Robust intent recognition for prosthesis control

A study of the stochastic nature of the processed electromyogram (EMG) and performance of Gaussian Bayesian spatial pattern recognition systems, for the control of an above-knee (A/K) prosthesis, was conducted. The primary focus of this study was to identify, derive, and implement myoelectric signal processing methods that are capable of recognizing the functional intent of an individual over as large a range of force and span of time as possible. A long term model of observed EMG pattern vectors was developed to account for the minute and hour scale nonstationary behavior. Patterns of processed EMG for isometric-isotonic and isometric-anisotonic, linearly, increasing force muscle states were collected. Processed EMG that spanned an hour was ensemble averaged to form a Gaussian Bayesian reference model which minimized the average expected loss in the wide sense. Signal to noise ratios (SNR) and percent correct classifications (%CC) were obtained and proved superior to shorter term models.