Nonlinear Flight Control Using Neural Networks and Feedback Linearization

Aircraft dynamics are in general nonlinear, time-varying, and may be highly uncertain. Current-generation controllers rely on approximate linearized models of the aircraft and use gain scheduling to accommodate changes in vehicle dynamics as the flight regime varies. The techniques of feedback linearization provide a means of developing invariant controllers that give a desired response in all flight modes. However, the implementation of these techniques involves intensive online computations. The structure imposed by feedback linearization proves an ideal setting for introducing neural networks to the flight-control loop. In this paper, a structure for the use of neural networks to represent the nonlinear inverse transformations needed for feedback linearization is proposed and evaluated. In order to compensate for unmodeled nonlinearities and parameter drifg a second network is introduced which permits online learning. In addition, the paper addresses the robust stability problem in the context of neural-network representation error.