Walking control method and system of photovoltaic robot

Abstract

The invention relates to the technical field of robot walking control, in particular to a walking control method and system for a photovoltaic robot, and the method comprises the steps: respectively collecting the point cloud data of two photovoltaic panels; fitting a photovoltaic panel surface equation based on the collected point cloud data of the two photovoltaic panels, and calculating to obtain a first height difference of the two photovoltaic panels according to the fitted photovoltaic panel surface equation; and acquiring a second height difference of the adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel points acquired by the depth camera, acquiring spatial attitude data of the photovoltaic panels and biaxial balance data of the cleaning robot, inputting the spatial attitude data and the biaxial balance data into a pre-constructed first machine learning model, and outputting a cleaning robot adjustment strategy. According to the walking control method of the photovoltaic robot, the laser radar point cloud data and the depth camera data are utilized at the same time, and after the two sets of data are subjected to weighted fusion, errors of a single sensor under complex illumination and light pollution conditions can be effectively compensated.

Description

Technical Field

[0001] The present invention relates to the technical field of cleaning robot walking control, specifically to a walking control method and system for a photovoltaic robot. Background Art

[0002] Traditional photovoltaic panel cleaning methods mainly rely on manual labor. This method is not only inefficient and labor-intensive, but also in some large-scale photovoltaic power stations or areas with complex terrains, the difficulty and cost of manual cleaning increase significantly. In addition, manual cleaning may damage the photovoltaic panels due to improper operation, affecting their power generation efficiency and service life. To solve these problems, photovoltaic cleaning robots have emerged.

[0003] For existing photovoltaic cleaning robots, when walking between photovoltaic panels for dust cleaning, there may be a height difference between adjacent photovoltaic panels, and it is necessary to obtain the height difference to control the walking of the cleaning robot. In the prior art, relying solely on a depth camera or a camera for data collection to obtain the height difference, due to complex lighting and widespread light pollution on photovoltaic panels, the accuracy of the collected data is low, affecting subsequent processing. Summary of the Invention

[0004] To solve the above problems, the present invention provides a walking control method and system for a photovoltaic robot.

[0005] The present invention adopts the following technical solutions. The walking control method for a photovoltaic robot includes:

[0006] Install a lidar on the cleaning robot, scan the surfaces of two adjacent photovoltaic panels in the forward direction, and respectively collect the point cloud data of the two photovoltaic panels;

[0007] Based on the collected point cloud data of the two photovoltaic panels, fit the surface equation of the photovoltaic panel, and calculate and obtain the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel;

[0008] Install an RGB-D depth camera in front of the robot to obtain the depth map of the surfaces of two adjacent photovoltaic panels in the forward direction;

[0009] Calculate the depth gradient of the photovoltaic panel surface through the depth map to obtain a gradient matrix, find the position of the pixel point with the largest gradient value, mark it as the target pixel point, and obtain the three-dimensional coordinates of the target pixel point through the depth camera;

[0010] Based on the three-dimensional coordinates of the target pixel point obtained by the depth camera, obtain the second height difference between adjacent photovoltaic panels, perform weighted summation on the obtained first height difference and second height difference to obtain the standard height difference, and collect the tilt angles of the two photovoltaic panels;

[0011] Obtain the spatial attitude data of the photovoltaic panel and the dual-axis balance data of the cleaning robot, input them into a pre-constructed first machine learning model, and output a cleaning robot adjustment strategy. The cleaning robot adjustment strategy includes controlling the height and angle adjustment of the wheels. The height adjustment is to adjust the wheel height to keep the robot chassis horizontal, and the angle adjustment is to adjust the tilt angle of the wheels to adapt to the angle change of the photovoltaic panel surface.

[0012] As a further description of the above technical solution: The method for fitting the photovoltaic panel surface equation based on the collected point cloud data of two photovoltaic panels includes:

[0013] Centralize the point cloud data of the first photovoltaic panel, calculate the mean value of the data points, and construct a centralized point matrix;

[0014] Calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix, take the eigenvector corresponding to the smallest eigenvalue, and denote this eigenvector as the normal vector of the fitting plane;

[0015] Generate a plane offset through calculation based on the normal vector of the fitting plane by a formula;

[0016] Obtain the fitting photovoltaic panel surface equation of the first photovoltaic panel based on the normal vector of the fitting plane and the plane offset. Similarly, obtain the fitting photovoltaic panel surface equation of the second photovoltaic panel.

[0017] As a further description of the above technical solution: The method for calculating and obtaining the first height difference between two photovoltaic panels according to the fitted photovoltaic panel surface equation is:

[0018] Obtain the absolute value of the difference in plane offsets between the two photovoltaic panels;

[0019] Divide the obtained absolute value of the difference in plane offsets by the modulus length of the normal vector of the second photovoltaic panel to obtain the first height difference.

[0020] As a further description of the above technical solution: The spatial attitude data of the photovoltaic panel includes the standard height difference and tilt angle of the two photovoltaic panels;

[0021] The dual-axis balance data of the cleaning robot includes the wheelbase and center of gravity distribution of the cleaning robot. The wheelbase includes the front and rear wheelbase and the left and right wheelbase.

[0022] As a further description of the above technical solution: The method for obtaining the center of gravity distribution includes:

[0023] Install pressure sensors on the four wheels of the cleaning robot to obtain the load borne by each wheel;

[0024] Taking the center of the robot chassis as the origin and the length of the cleaning robot as the horizontal axis, establish a rectangular coordinate system, and obtain the coordinates of the four wheels respectively, which are represented as ( , )、( , )、( , )、( , );

[0025] Based on the load borne by the wheels and the coordinates of the four wheels, calculate the abscissa and ordinate of the center of gravity;

[0026] Detect the tilt angle of the cleaning robot through an inertial measurement unit, calculate and obtain the vertical coordinate. The inertial measurement unit includes an accelerometer and a gyroscope, and the center of gravity distribution of the cleaning robot is represented by the abscissa, ordinate and vertical coordinate.

