Automated Layup Of Woven Composite On Nondevelopable Surface Based On Imitation Learning

Carbon fiber reinforced resin matrix composites are widely used in aerospace,railway transportation,automobile,watercraft,sports,medical and other fields due to their excellent properties.Among them,the two-dimensional woven fabric prepreg(referred to as woven)is more suitable for layup parts containing non-developable surfaces(referred to as curved parts)because of its good deformability.At present,there are some problems of low efficiency and poor quality consistency in manual woven layup of curved parts.Robot laying is helpful to solve these problems.Currently,the geodesics on the mould surface are taken as robot layup trajectory and the pressure is manually specified according to the shape of the mould.However,the layup path and pressure mentioned above fail to accurately describe the action characteristics of manual layup.This paper models the manual layup skill based on imitation learning,builds a woven layup robot platform and studies the teaching,modeling,reproduction and effectiveness evaluation of layup skill.The main work and conclusions are as follows:Firstly,a woven layup robot platform composes of UR5 e robot,ATI Gamma sixdimensional force/torque sensor,flexible roller,teaching handle,COMATRIX 3D camera and mould is built.Considering that only a single six-dimensional force/torque sensor can not meet the needs of dragging teaching and pressure recording simutaneously,a dual-sensor solution with the teaching handle installed between two sensors is proposed.Considering that manual teaching is difficult to ensure that the pressure is along the normal direction of the mould surface,a realtime adjustment of the teaching posture is realized based on the 3D point cloud.Then the teaching of layup skill is realized based on the admittance controller and the layup path and pressure are collected.Secondly,probability models are used to describe layup skill.Expectation maximization algorithm is adopted to establish Gaussian Mixture Model(GMM)of path and pressure versus time,and then Gaussian Mixture Regression(GMR)is used to obtain the expected output of the layup path and pressure.Finally,under the framework of force/position hybrid control,a pre-control quantity withdrawal method is proposed to reduce impact.For layup pressure tracking,Iterative Learning Control(ILC)is used to improve pressure control accuracy and the control quantity is sent to the robot several cycles in advance to solve the iterative nonconvergence caused by system delay.The effectiveness of the skill modeling method and the pressure controller is verified by woven layup experiment on a partial mould of an aircraft engine blade.The following conclusions are obtained: the root mean square error of pressure of each layup path is not more than 0.5N,and the maximum error of corresponding iterations is not more than 1.7N except for the second path,and the accuracy meets the requirements of use.Compared with the iterative learning pressure controller without considering delay,the maximum pressure error and root mean square pressure error are reduced by 50.0%and 55.6%,respectively.Compared with the PID controller with parameters tuning,the maximum pressure error and root mean square pressure error are reduced by 62.5% and63.6% respectively proving the effectiveness of the pressure controller.Through visual inspection of the surface profile of the molded parts,about 97.3% of the area meets the requirements,proving the effectiveness of layup skill.