The present Master’s Thesis proposes a methodology for the non–stationary time-series analysis for filtering and noise rejection purposes in pattern recognition. The methodology is divided into two different approaches: the analysis of periodic non–stationary behavior that relies into a cyclic process and how additional non–cyclic non–stationarities disrupt and affect the signal processing. Second approach deals with the problem of non–stationary signal extraction that affects inherent weak stationary processes. Both frameworks of analysis take base on (cyclo-)stationary constraints and subspace based representations in order to assess and characterize the signals dynamics to facilitate the identification of the undesired non–stationary components. Results are shown over each approach with different real and synthetic data, the obtained performances show high rejection, detection and extraction capabilities for noise and artifacts in (cyclo)–stationary signals using external and internal based constraints of analysis and high separation capability for stationary signals.
Download Full PDF Version (Non-Commercial Use)