[0027] As a further description of the above technical solution: The method for detecting the tilt angle of the robot through the inertial measurement unit and calculating and obtaining the vertical coordinate includes:

[0028] Measure the accelerations of the cleaning robot in the three directions of the horizontal axis, vertical axis and vertical axis through the accelerometer, and record them as ;

[0029] Measure the angular velocities of the cleaning robot around the horizontal axis, vertical axis and vertical axis through the gyroscope, and record them as ;

[0030] Calculate and obtain the tilting angles around the horizontal axis and vertical axis;

[0031] Input the tilting angles around the horizontal axis and vertical axis and the three-dimensional model data of the cleaning robot into the constructed second machine learning model, and output the vertical coordinate of the center of gravity distribution of the robot.

[0032] As a further description of the above technical solution: The method for collecting the training data of the second machine learning model includes:

[0033] When the cleaning robot is in different working states and postures, use an inclination sensor installed on the cleaning robot body to collect the inclination data around the horizontal axis and vertical axis in real time, collect different inclination combinations, including the inclinations of the robot on flat ground, climbing slopes and turning, so as to cover various postures that the robot may appear. Use three-dimensional modeling software or laser scanning technology to obtain the three-dimensional model data of the cleaning robot, and obtain the vertical coordinate of the center of gravity distribution of the cleaning robot in different working states and postures through actual measurement or calculation based on physical principles. Establish a mapping relationship based on the collected inclination data around the horizontal axis and vertical axis, the three-dimensional model data and the corresponding vertical coordinate of the center of gravity distribution as the training data.

[0034] As a further description of the above technical solution: The training method of the second machine learning model includes:

[0035] Select a convolutional neural network model;

[0036] Divide the collected data into a training set, a validation set, and a test set according to a preset ratio;

[0037] Use the training set to train the model, set the optimizer and learning rate, update the model's parameters through the backpropagation algorithm, minimize the loss function, and verify through the validation set to make the model achieve the best performance on the validation set;

[0038] Use the test set to evaluate the trained model, calculate the mean square error index of the model, evaluate the performance of the model, and deploy and apply the trained second machine learning model after the performance evaluation of the model meets the standard.

[0039] As a further description of the above technical solution: The method for calculating and obtaining the abscissa of the center of gravity is: multiply the abscissa of each wheel of the cleaning robot by the load it bears respectively, sum them up and then divide by the total load, and the total load is the sum of the loads borne by each wheel;

[0040] The method for calculating and obtaining the ordinate of the center of gravity is: multiply the ordinate of each wheel of the cleaning robot by the load it bears respectively, sum them up and then divide by the total load.

[0041] The walking control system of the photovoltaic robot is used to implement the walking control method of the photovoltaic robot, and the system includes:

[0042] The first data acquisition module installs a lidar on the cleaning robot to scan the surfaces of two adjacent photovoltaic panels in the forward direction and respectively acquire the point cloud data of the two photovoltaic panels;

[0043] The first data processing module fits the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels, and calculates and obtains the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel;

[0044] The second data acquisition module installs a depth camera in front of the robot to obtain the depth maps of the surfaces of two adjacent photovoltaic panels in the forward direction;

[0045] The second data processing module calculates the depth gradient of the photovoltaic panel surface through the depth map, obtains the gradient matrix, finds the position of the pixel point with the largest gradient value, marks it as the target pixel point, and obtains the three-dimensional coordinates of the target pixel point through the depth camera;

[0046] The data analysis module obtains the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel point obtained by the depth camera, performs weighted summation on the obtained first height difference and the second height difference to obtain the standard height difference, and collects the tilt angles of the two photovoltaic panels;

[0047] The walking control module obtains the spatial attitude data of the photovoltaic panel and the biaxial balance data of the cleaning robot, inputs them into a pre-built first machine learning model, and outputs a cleaning robot adjustment strategy, including controlling the height and angle adjustment of the wheels.

[0048] Beneficial effects:

[0049] The walking control method of the photovoltaic robot provided by the present invention can obtain height information from both global (fitting the surface equation of the photovoltaic panel to obtain the first height difference) and local (analyzing the depth gradient to obtain the three-dimensional coordinates of the target pixel point and calculating the second height difference) angles by simultaneously using lidar point cloud data and RGB-D depth camera data. After weighted fusion of the two sets of data, it can effectively compensate for the errors of a single sensor under complex lighting and light pollution conditions, thereby obtaining a more accurate and robust standard height difference.

[0050] Furthermore, by accurately obtaining the standard height difference and the tilt angle of the photovoltaic panel, and then combining the center of gravity distribution of the robot and the wheelbase, an adjustment strategy is generated through the first machine learning model to control the adjustment of the wheel angle and height, so that the forces on each wheel are balanced when the robot crosses the height difference. This not only prevents the wheels on one side from being suspended or overloaded, but also ensures that the cleaning components always maintain good contact with the surface of the photovoltaic panel, improving the cleaning efficiency and reducing the risk of damage to the photovoltaic panel caused by uneven force. Description of the drawings

[0051] The present invention will be further explained below with reference to the drawings and embodiments:

[0052] Figure 1 It is a flowchart of the walking control method of the photovoltaic robot provided in Embodiment 1 of the present invention;

[0053] Figure 2 It is a flowchart of the method for obtaining the first height difference between two photovoltaic panels provided in Embodiment 1 of the present invention;

[0054] Figure 3 It is a flowchart of the method for obtaining the center of gravity distribution provided in Embodiment 2 of the present invention;

[0055] Figure 4 It is a module connection diagram of the walking control system of the photovoltaic robot provided in Embodiment 3 of the present invention;

[0056] Figure 5 It is a structural schematic diagram of the photovoltaic cleaning device provided in Embodiment 4 of the present invention;

[0057] Figure 6 It is a schematic diagram of the structure of the walking component provided in Embodiment 4 of the present invention Figure 1 ;

[0058] Figure 7Schematic diagram of the walking component structure provided in Embodiment 4 of the present invention Figure 2 。

[0059] Reference numerals: 1, main body frame; 2, arc-shaped cleaning frame; 21, bearing seat; 22, arc-shaped moving hole; 31, base; 32, walking wheel; 33, rotating shaft; 34, limit nut; 35, driving hanging wheel; 36, limit guide wheel; 37, protective hook; 4, driving shaft; 5, lidar and depth camera. Detailed implementation manners

[0060] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below with reference to specific drawings. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0061] Embodiment 1:

[0062] Please refer to Figure 1 and Figure 2 , an embodiment of the present invention provides a technical solution: a walking control method for a photovoltaic robot.

