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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

144 CHAPTER 5 NN BASED SYSTEM IDENTIFICATION: RESULTS AND DISCUSSION<br />

5.5 MODEL PERFORMANCE COMPARISON USING OFF-LINE<br />

TRAINING<br />

<strong>The</strong> prediction performance among three NN architectures proposed in this work is<br />

compared by collecting the error statistics that have been generated in the Figure 5.6,<br />

5.11 and 5.15. <strong>The</strong> NN models performance comparison is made based on the optimal<br />

network structures obtained from these figures.<br />

Results point that the prediction<br />

from NNARX architecture using the MLP network produce prediction quality with a<br />

total RMSE percentage <strong>of</strong> 10.11%. <strong>The</strong> HMLP network <strong>of</strong>fers prediction with slight<br />

performance improvement than the MLP (9.82%) while the modified Elman network<br />

gives the lowest prediction quality (15.78%).<br />

Results from Table 5.3, 5.6 and 5.9 indicate that the average RMSE and confidence<br />

limit values from the weights are able to produce prediction that adequately fit with the<br />

test data. From the results, the prediction performance from the optimised structure<br />

<strong>of</strong> the MLP, HMLP and modified Elman networks are shown to be robust from large<br />

variation <strong>of</strong> weights values. <strong>The</strong>refore, it can be conclude that NN training from <strong>of</strong>f-line<br />

training method such as LM algorithm would produce weights set that can produce quite<br />

large latitude <strong>of</strong> value but still able to produce a satisfactory prediction performance.<br />

5.6 ON-LINE SYSTEM IDENTIFICATION<br />

In this section, the on-line training proposed in Section 4.3.5 is implemented to identify<br />

the attitude dynamics <strong>of</strong> the UAS helicopter model using the MLP network. <strong>The</strong> suitable<br />

regression vector structure and hidden neurons size for recursive NNARX model can<br />

be determined using the k-fold cross validation technique previously discussed. Using<br />

results from the <strong>of</strong>f-line system identification, network structure with n y = 3 and n u = 1<br />

is used as the basic model structure for comparing the generalisation performance <strong>of</strong><br />

the recursive Gauss-Newton (rGN) method with <strong>of</strong>f-line training method.<br />

<strong>The</strong> comparison also utilises the k-fold cross validation method to identify the<br />

efficiency <strong>of</strong> the selected neural network training methods in estimating the attitude<br />

dynamics <strong>of</strong> the helicopter. To compare the generalisation performance <strong>of</strong> the rGN

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

Saved successfully!

Ooh no, something went wrong!