| Summary: | The success of Unmanned Combat Aerial Vehicles (UCAVs) requires further
developments in the field of automated aerial refuelling (AAR) and control systems. AAR
aircraft models identified thus far do not take the centre of gravity (cg) position
movement into account during refuelling. A six-degree-of-freedom aircraft model was
combined with a moving cg model for refuelling. The equations of motion for the aircraft
in flight refuelling showed the aircraft dynamics to be coupled in the longitudinal and
lateral-directional planes when the cg had moved away from the reference point.
Applying assumptions specific to the flight conditions, simplified equations of motion
were derived. Modal analysis of four cases for the linearised aircraft model during aerial
refuelling was conducted. This revealed that the increase in mass was favourable to the
stability of the Dutch Roll mode, but the mode did become more oscillatory initially as
mass was increased, but as the cg moved forward, the mode became less oscillatory. The
opposite was observed with the Phugoid mode. The Short Period Oscillation (SPO)
decomposed into two first order modes during refuelling and these remained unchanged
during the refuelling process. Three radial basis function (RBF) neural networks
(RBFNN) were developed and trained to approximate the inverse plant dynamics and
predicted commanded deflections of the elevator, aileron and rudder. Training data
required for the network was randomly generated and the desired rates and commanded
control surface deflections were computed. The training error was the smallest in the
elevator deflection required during refuelling. A basic nonlinear dynamic inversion (NDI)
controller without a neural network (NN) was designed for the aircraft. The performance
of this controller was not satisfactory. The RBF was combined with the NDI to form a
RBFNN-based controller. The longitudinal NDI RBFNN-based controller was less
sensitive to modelling errors than the base NDI controller. The lateral NDI RBFNN-based
controller’s performance was worse than the longitudinal controller, but showed potential
as a technique for future consideration. Including the variation of aircraft inertia in the
model has been recommended as further work, as well as exploring other neural network
topologies in the NDI NN controller.
|
|---|