[0063] Install a lidar on the cleaning robot, scan the surfaces of two adjacent photovoltaic panels in the forward direction, respectively collect the point cloud data of the two photovoltaic panels, and represent the point cloud data on the first photovoltaic panel as: , = 1, 2, ……, N; where represents the th point cloud data, is the abscissa of the th point cloud data, is the ordinate of the th point cloud data, is the Z coordinate of the th point cloud data;

[0064] Based on the point cloud data of the two collected photovoltaic panels, fit the surface equation of the photovoltaic panel, and calculate and obtain the first height difference of the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel;

[0065] The method for fitting the surface equation of the photovoltaic panel based on the point cloud data of the two collected photovoltaic panels includes:

[0066] Centralize the point cloud data of the first photovoltaic panel, calculate the mean of the data points, and construct a centralized point matrix H; where , where , and is the average value of the abscissa, ordinate, and Z coordinate of all point cloud data, calculate the covariance matrix, and the calculation formula of the covariance matrix is: , where is the transpose of the centered point matrix H, perform eigenvalue decomposition on the covariance matrix , take the eigenvector (a1, b1, c1) corresponding to the minimum eigenvalue, and denote this eigenvector as the normal vector of the fitting plane;

[0067] Generate the plane offset based on the normal vector (a1, b1, c1) of the fitting plane, and the calculation formula of the plane offset is: , where is , and are the mean values of all point cloud data. Based on the normal vector of the fitting plane and the plane offset , obtain the fitting photovoltaic panel surface equation of the first photovoltaic panel, expressed as: . Similarly, obtain the fitting photovoltaic panel surface equation of the second photovoltaic panel, expressed as: . In the formula is the coordinate of the normal vector of the second photovoltaic panel based on the fitting plane.

[0068] The method for calculating and obtaining the first height difference between the two photovoltaic panels according to the fitting photovoltaic panel surface equation is:

[0069] Obtain the absolute value of the difference in the plane offsets of the two photovoltaic panels;

[0070] Divide the obtained absolute value of the difference in the plane offsets by the modulus length of the normal vector of the second photovoltaic panel to obtain the first height difference .

[0071] In this embodiment, the calculation formula of the first height difference is: ; where and are the plane offsets of the two photovoltaic panels, is the normal vector of the second photovoltaic panel, is for normalizing the normal vector to keep the scale of height calculation consistent.

[0072] It should be noted that by the difference in the offsets of the two planes, combined with the direction of the normal vector of the second plane, the height difference perpendicular to this plane is calculated. Since the modulus length of the normal vector will affect the coefficients of the plane equation, after normalization, this influence can be eliminated, ensuring that the height difference is based on the unit normal vector, so that the calculation result is more accurate and not affected by the scaling of the plane equation coefficients.

[0073] It should be understood that in three-dimensional space, the height difference of a plane depends not only on the difference in offsets but also on the direction of the plane (normal vector). Normalizing the normal vector is to eliminate the scale changes caused by different lengths of the normal vector and ensure that the height difference calculates the actual vertical distance in geometric terms, rather than a value affected by the coefficients of the equation. For example, if the normal vector is scaled, the coefficients in the plane equation will also change, but the actual geometric relationship remains the same. Normalization can maintain the consistency of the calculation.

[0074] In this embodiment, by fitting the surface equation of the photovoltaic panel based on the point cloud data of two collected photovoltaic panels and calculating the height difference of the photovoltaic panel surface through the surface equation of the photovoltaic panel, rather than directly using discrete point data, the calculation accuracy is improved, the influence of noise points on height calculation is avoided, and the height difference calculation becomes smoother and more stable.

[0075] Install an RGB-D depth camera (such as Intel RealSense or ZED) in front of the robot to obtain the depth maps of the surfaces of two adjacent photovoltaic panels in the forward direction. The depth maps provided by the depth camera are expressed as: , is the pixel coordinate of the depth image, is the height of the photovoltaic panel surface relative to the depth camera;

[0076] Calculate the depth gradient of the photovoltaic panel surface through the depth map, obtain the gradient matrix, find the position of the pixel point with the maximum gradient value, mark it as the target pixel point, and obtain the three-dimensional coordinates of the target pixel point through the depth camera, which are expressed as: , . are the three-dimensional coordinates of the target pixel point of the first photovoltaic panel, that is, the abscissa, ordinate, and Z coordinate, are the three-dimensional coordinates of the target pixel point of the second photovoltaic panel, that is, the abscissa, ordinate, and Z coordinate;

[0077] The method of calculating the depth gradient of the photovoltaic panel surface through depth information and obtaining the gradient matrix includes:

[0078] The depth gradient The calculation formula is: ; where the gradient intensity at point represents the degree of depth change, is the pixel coordinate of the depth image, that is, the depth value of each pixel point, represents the depth change rate along the x direction and represents the gradient in the horizontal direction, represents the depth change rate along the y direction and represents the gradient in the vertical direction.

[0079] Specifically, it should be noted that in the depth image, edge points are usually accompanied by large gradient changes because there are obvious changes in the depth values between the photovoltaic panel and the background or adjacent photovoltaic panels. Therefore, the edges can be found using the depth gradient.

