This paper presents a principled method for detecting the ‘abnormal’ content in vibration spectra obtained from rotating machinery. We illustrate the use of the method in detecting abnormalities in jet engine vibration spectra corresponding to unforeseen engine events. We take a novelty detection approach, in which a model of normality is constructed from the typically large numbers of examples of ‘normal’ behaviour that exist when monitoring jet engines. Abnormal spectral content is then detected by comparing new vibration spectra to the model of normality. The use of novelty detection allows us to take an engine-specific approach to modelling , in which the engine under test becomes its own model rather than relying on a model that is generic to a large population of engines. A probabilistic approach is taken that employs extreme value theory to determine the boundaries of normal behaviour in a principled manner. We also describe a novel visualisation technique that highlights significant spectral content that would otherwise be too low in magnitude to see in a standard plot of spectral power. 1. I ntroduction Vibration spectra obtained from rotating systems (such as gas turbine engines, combustion engines or machining tools) are characterised by peaks in spectral power at the fundamental frequency of rotation, and smaller peaks at harmonics of that fundamental frequency. In jet engine terminology, these peaks are conventionally called tracked orders. Methods exist [1,2,3] for the principled analysis of information pertaining to these tracked orders, such that precursors of system failure can be identified and preventative maintenance action taken. However, many modes of failure manifest themselves as changes in vibration spectra that are not related to the energy of the tracked orders. An example of this is the failure of engine bearings, which are small ball bearings enclosed within fixed cages such that they may rotate freely. These are used to form load-bearing contacts between the various rotating engine shafts and they maintain the position of the shafts relative to one another. Damage to the surfaces of these bearings may result in previously unobserved vibration energy at high frequencies, significantly removed from the narrow frequency bands of the tracked orders observed under normal conditions. A failure of the cages in which the bearings are mounted can result in constant peaks in spectral energy at previously unseen multiples of the fundamental tracked orders [4,5] . The latter could be described as novel tracked orders (NTOs), because they are peaks in vibration energy within narrow frequency bands and are thus tracked orders, but occur at frequencies for which tracked orders are not observed under normal conditions. This paper describes a method for identifying NTOs and other abnormal content in spectral data, allowing the identification of modes of failure that methods based on the modelling of tracked orders cannot detect. Principled methods are used for modelling the time-series of spectral data observed under normal conditions. The goal is to learn an engine-specific model of normality online in order to provide sensitive novelty detection without the need for tuning heuristic parameters. A model of normality is introduced in Section 2 and in Section 3 the principled methods for identifying which components of a vibration spectrum are significant with respect to background noise are discussed. We use these models to transform the problem into probability space in Section 4, describe how to perform novelty detection in Section 5 and present results from jet engine vibration data in Section 6. Throughout this paper, absolute values of vibration amplitude and frequency are not provided for purposes of commercial confidence, and units of measurement have been omitted from some Figures.
Read full abstract
Jan 1, 2024
M Abul Kawser + 2
Open Access
