Posture control method based on photovoltaic cleaning robot

Abstract

The present invention discloses a posture control method based on a photovoltaic cleaning robot, which belongs to the technical field of photovoltaic cleaning robots, comprising: obtaining the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, setting a preliminary adjustment method, and calculating a preliminary torsion angle; obtaining a preset cleaning path and a robot kinematic model, setting a simulation algorithm, optimizing the preliminary torsion angle, and generating an actual torsion angle applied to actual cleaning; driving the photovoltaic cleaning robot to adjust to the actual torsion angle, and performing a cleaning operation on the photovoltaic panel to be cleaned. After the cleaning is completed, setting a blind spot compensation method, calculating the blind spot compensation angle, judging the dirt in the blind spot, determining a specific cleaning posture, and performing compensatory cleaning; recording the cleaning data of the photovoltaic panel to be cleaned, setting an adjustment optimization method, optimizing the adjustment coefficient used to calculate the torsion angle, and applying it to the calculation of the next torsion angle, thereby improving the overall cleaning effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic cleaning robots and relates to a posture control method based on a photovoltaic cleaning robot. Background Art

[0002] Against the backdrop of the global push for clean energy development, photovoltaic power generation, as an important form of renewable energy utilization, has experienced rapid development in recent years. The scale of photovoltaic power stations continues to expand, and the number of both large-scale ground-based centralized photovoltaic power stations and distributed rooftop photovoltaic systems widely distributed in cities is increasing. However, the presence of dust, dirt, and obstacles on the surface of photovoltaic panels seriously affects the power generation efficiency of photovoltaic power stations.

[0003] Patent application CN117289696A discloses a system and method for balancing a photovoltaic cleaning robot based on posture information. The method includes the following steps: providing a torsion-type photovoltaic cleaning robot, comprising a robot body, a cleaning device, and a posture adjustment system; using a 9-axis posture sensor in the posture adjustment system to detect the robot's posture and motion state; receiving and processing the posture sensor data via a control circuit to calculate balance adjustments; and based on the calculated results, using an actuator to balance the robot body to maintain the robot's stability during the cleaning process. This technical solution utilizes a 9-axis posture sensor to monitor the robot's posture and motion state in real time, achieving balance adjustments and ensuring the robot's stability and safety during the cleaning process.

[0004] Although the existing technology can adapt to different terrains and inclinations through autonomous balance adjustment, thereby improving cleaning efficiency and coverage, the robot lacks precise control of the torsion angle, speed, and force when adjusting the cleaning posture, and does not consider the cleaning feedback after the torsion operation, resulting in limited cleaning effect and robot stability. In particular, it is impossible to evaluate the cleaning effect after the torsion adjustment in real time, and it is impossible to determine whether the cleaning device is completely in contact with the surface of the photovoltaic panel, or whether there are blind spots in cleaning. For example, during the cleaning process, if there are local depressions or protrusions on the surface of the photovoltaic panel, the robot may not be able to perceive and adjust the cleaning posture in time, resulting in these areas not being thoroughly cleaned. Therefore, the present application provides a posture control method based on a photovoltaic cleaning robot. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a posture control method based on a photovoltaic cleaning robot. By sensing the terrain changes on the surface of the photovoltaic panel in real time, the torsion angle, speed and force are dynamically adjusted to ensure that the cleaning device is completely fitted with the surface of the photovoltaic panel. Based on the fusion of visual, pressure and temperature data, the cleaning effect is comprehensively evaluated to avoid cleaning blind spots. A dynamic feedback mechanism is established to adjust the cleaning strategy in real time to ensure the optimization of the cleaning effect.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A posture control method based on a photovoltaic cleaning robot is applied to the photovoltaic cleaning robot. The photovoltaic cleaning robot is provided with an upper component and a lower component, and the upper component and the lower component are provided with a rotating shaft. The posture control method based on the photovoltaic cleaning robot includes:

[0008] Obtain the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, including the relative angle difference and height difference, set the initial adjustment method, and calculate the initial torsion angle;

[0009] Obtain the preset cleaning path and robot kinematic model, set up a simulation algorithm, optimize the preliminary twist angle, and generate the actual twist angle used for actual cleaning;

[0010] The photovoltaic cleaning robot is driven to adjust to the actual twisting angle, and performs a cleaning operation on the photovoltaic panel to be cleaned. After the cleaning is completed, a blind spot compensation method is set, a blind spot compensation angle is calculated, and dirt in the blind spot is judged, a cleaning posture is determined, and compensation cleaning is performed;

[0011] Recording cleaning data of the photovoltaic panel to be cleaned, setting an adjustment optimization method, optimizing an adjustment coefficient used to calculate a torsion angle, and applying the optimized adjustment coefficient to the calculation of the torsion angle for the next time;

[0012] When performing compensation cleaning, the blind spot compensation method includes:

[0013] Obtaining the round-trip time from the laser radar emitting the laser beam to receiving it, and calculating the distance between the laser radar and each point on the photovoltaic panel surface;

[0014] Repeating the process of emitting and receiving the laser beam to obtain distance data and construct three-dimensional topographic information of the photovoltaic panel surface;

[0015] Synchronously photographing the photovoltaic panels, acquiring images and performing pre-processing;

[0016] Analyzing the pre-processed image, identifying features on the photovoltaic panel surface, and marking abnormal areas;

[0017] Registering the three-dimensional terrain information with the abnormal area, and performing terrain analysis on the abnormal area to identify clearing blind spots;

[0018] The blind spot depth of the blind spot clearing is extracted from the three-dimensional terrain information, and the blind spot compensation angle is calculated based on the actual torsion angle.

[0019] Furthermore, when calculating the preliminary torsion angle, the preliminary adjustment method includes:

[0020] Obtain the relative angle difference and height difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, and calculate the angle θ to be adjusted for the lower component adj ;

[0021] The theoretical adjustable threshold of the lower component is set to θ lower.th , and determine whether the angle to be adjusted is executable;

[0022] If θ adj ≤θ lower.th , the angle to be adjusted can be executed, and the torsion angle of the lower end component is set to the angle to be adjusted; if θ adj >θ lower.th , the angle to be adjusted is not executable, setting the torsion angle of the lower end component to the theoretical adjustable threshold;

[0023] Obtaining the upper component direction vector and the lower component desired direction vector, and calculating the relative torsion angles of the upper and lower components; wherein the lower component desired direction vector is the direction vector of the lower component after the torsion angle adjustment is completed;

[0024] Generate a preliminary torsion angle, including the torsion angle of the lower end component and the relative torsion angles of the upper and lower end components.

