Semi-active control architecture for MRE-based structure vibration isolation systems via combined Long–Short-Term-Memory and predictive control

Semi-active vibration control of Magneto-Rheological Elastomers (MREs) based isolators plays a crucial role in protecting structure from ground vibrations. Different hierarchical and modular control architectures have been proposed in the technical literature and several methods have been enacted for deriving the best suitable inverse nonlinear MRE model to be embedded into the vibration isolation control systems. In this direction, the prediction of the MRE magnetization current may be more accurate when leveraging the Long–Short Term Memory (LSTM) networks as they are able to capture long-term dependencies and complex temporal patterns in sequential data. Thus, vibration isolation performance of structures could be further improved. Accordingly, this work aims at proposing a novel hierarchical control architecture which exploits the combined advantages of MPC approaches at higher layer and the predictive ones of the LSTM at lower layer. In doing so, the proposed solution is able to compute the proper control force to be imposed to the structure, also adapting its value to the effective varying external conditions, while supplying the isolation system with the more appropriate magnetization current thanks to the improved prediction capabilities of LSTM network. The effectiveness of the proposed semi-active control system in guaranteeing structure vibration isolation is experimentally evaluated by considering an own developed test rig which consists of a vibrating structure base, a MRE plus Ball Transfer Units (BTUs) isolation systems and a flexible beam with two masses. Experiments, carried out in different scenarios and several operative conditions, confirm and disclose the efficacy and the advantages of the proposed approach