Research On Automatic Solar Panel Cleaning System Based On Vision Control Point

Nowadays,in the environment of energy scarcity and the pursuit of pollution-free energy,renewable and clean energy is getting more and more attention.Among many renewable energy sources,the proportion of solar power generation is increasing year by year,but the photovoltaic power station is in a windy and sandy area,and the dust falling on the surface of solar panels will reduce the power generation efficiency,so the cleaning problem for the dust on the surface of solar panels needs to be solved urgently.In this paper,we choose the cleaning method of vehicle-mounted robotic arm from the actual environment of PV power station in China,develop vision algorithm to calculate the solar panel position information and control the actuator to complete the cleaning task.The main work of this paper is as follows.Firstly,design the robotic arm end cleaning device to clean the surface dust using wet type;then analyze the rotating torque of the roller brush and reasonably select the motor power.Next,the force existing between dust particles and solar panel is calculated,and the value of the force is calculated by taking the dust particles with the largest radius,and the finite element static analysis of the roller brush is performed to verify that the cleaning force on the surface of the roller brush is greater than the force between the dust particles and the solar panel.The Yaskawa industrial robot is selected as the actuator,and the end attitude of the robot arm and the transformation relationship between different joint coordinate systems are described by a matrix.The robot is modeled with reference to the D-H model,and the forward and inverse kinematics of the robot arm is analyzed and the solution process is derived.The interpolation methods in joint space and Cartesian space are analyzed,and linear and circular interpolation are simulated with the help of MATLAB.An algorithm based on digital image processing is designed to extract solar panel contours.The correspondence between a point in space and a pixel point in the image is explained using the small-aperture imaging model and the matrix transformation relationship is given.The principle of no-pass filtering algorithm implementation is analyzed,and the filtering algorithm is selected to suit the image processing of this paper.Through threshold segmentation and color space conversion,the approximate segmentation of solar panel outline is realized,and then the final complete outline is obtained through morphological processing.After experimental verification,it can meet the real-time requirements.Finally,camera calibration is performed on the vision system to obtain the internal parameter matrix of the camera.For the problem that digital image processing performs poorly in segmentation when facing the pixel points in the image with similar edges to the solar panel contour,the use of semantic segmentation in deep learning for contour extraction is further proposed.The first is to use Res Net-50 as the dust classification model backbone network to determine whether dust exists on the surface of solar panels,and change the Conv-1 convolutional layer of Res Net-50 to introduce the attention mechanism to improve the accuracy of classification;secondly,for solar panels with dust,Deep Lab V3+ is selected as the model for semantic segmentation,based on which Vo VNet27-slim replaces the backbone network of the original model,uses GSConv convolution to replace the 3×3 convolutional layers,and changes the convolutional module in the Decoder structure to MBConv to improve the detection rate.The datasets are produced by laboratory photography as well as supplementing similar images on the network,and the number of enhanced datasets is distributed as 5600 for the training set,700 for the validation set and 100 for the test set.Finally,the depth measurement model based on the center projection is proposed,and the maximum error is only 32 mm,which can meet the actual working conditions.