Exploring Granger causality in dynamical systems modeling and performance monitoring

Abstract

Data-driven approaches are becoming increasingly crucial for modeling and performance monitoring of complex dynamical systems. Such necessity stems from complex interactions among sub-systems and high dimensionality that render majority of first-principle based methods insufficient. This paper explores the capability of a recently proposed probabilistic graphical modeling technique called spatiotemporal pattern network (STPN) in capturing Granger causality among observations in a dynamical system. In this context, we introduce the notion of Granger-STPN (G-STPN) that leverages the concept of transfer entropy computed in a symbolic domain that can capture Granger causality. However, G-STPN can become significantly more computationally expensive compared to STPN while considering larger memory for a dynamical system. We numerically compare the two frameworks for a real-life anomaly detection problem involving an industrial robot platform.