - strength carbon fiber skeleton provides the basic support force, a high - elastic polyurethane matrix gives the bristles appropriate flexibility, and the nano - Ag + ion - modified microporous surface layer has antibacterial and self - cleaning properties. This composite structure design endows the bristles with the intelligent variable stiffness characteristic. When the environmental temperature is lower than 5°C, the material stiffness automatically increases by 10% to adapt to the low - temperature environment. When the environmental temperature is higher than 40°C, the material stiffness automatically decreases by 15% to prevent scratching the photovoltaic panel. A dirt collection channel running through the whole is designed inside the brush head main body, which is closely connected to the self - cleaning system to form a complete dirt collection and recycling system.

[0172] The multi - angle adjustment mechanism device realizes the flexible attitude adjustment of the brush head in three - dimensional space. It is composed of a horizontal rotation joint, a vertical pitching joint, and a torsion joint, which are driven by a rotation motor, an electric push rod, and a harmonic reducer respectively. The coordinated movement of these three joints enables the brush head to precisely adapt to photovoltaic panels installed at different inclinations and maintain the best cleaning angle. The joint movement speed and acceleration are intelligently adjusted according to the urgency of the cleaning task. In the normal cleaning mode, the maximum rotation speed of the horizontal rotation joint is 45 degrees per second, the maximum rotation speed of the vertical pitching joint is 30 degrees per second, and the maximum rotation speed of the torsion joint is 60 degrees per second. In the emergency cleaning mode, the maximum rotation speed of each joint increases by 50% to meet the emergency cleaning needs before sudden bad weather. Each joint is equipped with a high - precision angle sensor and a torque sensor to monitor the joint state in real time, forming a closed - loop control system to ensure the accuracy and safety of the movement.

[0173] The self - cleaning system device adopts an efficient circulation design. It pumps clean water from the water tank through a high - pressure water pump and sprays it onto the surface of the bristles through a micro - atomizing nozzle array for cleaning. These micro - atomizing nozzles adopt advanced ultrasonic atomization technology to atomize water into tiny particles with a particle size between 50 microns and 150 microns, and the atomized particle size can be automatically adjusted according to the degree of dirt adhesion, realizing the efficient utilization of water resources. The centrifugal fan, through intelligent variable - frequency control technology, automatically adjusts the rotation speed within the range of 800 revolutions per minute to 3000 revolutions per minute according to the real - time feedback of the flow sensor and the turbidity sensor, optimizing energy consumption while ensuring the dust - suction effect. The dirt separation system adopts cyclone separation technology and multi - stage filter screens for filtration to achieve the efficient separation of dirt in the sewage. The separated clean water flows back to the water tank for recycling, greatly reducing water resource consumption and improving the environmental protection performance of the system.

[0174] The environmental perception device consists of a variety of high-precision sensors, including an infrared thermal imager, a humidity sensor, a light sensor, a wind speed sensor, and a dust concentration sensor, comprehensively perceiving changes in environmental parameters and providing accurate decision-making basis for the system. The communication device adopts dual guarantees of industrial Ethernet and 5G wireless communication to ensure the reliability and real-time nature of data transmission. The data encryption processor performs multi-layer encryption on all transmitted data, effectively preventing data leakage and unauthorized access, and ensuring system security. The power management device and the fault diagnosis device work together to comprehensively monitor the system status, realize fault warning and automatic protection, and greatly improve the reliability and service life of the system.

[0175] To better understand and implement the present invention, the following provides an embodiment 2 of a specific application scenario of the present invention: In the application scenario of a large-scale photovoltaic power station, a cleaning and maintenance operation is carried out on a photovoltaic power station with an installed capacity of 20 megawatts. The photovoltaic power station is located in the northwest region, with an annual average dust deposition of 8.5 grams per square meter per month. It is equipped with 32,560 monocrystalline silicon photovoltaic modules of 60 cm × 100 cm, with an installation inclination of 30 degrees and an azimuth of due south (0 degrees). Conventional manual cleaning requires 16 people to work for 5 days to complete the full-site cleaning, and the cleaning quality is uneven. The average increase in power generation efficiency after cleaning is only 6.2%. Researchers applied the multi-angle self-cleaning photovoltaic module cleaning robot brush head system of the present invention to conduct cleaning operation tests. First, the system was initialized according to the photovoltaic module parameters and environmental conditions. The environmental parameter acquisition results are shown in Table 1:

