Active control of wake-induced vibration using deep reinforcement learning

Abstract

Wake-induced vibration (WIV) is a typical type of flow-induced vibration. Effectively controlling such vibration is of significant value in engineering fields. In this study, we focus on the feasibility, effectiveness, and efficiency of the deep reinforcement learning (DRL)-guided active flow control for WIV control. Here an elastically mounted circular cylinder is interfered by the wake of an upstream equal-size cylinder at Reynolds number 100. With different center-to-center in-line distances, the unwanted vibration is noted to be more complicated than the vortex-induced vibration, which is then controlled by the rotary control with sensory motor cues as feedback signals. The control strategy is established by the DRL and is trained in the numerical environment built upon the lattice Boltzmann solver. For the tandem configuration, the DRL learns effective control strategies that can control the vibration amplitude by 99.7%, 99.2%, and 95.7%, for the cases with nondimensionalized gap length of 2, 6, and 8, respectively. Both time-averaged flow fields and vortex dynamics are discussed, revealing that the DRL-guided control learns different control strategies for different gap spacing. With the successfully learned strategy in tandem configuration, the WIV in staggered configuration is further explored based on the transfer learning. The vibration amplitudes of all cases in the staggered configuration are mitigated by more than 97%. To conclude, this study confirms that the DRL is effective in situations involving strong wake interference. It is anticipated that the DRL can provide a general solution for controlling flow-induced vibration.