Semi-active vibration control of structural systems with a neural network that trained preview optimal control for MLD model

Abstract

We propose a new semi-active vibration control method of structural systems subject to seismic disturbances based on the optimal control for MLD models incorporating future seismic waveforms. In the proposed preview control method, we aim to realize the higher control performance than the conventional predictive control for the MLD model by considering future seismic waveforms. In the preview control law for the MLD model, a future data series of the seismic waveform is necessary. Because the future seismic waveform is not available, the method is generally impossible. In this paper, the control input data of the preview optimal control of the MLD model using future seismic waveforms are collected in advance subject to the recoded seismic waves, and the control inputs are trained with a multi-layered artificial neural network by using the past seismic waveforms. By replacing the preview optimal control for MLD model with the neural network, we are able to construct an implementable preview control system that shows high control performance using past seismic waveforms. Besides, the conventional method to obtain the control inputs of optimal control for MLD model has a problem on the high computational load, because it is MIQP that is NP-hard problem to obtain control inputs. Therefore, there are two difficulties in the real time implementation of the control method. By using the proposed neural network as the control law, the problem on the computational load can also be solved. The effectiveness of the proposed method is confirmed by a simulation using seismic waveforms observed in Japan. In the simulation study, the high control performance and the reduction of the computational load are demonstrated in comparison with implementable conventional methods.