Abstract
The application of artificial neural network (ANN) models in magnetorheological (MR) damper has gained interest in various studies because of the high accuracy in predicting the damping force, especially for control purposes. However, the existing neural network models have apparent drawbacks such as relatively long training time and the possibility to be trapped in local solutions. Therefore, this paper aims to propose a new method to deal with a highly nonlinear behavior of MR damper using an extreme learning machine (ELM) method. The ELM method is applied to a meandering valve-based MR damper for damping force prediction, which has been recently developed. A simulation scheme is selected with damping force as the output, and current, velocity, and displacement as the inputs. The simulations are then carried out based on fatigue dynamic tests data in various frequencies and currents. The training times for more than nineteen thousand data points using the ELM method with 10, 100, 1000 hidden neuron numbers are less than 1.70 s, which is faster than the conventional ANN. Based on 50 times training processes, the ELM and ANN models have comparable average accuracies with R-2 values of more than 0.95. ELM also has shown less value R-2 standard deviation showing its advantage to reduce the possibility of being trapped in local solution compared to the conventional ANN. (C) 2020 Elsevier B.V. All rights reserved.