Bayesian Extreme Value Statistics for Novelty Detection in Gas-Turbine Engines

Abstract

We present a novel method for the identification of episodes in gas-turbine vibration data, in which we show 1) how a model of engine behaviour is constructed using signatures of normal engine vibration response; 2) how extreme value theory (EVT), a branch of statistics used to determine the expected value of extreme values drawn from a distribution, can be used to set novelty thresholds in the model, which, if exceeded, indicate an abnormal episode; 3) application to large data sets of modern gas-turbine flight data, which shows successful novelty detection results with low false-positive alarm rates. The advantages of this approach over previous work are 1) a very low false-positive alarm rate, while maintaining sufficient sensitivity to detect known events; 2) the use of a Bayesian framework such that uncertainty in the distribution of normal data is modelled, giving a principled, probabilistic interpretation of results; 3) an implementation that is sufficiently lightweight in processing and memory resources that real-time, on-line novelty detection is possible in an on-wing engine health-monitoring system.