NARX neural network modeling and robustness analysis of magnetorheological elastomer isolator

Due to the controllability of the stiffness and damping under the applied magnetic field, magnetorheological elastomer isolator has been proved effective in the field of vibration control. For the realization of vibration control application, an accurate MRE isolator model is a non-trivial task. However, the existing parametric modeling methods are required to identify too many parameters, which are difficult to implement. Moreover, the corresponding inverse dynamic model of the isolator cannot even be obtained by the identified model inversion. Therefore, this paper proposes a nonparametric neural network approach to approximate the dynamic behaviors of magnetorheological elastomer isolator with the characteristics of nonlinearity and hysteresis. Firstly, the dynamic characteristics of the isolator in shear-compression mixed mode are experimentally tested under different loading conditions. Secondly, based on the experimental data, a NARX neural network with three-layer structure is developed to approximate the functional relationship between inputs (displacement, velocity and current) and output (force) of magnetorheological elastomer isolator. Thirdly, the effectiveness of the network model is validated by comparing the predicted force and experimental force. Finally, considering the common occurrence of inputs with noise disturbance in real application, the robustness of the network is also verified for displacement and current inputs with noise disturbance, respectively. The results of the network generalization for experimental data show that the proposed NARX network is more robust and optimal than BP network.