12.4.   LIMIT DETECTION AND LIMIT AVOIDANCE

In addressing the limit detection and avoidance problem, actuator control
limits are constantly redefined, based on the predicted control limit margins
using neural networks [17, 23, 24]. These artificial limits on the actuators are
seen as ‘‘control limits’’ and are recognized as ‘‘artificial control saturations’’
in case of a violation of the ‘‘fly safe zone.’’ At the same time, commands to
the controller are modified accordingly in order to prevent the controller
from ‘‘seeing’’ and reacting to vehicle response errors. This is also essential
for controllers with adaptive features, in order to avoid ‘‘wrong adaptation’’
due to control saturation. A technique developed recently called ‘‘pseudo-
control hedging’’[25] is used to modify commands to the controller.


12.4.1.   Adaptive Nonlinear Controller

The baseline flight controller, referred to as the ‘‘low-level controller,’’ is a
model-inversion-based adaptive nonlinear controller [26, 27]. This controller
architecture has been developed under the Georgia Tech Center of Excellence
in Rotorcraft Technology (CERT) program with diverse applications.
Those include fighter aircraft, helicopters, tiltrotor aircraft, missiles, and
munitions [28-31]. This technology is leveraged by the SEC program for
further improvements and for integration with mid- and high-level controllers.
The low-level controller can be used for trajectory or velocity
tracking, but also can be reconfigured to receive angular rate commands for
aggressive maneuvering [32]. An adaptive neural net block is used in the
feedback path in the inner loop to account for inversion errors and to
guarantee closed loop stability. More details on the inner- and outer-loop
controllers along with control law derivations are given in references 28 and
29. Additional design details, derivation of the neural network update law,
and a proof of closed-loop stability can be found in references 30-32.

12.4.2.   Neural-Net-Based Limit Detection and Avoidance

The functional relationship between the set of limit parameters and the set
of measurable state variables and control variables is modeled through a
neural network. Data sets corresponding to dynamic trim solutions of the full
model are used to train the neural network. In dynamic trim the states of the
aircraft are divided into fast and slow variables. The slow states include flight
parameters that vary slowly with time. The fast states include states that vary
quickly with time and reach a steady-state value during a maneuver. The
dynamic trim condition is defined as a quasi-steady condition in which the
fast states have reached their steady state [17]. Estimates of limit parameters
in dynamic trim can effectively predict a limit exceedance as soon as the
control inputs are applied. The time where the transient dynamics occur
provides a time buffer to take precautions to avoid exceedance of a selected
limit. This technique is equally applicable for a vehicle with multiple control
inputs with multiple limits as has been demonstrated in reference 17. The
dynamic trim predictions are used for calculating the control margins between
the current control positions and the control deflections that will result
in a limit exceedance. The most critical control margin is then used as a
variable control limit inside the flight controller for limit avoidance.