Robust and Reconfigurable Flight Control by Neural Networks

A linear matrix inequalities (LMIs) framework for developing robust, adaptive nonlinear flight control systems is presented. The controller structure is that of a feedforward sigmoidal neural network that is both adaptive and reconfigurable, since the control law it approximates can be modified by updating its parameters during operation. The neural network controller is designed via LMIs to meet multiple control objectives, that include but are not limited to LQG and H∞ performance, pole placement, and closed-loop stability, as dictated by the application of interest. Prior knowledge of the linearized equations of motion is utilized in order to guarantee that the neural network controller meets these objectives when the aircraft is operating in its linear-parameter varying (LPV) regime, or steady-state flight envelope. However, should unexpected changes or failures occur during flight, the controller also is capable to reconfigure according to the new and, possibly, nonlinear dynamics. The adaptation consists of a constrained optimization problem that is computationally feasible because it takes place incrementally over time, accounting for the new dynamics only if and when they arise. The LPV performance of the controller is preserved and guaranteed throughout adaptation, by means of a novel constrained-training technique for neural networks.