Observation of deep reinforcement learning for power generation efficiency of variable tuned inertial mass electromagnetic transducers

Abstract

Various kinds of devices employing inerter technologies have been proposed for the purposes of structural control and energy harvesting. Especially, in recent years, considerable efforts have been made to develop variable inerter mechanisms to improve the performance; however, effective algorithms to control the variable inerter mechanism in real time have not been proposed so far. In this research, deep reinforcement learning is applied to the tuned inertial mass electromagnetic transducer (TIMET), one of the inerter devices taking advantage of resonance effect, to control the variable inerter mechanism in addition to the motor. Numerical simulation studies of the TIMET installed on a single-of-freedom oscillation model subject to disturbances are carried out, and the obtained results show that the deep reinforcement learning can be applicable to inerter devices in real time and has a great potential to improve the energy harvesting performance of variable inerter devices.