Engine monitoring using neural networks

The ability of neural networks to diagnose the performance of aero-engines is assessed. An turbojet engine simulation program is first used to generate engine data under different engine operating conditions. Engine temperatures, thrust, and TSFC are calculated for different efficiencies of the engine components: diffuser, compressor, burner, turbine, nozzle, and mechanical shaft. A feed-forward neural network model is then developed and used to predict the performance of the engine component efficiencies based on engine temperatures, thrust, and TSFC. The back propagation algorithm with delta learning rule and hyperbolic transfer function is used to train the network. The burner efficiency is predicted very well; however, the predictions of the diffuser and nozzle efficiencies are poor. The correct trend for the compressor, mechanical and turbine efficiencies are predicted The neural network, with the same architecture, is then applied to the analysis of actual vibration data of twin-spool turbofan. The neural network is able to adequately predict both normal and abnormal vibrations of the engine. The neural network approach thus has the potential to provide timely improvements in the diagnosis and monitoring of aero-engines. Introduction A current method used to monitor engine performance employs "pattern matching techniques," in which spectral data of engine vibrations are compared to predetermined fault signatures. A drawback to this approach is that a highly trained professional is needed to interpret the spectral data. The approach therefore does not lend itself well to increased productivity, improved quality, or lower 'Mechanical and Aerospace Engineering Department, Student Member AIAA. t Associate Professor, Mechanical and Aerospace Engineering Department, Senior Member AIAA. Copyright © 1998 by Melissa L. Cifaldi and Ndaona Chokani. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission. costs, which are all necessary in an increasingly competitive market place. It is possible that improvement in engine monitoring may be achieved through the use of neural networks. Denney examined the use of neural networks to assess the performance of the F-16's F-100 engine; five engine parameters from actual flight data were monitored. However, the limited number of flight data did not allow the utility of neural networks to be clearly demonstrated. Dietz et al presented a more comprehensive description of engine fault diagnosis using neural networks. However, they did not describe details of the neural network architectures required to accomplish the diagnosis. The present work examines the use of neural networks for the monitoring of aeroengine performance. The focus is on the architecture of the neural network, and the network's ability to predict the trends in the engine's performance. In the first part of the study the engine operation is simulated using a quasi, one-dimensional model. In the second phase, actual engine vibration data are used. Approach The simulated engine data are generated using a quasi, one-dimensional model of a turbojet engine, Figure 1. Further details on turbojet engines can be found in Hill and Peterson and Cohen et al. The efficiencies of the six engine components, shown in Figure 1 are used to simulate the off-design engine performance. These efficiencies are randomly varied between the typical operating ranges for each component. The thrust and the thrust-specific-fuelconsumption rate (TSFC) of the engine, in addition to the five temperatures indicated in Figure 1, are then calculated using these component efficiencies. Thus each combination of efficiencies, temperatures, thrust, and TSFC represents an engine operating condition. The neural network is trained using 10,000 operating conditions and 25 operating conditions are used to evaluate the neural network's performance. The Marine's Harrier V/STOL aircraft is powered by an F402-RR-406A/B engine. This engine is a twinspool turbofan, whose components include a highand low-pressure compressor-turbine pair. The engine test cell at the Naval Aviation Depot (NADEP), Cherry 1 American Institute of Aeronautics and Astronautics