Multiscale Complexity Analysis: A Novel Approach for Anomaly Detection in Multivariate Data

Abstract

In this paper, a novel method is presented for anomaly detection in multivariate data. The proposed method is based on computing multivariate entropy of input data at multiple scales, via the MMSE method, a technique recently proposed for the dynamical complexity analysis of multivariate data. In the proposed methodology, the anomalous behaviour is assumed to be generated by a constrained system and thus is easily differentiated from the established normal behaviour, in accordance with the “complexity loss” hypothesis, traditionally used for physiological systems. Simulations are provided to demonstrate the effectiveness of the approach on real world data sets in terms of anomaly detection.