Multiscale detection of nonstationary signals

Abstract

A statistical method for detecting and/or localizing nonstationarities in a process observed over a time interval T is presented. Stationarity is induced by taking a wavelet transform of the process. A parametric model is fitted to the result. The error incurred in fitting the model is shown to preserve the singularity manifested in the transform. The error is then used to establish a statistical detection test that is free of any prior knowledge about the class of signals being analyzed, and of any user input.<>