sEMG pattern recognition based on multi feature fusion of wavelet transform

In view of the poor characterization of single feature value, multi feature fusion based on different wavelet basis was adopted to extract the surface EMG signal according to multi resolution analysis of wavelet transform. The experiment was conducted on ten testers and collected signals for four basic lower limb movements in daily life. First of all, discrete wavelet transform was used to decompose the surface EMG signals in multi-scale with DB, Dmey and Bior wavelet basis respectively. After that, it was founded that the characterization effects of different muscle vary by different extraction way. In order to combine the characteristics of different features, features were fused to analyze and compare. At last, the feature values were input to the Elman neural network and BP neural network for pattern recognition and comparison analysis. Experimental results showed that the recognition rate obtained by fusing the eigenvalues is higher than single feature with the accuracy up to 98.7%, and the BP neural network is better than the Elman neural network.