[0025] Furthermore, when optimizing the preliminary torsion angle, the simulation algorithm includes:

[0026] Constructing a comprehensive simulation model, and inputting the preliminary torsion angle and cleaning parameter values ​​into the comprehensive simulation model;

[0027] Setting initial parameters and monitoring the operating status of the photovoltaic cleaning robot in real time to detect problems;

[0028] Set the adjustment step size to θ step ,Once the problem detection result is that a problem occurs, the torsion angle is updated immediately;

[0029] Inputting the updated torsion angle into the comprehensive simulation model and performing iterative calculations until stable operation is achieved;

[0030] The optimized torsion angle is output and defined as the actual torsion angle.

[0031] Furthermore, the cleaning parameters include cleaning speed and cleaning intensity;

[0032] Setting a basic cleaning force, a basic cleaning speed, and a basic temperature, and obtaining the real-time cleaning force, real-time cleaning speed, and real-time temperature of the current environment of the photovoltaic cleaning robot;

[0033] Calculate the cleaning force and speed based on the real-time data obtained and the adjustment coefficient.

[0034] Furthermore, the blind spot compensation method further includes:

[0035] Adjust the cleaning posture according to the blind spot compensation angle and monitor in real time;

[0036] Comparing the dirt features in the cleaning blind area with a pre-established dirt sample library to obtain sample similarity of the image and determine the dirt condition in the cleaning blind area;

[0037] If the sample similarity exceeds a preset similarity threshold, and the dirt sample corresponding to the sample similarity is a heavily soiled sample, it is determined that there is a lot of dirt in the cleaning blind area, and the cleaning intensity is increased;

[0038] Otherwise, it is determined that the dirt level in the blind area is relatively light, and the original cleaning force and speed are maintained;

[0039] Clean the blind spot according to the blind spot compensation angle and cleaning force.

[0040] Furthermore, when optimizing the adjustment coefficient for calculating the torsion angle, the adjustment optimization method includes:

[0041] Set the target cleaning score and calculate the difference ΔE between the actual cleaning score and the target cleaning score;

[0042] Setting an acceptable threshold Δδ for the score difference to determine whether the score difference is within an acceptable range;

[0043] If |ΔE|≤Δδ, the score difference is within the acceptable range and the next cleaning is performed directly;

[0044] If |ΔE|>Δδ, the score difference is not within the acceptable range, and the adjustment coefficient is optimized;

[0045] Collect historical cleaning data to generate training datasets for training machine learning models;

[0046] Setting a functional expression of the adjustment coefficient and the actual cleaning score, defining the score difference as the objective function, calculating the gradient of the objective function with respect to the adjustment coefficient, and updating the adjustment coefficient according to the gradient descent algorithm;

[0047] Repeat the process of calculating the gradient and updating the adjustment coefficient for multiple iterations until the objective function converges and the iteration is stopped;

[0048] The optimized adjustment coefficient is applied to the parameter calculation of the next cleaning task.

[0049] Furthermore, before obtaining the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, various sensors arranged on the surface of the photovoltaic cleaning robot frame work together to continuously collect data;

[0050] During the cleaning process, the adaptive optimization of sensor layout is introduced to dynamically adjust the position and angle of the sensor according to the terrain and dirt distribution on the photovoltaic panel surface, and the time interval of the adaptive optimization is set to t op , every t op Perform an adaptive optimization.

[0051] Furthermore, the specific steps of adaptive optimization include:

[0052] Based on the collected historical environmental data, the terrain, dirt and temperature distribution characteristics are obtained and environmental information is generated;

[0053] Define the environmental information as state, sensor position and angle adjustment as action, and data collection quality as reward;

[0054] Based on the Q learning algorithm, the Q value is continuously updated iteratively until the algorithm converges, and the converged Q value table is output;

[0055] When the algorithm converges, the corresponding action is selected from the Q value table according to the current environment information, and the position and angle are adjusted.

[0056] Furthermore, before performing compensation cleaning, visual cameras, pressure sensors, and infrared sensors are used to collect data in real time to evaluate the cleaning effect, including:

[0057] Acquiring multi-dimensional real-time data on the surface of the photovoltaic panel;

[0058] Preprocess the photovoltaic panel surface image, divide the image into multiple regions, and identify the residual dirt area by comparing the pixel features of each region before and after cleaning, and calculate the visual image score;

[0059] Based on the mean and standard deviation of pressure, the pressure distribution score was calculated;

[0060] Dividing the photovoltaic panel surface into multiple regions, and calculating a temperature uniformity score based on the average temperature and the temperature range of each region;

[0061] Based on the visual image score, the pressure distribution score, and the temperature uniformity score, an actual cleaning score is calculated.

[0062] Beneficial effects of the present invention:

[0063] By adaptively optimizing the sensor layout and dynamically adjusting the sensor position and angle at specific time intervals based on the complex terrain and dirt distribution on the surface of the photovoltaic panel, it ensures stable and high-quality data collection even in complex environments, thereby improving data collection efficiency and stability. At the same time, it integrates multiple key factors to calculate the torsion angle and cleaning parameters, optimizes the comprehensive simulation model, monitors the motion state, corrects the torsion angle based on the adjustment coefficient and step size, and generates the actual torsion angle, so that the robot can flexibly adapt to the condition of the photovoltaic panel, avoid incomplete cleaning, and effectively improve the comprehensiveness and thoroughness of cleaning. By calculating the actual cleaning score in multiple dimensions and detecting the cleaning blind spot, accurate and efficient compensatory cleaning is achieved, further improving the overall cleaning effect. After each cleaning is completed, the adjustment coefficient is optimized based on the machine learning algorithm and applied cyclically to subsequent cleaning tasks, continuously improving the robot's cleaning performance, ensuring that it can maintain an efficient and high-quality cleaning working state in different photovoltaic panel environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is the overall structural diagram of the photovoltaic cleaning robot of the present invention;

[0065] Figure 2 It is a structural diagram of the upper end component of the present invention;

[0066] Figure 3 This is a flow chart of the posture control method based on the photovoltaic cleaning robot;

[0067] Figure 4 A flow chart for generating actual twist angles for the present invention;

[0068] Figure 5 is a flow chart of the blind spot compensation method of the present invention;

[0069] Figure 6 Flowchart of the optimization method for adjusting the present invention.

