optimized time-frequency.pdfVIP

Classication of time series withoptimized time-frequency representationsC. HeitzCentre for Data Analysis and Model BuildingUniversity of Freiburg, Albertstr. 26-28, D-79104 Freiburg, GermanySummary: We address the problem of classication of time series with nitelength. In contrast to the usual feature-based classication schemes we use time-frequency representations (TFRs) as non-parametric representations of signals.Since there are innitely many dierent TFRs, each yielding a dierent repre-sentation of the same signal, it is possible to adopt the representation to thestructure of the signals under consideration. It is shown how, for the problem ofclassication, the optimum TFR can be found, if the signal classes are given bya set of realizations. Two examples show the advantage of using the optimumrepresentation in the time-frequency domain compared with the original time re-presentation.1. IntroductionIn this article we consider the problem of classication of nite time seriesof length L: f(t); t = 1; : : : ; L, for the case of two classes. The usual wayof dealing with this problem is the following: First a number of features1; : : : ; p; p L are extracted from the time series which are supposed todescribe the signal f(t). In general the selection of these features is not aneasy task and needs some a priori knowledge of the inherent structure of thetime series as well as the dierence between the classes to be distinguished.After feature extraction the classication itself is performed by means of amore or less sophisticated classier. Usually it holds that the better thefeatures 1; : : : ; p are chosen the simpler the classication algorithm canbe.Classication with TFRs follows a completely dierent strategy: A timeseries f(t) is not represented by a set of features, but by its TFR whichis a non-parametric representation of f(t). The classication is performedby a simple Euclidian distance classier in the time-frequency domain. Theprincipal idea is t