Dimensionality Reduction and Classification of Myoelectric Signals for the Control of Upper-Limb Prostheses

K. Buchenrieder (Germany)


Myoelectric Signal Processing, Guilin Hills Selection Method, Statistical Cluster Analysis, UpperLimb Prostheses Control


The classification of prehensile motions from myoelectric signals (MES) for the control of prostheses or as non verbal input to computers is receiving much attention. Numerous methods and devices have been developed for the classification of MES for different grip types. In this contribution, we extend the established “Guilin Hills Selection Method”, a statistical cluster analysis technique for MES earlier described in [1], with a “Spyglass” procedure. When the measurements taken during the operation of the prostheses lead to an ambiguous classification, an alternative muscle-feature combination is chosen to remove the ambiguity. This spyglass enhancement allows us to differentiate hand positions with a high accuracy and a very good repeatability. Experimental results, based on the analysis of several hundred data sets, recorded with different individuals, show a better discrimination of hand positions compared to our previously employed, standard neural-net solution [2,3]. Furthermore, we were able to reduce the number of myoelectric skin-surface sensors for the control of an electrically powered hand prostheses. In this contribution, we illustrate the extended Guilin Hills classifier method with a set of standard time-domain features, derived from the MES of two sensors and differentiate four distinct hand-positions as an example.

Important Links:

Go Back