Neuromuscular activity classification based on autoregressive modeling of time-frequency domain decomposed surface EMG signal

Abstract

Electromyography (EMG) measures electrical activity in the muscle due to neuromuscular activities. It is acquired either by needle (nEMG) or by surface (sEMG) electrodes. The sEMG is getting popularly because of its non-invasive acquisition technique and widely used in prosthetic control and human-machine interaction. However, due to noise like characteristics of sEMG signals, poor performance may be obtained for classifying similar types of neuromuscular actions, especially when only a single lead data are used. In this thesis, efficient schemes are proposed to classify neuromuscular activities based on autoregressive (AR) reflection coefficients extracted from single lead original and decomposed sEMG signals. At first a given frame of raw sEMG signal is divided into short duration sub-frames and considering AR modeling, from each sub-frame AR reflection coefficients are extracted. Instead of using the entire frame at a time, sub-frame based AR analysis is expected to provide consistent estimates and capture short duration variations. The advantages of using AR reflectioncoefficients over the conventional AR parameters are that their values are bounded (0 to 1) and they provide better consistency, noise immunity and lower computational complexity. The reflection coefficients obtained from each sub-frame are finally averaged to construct the proposed feature vector. In the second scheme, in view of investigating the effect of utilizing the decomposed sEMG data, singular value decomposition (SVD) is performed on each sub-frame of sEMG data. Next, instead of using the original sEMG data, decomposed data is used for extracting the AR reflection coefficients. In order to analyze the effect of time-frequency domain decomposition of the sEMG data on the extracted feature quality, in the third scheme, the discrete wavelet transform (DWT) is chosen. Each sub-frame of sEMG data is decomposed by using the DWT and then both the approximate and detailed coefficients are then used for extracting AR reflection coefficients. For the purpose of classification, the k-nearest neighborhood (KNN)classifier is applied in a hierarchical approach. The proposed method is tested on a publicly available sEMG dataset containing six different hand movements collected from three females and two males. It is observed that the proposed method offers consistently a very high accuracy in classifying the hand movements using a very low feature dimension.