Prediction of magnetorheological grease compositions using extreme learning machine methods

This paper presents a data-driven model to predict magnetorheological (MR) grease composition as a function of its rheological properties using several machine learning methods. The methods are Single Hidden Layer Feedforward Neural Networks (SLFNs) and Kernel Based-Extreme Learning Ma-chine (KELM). The approach provides high accuracy prediction and the easiness of changing the inputs or outputs as long as the data is available. While the model output is carbonyl iron particles weight percentage, the model in-puts are the slope of the magnetic field density-dependent-yield stress change over the magnetic fields and the off-state yield stress. The kernel functions are varied from radial basis function, wavelet, linear, and polynomial functions. The simulation results of KELM show that R-squared values are more than 90% for both training and testing data. The root mean square errors also show relatively small values. With a relatively lower number of parameters than SLFNs-ELM, KELM can show comparable performance with SLFNs-ELM and Back Propagation neural networks.