Swarm-based Extreme Learning Machine for Finger Movement Recognition

An accurate finger movement recognition is required in many robotics prosthetics and assistive hand devices. The use of a small number of Electromyography (EMG) channels for classifying the finger movement is a challenging task. This paper proposes a novel recognition system which employs Spectral Regression Discriminant Analysis (SRDA) for dimensionality reduction, kernel-based Extreme Learning Machine (ELM) for classification and the majority vote for classification smoothness. Particle Swarm Optimization (PSO) is used to optimize the kernel-based ELM. Three hybridizations with three kernels, radial basis function (SRBF-ELM), linear (SLIN-ELM), and polynomial (SPOLY-ELM) are introduced. The experimental results show that SRBF-ELM significantly outperforms SLIN-ELM but not too much different compared to SPOLY-LIN. Moreover, PSO is able to optimize the three systems by giving the accuracy more than 90% with the highest accuracy is similar to 94%.