System Identification Technique and Neural Networks for Material Lifetime Assessment Application - Complex System Modelling and Control Through Intelligent Soft Computations - page 760

and principal component analysis (PCA) networks, and compared the prediction
results to the experimental data. Specimens with ve fiber angle orientations of 0 ° ,

19 1, 0 and

0.5. Ninety two experiment data made up the application data for the networks.
They found that NN can be trained to model the nonlinear behaviour of composite
laminate subjected to cyclic loading and the prediction results were comparable to
other current fatigue-life prediction methods.
Freire Junior et al. ( 2005 ) followed different approach, by which NN was utilized
to build constant life diagrams (CLD) of fatigue. The researchers built CLD of a
plastic reinforced with

° ,45 ° ,71 ° and 90 ° were tested under three stress ratio-R conditions of −

45/0] S lay-up. Four
training data sets (each set consists of 3R,4R,5R and 6R values, respectively) were
set up from twelve stress ratio-R values. It was found that the use of NN to build
CLD was very promising where the NN model trained using only three S-N curves
could generalize and construct other remaining S-N curves of the CLD building. For
better generalization, however, six S-N curves should be utilized in NN training.
Vassilopoulos et al. ( 2007 ) criticized that the determination of six S-N curves
was a costly task for the NN prediction purpose. Instead, these authors used a small
portion, namely 40

fiberglass (DD16 material) with [90/0/ ±

50 %, of the experimental data. It was shown that it is possible
to build CLD using the small portion data and NN was proven to be a suf

-

cient tool

for modelling fatigue life of GFRP multidirectional laminates.
Further, Vassilopoulos et al. ( 2008 ) have employed genetic programming for
modeling the fatigue life of several

reinforced composite material systems. It
was shown that if the genetic programming tool is adequately trained, it can pro-
duce theoretical predictions that compare favorably with corresponding predictions
by conventional methods for the interpretation of fatigue data. It was also pointed
out that the modeling accuracy of this computational technique was very high. In
addition, the proposed modeling technique presented certain advantages compared
to conventional methods. The new technique was a stochastic process that led
straight to a multi-slope S-N curve following the trend of the experimental data,
without the need for any assumptions.
Bezazi et al. ( 2007 ) have investigated fatigue life prediction of sandwich com-
posite materials under

ber

-

cial neural net-
work. The authors noticed the good generalization of NN trained with Bayesian
technique in comparison to that with maximum likelihood approach in predicting
fatigue behaviour of the sandwich structure. Nonetheless, only one lay-up con

fl flexural tests using a Bayesian trained arti

g-

uration was considered in the work.
Freire Junior et al. ( 2007 , 2009 ), in their next attempts, showed that the use of
modular networks (MN) gives more satisfactory results than feed-forward (FF)
neural network. However, it was still necessary to increase the training sets for
better results.
Hidayat and Melor ( 2009 ) have noticed the potential use of limited number of
fatigue data in the NN modelling of fatigue life of composite materials with
Bayesian regularization. The authors have investigated E-glass/epoxy ([ ± 45/0 4 /

± 45/0] S ) composites under fatigue

loadings with various stress ratio values. It was found that although only two stress

45]) and DD16 or E-glass/polyester ([90/0/ ±