[0070] Reference numerals: 1. cleaning assembly; 2. lower assembly; 3. upper assembly; 4. rotating shaft. DETAILED DESCRIPTION

[0071] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0072] The photovoltaic cleaning robot used in the present invention refers to the photovoltaic cleaning device in a photovoltaic cleaning device supporting large slope walking disclosed in Chinese patent authorization announcement number CN222515284U, and the structure of the photovoltaic cleaning device itself is not described in detail.

[0073] Example 1

[0074] refer to Figures 1 to 4 As shown, this embodiment introduces a posture control method based on a photovoltaic cleaning robot, which is applied to the photovoltaic cleaning robot. The photovoltaic cleaning robot is provided with an upper component 3 and a lower component 2, which are used to twist and adjust the angle of the cleaning component 1 on the photovoltaic cleaning robot. The upper component 3 and the lower component 2 are provided with a rotating shaft 4, comprising the following steps:

[0075] The photovoltaic cleaning robot is driven to clean the photovoltaic panels. Various sensors arranged on the surface of the robot frame work together to continuously collect data and pre-process the collected data. The sampling interval is set to t s , in order to capture the dynamic changes of the photovoltaic panel status in real time, the sensors include high-precision laser radar, high-resolution visual camera, sensitive infrared sensor and pressure sensor. The laser radar is used to accurately measure the position, boundary and small terrain changes of the photovoltaic panel. The visual camera shoots the surface of the photovoltaic panel to obtain image data for identifying the location, shape and distribution range of dirt. The infrared sensor detects abnormal temperature on the surface of the photovoltaic panel. The pressure sensor senses the pressure changes between the cleaning component 1 and the photovoltaic panel in real time. The preprocessing includes denoising, filtering and normalization.

[0076] Combined with the preset cleaning path and the robot kinematic model, the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel is obtained, a preliminary adjustment method is set, and the preliminary torsion angle is calculated, including the torsion angle of the lower component 2, the relative torsion angle between the upper component 3 and the lower component 2, and the cleaning parameter value. The cleaning parameters include cleaning force and cleaning speed; and the mechanical properties and motion range of each component of the robot are comprehensively considered to ensure that the torsion angle meets the cleaning requirements without exceeding the robot's bearing capacity. A simulation algorithm is set to optimize the preliminary torsion angle and generate an actual torsion angle for actual cleaning; wherein, the upper component 3 and the lower component 2 of the photovoltaic cleaning robot are connected by a connecting plate and can be relatively twisted to meet the torsion requirements between the connecting bridges on the photovoltaic panels with large slopes;

[0077] The photovoltaic cleaning robot is driven to adjust to the actual twisting angle and clean the photovoltaic panels to be cleaned. After cleaning, the robot uses visual cameras, pressure sensors, and infrared sensors to collect data in real time to evaluate the cleaning effect. At the same time, the robot uses laser radar and visual cameras to collaboratively detect the blind spots on the surface of the photovoltaic panels, set the blind spot compensation method, calculate the blind spot compensation angle, and judge the dirt in the blind spots to determine the specific cleaning posture and perform compensation cleaning.

[0078] After each cleaning is completed, the cleaning data of the photovoltaic panel to be cleaned is recorded. According to the difference between the actual cleaning score and the target cleaning score, it is judged whether the adjustment coefficient is applicable, and the adjustment optimization method is set to optimize the adjustment coefficient used to calculate the torsion angle. The optimized adjustment coefficient is applied to the calculation of the next torsion angle to provide better parameters for the next cleaning until the cleaning task of the entire photovoltaic panel area is completed.

[0079] Furthermore, during the operation of a photovoltaic cleaning robot, the quality of the data acquired by the sensors directly affects the robot's assessment of the photovoltaic panel surface condition and the effectiveness of its cleaning tasks. Traditional sensors have a fixed layout and cannot adapt to the complex and changing photovoltaic panel surface environment, such as the differences in photovoltaic panel tilt angles under different terrains, the random distribution of surface dirt, and possible local occlusion. This makes it impossible to ensure stable and high-quality data collection.

[0080] By introducing adaptive optimization of sensor layout, the position and angle of the sensor are dynamically adjusted according to the complex terrain and dirt distribution on the photovoltaic panel surface to maximize the data collection efficiency. At the same time, the time interval of adaptive optimization is set as t op , every t op Performing adaptive optimization ensures that the sensor layout can periodically adapt to environmental changes. Even if the environmental changes are not obvious, regular optimization can maintain the efficiency of data collection. For some relatively stable photovoltaic panel areas, regular optimization can promptly adjust the sensor layout deviation caused by vibration and slight displacement, ensuring the stability of data collection quality.

[0081] The specific steps of adaptive optimization are as follows:

[0082] For lidar data, a Kalman filter algorithm is used to remove noise interference. A clustering algorithm is used to extract the terrain distribution characteristics of photovoltaic panels, such as average height and terrain standard. Visual image data is first grayscaled, and then the Canny edge detection algorithm is used to extract the edge contour of dirt. Image segmentation techniques (such as the Otsu algorithm) are used to determine dirt distribution characteristics, such as dirt area percentage and center of gravity location. The data from infrared sensors and temperature sensors are normalized to unify the dimensions and range, and temperature distribution characteristics, such as maximum temperature and temperature gradient, are generated. The processed terrain, dirt, and temperature distribution characteristics are integrated to construct the environmental information required by the reinforcement learning algorithm.

