Discovering Precursors to Adverse Events in Time Series

In this paper, we consider the problem of discovering candidate precursors to anomalies in a set of time sequenced data. Typical scenarios involving time sequential data include dynamical systems and general monitoring systems. In such scenarios, a precursor could be any event that frequently precedes a given event of interest. Anomalies are rare but significant events in time series data and identifying precursors to anomalies is vital in proactive management. In this work, an inverse reinforcement learning (IRL) based method is formulated to succinctly represent the nominal behavior and identify sequences that preceded the anomalous events. A preliminary evaluation is performed on flight recorded data identifying challenges and future directions for application.