Classification of 17 voluntary movements using principal component analysis for myoelectric prosthetic hand

Abstract

This paper proposes a signal processing technique to classify 17 voluntary movements from electromyographic (EMG) signals. In the proposed method, EMG signals are acquired from six EMG sensors. The features of the voluntary movements are extracted from these EMG signals using principal component analysis and later classified using artificial neural network (ANN). To evaluate the validity of the proposed method, online classification experiments are conducted on one male and one female participants. A total of 15 data sets, where each set consists of EMG signals characterizing the 17 motions, are acquired from each participant. From this total, five data sets are used as training data, while the other 10 data sets are acquired for online testing. The same experiments are repeated on a different day. The validity of the algorithm, evaluated based on the mean correct and incorrect classification rates of both days are calculated from testing data. Results show that using all five training data to train the ANN yields higher accuracy than using only one training data. The best classification result shows that there are 10 out of 17 motions with an accuracy of over 50% and mean incorrect rate of 2%. Furthermore, classification where ANN is trained using training data of a different day is also conducted. The results show that the proposed algorithm can achieve an overall correct rate of 46% at best. Based on the above results and considering the fact that users can promptly modify any erroneous actions by looking at the actual output of the prosthesis, the proposed algorithm has demonstrated the potential to classify 17 voluntary movements from 6 EMG sensors.