Slovenian-France "COTUTELLE" Thesis

Author: Dean Korosec
Mentor: Prof. Dr.  Damjan Zazula
Co mentor: Prof. Dr. Christian Doncarli
Date: May. 29, 1999

Analysis of one-dimensional signals by processing of their time-frequency representations

Keywords: non-stationary signal analysis, linear models, time-varying parametric modelling, surface electromyographics, transient wide-band signals, ARMA distances, spectrum dynamics, time-frequency representations, ambiguity plane classification

UDK: 621.391:681.32

Abstract: This thesis is about the analysis and classification of non-stationary signals based on estimation and comparison of their time-varying frequency spectra. Two approaches are presented and compared: parametric models and time-frequency representations. Both have been successfully applied to analysis of various bio-signals and a brief overview of those applications is included. Our application to surface electrmyographics signals is concentrated arounbd the comparison of spectrum characteristics with respect to different signals leads in multichannel SEMG recordings. Parametric sliding window estimation and coefficient decomposition techniques are used to evaluate the spectrum changes due to the muscle fatigue. The relative decrease of mean frequency in confirmed  to be a stable parameter which can be detected from SEMG during the constant level isometric contractions with only a minor influence of the chosen parametric model order and electrode placement, although several signal characteristics (including the absolute mean frequency) differ greatly regarding different muscle physiology zones.

From the observation of the real SEMG spectrum behavior we constructed a model of transient wide-band signals and we derived the analysis and comparison tools in both parametric and non-parametric framework. Three parametric classification approaches (two estimation methods in combination with several generalised ARMA  distances and linear discriminant analysis of decomposition coefficients) proved to be adequate solutions to the presented problem. We also propose the generalised ARMA distances as a measure for expressing the changes of time-varying spectrum. Continuos spectrum needed for such purpose can be conveniently obtained by the coefficient decomposition estimation.

Finnaly, we propose a non-parametric classification scheme, based on the selected information in the ambiguity domain. In the four step procedure of a supervised classification we developed two approaches of dealing with several classes. Although this kind of classification yields the results inferior to parametric methods in case of wide-band signals, it is advantageous for multicomponent narow-band signals. Especially the enhanced probabilistic scheme of using several class-to-class comparisons to bring a final decision should be considered as a novel approach in data-driven time-frequency and ambiguity domain classification.