12.5.   ADAPTIVE LIMIT DETECTION

As the method described above relies on accurate dynamic trim data, an
online technique for predicting dynamic trim parameters and calculating the
corresponding control margins is essential for developing robust automatic
limit detection and avoidance algorithms. The online technique utilizes an
observer-type adaptive neural network loop for the estimation of the correct
aircraft model. The constructed aircraft model is then used to predict the
quasi-steady response behavior of the limit parameters and the corresponding
control margins using a second adaptive neural network loop. Though the
approach does not require any offline training of the neural networks,
existing offline trained neural network data maps can be accommodated in
the procedure. Only standard sensor measurements are used for adaptation.
A detailed development of the method along with a simulation evaluation are
included in reference 24. As offline training of the network weights is not
required, the system has the advantage of adapting to varying flight condition
and different vehicle configurations.

The proposed approach consists of two loops. The first loop is similar to
an observer loop utilizing adaptive neural networks (ANN). This is used to
capture the uncertain plant dynamics. Similar adaptive observers have been
introduced previously [33], mainly for control purposes [34, 35]. The observer
in our case is used to approximate the local behavior of the limit parameter.
The second loop is used to obtain a dynamic trim solution using the dynamic
model generated in the first loop. A neural network is also utilized in the
second loop to speed up the solution convergence. In addition to finding the
dynamic trim solution for the limit parameter, the second loop is used for
finding the corresponding limit control margin as well. The limit control
margin predictions can be used as artificial control limits to limit control
commands from the low-level controller as described in the previous section.
Simulation evaluations of the proposed adaptive limit detection and avoidance
algorithms are carried out using a linearized helicopter model and the
nonlinear 6-DOF Generic Tilt-Rotor Simulation, GTRSIM. High-g pull-up
and push-over maneuvers are included as examples, to demonstrate satisfactory
performance of the proposed technique.