A Recurrent Neural Network Controller for Gust Load Alleviation on a Transport Aircraft

A Recurrent Neural Network controller for the alleviation of gust loads on a regional transport aircraft is designed. The aerodynamic and elastic properties of the aircraft are modeled using a linear state-space system, while a nonlinear model is used for the actuation of the control surfaces. Two neural networks are employed to obtain a model predictive controller; the first one is an identification network which predicts future outputs, the second one is a controller network which computes the control action by optimizing a quadratic cost function, based on the amplitude of both the reconstruction error and of the control input. Both networks are trained on-line, thus providing an adaptive control. Two different controllers acting on the aileron are designed, the first one is based on the measure of the vertical acceleration at the center of mass and at the wing tip, while the second is based on the measure of the angle of attack by means of a sensor located on the aircraft nose. The evaluation of the controller performances under both stochastic turbulence and deterministic gust proved the superiority of the controller based on the angle of attack sensor.