28.02.2014 Views

The Development of Neural Network Based System Identification ...

The Development of Neural Network Based System Identification ...

The Development of Neural Network Based System Identification ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

30 CHAPTER 2 LITERATURE REVIEW<br />

prediction and past input and output measurements [Norgaard, 2000].<br />

In the NNARX model structure, the variables to be estimated and other influencing<br />

variables including their time lags are typically fed into a static feed-forward network<br />

such as the multi-layer perceptron (MLP) network [Samarasinghe, 2007]. Studies have<br />

shown that the NN based approach using the NNARX architecture is effective in<br />

modelling the dynamic response <strong>of</strong> the helicopter based UAS [Suresh et al., 2002, Samal<br />

et al., 2008, San Martin et al., 2006, Rimal et al., 2011, Chetouani, 2010]. Another<br />

interesting finding following the work <strong>of</strong> Rahideh et al. [2008] suggests that the NNARX<br />

networks trained with Levenberg-Marquardt (LM) algorithm is capable to exceed the<br />

prediction performance <strong>of</strong> first principle model.<br />

Different neural network structures such as Radial Basis Function (RBF) and<br />

Recurrent <strong>Neural</strong> <strong>Network</strong>s (RNN) have been used in past research works with success<br />

in identifying the helicopter dynamics [Ahmad et al., 2002, Kumar et al., 2010, Pedro<br />

and Kantue, 2011]. An example <strong>of</strong> RBF network application for system identification <strong>of</strong><br />

a twin rotor helicopter system was proposed in Ahmad et al. [2002]. <strong>The</strong> RBF network<br />

typically uses the orthogonal least square (OLS) algorithm to systematically find a set<br />

<strong>of</strong> weights and radial basis centres to ensure that the desired input-outputs mapping is<br />

performed [Chen et al., 1991]. However, the RBF network usually requires more hidden<br />

neurons compared with the conventional NNARX network with sigmoid or tangent<br />

hyperbolic activation functions in the hidden layer. <strong>The</strong> reason behind such an increase<br />

in the number <strong>of</strong> neurons used in the RBF network contributes to the fact that the<br />

outputs <strong>of</strong> the radial basis neurons only cover a relatively small input space compared<br />

to the outputs <strong>of</strong> typical activation functions in the NNARX networks [Shaheed, 2005].<br />

Thus, in order to adequately represent broader input space, a much larger number<br />

<strong>of</strong> neurons need to be included in the network to achieve more reliable identification<br />

results.<br />

Many parameters in a RBF network such as the network weights, biases, centre<br />

vectors and spread constant need to be optimised compared with a much simpler<br />

NNARX network that only require optimisation <strong>of</strong> weights and biases [Shaheed and<br />

Tokhi, 2002]. In terms <strong>of</strong> generalisation performance, NNARX network produces a much

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!