Time-Frequency Learning Machines For NonStationarity Detection Using Surrogates

Testing stationarity is an important issue in signal analysis and classification. Recently, time-frequency analysis has been investigated to detect the nonstationarity of a given signal, by constructiing from it a set of surrogate, stationarized signals. Time-frequency features are extracted to test the stationarity. Our paper is a further contribution by exploring the powerful framework of time-frequency learning machines. We show that one can relate stationarity to the structure of surrogates spectrograms and detect nonstationarity using a one-class classification approach. The proposed method does not suffer from any prior knowledge for extracting features, since it uses the entire time-frequency information. Using spherical multidimensional scaling technique, we illustrate the relevance of the proposed approach with simulation results.