Suboptimal partitioning of time-series data for anomaly detection

Abstract

The concepts of symbolic dynamics and partitioning of time series data have been used for feature extraction and anomaly detection. Although much attention has been paid to modeling of finite state machines from symbol sequences, similar efforts have not been expended for partitioning of time series data to optimally generate symbol sequences. This paper addresses this issue and proposes a partitioning method based on maximum migration of data points across cell boundaries. Various aspects of the proposed partitioning tool, such as identification of evolution characteristics of dynamical systems and adaptive selection of alphabet size, are discussed. Experimental results on an electronic circuit apparatus implementing the Duffing equation show that maximum-migration partitioning yields significant improvement over existing partitioning methods (e.g., maximum entropy partitioning) for the purpose of anomaly detection.