Rapid Controller Generation for Vibration Suppression of Structures Using Direct Excitation with Machine Learning

Abstract

Active mass dampers (AMDs) have been used to suppress vibration of structures subjected to dynamic loading. Generally, controller design of AMD requires an identified numerical model of the structure. However, unmodeled dynamics and system uncertainties could lead to mediocre control performance. In this study, a novel controller synthesis method of AMD named direct excitation with machine learning (DEML) is proposed and verified. In DEML, the AMD on top of the structure generates trivial excitation with sufficient bandwidth. Then the corresponding structural acceleration response is collected and used for training a controller formulated in an artificial neural network. Accordingly, the associated controller can be realized and implemented to generate control force from the structural acceleration response directly. Seismic control performance of the controller synthesized by DEML was verified numerically in which 9-story and 27-story building models were considered. Moreover, shake table testing of an AMD on top of a 3-story structural specimen was conducted to further validate the effectiveness of the proposed DEML. Experimental results demonstrate that the controller synthesized by DEML achieves competitive seismic control performance compared with a conventional linear-quadratic Gaussian controller.