Extreme event detection

With the increasing availability of data, machine/deep learning methods are becoming an important tool for modeling, analysis and prediction of several phenomena. In this work, we develop a new anomaly detection method to identify anomalous/extreme events in scientific phenomena, which are often described by multi-scale, multi-variate, smoothly varying data. The method leverages the statistical signature of anomalies hidden in the higher order statistical moments. For multi-variate data, this translates to an examination of the higher order joint moments and their association with anomalies or extreme events. Specifically, we use the direction of the principal vectors obtained from the decomposition of the fourth order joint moment tensor, in characterizing the occurrence of anomalous events. We use this method to identify the inception of auto-ignition kernels in turbulent premixed combustion and the extreme intermittent events in isotropic turbulence.