Clustering of sEMG signals on real-life activities using fractal dimension and self-organizing maps

Recent advances in hand classification using noninvasive sensors permit the adequately recognition of movements with high precision. However, these applications in prosthesis are far from reality, since the acquired muscle signals does not meet real-life conditions. As recent databases incorporate these real conditions into their data acquisition protocol, it is necessary to analyze the muscle signal characteristics and evaluate if these could be separated. This paper applies the Higuchi's fractal dimension in two activities of daily living using real-life signals of the triceps brachii from the NinaPro database. The characteristics are first obtained from a feature extraction technique, then clustered using a two-level approach of k-means in a self-organizing map (SOM). The results from intra-subject analysis in 15 individuals show clusterization of the fractal dimension for sEMG signals using three K
_max
values. The clusters selection are analyzed using a cluster score based on a similarity index for task identification.