Analysis of adaptive neural networks for helicopter flight controls

This paper treats the design of online adaptive neural networks for use in a nonlinear helicopter flight control architecture. Emphasis is given to network architecture and the effect that varying the adaptation gain has on performance. Conclusions are based on a nonlinear evaluation model of an attack helicopter, and a metric that measures the networks abilityto cancel the effect of modeling errors for a complicated maneuver. The network is shown to provide nearly perfect tracking in the face of significant modeling errors and additionally to cancel the model inversion error after a short initial period of learning. Furthermore, it is shown that the performance varies gracefully and monotonically improves as the adaptation gain parameter is increased. The effect on control effort is modest, and is mainly perceptable only during a short training episode that can be associated with transition from hover to forward flight.