Cepstral coefficients as the new features for electromyography(EMG) pattern rrcognition

The EMG signature discrimination of head and shoulder movements is performed in three normal volunteers. The cepstral distance measure is described and applied to modify the coefficients computed from the Autoregrcssive (AR) Model. A significant performance improvement is seen in the new features over the conventional AR coeficients. The results promote the feasibility for myoelcctric contort in the quadriplegia and amputees. recorded on the tape recorder, then digitized by the A/D converter and stored on the disks. INTRODUCTION EMG Processing EMG picked up from the bipolar electrodes during volitional motion is thought to be suitable as a contorl signal far FES or artificial arm. Function discrimination is basically a pattern recognition problcm. Thcrefore. the selection of features is an important process. The most popular features used in EMG signature discrimination are AR coeficients [l]. The fourth order model is well adopted among various investigators. The higher order (3rd and 4th) coeflicients are seriously variant and less significant in estimating the Euclidean distance between two different classes. This will degrade the performance of the classifier. Ccpstral parameters were found to be more effective than linear predictive coefficients (LPCs) in speech recognition[2]. The LPCs in speech processing are equivalent to the AR coefficients in time-series analysis. Here we use the cepstral coefficients instead of the AR coefficients for distance measure in function classification. The performance improvement on three normal subjects is shown.