Métodos de dados sub-rogados aplicados a séries temporais

• Ano de defesa: 2008
• Autor(a) principal: Cassio Pascoal Costa
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Ruídos; mecanismo; Engenharia elétrica
• Link de acesso: http://hdl.handle.net/1843/BUOS-8D2FQH

In recent years the method of surrogate data has been widely applied to the analysis of determinism in time series, mainly those time series generated by a nonlinear mechanism (Theiler et al., 1992). Since the techniques available in the literature are not always prepared to deal with the problem of determining and analyzing whether underlying dynamics of the generating mechanism of the time series is deterministic or not, the study of new methods is necessary. However, little attention has been given to the development of specific tools to deal with this problem. This work proposes the investigation of the dynamic behavior from a time series contaminated with Gaussian or Non-Gaussian noises by using the method of hypothesis testing based on the surrogate data methodology. The first attempts to find the generating mechanism of such time series are performed by using statistical tests, also called statistical discriminate analysis that can be parametric and non-parametric. In this work, several techniques available in the literature and used for generating surrogate time series are tested. From analyzing the results, using simulated time series with different dynamic behavior and also real time series, it is determined which surrogate data algorithm - algorithm 0, algorithm 1, algorithm 2, algorithm SS, algorithm CS or algorithm PPS - is more appropriate for the test of determinism. Based upon the detection or not of a deterministic behavior in the time series under analysis, the results obtained here reveal that such techniques for generation of the surrogate data are not always well tuned and that there are particular situations that lead to misleading results for certain model structures of the mechanism generator of such time series. As the major result, a procedure for general analysis is given using the surrogate method and appropriate null hypotheses that can determine randomness (pure noise) and determinism (linear, nonlinear and etc.) for each case illustrated in this work.