Specific and Generic Modelling for Jet Engine Novelty Detection

This paper describes a method for modelling the normal behaviour of a modern military gas-turbine engine, for which fault data are particularly scarce. Signatures of fundamental tracked order vibration amplitude are used to characterise the normal operation of the engine, with models formed in high-dimensional space trained on a fleet of engines of the same class. These models are shown to be able to identify fault behaviour in engine tests of the same type, using well-defined “PAT” engine tests. These generic class-wide models are then applied to a single engine undergoing repeated cyclic testing (“ASMET” cycles), to evaluate their performance with respect to the fleetbased models. Initially proven using a retrospective analysis, identification of potential engine events is shown to be possible using the visualisation models developed using the original fleet data, in which such “events” are shown as step-changes in signature visualisation. Models of normality formed from pass-off tests are shown to be inapplicable for endurance tests, due to the differing nature of each test type.