Toward reducing the effect of force variations on electromyography pattern recognition by Mel-frequency spectrum

Currently, the electromyography pattern recognition (EMG-PR) is considered as a promising approach to control the human-machine interaction systems such as multifunctional prostheses. However, the robustness of EMG-PR method is still not strong enough to against some issues such as different arm positions, electrode shift, muscle fatigue and force variation in the clinical application. And among these issues, the force variation is an important problem that greatly affects the performance of EMG-PR based systems. In this study, a feature of log-Mel-frequency spectrum (log-MFS) was proposed to reduce the effects of force variations on the classification performance of the EMG-PR method. Eight channels of EMG signals were recorded from the upper limbs of eight subjects when performing different hand motions at low, medium and high force levels, respectively. Then the proposed feature of log-MFS was extracted from the EMG signals and used to classify the motions. Compared with the commonly used time domain feature set, the feature of log-MFS achieved the higher classification accuracies for all the three force levels. Especially for the un-trained high and low force levels, the average classification accuracies increased by about 27% and 11%. These results demonstrated that the feature of log-MFS is effectiveness to enhance the robustness of the EMG-PR based systems to against force variations in practical application.