Novel model-free optimal active vibration control strategy based on deep reinforcement learning

Abstract

Neural networks (NNs) can provide a simple solution to complex structural vibration control problems. However, most past NN-based control strategies cannot guarantee an optimal policy in structural vibration control. In this study, a novel active vibration control strategy based on deep reinforcement learning is proposed, which utilizes the learning ability of NN controllers and simultaneously provides control performance comparable to traditional model-based optimal controllers. The proposed learning algorithm can determine the control policy through interaction with the environment without knowing dynamic system models. This study shows that the proposed model-free strategy can provide optimal control performance to various systems and excitations. The proposed control strategy is first verified on a single-degree-of-freedom model and subsequently extended to a multi-degree-of-freedom shear-building model. Its control performance with full-state feedback is nearly the same as that of a classical linear quadratic regulator. Moreover, the learned policy can outperform a traditional output feedback controller in a partially observed system. The robustness of the proposed control strategy against measurement noise is also tested.