Robert Knox

Auto-regressive Based Features for Upper Limb Pattern Recognition

May 27, 1998
4:00 PM
306 Egan

Abstract

The contraction of skeletal muscles is controlled by coordinated changes in electrical potentials across the muscle cell membranes. Aggregates of these changes over many cells can be measured by electrodes placed near the muscle or even on the skin surface. The resulting signal is known as an electromyogram (EMG). There has been considerable interest for many years in being able to interpret intended muscle function by processing these measured signals. One possible application for such processing would be to control a prosthesis for an amputee, using the EMG generated by the residual musculature. There have been a number of attempts to accomplish this goal over the past two decades. However, due the a number of factors reflecting the physiological complexity involved, EMG signals have a stochastic-like structure, and as a result this problem is still not completely solved.

Many of these attempts have used autoregressive (AR) models as features which are then used as inputs to a pattern recognition system. AR models are also commonly used as features for speech recognition, but generally they are first transformed to one of several alternate model parameterizations which have been found to have better properties for pattern recognition. In this work we tested AR models along with three alternate representations (reflection coefficients, log-area ratios (LAR's), and cepstral coefficients) as feature inputs to both parametric and non-parametric linear classification routines. The schemes were tested with measured EMG signals

Thesis Committee:
Prof. D.H. Brooks (Advisor)
Prof. E.S. Manolakos
Prof. M. Salehi