Detection of changes of signal structure by using the Wigner-Ville spectrum

Abstract

An efficient estimation of the Wigner-Ville spectrum of non-stationary processes requires the segmentation of the observed signals into locally stationary signals. We present a detection procedure of such a segmentation based on the pseudo-Wigner estimates. These estimates are known to be uncorrelated estimates of the Wigner-Ville spectrum for neighboured, appropriately spaced frequencies. Therefore, a detection using the pseudo-Wigner estimates can be designed for detection of non-stationarities in any specified band of frequencies. The procedure is based on a subset regression approach and comes up with an informal Akaike type of criterion as a detector of changes of the signal structure. The performance of this detector is evaluated by a simulation study. It works especially well in the case of amplitude and frequency changes of deterministic signals in white noise. Examples dealing with the analysis of biological clocks and their non-stationary properties indicate the successful application of the method.