[0080] Obtain the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel points obtained by the depth camera , the second height difference The calculation formula is: .

[0081] Sum the obtained first height difference and the second height difference through weighted summation to obtain the standard height difference;

[0082] Preferably, the calculation formula for summing the obtained first height difference and the second height difference through weighted summation to obtain the standard height difference is:

[0083] ;

[0084] In the formula, is the standard height difference, is the weight factor. Optionally , when the lidar signal is stable or the complex illumination affects the depth camera, assign 0.8 > > 0.7. Conversely, when the lidar signal is unstable or the illumination is well adjusted, assign 0.7 > > 0.6.

[0085] Specifically, by dynamically adjusting the size of the weight factor based on the real-time environmental state, automatically adjusting the weight under different illumination conditions, making the system more robust, avoiding the long-term impact of incorrect camera data on the fusion result, and allowing more reliable data sources to dominate the final calculation, thereby improving the overall measurement accuracy.

[0086] In this embodiment: By using the point cloud data of two photovoltaic panels collected, the surface equations of each are fitted by the least squares method, thereby calculating the first height difference. This method is not affected by a single lighting condition and can more accurately reflect the true geometric shape of the photovoltaic panel. Then, a depth map in the forward direction is obtained through an RGB-D depth camera, the depth gradient is calculated, and the pixel point with the maximum gradient value is found. Then, the three-dimensional coordinates of this pixel point are obtained, thereby obtaining the second height difference, that is, by using the local depth change information, it can supplement and verify the result of point cloud fitting. By simultaneously obtaining two sets of height difference data and then using weighted summation, the data deviation caused by environmental changes or local errors can be compensated. This redundant design makes the overall system less sensitive to external interference and ensures the accuracy of the collected data, thereby overcoming the problem that relying solely on a depth camera or a camera is easily affected by complex lighting and light pollution (such as solar reflection, shadows, etc.), resulting in large data noise and low acquisition accuracy. By fusing the results of two different data acquisition methods (point cloud and depth map), they correct each other to a certain extent and reduce the influence of single-sensor errors.

[0087] Embodiment 2:

[0088] Refer to Figure 3 , on the basis of the above embodiment, this embodiment discloses a method for generating an adjustment strategy for a cleaning robot:

[0089] Collect the inclination angles of two photovoltaic panels, and mark the standard height difference and inclination angles of the two photovoltaic panels as spatial attitude data;

[0090] Obtain the wheelbase and center-of-gravity distribution of the cleaning robot, and record the wheelbase LJ and center-of-gravity distribution of the cleaning robot as biaxial balance data; the wheelbase of the cleaning robot includes the front and rear wheelbase and the left and right wheelbase.

[0091] The method for obtaining the center-of-gravity distribution includes:

[0092] Install pressure sensors on the four wheels of the cleaning robot to obtain the loads F1, F2, F3, and F4 borne by each wheel;

[0093] Taking the center of the chassis of the cleaning robot as the origin and the length of the cleaning robot as the horizontal axis, establish a rectangular coordinate system, and respectively obtain the coordinates of the four wheels, expressed as ( , ), ( , ), ( , ), ( , ); where , , and are the abscissas of the four wheels, , , and are the ordinates of the four vehicles.

[0094] Based on the loads borne by the wheels and the coordinates of the four wheels, calculate and obtain the abscissa and ordinate of the center of gravity;

[0095] Detect the tilt angle of the cleaning robot through an inertial measurement unit, calculate and obtain the vertical coordinate. The inertial measurement unit includes an accelerometer and a gyroscope. The center of gravity distribution of the cleaning robot is represented by the abscissa, ordinate, and vertical coordinate.

[0096] Specifically, in this embodiment, the function of obtaining the center of gravity distribution of the cleaning robot is that according to the change of the center of gravity distribution, the system can adjust the suspension system and the angles of the wheels to ensure that the robot always maintains balance when crossing a height difference or an inclined photovoltaic panel, prevent tipping over or the wheels from hanging in the air, and the center of gravity distribution data can reflect the force conditions of each wheel. If the load borne by a certain wheel is too large, it may cause slipping or insufficient grip. The control system can adjust the driving force distribution and the wheel height based on these data to make the forces on each wheel more uniform, thereby improving the passability and stability.

[0097] The method for calculating and obtaining the abscissa and ordinate of the center of gravity based on the loads borne by the wheels and the coordinates of the four wheels includes:

[0098] The method for calculating and obtaining the abscissa of the center of gravity is: multiply the abscissa of each wheel of the cleaning robot by the load it bears respectively, sum them up and then divide by the total load, where the total load is the sum of the loads borne by each wheel.

[0099] The method for calculating and obtaining the ordinate of the center of gravity is: multiply the ordinate of each wheel of the cleaning robot by the load it bears respectively, sum them up and then divide by the total load.

[0100] The calculation formula for the abscissa of the center of gravity is: ;

[0101] The calculation formula for the ordinate of the center of gravity is: .

[0102] The method for detecting the tilt angle of the robot through the inertial measurement unit and calculating and obtaining the vertical coordinate includes:

[0103] Measure the accelerations of the cleaning robot in the three directions of the horizontal axis, vertical axis, and vertical coordinate axis through the accelerometer, and record them as ;

[0104] Measure the angular velocities of the cleaning robot around the horizontal axis, vertical axis, and vertical coordinate axis through the gyroscope, and record them as ;

[0105] Calculate and obtain the inclination angles around the horizontal axis and the vertical axis;

[0106] The inclination angle around the horizontal axis The calculation formula is: ;

[0107] The inclination angle around the vertical axis The calculation formula is: ;

[0108] Input the inclination angles around the horizontal axis and the vertical axis and the three-dimensional model data of the cleaning robot into the constructed second machine learning model to output the vertical coordinate of the center-of-gravity distribution of the robot.