[0083] Define environmental information as state, sensor position and angle adjustment as action, including the position translation of each sensor in three-dimensional space and the rotation angle around different axes, and data collection quality as reward. For example, when the clarity of the image collected by the sensor improves (measured by image information entropy), the data coverage range increases (calculating the coverage rate of collected data points on the photovoltaic panel surface), or the data consistency improves (analyzing the degree of difference between multiple collected data), a higher reward is given, and vice versa.

[0084] Using the Q-learning algorithm for learning, we first create a two-dimensional table with rows representing different states and columns representing different actions. We then initialize all Q(s,a) values ​​to 0 or a small random value to initialize the Q-value table. Q(s,a) is the expected cumulative reward after performing action a in state s. The purpose of initializing the Q-value table is to provide a starting point for the algorithm, allowing it to gradually update and optimize these values ​​during the subsequent learning process.

[0085] At each time step t, according to the current state s t , adopt greedy strategy to select action a t ; Randomly select an action with probability ∈, and select the action with the largest Q value in the current state with probability 1-∈; As learning progresses, ∈ gradually decreases, and ∈=∈0×0.99 t , where ∈0 is the initial value;

[0086] Execute action a t , to adjust the position and angle of the sensor, perform data acquisition operations, and observe the new state s after the acquisition is completed t+1 , and update the Q value, calculate the maximum change ΔQ of all elements in the Q value table max,t , the expression is as follows:

[0087] Q new (s t ,a t )=Q old (s t ,a t)+α[r t+1 +γmax a Q(s t+1 ,a)-Q(s t ,a t )];

[0088]

[0089] Where Q old (s t ,a t ), Q new (s t ,a t ) are the Q values ​​before and after the update, α is the learning rate, which controls the influence of new information, r t+1 To take action a t The immediate reward obtained after, γ is the discount factor, which controls the current value of future rewards, max a Q(s t+1 ,a) is in the new state s t+1 The maximum Q value of all possible actions under the given conditions represents the maximum expected future rewards; Q(s t ,a t ) is in state s t Next, perform action a t The expected cumulative reward after For all states s t and action a t , calculate the absolute value of the difference between the Q values ​​before and after the update, and take the maximum value of these absolute values;

[0090] Repeat the above process of selecting actions, executing actions, observing results and updating Q values ​​until the algorithm converges. The judgment condition for algorithm convergence is the maximum change ΔQ in multiple consecutive iterations. max,t Less than the preset convergence threshold;

[0091] After the algorithm converges, it obtains the current environmental information of the photovoltaic panel surface, searches for the corresponding row in the converged Q value table based on the state vector, and selects the action corresponding to the column with the largest Q value. This action is the optimal action for adjusting the sensor layout in the current environment, including the position translation and angle rotation operations that each sensor should perform.

[0092] Analyze the adjustment information of each sensor during the action, drive the sensor's movement and rotation mechanism according to the respective adjustment instructions, and make precise adjustments to the position and angle;

[0093] After the sensor layout adjustment is completed, the current position and angle information of each sensor is obtained, and this information is encoded in a certain format, such as using coordinate values ​​to represent the position and angle values ​​to represent the rotation angle, and stored in the robot's local storage device.

[0094] Furthermore, the specific steps of the preliminary adjustment method include:

[0095] In photovoltaic power stations, different photovoltaic panels generally have different tilt angles, irregular arrangements, and uneven distribution of obstacles. These problems make it impossible for the photovoltaic cleaning robot to flexibly adjust its posture during the cleaning process, and the cleaning parts will not be able to effectively contact the surface of the photovoltaic panel, resulting in incomplete cleaning or missed areas, which seriously affects the cleaning effect and the power generation efficiency of the photovoltaic panel. According to the preset cleaning path, the robot starts to clean the photovoltaic panel. During the cleaning process, the sensor obtains the tilt angle and height change information of the photovoltaic panel surface, and uses the laser radar and visual camera to collect the relative angle difference Δθ and height difference Δh between the photovoltaic panel at the next moment and the photovoltaic panel at the current moment in real time to reflect the real-time status of the photovoltaic panel and calculate the angle θ to be adjusted of the lower component 2. adj , taking into account the height difference and angle difference of the photovoltaic panels, the robot can accurately adjust the posture of the lower component 2 according to the actual situation to better adapt to the surface condition of the photovoltaic panels; the expression of the angle to be adjusted is as follows:

[0096]

[0097] Where L is the wheelbase between the lower assembly 2 and the upper assembly 3, k1 is the adjustment coefficient, which is used to fine-tune the angle to be adjusted based on the relative angle difference Δθ of the photovoltaic panels, and the height difference Δh is the vertical position difference of the photovoltaic panels, which is caused by uneven installation of the photovoltaic panels, undulating terrain, or tilted connecting bridges.

[0098] Assume the theoretical adjustable threshold of the upper component 3 is θ upper.th The theoretical adjustable threshold of the lower component 2 is θ lower.th ; Among them, the theoretical adjustable threshold is related to the corresponding arc and preset angle of the connecting chute in the upper component 3 and the lower component 2. If the arc of the chute is The theoretical adjustable threshold is

[0099] The calculated angle to be adjusted θ adj Theoretically adjustable angle θ lower.th Comparison is performed to determine whether the angle to be adjusted is executable, ensuring that the cleaning component 1 does not exceed the structural limit during the adjustment process to avoid damage to the equipment, while maintaining the parallel posture of the cleaning roller brush to ensure the cleaning effect;

[0100] If θadj ≤θ lower.th , indicating that the angle to be adjusted does not exceed the maximum adjustable angle of the lower component 2, then the angle to be adjusted can be executed, let θ lower =θ adj , at this time the torsion angle of the lower component 2 is θ adj ;

[0101] If θ adj >θ lower.th , indicating that the angle to be adjusted exceeds the maximum adjustable angle of the lower component 2, the angle to be adjusted cannot be executed, let θ lower =θ lower.th , at this time the torsion angle of the lower component 2 is θ lower.th ; where θ lower is the torsion angle of the lower end component 2;

[0102] The robot's kinematic model, path planning information, and perception of the photovoltaic panel surface are used to determine the three-directional vectors of the upper component. and the desired direction vector of the lower component 2 Calculate the relative torsion angle θ between the upper component 3 and the lower component 2 using the vector angle formula rel , where, in order to make the cleaning component in good contact with the surface of the photovoltaic panel, the desired direction vector of the lower component 2 is the direction vector after the sliding groove of the lower component 2 completes the torsion angle adjustment, and the expression is as follows:

