TIME-FREQUENCY LEARNING MACHINES FOR NONSTATIONARITY DETECTION USING SURROGATES

An operational framework has recently been developed for testing stationarity of any signal relatively to an observation scale. The originality is to extract time-frequency features from a set of stationarized surrogate signals, and to use them for defining the null hypothesis of stationarity. Our paper is a further contribution that explores a general framework embedding techniques from machine learning and time-frequency analysis, called time-frequency learning machines. Based on one-class support vector machines, our approach uses entire time-frequency representations and does not require arbitrary feature extraction. Its relevance is illustrated by simulation results, and spherical multidimensional scaling techniques to map data to a visible 3D space.