Blind Source Separation for Surface Electromyograms Using a Bayesian Approach

This paper presents a blind source separation algorithm to identify binary and sparse sources from convolutive mixtures with linear and time-invariant finite impulse responses. Our approach combines Bayesian algorithms for detecting source activity with a linear minimum mean-square error estimator to identify all the time samples when each source is active. The algorithm was implemented on simulated electromyo-grams to identify neural commands. Our algorithm identified more than 96% of the sources on average with 16 or more measurement channels and
$\text{SNR}\geq 14\text{dB}$
. For the detected sources, this algorithm correctly identified more than 94% of the samples on average. This performance was significantly better than that of a competing algorithm available in the literature.