Magnetorheological elastomer(MRE), the smart rubber, is easy to change its shape and has characteristic of changing its stiffness according to the amount of magnetic field. Due to this characteristic, it is usually applied to many industries such as automobile, drone, robot, and so on. However, because of nonlinear-time varying characteristic of MRE, it is not practically used in the fields.
This paper proposes nonlinear-time varying characteristic control of the smart rubber and control system design of the smart rubber based tunable vibration absorber using reinforcement learning. Reinforcement learning is used to optimize and to solve complex problems in various fields and is applied in nonlinear and complex system optimization. Reinforcement learning and control system are verified by using the smart rubber based tunable vibration absorber. The tunable vibration absorber, developed in the previous research, is able to generate magnetic field from 0 to 340mT. The reinforcement learning model is designed to find the magnetic field that minimizes the vibration of the system when the disturbance vibration is applied to the target system. In order to verify the learning performance depending on the domain, the learning is performed in both frequency and time domain.
As a result, learning using the frequency spectrum in frequency domain learning converges very strongly and is distributed in the main magnetic field band for vibration isolation. However, characteristic of converging only in the main magnetic field prevents precise vibration control for various vibrations. In the time domain, the reinforcement learning used the most recently acquired data instead of frequency spectrum and had an opposite result with the learning in frequency domain. While the frequency domain learning had rapid convergence, the time domain learning did not converge for the various vibration. However, time domain learning converges not only to the main magnetic field but also to other magnetic fields for vibration isolation. The reason of different learning convergence rate between the frequency and time domain is analyzed as diversity of input data. Signals of the same amplitude have the same spectrum in the frequency domain, but are changed in various forms by drawing a sine wave in the time domain. Therefore, the result is analyzed that relatively static input data in the frequency domain allows the learning to be converged only to the main magnetic field at high, while it converges slowly but exactly in the time domain where the input data change is relatively large. it is necessary to find proper convergence rate and direction for reinforcement learning based control system for a Tunable Vibration Absorber. The rate and direction could be found through modifying network structure and hyperparameter of reinforcement learning.