SVD and higher-order statistics applied in the detection of human nerve signal activity

Abstract

Due to the very low signal levels (/spl mu/V) and the significantly higher levels of interference from adjacent muscles (and other noise sources), the overall signal-to-noise ratios (SNRs) of human nerve signals recorded from whole-cuff electrodes is very poor. Typically, non-real-time methods (ensemble averaging) are used to contend with these poor SNRs. However, if these signals are to be useful in providing real-time information (in a closed-loop control system), other methods must be employed to this end, subspace analysis methods using the eigenvalues of the autocorrelation (a 2/sup nd/ order statistic) and cumulant (a 3/sup rd/ order statistic) matrices of time-series samples were evaluated.