[0103]

[0104] Where MOD(·) is the formula for calculating the modulus of the direction vector;

[0105] Taking into account the uneven distribution of dirt on the surface of the photovoltaic panel and the differences in cleaning requirements of the robot at different positions and motion states, it is necessary to dynamically adjust the cleaning force and speed; set the basic cleaning force to F base , basic cleaning speed is v opt and base temperature is T opt , and use the pressure sensor to obtain the real-time cleaning force F during the current cleaning cu , the infrared sensor obtains the real-time temperature T of the current environment cu , and obtain the robot's current real-time cleaning speed v cu , calculate the cleaning force F based on the acquired data adj and sweeping speed v adj , in order to achieve dynamic adjustment of the cleaning force, the expression is as follows:

[0106] F adj =F base +k2(F target -F cu)+k3(v opt -v cu )+k4(T opt -T cu );

[0107] v adj =v opt +k5(F adj -F base );

[0108] In the formula, k2, k3, k4, k5 are adjustment coefficients, F target The target pressure value is determined based on the dirt level. The adjustment of the cleaning speed is closely related to the cleaning force, and the robot's power performance, energy consumption, and the tolerance of the photovoltaic panel are also comprehensively considered. When the cleaning force increases, the cleaning speed needs to be appropriately reduced to ensure the cleaning effect and robot stability. Conversely, in areas with less dirt, the cleaning speed can be increased to improve work efficiency.

[0109] Output the cleaning parameter value and the preliminary torsion angle, including the torsion angle of the lower end component 2 and the relative torsion angle between the upper end component 3 and the lower end component 2.

[0110] Furthermore, the specific steps of the simulation algorithm include:

[0111] In order to reproduce the motion state of the robot during the actual cleaning process in a virtual environment, various actual parameters are input to comprehensively and intuitively observe the performance of the robot under different working conditions. Computer simulation technology is used to integrate the robot's kinematic model, mechanical model and photovoltaic panel environment model to construct a comprehensive simulation model. The expected cleaning path, calculated torsion angle and cleaning parameter adjustment amount are input into the comprehensive simulation model. Among them, in the kinematic model, the robot is divided into an upper component 3, a lower component 2 and a cleaning component. The motion relationship between the various components of the robot is determined based on the robot's structural parameters (such as wheelbase, component dimensions, and joint range of motion). The mechanical model analyzes the stress and strain of each component when subjected to force based on material properties and structural design. The photovoltaic panel environment model covers the layout, tilt angle, row and column spacing of the photovoltaic panels, and possible obstacle information.

[0112] The torsion angle when a problem occurs in the current simulation (such as unstable center of gravity, excessive force on components, or collision with the photovoltaic panel frame) is used as the initial parameter, and the robot's motion state is monitored in real time. With the help of the robot's kinematic model and mechanical model, key indicators such as the center of gravity position (calculated through the mass distribution and position coordinates of each component), the force on the component (such as the torque of the rotating axis 4, the tension of the connecting component), and the distance from the photovoltaic panel frame are calculated to detect the problem. The problem detection monitors the key indicators of the photovoltaic cleaning robot. When the value of any of the above indicators exceeds the preset threshold, it is determined that the key indicator is abnormal (such as unstable center of gravity, excessive force on components, too close distance to the photovoltaic panel frame), and at this time, there is a problem with the motion state of the photovoltaic cleaning robot.

[0113] Set the adjustment step size to θ step , once a problem occurs, immediately update the torsion angle, the expression is as follows:

[0114] θ lower.new =θ lower -θ step ;

[0115] θ rel.new =θ rel +k6×θ step ;

[0116] Where k6 is the adjustment coefficient, which is determined according to the robot structure and center of gravity offset and can be obtained through multiple simulation tests; θ lower.new is the torsion angle of the lower end component 2 after adjustment, θ rel.new is the relative torsion angle between the upper component 3 and the lower component 2 after adjustment;

[0117] The adjusted torsion angle θ lower.new and θ rel.new Input the comprehensive simulation model, re-simulate the motion state of the robot during the entire cleaning process, and perform iterative calculations. When the simulation results satisfy the robot's stable operation during the entire cleaning process, and the mechanical properties and motion range of each component are within the safe range, that is, the center of gravity is within the support range, the component force is within the allowable range, and the distance from the photovoltaic panel frame is greater than the safety threshold, then stop the iteration; Among them, if the effect of multiple consecutive adjustments is not good, reduce the adjustment step size and let θ step =0.8×θ step , continue to adjust the torsion angle parameters and re-simulate;

[0118] Output the optimized torsion angle, including the torsion angle of the lower component 2 that meets the requirements, and the relative torsion angle θ between the upper component 3 and the lower component 2 rel.new , and is defined as the actual torsion angle, which is used in subsequent actual cleaning.

[0119] Furthermore, the specific steps for evaluating cleaning effectiveness include:

[0120] Using visual cameras, pressure sensors, and infrared sensors, real-time data collection begins, obtaining multi-dimensional, real-time data on the cleaning status of the photovoltaic panel surface. This provides rich data support for the cleaning effect evaluation model, enabling more accurate evaluation of cleaning results. The visual camera continuously captures images of the photovoltaic panel surface, recording changes in the panel surface state before and after cleaning. The pressure sensor measures the pressure distribution of the cleaning roller brush when it contacts the photovoltaic panel surface, obtaining pressure data for different areas. The infrared sensor monitors temperature changes on the photovoltaic panel surface and collects temperature data.

