A System for the Analysis of Jet Engine Vibration Data

A system has been developed to extract diagnostic information from jet engine carcass vibration data. The system consists of a number of modules, each of which focuses on particular subsets of the data known to hold valuable information. Two of these modules, based on neural network techniques, are described in detail in this paper. In the first module, novelty detection provides a measure of how unusual the shape of a vibration signature is, by learning a representation of normality based entirely on normal examples. The low-dimensional vectors which encode vibration signatures are normalised by an appropriate transform before their distribution is modelled by a few kernels, whose placement is optimised by clustering techniques. Novelty is then measured as the local distance from the nearest kernel centre. This method provides good separation between usual and unusual vibration signatures but, given the small number of examples of normal engines, the resulting representation of normalitymay be overfitting the training data. The severity of this effect is investigated for two different normalising transforms. The second module detects sudden transitions in vibration signature curves. A multi-layer-perceptron is trained to predict one step-ahead for curves without these unexpected transitions. Sudden transitions in the test engine data are then reported whenever the prediction error exceeds a predetermined threshold.