Sliding Window Nonnegative Matrix Factorization (SW-NMF) for Robustness Low-Density Myoelectric Signals Decoding Against Electrodes Shift

Abstract

In this paper, the problem of electrodes shift is studied in low-density surface electromyographic (sEMG) based prosthetic control with the proposed Sliding Window Nonnegative Matrix Factorization (SW-NMF) algorithm. By artificially switching the electrode positions clockwise for π/8, the 8 channel sEMG signals of 10 gestures were recorded before and after the electrodes shift. It is found that electrodes shift makes the feature space of the sEMG signal non-stationary, which has a great influence on the classify accuracy. Besides, all kinds of existing algorithms for electrodes shift in the high-density electrode environment have limited effect in the low-density electrode environment. In the proposed SW-NMF method, we firstly place the sum constrain on the coefficient matrix H to reduce change of the sample distribution in the feature space. Secondly, a sliding widow strategy is applied accompany with the sum constrain on H to make the algorithm can be run online. Finally, a self-enhanced version of Linear Discriminant Analysis (LDA) is included in the SW-NMF algorithm to make the classifier be able to follow the change of sample distribution in the feature space for further improve the decoding accuracy. Compared with the traditional TD+LDA, and the other type of NMF based methods, the result of the proposed SW-NMF shows a high robustness for electrodes shift.

Gan Huang

My research interests include Neural Modulation, Brain Computer Interface and Neural Prosthetics.