[0121] The data collected by the visual camera, pressure sensor and infrared sensor are integrated to calculate the actual cleaning score E clean , the expression is as follows:

[0122] E clean =w1E image +w2E pre +w3E temp ;

[0123] Where, E image To score visual images, the surface images of photovoltaic panels are analyzed through image recognition algorithms to identify indicators such as the degree of dirt residue and cleaning coverage and perform quantitative scoring. pre To score the pressure distribution, we analyze the uniformity and stability of the pressure distribution based on the pressure data collected by the pressure sensor. temp The temperature uniformity score is based on the temperature data obtained by the infrared sensor to evaluate the uniformity of the photovoltaic panel surface temperature. w1, w2, and w3 are weight coefficients, and w1+w2+w3=1;

[0124] Furthermore, the image recognition technology is used to pre-process the surface image of the photovoltaic panel taken by the visual camera, and the image is divided into multiple small areas. By comparing the pixel features of each area before and after cleaning, the dirt residue area is identified and quantitatively scored. The area of ​​the dirt residue area in the image is set as S dirt , the area of ​​the entire photovoltaic panel image is S total , the dirt type influence coefficient is k type , calculate the visual image score E image , the expression is as follows:

[0125]

[0126] Where, represents the dirt coverage ratio, k type Used to adjust the influence weight of different types of dirt, E imageThe rating range is 0-1, with 1 representing completely clean and 0 representing completely unclean;

[0127] Obtain pressure data from pressure sensors at different locations and calculate the average pressure value and standard deviation σ P The better the pressure distribution uniformity and the smaller the standard deviation, the more uniform the contact between the cleaning roller brush and the photovoltaic panel surface, and the better the cleaning effect may be; and the pressure distribution score E is calculated. pre , the expression is as follows:

[0128]

[0129] Where, reflects the discrete degree of pressure distribution, The smaller the E pre The closer it is to 1, the more uniform the pressure distribution is and the better the cleaning effect is; conversely, The larger the E pre The closer it is to 0, the more uneven the pressure distribution is, and there are areas of insufficient cleaning;

[0130] Use infrared sensors to obtain temperature data on the surface of photovoltaic panels, divide the surface of photovoltaic panels into multiple areas, and calculate the average temperature of each area and temperature extreme difference ΔT max , ΔT max The temperature uniformity score is the difference between the highest temperature and the lowest temperature. It mainly measures the consistency of the surface temperature of the photovoltaic panel. The smaller the temperature difference, the better the temperature uniformity. The temperature uniformity score E is calculated. temp , the expression is as follows:

[0131]

[0132] Where, is the relative change of temperature, The smaller the E temp The closer it is to 1, the better the temperature uniformity of the photovoltaic panel surface is. There is no temperature anomaly caused by excessive local friction or other abnormal conditions during the cleaning process. On the contrary, The larger the E temp The closer it is to 0, the worse the temperature uniformity is and the presence of cleaning problems.

[0133] Further, see Figure 5 ,The specific steps of the blind spot compensation method include:

[0134] During the cleaning process of the photovoltaic cleaning robot, the laser radar works continuously, emitting laser beams at a preset frequency. The laser beams are reflected back after hitting the surface of the photovoltaic panel, and the laser radar receives the reflected signals. Since the propagation speed of light in the air is constant, the round-trip time t from the laser emission to the reception is accurately measured. trip , calculate the distance d between the lidar and each point on the photovoltaic panel surface. The expression is as follows:

[0135]

[0136] Where c is the speed of light;

[0137] The process of emitting and receiving laser beams is repeated continuously. As the number of measurements increases, a large amount of distance data is obtained. A global coordinate system is constructed with the position of the lidar as the origin. The positioning sensor equipped with the lidar is used to obtain the position and posture information of the robot in the global coordinate system. The relative coordinates (distance) measured by the lidar are converted into absolute coordinates. These absolute coordinates are integrated to form point cloud data. The point cloud data is interpolated and fitted. The point cloud data is converted into a 3D model using 3D modeling software. The 3D terrain information of the photovoltaic panel surface is constructed and presented in the form of point cloud data.

[0138] Use a visual camera to synchronously shoot the surface of the photovoltaic panel, obtain a series of images, and pre-process the images;

[0139] A deep learning-based target detection algorithm is used to analyze pre-processed images, identify various features of the photovoltaic panel surface, and compare each area in the image with pre-set normal photovoltaic panel surface features. Areas that do not conform to normal photovoltaic panel surface features are marked as abnormal areas, such as areas with significant color or texture differences from normal areas. The target detection algorithm learns from a large number of labeled photovoltaic panel surface images (including normal areas and various abnormal areas) to acquire the ability to identify features of different areas. The visual camera can capture rich image details on the photovoltaic panel surface, and image recognition technology can use these details to quickly detect surface anomalies. Compared with other detection methods, it can intuitively determine surface abnormalities and provide more intuitive information for clearing blind spots.

[0140] To effectively integrate the 3D terrain information acquired by the LiDAR with the abnormal areas identified by the visual camera, the LiDAR point cloud data is registered with the visual image. By finding common feature points in the point cloud data and the visual image (such as the edges of the photovoltaic panel and specific marking points), a spatial transformation algorithm (such as the ICP algorithm) is used to make the two correspond in spatial position.

[0141] After completing data registration, the system combines information from the LiDAR and visual cameras to perform terrain analysis on abnormal areas marked in the visual image. If the LiDAR data shows a local depression or bulge in the abnormal area of ​​the visual image, and the depth or height exceeds a certain threshold, the abnormal area is determined to be a clearing blind spot.

[0142] Obtain the actual torsion angle, extract the blind spot depth Δd of the cleared blind spot from the point cloud data measured by the lidar, and calculate the blind spot compensation angle. The expression is as follows:

[0143] θ lower.comp =θ lower.new +k7×Δd;

[0144] θ rel.comp =θ rel.new +k8×k7×Δd;

[0145] Where k7 and k8 are compensation coefficients, which are determined by comprehensive analysis of simulation test and actual test data; blind area depth refers to the height difference between the blind area and the normal photovoltaic panel surface calculated based on point cloud data, which is obtained by analyzing and calculating the point cloud coordinates of the blind area and the normal area; θ lower.comp is the compensation angle of the lower component 2, θ rel.comp is the relative compensation angle between the upper component 3 and the lower component 2;

[0146] According to the calculated blind spot compensation angle, the robot sends instructions to adjust its cleaning posture. During the adjustment process, the robot's posture changes are monitored in real time to ensure that the cleaning roller brush can accurately move toward the blind spot, making it closer to or covering the blind spot.

