Learning Shape for Jet Engine Novelty Detection

摘要

Application of a neural network approach to data exploration and the generation of a model of system normality is described for use in novelty detection of vibration characteristics of a modern jet engine. The analysis of the shape of engine vibration signatures is shown to improve upon existing methods of engine vibration testing, in which engine vibrations are conventionally compared with a fixed vibration threshold. A refinement of the concept of "novelty scoring" in this approach is also presented.