Deep learning-based classification using Cumulants and Bispectrum of EMG signals

OROSCO, EUGENIO C.; SORIA, CARLOS; GIMENEZ ROMERO, JAVIER ALEJANDRO; GAIA, JEREMÍAS

IEEE LATIN AMERICA TRANSACTIONS

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

Lugar: New York; Año: 2019 vol. 17 p. 1937 - 1944

1548-0992

Surface electromyographic signals (EMG) historically have been used to classify tasks in basis of a feature extraction scheme and low complexity classifiers. Deep networks, as Multilayer Perceptron and Convolutional Neural Network (MLP and CNN, respectively), avoid the traditional, complex and heuristic (handcrafted) process of feature extraction. Today, it is possible to face the computational cost that these automatic techniques require due to the technology advancement. This allowed deep learning techniques to be quickly generalized to countless applications. This paper proposes to use the third order cumulants and their 2D Fourier transform (Bispectrum) to directly feed CNN and MLP deep learning networks. The classifier is not user-dependent (same classifier for all users) and obtains better results than the classical scheme according to several metrics.