Synthesis of a DQN-Based Controller for Improving Performance of Rotor System with Tribotronic Magnetorheological Bearing

Abstract

Journal bearings with magnetorheological fluids as tribotronic devices allow to adapt the parameters of the rotor-support system to changing operating modes. Multiple nonlinearities make them relatively complex objects for control. It is also necessary to take into account a number of restrictions at once, e.g., on the control action, friction, vibration amplitude. Such cases usually require applying advanced control techniques, such as based on machine learning methods. A controller based on deep Q-learning networks (DQN) was designed and tested for solving the task of reducing vibration and friction in the considered magnetorheological journal bearing. A nonlinear dynamic model of the rotor on such bearings was developed and validated by experimental study to implement the DQN controller. The bearing model is based on the magnetohydrodynamics equations and allows determining its dynamic parameters using the Multi-Objective Genetic Algorithm. The linearized dynamic bearing model takes into account the variability of the rotational speed and the control electromagnetic field. The learning process of the DQN agent was carried out with the establishment of a threshold on the vibration amplitude, which corresponds to the passing the resonant frequency of the rotor-bearing system. Testing the trained controller on the simulation model showed a decrease in the maximum amplitude by 15% in absolute values of vibration displacements, and by 32% in the time-averaged values comparing to a passive system. Also, the control current was not applied to the bearing outside the resonance zone, without increasing the coefficient of friction.