Detection and classification of multiple events in piecewise stationary signals: Comparison between autoregressive and multiscale approaches

Abstract

In this paper, we present methods of detection and classification of events in nonstationary signals which are well adapted to uterine EMG processing. Two sequential methods of detection are presented: the first one is monodimensional and based on AR modelling, the second is multidimensional and achieved by decomposing the signal onto scales using wavelet transform. Hypothesis rejection is achieved using either AR coefficients or a variance covariance matrix computed from the scales. Both methods are adaptive and allow event detection without necessarily returning to the null hypothesis H 0. They have been applied to simulated data and uterine EMG. Their performances have been compared together.