A novel spectral analysis method of EMG feature extraction for wrist motions recognition

This paper presents a new time-dependent spectral analysis, smooth localized complex exponential (SLEX), for extracting the surface EMG features of the wrist motions. Different from conventional Fourier method, SLEX-based spectral analysis applies two special smooth windows on Fourier basis function and can be simultaneously orthogonal and localized. In wrist motions recognition, we conduct SLEX transform on 4 selected channels with dimensionality reduction projection -LDA (linear discriminant analysis) in order to obtain EMG features of wrist motions. Then we evaluate separation of SLEX-based feature vector in LDA-subspace by the use of visualization 3-dimension plot and quantitative measurement, Davies-Boulder clustering Index. Finally, MLP (multiple layers perceptron) classifier is used to evaluate the classification accuracy of the SLEX-based feature extraction. Compared to commonly used methods- AR model and power spectral estimation (PSE), the SLEX-based model has better performance in classifying wrist motions and the classification accuracy is up to 98%.