Processing of Myoelectric Signals by Feature Selection and Dimensionality Reduction for the Control of Powered Upper-Limb Prostheses

Abstract

The extraction of features from myoelectric signals (MES) for the classification of prehensile motions is difficult to achieve. The optimal selection of features, extracted from a MES and the reduction of dimensions is even more challenging. In the context of prosthetic control, dimensionality reduction means to retain MES information, that is important for class discrimination and to discard irrelevant data. Dimensionality reduction strategies are categorized into feature selection and feature projection methods according to their objective functions. In this contribution, we bring forward a statistical cluster analysis technique, which we call the “Guilin Hills Selection Method”. It combines selection plus projection and can be applied in the time- and in the frequency-domain. The goal is to control an electrically-powered upper-limb prostheses, the UniBw-Hand, with a minimum number of sensors and a low-power processor. We illustrate the technique with time-domain features derived from the MES of two sensors to clearly differentiate four hand-positions.