[0176] Table 1 Environmental parameter acquisition data at the start of the cleaning operation

[0177] Environmental parameter Numerical value Weight coefficient Membership function value Surface temperature of photovoltaic panel 38℃ 0.3 100 Ambient relative humidity 25％ 0.2 85 Ambient wind speed 4.5 m / s 0.3 92 Dust concentration in air <![CDATA[75μg / m 3 > 0.2 78

[0178] Substituting these environmental parameters into the environmental suitability score calculation formula, we get S e = 90.1 points, higher than the threshold of 75 points. The system determines that the environmental conditions meet the requirements for the cleaning operation. Then, entering the pre-cleaning preparation stage, the system detects that the environmental temperature is 32°C, while the average temperature on the surface of the photovoltaic panel is 38°C. According to the water temperature calculation formula, the optimal cleaning water temperature should be 35°C (38°C - 3°C). At the same time, according to the measured dust concentration of 75 μg / m 3 , the system sets the brush rotation speed to 2,200 revolutions per minute, the cleaning unit pressure to 8.93 N, and the water injection intensity to 0.128 liters per minute per square meter. In the brush head attitude pre-adjustment stage, based on the installation inclination of 30 degrees and azimuth of 0 degrees of the photovoltaic module, the system uses the coordinate transformation matrix to calculate the theoretical optimal contact angle vector θ best = [0, 30, 0] T, then control the multi-angle adjustment mechanism device for precise positioning. During the adjustment process, the angle sensor array monitors the angles of each joint in real time, and the initial measurement value is θ actual =[1.2, 28.5, 0.5] T , and the error vector e = [1.2, -1.5, 0.5] is calculated T . The system starts the PID controller for error compensation, and uses the parameters K p =2.0, K i =0.1, K d =0.5. After 0.8 seconds of adjustment, the final measured angle is θ actual =[0.1, 29.8, 0.2] T , and the error is less than the 0.5-degree threshold, meeting the accuracy requirements.

[0179] In the intelligent cleaning operation stage, the system first uses the multi-parameter cleaning resistance prediction function to generate a heat map of the cleaning resistance distribution. The prediction results show that the resistance in the area below the photovoltaic module is significantly higher than that in the upper area. The numerical comparison is shown in Table 2:

[0180] Table 2 Comparison of the prediction results of the cleaning resistance distribution

[0181] Location area Predicted cleaning resistance (N) Cleaning frequency setting Area above panel 0.65-0.85 1 Middle area of panel 0.90-1.35 2 Area below panel 1.40-1.80 3

[0182] The system starts the independent drive motor array, and each cleaning unit starts to rotate for operation. The current sensing circuit monitors the running current of the motor in real time. The initial average value is 0.32 amperes, and the no-load current is 0.12 amperes. According to the cleaning resistance calculation formula, the actual initial resistance F r =0.25 N, which is significantly lower than the predicted value. The system automatically increases the pressure of the cleaning unit to 10.72 N. In the real-time cleaning effect monitoring stage, the system uses the bilinear interpolation algorithm to generate a heat map of the cleaning resistance distribution, with a resolution of 5 cm × 5 cm. After 15 minutes of cleaning, it is found that the resistance in a local area suddenly increases to 1.68 N, suspected of having bird droppings stains. The system automatically increases the cleaning times in this area to 3 times, and at the same time tracks and records the change of the cleaning resistance. The results are shown in Table 3:

[0183] Table 3 Record of the cleaning process in the local high-resistance area

[0184] Cleaning frequency Resistance before cleaning (N) Resistance after cleaning (N) Resistance reduction rate (%) The 1st time 1.68 1.25 25.6 The 2nd time 1.25 0.86 31.2 The 3rd time 0.86 0.42 51.2

[0185] During the cleaning process, when the average cleaning resistance increases to 0.83 N, exceeding 1.3 times the initial resistance (0.25×1.3 = 0.325), and the turbidity sensor detects that the water turbidity reaches 215 NTU, the system automatically triggers the self-cleaning program of the brush head. During the self-cleaning process, the high-pressure water pump starts and raises the water outlet pressure to 0.8 MPa, the cyclone separation efficiency reaches 92.4%, and the turbidity of the separated clear water drops to 23 NTU, meeting the requirements for recycling.

