System Identification Technique and Neural Networks for Material Lifetime Assessment Application - Complex System Modelling and Control Through Intelligent Soft Computations - page 782

Fig. 23 S-N curves obtained
by the NN-NARX model and
the experimental data for
R = 0.5: ʸ =90 ° of
E-glass fabrics/epoxy

In the form of S-N curves, it can be seen that the discrepancies observed in fatigue
lives predicted by the NN-NARX model and those of experimental data were not
excessively large. The results may justi

cient of determination (R 2 )
of the fatigue life prediction can be considered high for the materials examined. The
values of R 2 was ranging from 0.7445 to 0.9504 and from 0.8737 to 0.9788 for E-
glass/polyester and E-glass fabrics/epoxy, respectively. In addition, it is interesting
to note that the highest R 2 values for the materials were coming from the fatigue data
related to the non-on-axis orientations i.e.

ed that the coef

for E-glass/polyester
and E-glass fabrics/epoxy, respectively. Furthermore, it is also obvious that the
fatigue life assessment of multivariable and multiaxial loadings is fashioned in an
ef

ʸ =30 ° and ʸ =90 °

cient manner. This emphasized again the applicability and the feasibility of the
present NN-NARX model and the procedure developed in the present study for
multivariable and multiaxial fatigue life assessment of the composite materials.
For remarks on the presented fatigue life prediction results, the following dis-
cussions are highlighted. Firstly, because fatigue life assessment is realized as one-
step ahead prediction with respect to each stress level-S corresponding to the related
R and values, the accuracy of the NN prediction results hence will depend on the

accuracy of the NN prediction result on each stress level-S examined. The obtained
NN fatigue life prediction for a stress level will affect that of the next stress level.
With respect to this matter, two aspects may be considered that the change in the
selection of training fatigue data will also change the NN fatigue life prediction
results. Also, the sequence of fatigue data of the testing sets examined can affect the
NN fatigue life prediction results obtained. To be consistent, in the present study,
stress ratio-R values were arranged according to their positions and transitions in the
CLD, while on-axis/off-axis orientation-

ʸ

values were arranged based on the
magnitude value from longitudinal to transverse direction, as shown in Table 2 .
Secondly, only one value of

ʸ

.It
is intended here to employ the value to examine the feasibility of the present
approach. Looking at the NN prediction results obtained, it appears that the NN-
NARX model was able to perform the fatigue life prediction of the material fairly
well using the starting point of limited training fatigue data. Moreover, different

ʸ was employed in the training set here i.e. ʸ =0 °