ANALYSIS AND CONTROL OF MAGNETIC BEARINGS USING NEURAL NETWORKS FOR TURBINE ENGINE APPLICATIONS

The Air Force has identified a need for more advanced digital control methodologies to make feasible the application of active magnetic bearings to turbine engines. Achieving the goals for enhanced turbine rotor performance necessitates improved bearing control with associated compact size/weight, fault tolerance, and minimum power consumption. Conventional control methods are inadequate because of the wide and highly nonlinear operating regimes which can occur due to unbalance or shaft critical speeds. To address this need, American Research Corporation of Virginia proposes the development and training of an innovative hardware based multiple layer, feed-forward neural network capable of addressing nonlinear operational conditions. The program is innovative in combining neural networks and digital systems into a hybrid adaptive control package to provide robust control of magnetic bearings under nonlinear conditions. Technical objectives include selection of neural network configurtions, development of a training set, evaluation of hybrid controller response, and demonstration of the proof-of-concept system. The proposed effort will demonstrate the feasibility of a neural network-based controller to provide superior engine performance. Results anticipated include neural network definition, system implementation, stability margins, fault tolerance and compactness. The significance of the program is related to the improved control of advanced gas turbine and other rotating machinery systems made possible by neural network adaptive control.