A WAVELET PACKET APPROACH TO TRANSIENT SIGNAL CLASSIFICATION

Time-frequency transforms, including wavelet and wavelet packet transf orms, are generally acknowledged to be useful for studying non-station ary phenomena and, in particular, have been shown or claimed to be of value in the detection and characterization of transient signals. In m any applications time-frequency transforms are simply employed as a vi sual aid to be used for signal display. Although there have been sever al studies reported in the literature, there is still considerable wor k to be done investigating the utility of wavelet and wavelet packet t ime-frequency transforms for automatic transient signal classification . This paper contributes to this ongoing investigation through the dev elopment of a non-parametric wavelet packet feature extraction procedu re which identifies features to be used for the classification of tran sient signals for which explicit signal models are not available or ap propriate. Recent literature in this area is devoted to truly ad-hoc, high-dimensional, non-parametric types of classification in which one or more time-frequency transform forms the base from which a large num ber of features are determined by trial and error. In contrast, the wa velet-packet-based procedure presented in this paper was formulated to systematically adapt to any data dictionary within which several clas ses must be distinguished. This method is aimed at focusing the inform ation in the data set to find the smallest number of features for robu st, reliable classification. The promise of our method is illustrated by performing our procedure on a set of biologically generated underwa ter acoustic signals. For this example the wavelet-packet-based featur es obtained by our method yield excellent classification results when used as input for a neural network and a nearest neighbor rule. (C) 19 95 Academic Press, Inc.