Robust Parametric Empirical Bayes based anomaly detection for flight safety events

▪ Monitoring needs to produce a small number of alerts that are relatively intuitive to interpret, and likely to be actionable by experts in flight operations ▪ Cauchy — Poisson model is robust enough to: — Apply to all types of safety events — rare or otherwise — Account for over-dispersion in the data as well as deal with spikes in the training period — Effectively filter out nuisance alarms and at the same time be sensitive enough to detect true outliers ▪ Reaction from the customer, CAST Working Group (WG), has been largely positive — Use control charts to monitor the effectiveness of their safety enhancements each quarter ▪ The control charts have been annotated with the Westinghouse Electric alerting rules so a change in the mean could be detected as well ▪ The methodology helped the WG to focus their attention on a limited set of airports with anomalies — Facilitating the understanding of the causal factors and operational procedures causing the anomaly.