[0109] The method for collecting the training data of the second machine learning model includes:

[0110] When the cleaning robot is in different working states and postures, use an inclination sensor installed on the cleaning robot body to collect the inclination angle data around the horizontal axis and the vertical axis in real time (it should be noted that high-precision inclination sensors are relatively expensive. If high-precision inclination sensors are directly installed on each cleaning robot for inclination angle data collection, the cost is too high. Only when obtaining experimental data can they be installed and used). Collect a variety of different inclination angle combinations, including the inclination angles of the robot in various situations such as on flat ground, climbing slopes, and turning, so as to cover all possible postures of the robot; The reason for doing this is that different working scenarios and actions will cause changes in the inclination angle of the robot, and these inclination angle changes are closely related to the center-of-gravity distribution of the robot. By widely collecting the inclination angle data in different working states and postures, the model can learn the complex relationship between the inclination angle and the center-of-gravity distribution;

[0111] Use 3D modeling software or laser scanning and other technologies to obtain the accurate three-dimensional model data of the cleaning robot. Include the size, shape, and mass distribution information of each component of the robot; It should be noted that the structure and mass distribution of the robot are important factors determining its center-of-gravity position. Accurate three-dimensional model data can provide accurate information about the physical characteristics of the robot for the model and help the model better understand the internal relationship between the structure of the robot and the center-of-gravity distribution;

[0112] Obtain the vertical coordinate of the center-of-gravity distribution of the cleaning robot in different working states and postures through actual measurement or calculation based on physical principles. For example, a weighing sensor can be used to measure the force on the robot at different support points, and combined with the geometric parameters of the robot, calculate the vertical coordinate of the center of gravity, or directly measure the position of the center of gravity through a balance test of the robot on a specific experimental device; Establish a mapping relationship based on the collected inclination angle data around the horizontal axis and the vertical axis, the three-dimensional model data, and the corresponding vertical coordinate of the center of gravity as the training data.

[0113] The training method of the second machine learning model includes:

[0114] Select a convolutional neural network model;

[0115] Divide the collected data into a training set, a validation set, and a test set according to a preset ratio (e.g., 7:2:1); the training set is used for model training, enabling the model to learn the patterns and features in the data; the validation set is used to monitor the model's performance during training, adjust hyperparameters, and prevent overfitting; the test set is used to finally evaluate the model's generalization ability and accuracy. Such data set division can ensure that the model can fully learn the features of the data during training, while evaluating the model's performance and generalization ability through the validation set and the test set, avoiding overfitting on the training data and poor performance in actual applications;

[0116] Use the training set to train the model, set an optimizer (such as Adam, SGD, etc.) and a learning rate, and update the model's parameters through the backpropagation algorithm to minimize the loss function. The loss function selects the mean squared error (MSE) function, which is used to measure the difference between the model's predicted value and the true value. During training, regularly record the training loss and validation loss of the model to observe the model's convergence. If overfitting is found in the model, regularization methods can be used for processing. By continuously adjusting hyperparameters and optimizing the model structure, and validating through the validation set, the model can achieve the best performance on the validation set.

[0117] Specifically, training the model in this way can enable the model to gradually learn the mapping relationship between the input data and the output data, continuously adjust the model parameters through the optimization algorithm, so that the model can accurately predict the vertical coordinate of the center of gravity distribution of the cleaning robot, while preventing overfitting through various strategies, and improving the model's generalization ability and stability.

[0118] Use the test set to evaluate the trained model, calculate the mean squared error index of the model, evaluate the model's performance. After the evaluation meets the standard, it can be used. If the model's performance does not meet the requirements, the model's prediction results can be further analyzed to find out the problems existing in the model, such as in which cases the model's prediction is inaccurate, whether it is due to data feature problems or model structure problems. According to the analysis results, further optimize and adjust the model, such as reselecting features, adjusting the model structure, increasing the data volume, etc., until the model's performance reaches a satisfactory effect. By evaluating and optimizing the model, the accuracy and generalization ability of the model can be continuously improved, enabling it to better apply to actual scenarios, accurately predict the center of gravity distribution of the cleaning robot, and provide strong support for the control and stability of the robot.

[0119] Input the spatial attitude data and the biaxial balance data into a pre-built first machine learning model, and output the adjustment strategy for the cleaning robot;

[0120] It should be noted that the adjustment strategy for the cleaning robot includes the adjustment of the height and angle of the wheels; Height adjustment: The adaptive suspension system adjusts the wheel height in real time by receiving the adjustment instruction, keeping the robot chassis level, ensuring that the cleaning components are closely attached to the surface of the photovoltaic panel, and making the force on the photovoltaic panel more uniform; Angle adjustment: When there is an inclination or unevenness on the surface of the photovoltaic panel, the control system of the cleaning robot will adjust the inclination angle (camber or caster angle) of the wheels to adapt to the angle change of the surface, ensuring that the robot remains stable during driving and the cleaning components apply force evenly.

[0121] The method for collecting the training data of the first machine learning model includes:

[0122] In different photovoltaic power stations, select representative areas, use high-precision measuring instruments (such as laser rangefinders) to measure the height difference between adjacent two photovoltaic panels, and use inclinometers to measure the inclination angle of the photovoltaic panels. The inclination angle of the photovoltaic panel is designed according to the local geographical latitude and lighting conditions, and the inclination angles of the photovoltaic panels in different regions are different.

[0123] Measure and obtain the wheelbase and center-of-gravity distribution of the cleaning robot. For example, when the cleaning robot is unloaded, fully loaded, and performing different cleaning tasks, measure its center-of-gravity position respectively. It is also possible to establish a physical model of the cleaning robot and calculate the center-of-gravity distribution according to the mass and layout of its components to obtain the center-of-gravity distribution data under different working conditions.

