Single-Channel Separation Between Stationary and Non-stationary Signals Using Relevant Information

Abstract

AbstractWe propose a novel unsupervised single-channel approach to separate between stationary and non-stationary signals. To this, we enhance data representation through its time-frequency space, where stationarity is defined based on information theory. Then, we search for the projection of the time-frequency representation that is as stationarity as possible, but preserving most of the data information. The proposed approach validated on synthetic data. As performance measure, we use the correlation coefficient and the mean squared error between the original and the estimated stationary composing signals. Obtained results are compared against the baseline non-negative matrix factorization that separates dynamics from the time-frequency representation. As a result, our approach gets better performance even if assuming low power ratios, i.e., non-stationary signal power is higher or even equal than the stationary signal power.KeywordsStationary signal separationInformation theoretic learningTime frequency analysis