A Novel Feature Extraction Scheme for Myoelectric Signals Classification Using Higher Order Statistics

Abstract

We present a novel feature extraction scheme for surface myoelectric signal (sMES) classification. We employ a multilayer perceptron (MLP) in which the feature vector is a mix of the second-, the third-, and the fourth order cumulants of the sMES stationary segments obtained from two recording channels. To reduce the number of features to a sufficient minimum, while retaining their discriminatory information, we employ the method of principle components analysis (PCA). The detected sMES is used to classify four upper limb primitive motions, i.e., elbow flexion (F), elbow extension (E), wrist supination (S), and wrist pronation (P). Simulation results indicate a substantial reduction in the required computations to achieve similar results as compared to existing methods