Abstract
Dimensions of nonstationary time series is studied. The nonstationarity is considered to be due to multiple episode where an episode is a piece of stationary time series. The dimension estimation algorithms in the literature can be naturally extended to study multi-episode time series by restricting the calculation on data segment of pre-determined length. Inevitably, more than one episode will be included in the segment. This work focuses on finding when such dimension estimate has a local interpretation as the dimension of the episode. It was found that the local interpretation is valid if there is a large enough difference in the autocorrelation time of the episodes. This is termed EES. In practice, the average first passage time of the reconstructed “orbit” can be used to determine EES. Numerical evidence of these results are given and the application to the mechanical gearbox signal are shown.
