An optimum tuning of machine learning methods for predicting magnetorheological damper damping force for shock and vibration mitigation

Abstract

The control of vibration in vehicles can be effectively achieved through the use of magnetorheological (MR) dampers. Feedforward neural network (FFNN) models offer a flexible and powerful alternative for predicting the damping force of MR dampers. However, traditional backpropagation-based neural network models often suffer drawbacks such as long training durations and the risk of getting stuck in local solutions. To address these challenges, this paper proposes a hybrid approach that combines particle swarm optimization (PSO) for tuning hyperparameters with the extreme learning machine (ELM) method to predict the highly nonlinear force behavior of MR dampers. The ELM method is applied to a newly developed MR damper with a meandering valve, using a simulation scheme that takes velocity, electrical currents, and strokes/displacements as inputs and damping force as the output. The model is constructed based on real data obtained from tests conducted on the magnetorheological damper under different operating conditions. The results demonstrate that combining the ELM method and PSO enables accurate prediction of the MR damper's behavior while maintaining a relatively low number of hidden nodes.