[0147] When performing compensatory cleaning on blind spots, the dirt features in the blind spots are compared with a pre-established dirt sample library. The similarity between the current image and the images in the sample library is calculated, and the calculated similarities are sorted from largest to smallest. The largest similarity value is defined as the sample similarity of the image to determine the dirt situation in the blind spots. The dirt sample library contains image data of different types and degrees of dirt and their corresponding feature parameters.

[0148] If the sample similarity exceeds the preset similarity threshold, and the dirt sample corresponding to the sample similarity is a heavily soiled sample, it is determined that there is a lot of dirt in the cleaning blind area, and the cleaning intensity is increased;

[0149] Otherwise, it is determined that the dirt level in the blind area is relatively light, and the original cleaning force and speed are maintained;

[0150] Clean the blind spot according to the adjusted cleaning posture, blind spot compensation angle and cleaning force.

[0151] Further, see Figure 6 , the specific steps of adjusting the optimization method include:

[0152] Taking into account the power generation efficiency requirements of photovoltaic panels, cleaning industry standards and past cleaning experience, the target cleaning score is set to E. target , reflects the cleaning effect standard that the photovoltaic cleaning robot is expected to achieve after each cleaning, and calculates the score difference ΔE between the actual cleaning score and the target cleaning score, which reflects the degree of deviation between the actual cleaning effect and the expected effect. The expression is as follows:

[0153] ΔE=E target -E clean ;

[0154] The acceptable threshold of the score difference is set as Δδ, which is used to determine whether the gap between the current cleaning effect and the target effect is within an acceptable range;

[0155] Compare the calculated score difference ΔE with the acceptable threshold Δδ. If |ΔE| ≤ Δδ, the cleaning effect is close to the target cleaning effect, and the current adjustment coefficient can meet the cleaning needs to a certain extent and is relatively suitable, and the next cleaning can be carried out directly. If |ΔE| > Δδ, it indicates that there is a large gap between the cleaning effect and the target, and the adjustment coefficient needs to be optimized. Then proceed to the next step to optimize the adjustment coefficient. Among them, the adjustment coefficients include k1, k2, k3, k4, k5, and k6. |ΔE| represents the absolute value of the score difference ΔE.

[0156] Collect historical cleaning data, including the actual cleaning score, target cleaning score and corresponding adjustment coefficient of each cleaning, and organize the historical cleaning data into a training data set {A1,…,A n}, used to train machine learning models; where A n =(E target.n ,E clean.n ,k 1.n ,k 2.n ,k 3.n ,k 4.n ,k 5.n ,k 6.n ), A j is the cleaning data set of the jth cleaning in the historical cleaning data, E clean.j 、E target.j are the actual and target cleaning scores for the jth cleaning, k i.j is the i-th adjustment coefficient in the j-th cleaning, i = 1,…, 6, j = 1,…, n;

[0157] For the adjustment factor k i , set the function expression of the adjustment coefficient and the actual cleaning score to Eclean =F(k i ), the specific form of the function expression is determined by analyzing historical data and training the machine learning model, the score difference is defined as the objective function, and the objective function is calculated with respect to the adjustment coefficient k according to the principle of calculus i Gradient According to the gradient descent algorithm, the adjustment coefficient is updated and the expression is as follows:

[0158]

[0159] Where τ is the learning rate; k i.new is the updated adjustment coefficient, k i.old is the adjustment coefficient before updating;

[0160] Repeat the process of calculating the gradient and updating the adjustment coefficient for multiple iterations. Each iteration adjusts the adjustment coefficient in the direction of reducing the objective function until the score difference is less than the set convergence threshold, and then stop the iteration.

[0161] The optimized adjustment coefficient is applied to the parameter calculation of the next cleaning task.

[0162] In summary, the present invention drives the robot to clean, and multiple sensors collaborate to collect and pre-process data, improve data quality through adaptive optimization layout, combine the cleaning path and kinematic model, calculate the torsion angle and cleaning parameters, and iteratively optimize the torsion angle through simulation to generate the actual torsion angle; when performing cleaning, multi-sensor data is collected in real time, and the actual cleaning score is evaluated through multi-dimensional calculation. The laser radar and visual camera are used to collaboratively detect blind spots and compensate for cleaning. After each cleaning, the actual and target cleaning scores are compared, and the coefficient is dynamically optimized and adjusted based on machine learning to form a closed-loop feedback mechanism. Through continuous cycles and iterative operations, the robot's cleaning performance is continuously improved to complete the cleaning task of the entire photovoltaic panel area, ensuring that the robot can accurately and efficiently adapt to the complex photovoltaic panel environment and complete high-quality cleaning tasks.

[0163] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A posture control method based on a photovoltaic cleaning robot, applied to the photovoltaic cleaning robot, wherein the photovoltaic cleaning robot is provided with an upper component (3) and a lower component (2), and the upper component (3) and the lower component (2) are provided with a rotating shaft (4), characterized in that: include: Obtain the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, including the relative angle difference and height difference, set the initial adjustment method, and calculate the initial torsion angle; Obtain the preset cleaning path and robot kinematic model, set up a simulation algorithm, optimize the preliminary twist angle, and generate the actual twist angle used for actual cleaning; The photovoltaic cleaning robot is driven to adjust to the actual twisting angle, and performs a cleaning operation on the photovoltaic panel to be cleaned. After the cleaning is completed, a blind spot compensation method is set, a blind spot compensation angle is calculated, and dirt in the blind spot is judged, a cleaning posture is determined, and compensation cleaning is performed; Recording cleaning data of the photovoltaic panel to be cleaned, setting an adjustment optimization method, optimizing an adjustment coefficient used to calculate a torsion angle, and applying the optimized adjustment coefficient to the calculation of the torsion angle for the next time; When performing compensation cleaning, the blind spot compensation method includes: Obtaining the round-trip time from the laser radar emitting the laser beam to receiving it, and calculating the distance between the laser radar and each point on the photovoltaic panel surface; Repeating the process of emitting and receiving the laser beam to obtain distance data and construct three-dimensional topographic information of the photovoltaic panel surface; Synchronously photographing the photovoltaic panels, acquiring images and performing pre-processing; Analyzing the pre-processed image, identifying features on the photovoltaic panel surface, and marking abnormal areas; Registering the three-dimensional terrain information with the abnormal area, and performing terrain analysis on the abnormal area to identify clearing blind spots; The blind spot depth of the blind spot clearing is extracted from the three-dimensional terrain information, and the blind spot compensation angle is calculated based on the actual torsion angle.

