SEMG feature extraction methods for pattern recognition of upper limbs

Abstract

In this paper, a new feature of surface electromyo-graphy (sEMG) by using discrete wavelet transform (DWT) is proposed for motion recognition of upper limbs, and this method can be eventually used for rehabilitation robot control. Seven traditional features of sEMG are also extracted for comparative study, they are integral of absolute value (IAV), difference absolute mean value (DAMV), zero crossing (ZC), variance (VAR), mean power spectral density (MPSD), mean frequency (MF) and median frequency (MDF) respectively. For comparing the recognition rate of the different motions of the upper limb, each feature or their combination are used to construct the feature vectors, and the BP neural network with variable learning rate back propagation with momentum (GDX) algorithm is used to classify these motion modes. The experimental results summarize that the new feature extracted by using DWT presents a higher recognition rate (98.9%) than all of the traditional features, and the traditional features combination can also greatly improve the recognition rate (99%).