Reinforcement learning vibration control of moving three flexible coupled beams

Abstract

A multi-agent reinforcement learning (RL) vibration controller is designed to actively suppress the vibration of a three-flexible coupled moving beam system and optimize the translational motion trajectory to minimize the residual vibration after motion. The finite element method is used to model the dynamic equation of the system, and the parameters of the actual model are identified by wavelet transform and intelligent optimization algorithm. A structured light visual vibration detection system is constructed to measure the vibration of three flexible beams. Based on the modified model, a trajectory optimization method of the three-flexible beam coupling system based on neural network optimization algorithm is designed. The QMIX and phasic policy gradient (PPG) controller are trained to obtain excellent nonlinear controllers for vibration control of piezoelectric actuators. The simulation and experimental results show that the optimized trajectory greatly reduces the vibration excitation, and the vibration at the end of the translation and movement is quickly suppressed. For the dual flexible beam coupled system, the designed QMIX controller suppresses the vibration completely in about 7 s, while under large gain proportional and derivative (PD) control, there is still residual vibration in 20 s. For the vibration of the first bending mode of the single flexible beam system, the designed PPG controller completely suppresses the vibration in about 3 s, while the vibration attenuation takes about 10 s for the large gain PD control. For the vibration of the first two bending modes, the PPG controller completely suppresses the vibration in about 4 s, while it takes about 11 s for the large gain PD control. The control effects of the two RL methods are better than that of PD control, especially for suppressing small amplitude vibration.