[0124] Associate the collected height difference and inclination angle data of the photovoltaic panels with the wheelbase and center-of-gravity distribution data of the cleaning robot. For each set of data, jointly determine the corresponding adjustment strategy for the cleaning robot (i.e., the adjustment values of the height and inclination angle of the wheels) by the robot control experts and the operation and maintenance personnel of the photovoltaic power station. Keep the cleaning robot stable and make the photovoltaic panel receive force evenly when passing through two photovoltaic panels with a height difference, and record the corresponding adjustment strategy of the height and inclination angle of the wheels as the corresponding label for each set of data and use it as the training data.

[0125] The first machine learning model training method includes: selecting a decision tree model, such as CART (Classification and Regression Tree). The decision tree model has good adaptability to tasks of classifying or making decisions based on multiple features. In this scenario, there is a complex logical relationship between the input spatial attitude data (standard height difference and tilt angle of the photovoltaic panel) and the dual-axis balance data (front and rear wheel distances, left and right wheel distances, and center of gravity distribution of the cleaning robot) and the output cleaning robot adjustment strategy (height and tilt angle adjustment of the wheels). The decision tree can build a tree structure and gradually divide samples according to different feature values, intuitively showing the decision-making process from input features to output strategies, with strong interpretability, facilitating understanding and analysis.

[0126] Divide the collected training data into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%. The training set is used for parameter learning of the model. A large amount of data enables the model to fully learn the rules between input features and output adjustment strategies. The validation set is used to evaluate the model performance in real time during training. By adjusting the hyperparameters of the model (such as the maximum depth, minimum sample number, etc. of the decision tree), overfitting of the model can be prevented, ensuring that the model has good generalization ability on different data. The test set is used to finally evaluate the performance of the model on new data, and to judge whether the model can accurately output an effective cleaning robot adjustment strategy, thereby verifying the practicability and reliability of the model, and making full use of data resources for model training and optimization.

[0127] Use the training set to train the decision tree model. During the training process, the model will calculate the information gain index based on the input spatial attitude data and dual-axis balance data, select the best splitting feature and splitting point, and build a decision tree. As the training progresses, the decision tree gradually grows, continuously dividing the sample data into different child nodes until specific stopping conditions are met (such as reaching the maximum depth or the number of samples in the child node is less than the minimum sample number). By continuously adjusting the structure and parameters of the decision tree, the model can accurately predict the corresponding cleaning robot adjustment strategy according to the input features, improving the prediction accuracy on the training data.

[0128] Evaluate the model performance on the validation set using the accuracy metric. The accuracy reflects the proportion of samples correctly predicted by the model in the total predicted samples; according to the evaluation results, adjust the hyperparameters of the decision tree. For example, if the model shows overfitting on the validation set, the maximum depth can be appropriately reduced to limit the complexity of the decision tree; if the model is underfitting, the maximum depth can be increased or the minimum sample number can be reduced to enable the decision tree to better learn the features in the data. By repeatedly adjusting the hyperparameters and evaluating the model performance, the optimal model parameter configuration on the validation set can be found.

[0129] The optimized model is finally tested using the test set, and the accuracy metric of the model on the test set is calculated. If the test results meet the expectations, it indicates that the model has good generalization ability and can be applied to actual scenarios to effectively output adjustment strategies according to the characteristics of the photovoltaic panels and the cleaning robot. If the test results are not ideal, it is necessary to recheck steps such as data preprocessing, model selection, and hyperparameter tuning, identify the problems and make improvements. The test set is independent of the training set and the validation set, and it is used to evaluate the model to truly reflect the performance of the model in actual applications, ensuring that the model can accurately output adjustment strategies when facing unseen data and improving the stability of the cleaning robot when passing over photovoltaic panels with height differences.

[0130] In this embodiment, by accurately obtaining the standard height difference and the tilt angle of the photovoltaic panel, and then combining the center-of-gravity distribution of the robot and the wheelbase, an adjustment strategy is generated through the first machine learning model to adjust the wheel angle and height, enabling the forces on each wheel to be balanced when the robot crosses the height difference. This not only prevents the wheels on one side from being suspended or overloaded but also ensures that the cleaning component always maintains good contact with the surface of the photovoltaic panel, improving the cleaning efficiency and reducing the risk of damage to the photovoltaic panel caused by uneven forces. In summary, this control method can significantly improve the driving stability and cleaning efficiency of the photovoltaic cleaning robot when crossing height differences and photovoltaic panels with different tilt angles through multi-angle data fusion, intelligent decision-making, and adaptive adjustment, while enhancing the system's adaptability to complex environments.

[0131] Embodiment 3:

[0132] Please refer to Figure 4 , this embodiment discloses a walking control system for a photovoltaic robot, which is used to implement the walking control method of the photovoltaic robot. The system includes:

[0133] The first data acquisition module installs a lidar on the cleaning robot to scan the surfaces of two adjacent photovoltaic panels in the forward direction and respectively collect the point cloud data of the two photovoltaic panels;

[0134] The first data processing module fits the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels and calculates and obtains the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel ;

[0135] The second data acquisition module installs a depth camera in front of the robot to obtain the depth maps of the surfaces of two adjacent photovoltaic panels in the forward direction;

[0136] The second data processing module calculates the depth gradient of the surface of the photovoltaic panel through the depth map to obtain a gradient matrix, finds the position of the pixel point with the largest gradient value, marks it as the target pixel point, and obtains the three-dimensional coordinates of the target pixel point through the depth camera;

[0137] The data analysis module obtains the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel points obtained by the depth camera. The obtained first height difference and the second height difference are weighted and summed to obtain the standard height difference, and the tilt angles of the two photovoltaic panels are collected.

[0138] The walking control module obtains the spatial attitude data of the photovoltaic panel and the biaxial balance data of the cleaning robot, inputs them into a pre-constructed first machine learning model, and outputs a cleaning robot adjustment strategy, including controlling the height and angle adjustment of the wheels.

