en fr

Detection and classification for non stationary signals using the wavelets. Application to the uterine electromyographic signals

Détection et classification des signaux non stationnaires par utilisation des ondelettes. Application aux signaux électromyographiques utérins

Abstract : The global aim of this work was to develop methods of selection of the best basis from wavelet packet decomposition for the detection and the classification purposes. These methods were applied to uterine EMG in order to detect events useful for pre-term birth prevention. The global objective of the survey, the presentation of uterine EMG, their main features, and the various events contained in the signal, their content and their origin are presented. Two modelling approaches are described to put in evidence the frequency content of the events. The first one was based on the discrete wavelet transform, and the other was based on the wavelet packet transform. The Kullback Leibler distance is used like a criterion for choosing the most efficient packets (or best basis) for further detection of significant events in the original recordings. The detection is done on the selected wavelet packets coefficients. A delayed time position is generated differently on wavelet packet coefficients. A correction and fusion procedures are proposed to induce delay in order to get the right event location during the detection step, and to have only one change time corresponding to the original signal. The selection of the most discriminative packets for the classification is described. Detected events were "physiologically" identified using the methods of K Nearer Neighbors, the distance of Mahalanobis, neural networks and Support Vector Machine. More than 85% of events were well classified.