Modelling of a self-sensing magnetorheological damper using bayesian regularized NARX neural network

The magnetorheological (MR) damper has been demonstrated to be one of the most promising semiactive control devices to suppress structural vibration. Recently, a novel self-sensing MR damper has been fabricated by integrating an actuation-only MR damper with a piezoelectric force sensor. Possessing the sensing-while-damping function, the damper offers a cost-effective innovation for real-time semiactive structural vibration control. However, due to its intrinsic nonlinear characteristics, modelling of the damper to adequately describe its hysteresis dynamics has been one of the prerequisite and challenging tasks for fully exploring its capabilities in real-time control implementation. In this paper, forward and inverse dynamic models of the self-sensing MR damper are developed based on the combined NARX (nonlinear autoregressive model with exogenous inputs) and neural network techniques. Experiments are performed to collect training and validation data for the NARX neural networks. The Bayesian regularization is adopted in the training phase to prevent over-fitting. Validation results indicate that the trained NARX neural network models accurately represent the forward and inverse dynamics of the damper, exhibit good generalization capability, and are adequate for control design and analysis.