Wavelet Selection for EMG Based Grasp Recognition through CWT

This paper details a strategy of discriminating grasp types using surface electromyogram (EMG) signals, which has the potential to be applied for controlling advanced prosthesis for extreme upper limb amputees. We have investigated the classification of six basic grasp types used during 70% of daily living activities. The feature vector for EMG based grasp recognition was derived using continuous wavelet transform (CWT). The proper wavelet basis function was selected through computation of entropy of the preprocessed EMG signals and wavelet transform coefficients of six different wavelet families: Gaussian, Daubechies, Morlet, Mayer, Mexicanhat and Symlet. Based on this, Gaussian wavelet function has been concluded to be possessing maximum informations about grasp types. Experimental results have validated our hypothesis that the CWT coefficients having entropy values close to the entropy of preprocessed EMG signals possesses maximum informations about the grasp types. Classification was through one vs. all multi-class support vector machine with linear kernel following preprocessing and maximum voluntary contraction normalization of EMG signals. We have achieved an average recognition rate of 80% (using the Gaussian wavelet function) cross validated through 10-fold cross validation.