A new feature extraction method based on autoregressive power spectrum for improving sEMG classification

The feature extraction is an important step to achieve multifunctional prosthetic control based on surface electromyography (sEMG) pattern recognition. In this study, we propose a new sEMG feature extraction method which is based on autoregressive power spectrum (ARPS). An experiment with a task containing thirteen motion classes was developed to examine the effectiveness of this method. The results show that the new feature, ARPS, has better performance comparing with other two frequently used features, the time domain set (TDS) and autoregressive coefficients (ARC). The ARPS obtains the highest separability index (SI)-a metric measuring the discriminative ability of the sEMG feature. And the average classification errors of ARPS, TDS and ARC are 5.00%, 8.43% and 6.39% respectively. This suggests that the ARPS is suitable for the sEMG pattern recognition.