CBi-AM: A fusion model for predicting inverse nonlinear dynamics of magnetorheological elastomer isolator

Abstract

Magnetorheological elastomers (MREs) have attracted considerable attention due to their adaptive stiffness and damping properties, which make them highly effective for vibration isolation applications. Despite recent advancements in MRE isolation technology, significant challenges persist in accurately modeling the inverse nonlinear dynamics of MRE isolator. This paper proposes a novel CBi-AM model that integrates convolutional neural network (CNN), bi-directional long short-term memory (Bi-LSTM), and attention mechanism, to enable a refined understanding of feature importance across time-series data. Firstly, the dynamic performance of an MRE laminated isolator is test, showcasing the significant influence of excitation conditions. Then a robust dataset is generated through data augmentation on experimental signals, thereby enhancing the model training. Finally, several publicly models from both traditional machine learning and contemporary deep learning are constructed, and ablation experiments are conducted for comparative analysis. The results demonstrate that the CBi-AM model outperforms traditional and state-of-the-art baseline models in predicting the inverse performance of MRE isolator, achieving remarkable improvements in higher predictive accuracy ($R^2$=0.97802) and lower error metrics (MSE=0.04157, MAE=0.10665), which also exhibits its enhanced robustness and generalization capabilities. In addition to its intrinsic contribution to advancing inverse modeling techniques for MRE devices, this research lays the groundwork for future developments in modeling other mechanical systems characterized by strong nonlinearity.