2. The posture control method based on the photovoltaic cleaning robot according to claim 1 is characterized in that: When calculating the preliminary torsion angle, the preliminary adjustment method includes: Obtain the relative angle difference and height difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, and calculate the angle θ to be adjusted of the lower component (2) adj ; The theoretical adjustable threshold of the lower component (2) is set to θ lower.th , and determine whether the angle to be adjusted is executable; If θ adj ≤θ lower.th , the angle to be adjusted can be executed, and the torsion angle of the lower end component (2) is set to the angle to be adjusted; if θ adj >θ lower.th , the angle to be adjusted is not executable, and the torsion angle of the lower end component (2) is set to the theoretical adjustable threshold; Obtaining the direction vector of the upper component (3) and the desired direction vector of the lower component (2), and calculating the relative torsion angle between the upper component (3) and the lower component (2); wherein the desired direction vector of the lower component (2) is the direction vector of the lower component (2) after the torsion angle adjustment is completed; A preliminary torsion angle is generated, including the torsion angle of the lower end component (2) and the relative torsion angle of the upper end component (3) and the lower end component (2).

3. The posture control method based on the photovoltaic cleaning robot according to claim 2, characterized in that: When optimizing the preliminary torsion angle, the simulation algorithm includes: Constructing a comprehensive simulation model, and inputting the preliminary torsion angle and cleaning parameter values ​​into the comprehensive simulation model; Setting initial parameters and monitoring the operating status of the photovoltaic cleaning robot in real time to detect problems; Set the adjustment step size to θ step ,Once the problem detection result is that a problem occurs, the torsion angle is updated immediately; Inputting the updated torsion angle into the comprehensive simulation model and performing iterative calculations until stable operation is achieved; The optimized torsion angle is output and defined as the actual torsion angle.

4. The posture control method based on the photovoltaic cleaning robot according to claim 3 is characterized in that: The cleaning parameters include cleaning speed and cleaning intensity; Setting a basic cleaning force, a basic cleaning speed, and a basic temperature, and obtaining the real-time cleaning force, real-time cleaning speed, and real-time temperature of the current environment of the photovoltaic cleaning robot; Calculate the cleaning force and speed based on the real-time data obtained and the adjustment coefficient.

5. The posture control method based on the photovoltaic cleaning robot according to claim 4 is characterized in that: The blind spot compensation method further includes: Adjust the cleaning posture according to the blind spot compensation angle and monitor in real time; Comparing the dirt features in the cleaning blind area with a pre-established dirt sample library to obtain sample similarity of the image and determine the dirt condition in the cleaning blind area; If the sample similarity exceeds a preset similarity threshold, and the dirt sample corresponding to the sample similarity is a heavily soiled sample, it is determined that there is a lot of dirt in the cleaning blind area, and the cleaning intensity is increased; Otherwise, it is determined that the dirt level in the blind area is relatively light, and the original cleaning force and speed are maintained; Clean the blind spot according to the blind spot compensation angle and cleaning force.

6. The posture control method based on the photovoltaic cleaning robot according to claim 5, characterized in that: When optimizing the adjustment coefficient for calculating the torsion angle, the adjustment optimization method includes: Set the target cleaning score and calculate the difference ΔE between the actual cleaning score and the target cleaning score; Setting an acceptable threshold Δδ for the score difference to determine whether the score difference is within an acceptable range; If |ΔE|≤Δδ, the score difference is within the acceptable range and the next cleaning is performed directly; If |ΔE|>Δδ, the score difference is not within the acceptable range, and the adjustment coefficient is optimized; Collect historical cleaning data to generate training datasets for training machine learning models; Setting a functional expression of the adjustment coefficient and the actual cleaning score, defining the score difference as the objective function, calculating the gradient of the objective function with respect to the adjustment coefficient, and updating the adjustment coefficient according to the gradient descent algorithm; Repeat the process of calculating the gradient and updating the adjustment coefficient for multiple iterations until the objective function converges and the iteration is stopped; The optimized adjustment coefficient is applied to the parameter calculation of the next cleaning task.

7. The posture control method based on the photovoltaic cleaning robot according to claim 6, characterized in that: Before obtaining the difference between the photovoltaic panel to be cleaned and the current photovoltaic panel, various sensors arranged on the surface of the photovoltaic cleaning robot frame work together to continuously collect data; During the cleaning process, the adaptive optimization of sensor layout is introduced to dynamically adjust the position and angle of the sensor according to the terrain and dirt distribution on the photovoltaic panel surface, and the time interval of the adaptive optimization is set to t op , every t op Perform an adaptive optimization.

8. The posture control method based on the photovoltaic cleaning robot according to claim 7, characterized in that: The specific steps of adaptive optimization include: Based on the collected historical environmental data, the terrain, dirt and temperature distribution characteristics are obtained and environmental information is generated; Define the environmental information as state, sensor position and angle adjustment as action, and data collection quality as reward; Based on the Q learning algorithm, the Q value is continuously updated iteratively until the algorithm converges, and the converged Q value table is output; When the algorithm converges, the corresponding action is selected from the Q value table according to the current environment information, and the position and angle are adjusted.

9. The posture control method based on the photovoltaic cleaning robot according to claim 8, characterized in that: Before performing compensation cleaning, use visual cameras, pressure sensors, and infrared sensors to collect data in real time to evaluate the cleaning effect, including: Acquiring multi-dimensional real-time data on the surface of the photovoltaic panel; Preprocess the photovoltaic panel surface image, divide the image into multiple regions, and identify the residual dirt area by comparing the pixel features of each region before and after cleaning, and calculate the visual image score; Based on the mean and standard deviation of pressure, the pressure distribution score was calculated; Dividing the photovoltaic panel surface into multiple regions, and calculating a temperature uniformity score based on the average temperature and the temperature range of each region; Based on the visual image score, the pressure distribution score, and the temperature uniformity score, an actual cleaning score is calculated.

Images