Neural networks approach to AIAA aircraft control design challenge

Abstract

This paper focuses on designing a discrete-time lateral-directional control law for a high-performance aircraft using neural networks. The control law structure is composed of feedback and filter components formulated in the form of a three-layer feedforward neural network whose parameters are adjusted by a gradient descent algorithm to provide stabilization about the aircraft center of mass and asymptotic tracking of pilot command input. The number of parameters was chosen in an ad hoc manner. Only rate gyro and lateral accelerometer outputs are available for feedback, whereas rudder pedal and lateral stick commands are input signals to the filter. Linear simulation results at an operating point within the aircraft's envelope in the presence of atmospheric turbulence and actuator and sensor noises shed light on the ability of neural networks to serve as a practical tool for flight control law designers. UMEROUS classical and modern control law design methodologies have been proposed for designing continuous- or discrete-time longitudinal and lateral-directional control laws for aircraft.1'2 These time- and frequency-domain design techniques usually apply to a linear, time-invariant state-space dynamic model of the aircraft from either a single-input, single-output (SISO) or a multi-input, multi-output (MIMO) point of view. However, the aircraft's equations of motion are highly nonlinear in six degrees of freedom (6-DOF) throughout its operating envelope. Thus, those who design control laws for aircraft face some difficulties in extending these techniques to real-flight vehicles. Traditionally, gain scheduling has been employed to compensate for nonlinearity. In this approach, the resulting nonlinear control law for the full-flight envelope is interpolated between various linear controllers scheduled as a function of speed, altitude, and wing loading. However, the desire to enlarge the flight envelope of modern aircraft and to provide enhanced maneuver capabilities may make gain scheduling infeasible due to rapid changes in dynamics. A more sophisticated flight control system that could work well over a wide range of operating conditions is clearly needed. Adaptive control is intended for plant models that must operate in the presence of uncertainty. It utilizes the available a priori knowledge and adapts to the unknown part of the plant such as timevarying parameters. An adaptive flight control system is capable of adjusting on-line the controller's parameters in response to changing flight characteristics.3 However, instability in the adaptive control system in the presence of unmodeled dynamics, bounded input, and