ABSTRACT
A motivation for using the wavelets rather than spectral features of time series is that when using the frequency domain methods, it is not easy to make the connection to specific time points in the time domain. On the other hand, the wavelets approach enables discrimination between time series patterns using available information from both the time and frequency domains simultaneously. If the number of time series is such that it is too small to permit the estimation of the complete variance-covariance matrices, diagonal linear or diagonal quadratic discriminant procedures are used. Since they are also interested in the relationship between every pair of leads, more useful information can be obtained from the correlation between every pair of leads at each of the frequency bands, instead of a single correlation coefficient between a pair of leads in the time domain. All of this therefore provides a motivation for the use of these wavelet based features.
