Tensor Train Subspace Analysis for Classification of Hand Gestures with Surface EMG Signals.

Processing and classification of surface EMG signals is a challenging computational problem that has received increasing attention for at least two decades. When multichannel EMG signals are transformed into spectrograms, classification can be performed using multilinear features that can be extracted from a set of spectrograms by various tensor decomposition methods. In this study, we propose to use one of the most efficient tensor network models, i.e. the tensor train decomposition method and to combine it with the tensor subspace analysis to extract the more discriminant 2D features from multi-way data. Numerical experiments, carried out on surface EMG signals registered during hand gesture actions, demonstrated that the proposed feature extraction method outperforms well-known tensor decomposition methods in terms of classification accuracy.