Improvements in transfer entropy calculations for causality detection in time series

Abstract

The discovery of cause-effect relationships in signals from industrial processes is a challenging problem. A data-driven method to achieve this relation is the transfer entropy, a method based on the conditional probability density functions that measures directionality of variation. This method requires several parameters that must be properly chosen to avoid misleading results. In this work, the analysis of these parameters in the transfer entropy calculations is performed, and a methodology is pro-posed for their selection. The utility of the proposed approach is illustrated by several examples including the analysis of routine operating data in an industrial case study.