[0139] Embodiment 4:

[0140] Please refer to Figures 5 - 7 In this embodiment, on the basis of the above embodiment, the walking control system of the photovoltaic robot is applied to a photovoltaic cleaning device for use.

[0141] Among them, the photovoltaic cleaning device includes a main frame 1, two walking components, and an arc-shaped cleaning frame 2 installed on the lower surface of the main frame 1. Driving shafts 4 are correspondingly arranged on the two vertical ends of the arc-shaped cleaning frame 2. A bearing seat 21 is arranged at the connection between the driving shaft 4 and the arc-shaped cleaning frame 2. A cleaning roller is installed between the two driving shafts 4, and the rotation of the cleaning roller is driven by the driving shaft 4 to clean the surface of the photovoltaic panel.

[0142] A lidar and a depth camera 5 are installed on the main frame 1.

[0143] The walking component includes a base 31 and two walking wheels 32. A rotating shaft 33 is arranged on the base 31. Arc-shaped moving holes 22 are opened above the driving shafts 4 on the two vertical ends of the arc-shaped cleaning frame 2. The rotating shaft 33 protrudes and is placed in the arc-shaped moving holes 22, and one end of the rotating shaft 33 extends out of the arc-shaped moving holes 22, and a limit nut 34 is arranged on the extended part of the rotating shaft 33. When the walking component generates torsion, the arc-shaped moving holes 22 move in the arc-shaped cleaning frame 2 relative to the arc-shaped cleaning frame 2. The end of the arc-shaped moving hole 22, that is, the limit position where the rotating shaft 33 can move, corresponds to the limit position of the torsion of the walking component.

[0144] A driving hanging wheel 35 is also rotatably arranged on one of the bases 31. The driving hanging wheel 35 is arranged perpendicular to the walking wheel 32. There are two driving hanging wheels 35 in total, and a limit guide wheel 36 is arranged between the two driving hanging wheels 35. The limit guide wheel 36 is bolted to the base 31.

[0145] A protective hook 37 is arranged on the other base 31, and the protective hook 37 is located between the two walking wheels 32.

[0146] Further, the walking assembly is provided with a rotating shaft 33. When the device is in use, the rotating shaft 33 can rotate within the arc-shaped movable hole 22, realizing the movable connection between the main body frame 1 and the walking assembly. On this basis, the structural feature that the rotating shaft 33 rotates around the rotating shaft 33 is realized, and the angle of the walking assembly is adjusted, thereby realizing the adjustment of the angle of the walking wheel 32;

[0147] Further, the walking wheel 32 is slidably arranged on the base 31 and can be adjusted in height. A hydraulic telescopic rod (not shown in the figure) is cooperatively provided. By controlling the hydraulic telescopic rod, the walking wheel 32 is driven to slide on the base 31 to adjust the height of the walking wheel 32.

[0148] When applying the control system for the automatic walking of the photovoltaic cleaning robot to this photovoltaic cleaning device, a lidar is installed on the main body frame 1 to scan the surfaces of two adjacent photovoltaic panels in the forward direction, respectively collect the point cloud data of the two photovoltaic panels, fit the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels, and calculate and obtain the first height difference between the two photovoltaic panels based on the fitted surface equation of the photovoltaic panel ;

[0149] An RGB-D depth camera is installed in front of the robot to obtain the depth map of the surfaces of two adjacent photovoltaic panels in the forward direction. An RGB-D depth camera is installed in front of the photovoltaic cleaning device to obtain the depth map of the surfaces of two adjacent photovoltaic panels in the forward direction; calculate the depth gradient of the photovoltaic panel surface through the depth map to obtain a gradient matrix, find the position of the pixel point with the largest gradient value and mark it as the target pixel point, and obtain the three-dimensional coordinates of the target pixel point through the depth camera; obtain the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel point obtained by the depth camera , and the obtained first height difference and the second height difference , perform weighted summation to obtain the standard height difference, and collect the inclination angles of the two photovoltaic panels; obtain the spatial attitude data of the photovoltaic panel and the biaxial balance data of the cleaning robot, input them into the pre-constructed first machine learning model, output the adjustment strategy of the cleaning robot, and adjust the height and angle of the walking wheel 32 based on the generated adjustment strategy.

[0150] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The descriptions in the above embodiments and the specification only illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A walking control method for a photovoltaic robot, characterized in that, Including: Install a lidar on the cleaning robot, scan the surfaces of two adjacent photovoltaic panels in the forward direction, and respectively collect the point cloud data of the two photovoltaic panels; Fit the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels, and calculate the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel; Install a depth camera in front of the robot to obtain the depth maps of the surfaces of two adjacent photovoltaic panels in the forward direction; Calculate the depth gradient of the surface of the photovoltaic panel through the depth map, obtain the gradient matrix, find the position of the pixel point with the largest gradient value, mark it as the target pixel point, and obtain the three-dimensional coordinates of the target pixel point through the depth camera; Obtain the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel point obtained by the depth camera, perform weighted summation on the obtained first height difference and the second height difference to obtain the standard height difference, and collect the tilt angles of the two photovoltaic panels; Obtain the spatial attitude data of the photovoltaic panel and the biaxial balance data of the cleaning robot, input them into a pre-constructed first machine learning model, and output the adjustment strategy of the cleaning robot. The adjustment strategy of the cleaning robot includes controlling the height and angle adjustment of the wheels. The height adjustment is to adjust the height of the wheels to keep the robot chassis horizontal, and the angle adjustment is to adjust the tilt angle of the wheels to adapt to the angle change of the surface of the photovoltaic panel.

2. The walking control method of the photovoltaic robot according to claim 1, characterized in that The method for fitting the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels includes: Perform centering processing on the point cloud data of the first photovoltaic panel, calculate the mean of the data points, and construct a centered point matrix; Calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix, take the eigenvector corresponding to the smallest eigenvalue, and record this eigenvector as the normal vector of the fitting plane; Generate a plane offset through calculation based on the normal vector of the fitting plane using a formula; Obtain the fitted surface equation of the first photovoltaic panel based on the normal vector of the fitting plane and the plane offset. Similarly, obtain the fitted surface equation of the second photovoltaic panel.

