Aircraft Condition Monitoring - Nonlinear Approaches
The identification
of abnormal operating behaviour in gas-turbine jet engines is
of critical importance in the avoidance of hazard. Recently, engine
manufacturers are adopting a condition monitoring approach to maintenance, in
which embedded intelligent data analysis systems process data from
engine-mounted sensors to assess the "ealth" of engine components. These
health-monitoring systems can provide early warning of engine failure by
detecting abnormal operating behaviour.
This project [1] introduces a framework for the analysis of engine vibration data, providing on-line (in-flight) abnormality detection, and off-line (ground-based) engine monitoring, using data from the development of the Rolls-Royce Trent 900 engine (used by the Airbus A380, above figure) and the Rolls-Royce EJ200 (used by the Typhoon Eurofighter), using nonlinear techniques.
Case-mounted
sensors measure engine vibration from which FFTs are computed
(see figure to the left). Tracked orders are extracted from spectral data, here
shown in the figure are spectrograms (y-axis) against time (x-axis), were the
orders appear as approximately straight lines. The spectral data is defined
to be vibration amplitude and phase within a narrow frequency band centred on
the fundamental or harmonic of the rotational frequency of an engine shaft.
Previous work [2-4] has examined speed-based signatures of tracked order vibration and phase, which are constructed for each engine shaft [6], and used for event detection. Examples of abnormal engine behaviour are rare [5], and so a novelty detection approach is taken, in which departures from a model of normality (constructed from normal data) are identified.
The figure to the
right shows the speed-based signature for 1LP tracked order
(vibration amplitude). Such vibration signatures have been shown to provide early
warning of engine failure, in both flight and ground-based analyses, providing
protection against hazard during engine development programmes, and in service.
Future research will focus on the development of models of normality that integrate vibration data with performance data (engine temperatures, pressures, etc.) Time- and time/frequency-domain approaches to on-line novelty detection will also be investigated, providing complementary models to the frequency-domain analyses performed to date.
[1] D.A. Clifton (2006), Condition Monitoring of Gas-Turbine Engines. D.Phil. Transfer Report, University of Oxford, Oxford, UK
[2] D.A. Clifton, P.R. Bannister, L. Tarassenko (2006), Learning shape for jet engine novelty detection in Advances in Neural Networks - Isnn 2006, Pt 3, Proceedings. pp. 828-835.
[3] D.A. Clifton, P.R. Bannister, L. Tarassenko (2006), Application of an Intuitive Novelty Metric for Jet Engine Condition Monitoring in Advances in Applied Artificial Intelligence, 19th International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems, IEA/AIE 2006, A. Moonis and R. Dapoigny, Editors, Springer-Verlag: Annecy, France.
[4] P.R. Bannister (2005), QUICK(TM) Shaft Model Feature Detector: A Re-Examination of Training Methods and Speed Dependencies. Signal Processing and Artificial Neural Networks Research Group Technical Reports, SPANN-05-PB01, University of Oxford, Oxford, UK.
[5] P.R. Bannister, D.M. King, L. Tarassenko (2004), Rolls-Royce Jet Engine Events. Signal Processing and Artificial Neural Networks Research Group Report, SPANN-04-PB02, University of Oxford, Oxford, UK, 2004.
[6] P.R. Bannister, L. Tarassenko, I. Shaylor (2004), Rolls-Royce Trent 500 Shape Analysis Tools. Signal Processing and Artificial Neural Networks Research Group Report, SPANN-04-PB01, University of Oxford, Oxford, UK.