No abstract

No abstract

A major goal in flight control over the past decade has been the development of reconfigurable flight control systems which can adapt their gains in real-time to compensate for aircraft damage and in-flight system failures. The purpose of this paper is to describe the controller developed for the LoFLYTEQ aircraft, which is a testbed for neural networks research. The LoFLYTEQ control system is based on the Accurate Automation Corp. Neural Adaptive Controller VACTM) which is designed to achieve this goal. The LoFLYTEQ program is an active flight test program at the Air Force Flight Test Center at Edwards Air Force Base, with the objective of demonstrating a neural network control system for a waverider vehicle. The AAC control system has two innovative components: an adaptive actuatorklight surface controller, and a leaming/adaptive stability augmentation system designed with neural network and reinforcement learning techniques. AAC NEURAL ADAPTIVE CONTROLLER (NACTM)The NACTM, whose architecture is illustrated in Figure 1, is an n* order multivariate adaptive controller derived fiom Seraji's , 2"d order single variate robotic joint controller, to which a neural outer loop has been added to improve the short and intermediate time response of the controller [Cox, et a1 (1994)] [Saeks, et a1 (1997)l. The NACTM is applicable to an essentially arbitrary n* order multivariate linear or nonlinear system that satisfies an appropriate "slow-varying" assumption. [In the continuous-time situation, this assumption is straightforward to satisfl. It turns out, however, when discretization is performed for computational reasons, the discretization itself injects "high speed" variations, and theoretically it can be shown that terms of the "wrong sign" enter the discrete equations. Theoretically, this computation-induced condition may be alleviated by using sufficiently high sampling rates.] Stability and robust tracking are guaranteed by the underlying adaptive controller. For some applications, a neural network "gain scheduler'' is used to speed up the adaptation process and to enhance the short and intermediate time response of the system. The N A P M deals equally well with linear and nonlinear systems and requires no explicit aircraft model. Rather, it implicitly models the aircraft dynamics via the observed error between the actual and reference trajectories. It can adapt in real-time to changes in the plant dynamics caused by system failures,

n e~r a l l~* '~~' ' and fuzzy leamhg laws15, sliding mode techniques, and input-output techniques. A globally convergent nontinear Approximate Dynamic Programming algorithm is described, and an implementation The centerpiece of Dynamic Programming is the Hamilton of the algorithm in the linear case is developed. The resultant Jacobi Bellman (HJB) Equation3** which one solves for the linear Approximate Dynamic R0-g algorithm is cptimal costfinctional, ~(x o '3. his equation characterizes illustrated via the design of an autolander for the NASA X-43 research without a priori knowledge of the x-43.s the cost to drive the initial state, xb at time to to a prescribed fmal state using the optimal control. Given the optimal cost flight dynamics. functional, one may then solve a second partial differential 1. On Sabbatical Leave fiom Portland State University. 2. This research performed in part on NSF SBIR contract DMI-9983287and NASA Ames SBIR Contract NAS2-99047.

No abstract