3. The walking control method of the photovoltaic robot according to claim 2, characterized in that, The method for calculating the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel is: Obtain the absolute value of the difference between the plane offsets of the two photovoltaic panels; Divide the obtained absolute value of the difference between the plane offsets by the modulus of the normal vector of the second photovoltaic panel to obtain the first height difference.

4. The walking control method of the photovoltaic robot according to claim 1, wherein, The spatial attitude data of the photovoltaic panel includes the standard height difference and tilt angle of the two photovoltaic panels; The biaxial balance data of the cleaning robot includes the wheelbase and center of gravity distribution of the cleaning robot. The wheelbase includes the front and rear wheelbase and the left and right wheelbase.

5. The walking control method of the photovoltaic robot according to claim 4, wherein, The method for obtaining the center of gravity distribution includes: Install pressure sensors on the four wheels of the cleaning robot to obtain the load borne by each wheel; Establish a rectangular coordinate system with the center of the robot chassis as the origin and the length of the cleaning robot as the horizontal axis, and respectively obtain the coordinates of the four wheels; Based on the load borne by the wheels and the coordinates of the four wheels, calculate and obtain the abscissa and ordinate of the center of gravity; Detect the tilt angle of the cleaning robot through an inertial measurement unit, calculate and obtain the vertical coordinate. The inertial measurement unit includes an accelerometer and a gyroscope, and the center of gravity distribution of the cleaning robot is represented by the abscissa, ordinate, and vertical coordinate.

6. The walking control method of the photovoltaic robot according to claim 5, characterized in that, The method for detecting the tilt angle of the robot by the inertial measurement unit and calculating the vertical coordinate includes: Measuring the accelerations of the cleaning robot in the three directions of the horizontal axis, vertical axis, and longitudinal axis by an accelerometer; Measuring the angular velocities of the cleaning robot around the horizontal axis, vertical axis, and longitudinal axis by a gyroscope; Calculating the tilting angles around the horizontal axis and the vertical axis; Inputting the tilting angles around the horizontal axis and the vertical axis and the three-dimensional model data of the cleaning robot into the constructed second machine learning model, and outputting the vertical coordinate of the center-of-gravity distribution of the robot.

7. The walking control method of the photovoltaic robot according to claim 6, characterized in that The method for collecting the training data of the second machine learning model includes: When the cleaning robot is in different working states and postures, using an inclinometer sensor installed on the cleaning robot body to collect in real time the tilting angles around the horizontal axis and the vertical axis, collecting different combinations of tilting angles, including the tilting angles of the robot on a flat ground, climbing a slope, and turning, obtaining the three-dimensional model data of the cleaning robot by using three-dimensional modeling software or laser scanning technology, obtaining the vertical coordinates of the center-of-gravity distribution of the cleaning robot in different working states and postures through actual measurement or calculation based on physical principles, and establishing a mapping relationship based on the collected tilting angles around the horizontal axis and the vertical axis, the three-dimensional model data, and the corresponding vertical coordinates of the center-of-gravity distribution as the training data.

8. The walking control method of the photovoltaic robot according to claim 7, characterized in that The training method of the second machine learning model includes: Selecting a convolutional neural network model; Dividing the collected data into a training set, a validation set, and a test set according to a preset ratio; Using the training set to train the model, setting an optimizer and a learning rate, updating the parameters of the model through the backpropagation algorithm, minimizing the loss function, validating through the validation set, and enabling the model to achieve the best performance on the validation set; Using the test set to evaluate the trained model, calculating the mean square error index of the model, evaluating the performance of the model, and deploying and applying the trained second machine learning model after the performance evaluation of the model meets the standard.

9. The walking control method of the photovoltaic robot according to claim 5, characterized in that, The method for calculating the horizontal coordinate of the center of gravity is: multiplying the horizontal coordinate of each wheel of the cleaning robot by the load it bears respectively, summing them up and then dividing by the total load, where the total load is the sum of the loads borne by each wheel; The method for calculating the vertical coordinate of the center of gravity is: multiplying the vertical coordinate of each wheel of the cleaning robot by the load it bears respectively, summing them up and then dividing by the total load.

10. A walking control system for a photovoltaic robot, which is used to implement the walking control method of the photovoltaic robot described in any one of claims 1-9, characterized in that, The system includes: A first data acquisition module, installing a lidar on the cleaning robot to scan the surfaces of two adjacent photovoltaic panels in the forward direction, and respectively collecting the point cloud data of the two photovoltaic panels; A first data processing module, fitting the surface equation of the photovoltaic panel based on the collected point cloud data of the two photovoltaic panels, and calculating and obtaining the first height difference between the two photovoltaic panels according to the fitted surface equation of the photovoltaic panel; A second data acquisition module, installing a depth camera in front of the robot to obtain the depth maps of the surfaces of two adjacent photovoltaic panels in the forward direction; A second data processing module, calculating the depth gradient of the photovoltaic panel surface through the depth map, obtaining a gradient matrix, finding the position of the pixel point with the maximum gradient value, marking it as the target pixel point, and obtaining the three-dimensional coordinates of the target pixel point through the depth camera. The data analysis module obtains the second height difference between adjacent photovoltaic panels based on the three-dimensional coordinates of the target pixel points obtained by the depth camera, performs weighted summation on the obtained first height difference and the second height difference to obtain the standard height difference, and collects the tilt angles of the two photovoltaic panels. The walking control module obtains the spatial attitude data of the photovoltaic panel and the biaxial balance data of the cleaning robot, inputs them into a pre-constructed first machine learning model, and outputs a cleaning robot adjustment strategy, including controlling the height and angle adjustment of the wheels.

Images