Mining Multiple Periodic Time Series for Detecting Unusual Sub-Sequences

Advances in computer and information technology have opened a new avenue in the analysis of large and more detailed datasets that has become possible to observe. Most of the classical methodologies and techniques have become obsolete and fresh approaches of data analysis are overdue. In our previous articles Ameen JRM & Basha R (2006) and Basha R & Ameen JRM (2007), we have introduced algorithms for observing sub sequence discords for time series data with periodicity making it possible to save considerable processing time using logic and borrowing ideas from classical statistics. This paper deals with time series with multiple periodicities. Such datasets are naturally occurring when data are observed at higher frequencies and at a near continuous level from real life processes. Our approach and methodology will be applied to data that have been observe from daily consumption of electricity for more than two years indicating periodicity at weekly, monthly and yearly levels having multiple periodicity. We will demonstrate our approach´s efficiency for situations of a similar nature in identifying subsequence discords efficiently together with ´motifs´ as defined by Keogh et al.