[0186] In the adaptive cleaning parameter optimization stage, the system detects that the temperature of the photovoltaic panel rises to 45°C, the ambient humidity drops to 18%, and the wind speed increases to 7.8 m / s. Based on the Bayesian network model, the system automatically adjusts the cleaning parameters: the water temperature is reduced by 5°C to 30°C, the water injection intensity is increased by 40% to 0.179 L / min·m², the pressure is increased by 18% to 10.72 N, and the brush rotation speed remains unchanged. At the same time, the CleanFormer deep learning model calculates that the dirt distribution uniformity index U = 0.68, the standardized temperature value T n = 0.57, the standardized humidity value H n = -0.64, thereby calculating the attention sparsity s = 0.77, making the model pay more attention to the areas with uneven dirt distribution.

[0187] During the cleaning operation, the fault diagnosis device detects that the temperature of the motor of the 8th cleaning unit reaches 72°C, exceeding the warning threshold of 70°C, and the vibration sensor detects that the vibration frequency energy ratio of this unit is 175% higher than the normal value. The support vector machine fault classifier determines it as a cleaning unit blockage fault, and the system automatically raises the water pressure of the corresponding micro-atomizing nozzle to 1.2 MPa and the suction of the centrifugal fan to 10 kPa. After 10 seconds, the blockage is successfully removed and the motor temperature drops to the normal range.

[0188] After the cleaning operation is completed, the system conducts a cleaning quality assessment. Calculated based on the resistance change before and after cleaning: the area of the excellent cleaning area (resistance reduction rate > 80%) is 17,850 m², the area of the good cleaning area (resistance reduction rate 60% - 80%) is 1,425 m², the total cleaning area is 19,530 m², and the cleaning coverage rate index C = 95.8%. The energy consumption index E = 24, and the dirt recovery rate R = 92.5%. Using the weight coefficients w c = 0.5, w e = 0.3, w r = 0.2, the comprehensive cleaning quality score Q = 72.95 is calculated, reaching a good level.

[0189] Traditional cleaning of photovoltaic modules mainly relies on manual operation or simple mechanical equipment. These methods generally have problems such as low cleaning efficiency, uneven cleaning quality, large waste of water resources, and easy damage to the panels. Traditional cleaning equipment usually adopts a brush head structure with a fixed angle, which cannot be optimized and adjusted according to the installation angle of the photovoltaic panel and the dirt distribution, and lacks a self-cleaning function, which is easy to cause secondary pollution. In terms of cleaning parameter settings, traditional equipment mostly uses fixed parameters and cannot adapt to complex and changeable environmental conditions, resulting in unsatisfactory cleaning effects or damage to the panels. The multi-angle self-cleaning brush head system of the photovoltaic module cleaning robot in the present invention has brought significant progress compared with traditional means:

[0190] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Table 4 below.

[0191] Table 4 Variables and Their Explanations

[0192]

[0193]

[0194] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered by the protection scope of the present invention.

Claims

1. A multi-angle cleaning brush head device for a photovoltaic module cleaning robot, characterized in that, It includes a control chip, a brush head main body device, a multi-angle adjustment mechanism device, a self-cleaning system device, an environmental perception device, a communication device, a power management device, a fault diagnosis device, and a human-computer interaction device. The control chip is electrically connected to other devices respectively. An automatic cleaning control module is provided in the control chip. The automatic cleaning control module is used to execute steps such as system initialization, preparation before cleaning, pre-adjustment of brush head posture, intelligent cleaning operation, real-time monitoring of cleaning effect, execution of brush head self-cleaning, optimization of adaptive cleaning parameters, fault warning and self-recovery, evaluation of cleaning quality, and system self-learning optimization. The cleaning resistance evaluation optimization is carried out by introducing a multi-parameter cleaning resistance prediction function, and the parameter optimization and self-learning process optimization are carried out by a pre-trained CleanFormer deep learning model.

2. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 1, characterized in that, The execution steps of the automatic cleaning control module include: system initialization, receiving a cleaning task instruction and judging whether the environmental conditions meet the requirements of the cleaning operation; preparation before cleaning, calculating the optimal cleaning water temperature and presetting the cleaning intensity parameters; pre-adjustment of brush head posture, adjusting the initial posture of the brush head; intelligent cleaning operation, performing cleaning and dynamically adjusting the pressure of the cleaning unit; real-time monitoring of cleaning effect, drawing a heat map of the cleaning resistance distribution; execution of brush head self-cleaning, triggering the brush head self-cleaning program; optimization of adaptive cleaning parameters, dynamically adjusting the cleaning parameters; fault warning and self-recovery, detecting abnormalities and performing fault handling; evaluation of cleaning quality, calculating the cleaning efficiency and quality score; system self-learning optimization, optimizing the cleaning strategy model.

3. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 2, characterized in that, The pre-adjustment of the brush head posture is specifically that, according to the actual installation angle of the photovoltaic module and the cleaning path planning, the initial posture of the brush head is adjusted through the multi-angle adjustment mechanism device. Specifically, through the coordinated movement of the horizontal rotation joint, the vertical pitching joint, and the torsion joint, the brush head main body device is kept at the best contact angle with the surface of the photovoltaic panel. At the same time, the joint angle data is fed back in real time through the angle sensor array to ensure accurate posture adjustment.

4. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 3, characterized in that, The brush head main body device includes a cleaning unit array, an independent drive motor array, nano-composite material bristles, a dirt collection channel, and a current sensing circuit. The nano-composite material bristles are composed of a high-strength carbon fiber skeleton, a high-elasticity polyurethane matrix, and a nano-Ag + ion-modified microporous surface layer.

5. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 4, characterized in that, The intelligent cleaning operation is specifically that the independent drive motor array in the brush head main body device is started to drive the nano-composite material bristles to rotate for cleaning. At the same time, the cleaning resistance is calculated according to the change of the motor running current collected by the current sensing circuit. Combining the joint load torque data monitored by the torque sensor array, the pressure applied by the cleaning unit on the photovoltaic panel is dynamically adjusted. If the cleaning resistance suddenly increases, the system automatically reduces the pressure of the corresponding cleaning unit and slows down the rotation speed of the bristles to prevent damage to the photovoltaic panel.

6. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 5, wherein, The real-time monitoring of the cleaning effect is specifically that, according to the motor running current data collected by the current sensing circuit, a heat map of the cleaning resistance distribution is drawn to identify the areas with uneven dirt distribution, and the cleaning times are increased in the areas with more dirt. At the same time, the contact situation between the bristles and the panel surface is analyzed through the bristle deformation data to ensure that the cleaning coverage rate reaches 100%.

7. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 6, characterized in that, Specifically, for the self-cleaning of the brush head, when the data collected by the current sensing circuit indicates that the cleaning resistance continues to increase or the turbidity sensor monitors that the turbidity of the water after separation exceeds the threshold, the system automatically triggers the self-cleaning program of the brush head, including stopping the cleaning operation and moving the brush head main body device to a safe position, starting the high-pressure water pump and the centrifugal fan, separating the sewage and dirt, collecting the dirt in the sewage collection box, and recycling the clean water back to the water tank for reuse.

8. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 7, characterized in that, Specifically, for the optimization of the adaptive cleaning parameters, by comprehensively analyzing the environmental parameters collected by the environmental perception device, the cleaning resistance data collected by the current sensing circuit, and the water quality cleanliness data monitored by the turbidity sensor, a multi-parameter correlation model is established to dynamically adjust the cleaning water temperature, water injection intensity, cleaning time arrangement, cleaning unit pressure, and brush rotation speed.

9. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 8, characterized in that Specifically, for the intelligent cleaning operation stage, a multi-parameter cleaning resistance prediction function is introduced for cleaning resistance evaluation and optimization. The multi-parameter cleaning resistance prediction function is used to predict the cleaning resistance distribution before the start of the intelligent cleaning operation. The inputs include the surface temperature of the photovoltaic panel, environmental humidity, dirt adhesion time, dust particle size distribution data, and the contact angle data between the brush and the panel. The output is the predicted cleaning resistance value and its spatial distribution heat map.

10. The multi-angle cleaning brush head device of a photovoltaic module cleaning robot according to claim 9, characterized in that, The CleanFormer deep learning model is a multi-modal deep learning network based on the Transformer architecture, including four main components: an environmental perception encoder, a cleaning parameter encoder, a timing information encoder, and a parameter optimization decoder. The environmental perception encoder is responsible for processing multi-dimensional environmental data from an infrared thermal imager, a humidity sensor, a light sensor, a wind speed sensor, and a dust concentration sensor and mapping it to a unified feature space; the sparsity parameter in the adaptive sparse attention mechanism of the CleanFormer deep learning model is determined by three key parameters: the surface temperature of the photovoltaic panel, environmental humidity, and the uniformity of dirt distribution.

Images