Surface electromyography classification of hand motions using time domain features for real time application

Surface electromyography is a technique of analyzing muscle functions through signals emanating from the physiological variations of muscles states. Through this technique, various applications such as prosthetic hands control have been made for purposes of giving basic functionality for people that are unable to do daily tasks. Many researches have been made over the years to develop the prosthetic hand control system by using the surface electromyography signals through pattern recognition. Recent researches have shown that various method have been able to achieve above 90% accuracy. However, the challenge of developing a control system that is both accurate while being suitable for real time application with less than 300 ms delay still remains. In addition, no literatures have been reported classifying hand motions with stages of contraction despite patterns being observed. The objective of this study is to investigate the accuracy and real time suitability of using time domain features and artificial neural network, to characterize and classify different hand motions and stages of the contraction. To achieve this goal, the signal is first segmented into windows of two sizes, which are 132.5ms and 165 ms, and then full wave rectified. Then the signal is separated into raw and normalized signal. Five time domain features, namely mean absolute value, variance, root mean square, integral absolute value and waveform length were extracted from the segmented windows to characterize three different hand motions of wrist flexion, wrist extension and co-contraction using raw signal and three different stages of contraction of start, middle and end using normalized signal. From the characterization obtained and t-test made, all raw features, waveform length normalized, and the 132.5 ms segmented window size were selected for classification. The features are then used by artificialneural network to be trained offline and evaluated for performance in terms of classification accuracy. Computational times have been recorded to determine nreal time suitability at all steps. It is determined that during feature extraction stage, the features were able to differentiate hand motions as the mean values were different. However, for stages of contraction, although patterns were observed, only waveform length features could differentiate the different stages for all three motions. Overall, it is determined that an artificial neural network can be used with time domain features to achieve 98.5% accuracy when differentiating three different hand motions but not with stages of contraction achieving only 80.4% accuracy. Meanwhile, in terms of computational time, although artificial neural network is considered less suitable for real time application, when using time domain features, 245.8 ms delay is achieved which is below 300 ms, thus making it suitable for real time applicatio