The Development of Neural Network Based System Identification ...

1.2 PROBLEM STATEMENT 7
relationship between inputs and outputs. Several recommendations have been made
in this thesis in terms of model structure selection, validation and training methods
to overcome the problems presented by the neural network based modelling. Better
selection of neural network model structure should improve the prediction of the model
while using more advanced neural network architectures other than standard multilayered
perceptron should reduce the prediction model training time. This should be
beneficial to the real-time implementation of the adaptive flight control system. The
task to explain how the neural network solves the modelling problems is outside the
scope of this thesis. Nevertheless, recent research has emerged in NN literature to
extract regression rules that explain knowledge gained by the NN model [Setiono et al.,
2002, Jianguo et al., 2011, Kamruzzaman and Islam, 2010].
The ability to model the time varying dynamics of a UAS helicopter is also important
in the development of adaptive type flight controller. Typically, a neural network model
derived from the off-line based training (batch training) will not be able to represent all
the operating points of the flight envelope very well [Samal, 2009, Ljung and Soderstrom,
1983]. Several attempts have been made in previous studies to update the neural
network prediction model during flight using mini-batch off-line training on a smaller
number of data samples [Samal, 2009, Samal et al., 2008, 2009, Puttige, 2009, Puttige
and Anavatti, 2006]. However, the proposed method can only be employed to reasonably
small networks and is limited to model uncoupled helicopter dynamics due to high
computation cost. In order to accommodate the time-varying properties of helicopter
dynamics which change frequently during flight, a recursive based learning algorithm
is required to properly track the dynamics of the system under consideration. The
application of a recursive type neural network should further improve the prediction
and adaptability of the dynamic model.
This thesis attempts to overcome the problem by presenting neural network based
modelling and control frameworks that greatly reduce the development time, cost and
resources needed to design a high performance control system for helicopter based UAS.
The neural network based approach to the system identification of unmanned helicopter
dynamics has shown promising capability to facilitate the development of accurate flight