XVI Portuguese Conference on Fracture 2018 23-24th April 2018

IMPROVEMENT OF FATIGUE LIFE ESTIMATION OF COMPOSITE

MATERIALS UNDER CONSTANT AMPLITUDE LOADING

Joelton F. Barbosa1,2(*), J.A.F.O. Correia3, R.C.S.F. Júnior1, A.M.P. De Jesus3

1
UFRN, Postgraduate in Mechanical Engineering, Federal University of Rio Grande do Norte, Natal, Brazil
2
UFERSA, Department of Engineering, Federal Rural University of the Semi-Arid, Mossoró, Brazil
3
INEGI and CONSTRUCT, Faculty of Engineering, University of Porto, Campus FEUP, Portugal
(*)
Email: joeltonfb@gmail.com

ABSTRACT

Knowledge of the stochastic nature of fatigue life of composite materials can be modeled by the
failure time with the Weibull distribution. This task becomes complex when the samples are
small and scattered. In this way, it is necessary to know and to improve, robust models of
estimation of the parameters of the distribution of Weibull. The aim of this work is to compare
the performance of least squares (LS), least squares weighted (WLSE), maximum likelihood
(MLE) and momentum method (MOM), and to suggest the method that obtains better
performance in life behavior to fatigue with small samples. Monte Carlos simulations were
performed to estimate the distribution parameters with different sample sizes and an application
with real fatigue data that compares performance using goodness-of-fit. The results of the
simulations showed that the WLSE was able to generate more reliable estimators for fatigue
behavior during its useful life. With this, it is possible to conclude that small samples make the
real representation of life difficult to the material fatigue but using the WLSE method it is
possible to obtain more estimates.

Keywords: Fatigue; Composite; Statistical properties/methods; Weibull parameter estimation.

INTRODUCTION

The stochastic nature is very important to evaluate the fatigue life of composite materials and can
be modelled by the failure time with the Weibull distribution, being a complex task when the
samples are small and scattered. The necessity of the study of fatigue to ensure that it does not
suffer failures before the end of the useful life, because when subjected to fluctuating or
repetitive tensions [1] under cyclic loadings, stresses much lower than those subjects to the
monotonic loadings (tensile or compression) [2].
Some papers present studies considering small samples of size 10 and 20, but for the reality of
the fatigue tests it is common to observe samples of size equal to 5 for a test with constant
cyclical load. It is possible to see this from the work of Khalili and Kromp [3], who analysed
different techniques of estimation of fragile materials and concluded that at least 30 samples
were necessary for the characterization of this fragile material.
More robust regression models, known as Weighted Least Squares Estimation (WLSE), capable
of giving each point of the data its ideal weight of influence on the estimate. The work of Hung
[4] proposed using the WLSE to estimate the shape parameter of the Weibull distribution and its
results, by simulation, presented better performance than the techniques already mentioned.
Some publications, such as, Yavuz [5] and Kantar et al. [6] compare and analyse the WLSE
performance, but nothing was applied to actual data and small samples. A WLSE model with
adjusted weights that considerably improves the performance of the estimate for a small set of
data, was also proposed by Zhang [7].
http://www.pcf2018.ubi.pt PCF 2018
UBI, Covilhã, Portugal
XVI Portuguese Conference on Fracture 2018 23-24th April 2018

The purpose of this work is to perform a comparative research on the performance of least
squares (LS), maximum likelihood (MLE), and momentum method (MOM) methods with the
new Weighted Least Squares Estimation (WLSE) model of Weibull distribution for a small data
set. To evaluate the performance of these estimators will use data from the high cycle fatigue
tests to prove the performance of these estimators.

MATERIAL AND METHODS

With aims to be evaluated the performance of the estimation methods for reduced samples, it is
necessary to compare the different methods through the Monte Carlo simulation and the
performance is evaluated using the criterion of goodness of fit of the distribution estimated with
the experimental data. A set of experimental data is used, in which the results of the work of Lee
[8] which tested a 98 Layer epoxy matrix composite reinforced with 16-layer HTA carbon fiber
and with unidirectional fiber loaded with a strain amplitude of 1.6 GPa (73% of tensile strength)
and fatigue ratio of 0.1. The reason for choosing this data is because it has a large amount of
experimental data (sample size 50), which helps us in comparisons in goodness-of-fit tests.

This estimation model considers for a data set that each element has a specific weight, that is,
each element of a given sample will have a greater or lesser contribution in the calculation of
probability. The idea of the WLSE is to consider that each observation point, used in the
regression analysis, has an influence on each of these points through a weight Wi. The choice of
weights is maximized to improve parameter estimation. The objective function of the WLSE
estimator is given by:
n

  y    x   ln  
2
min SS '  i i (1)
i 1
Taking the partial derivatives of SS' relative to α and β, respectively, and set the resulting
expressions equal to zero. The estimated weighted least squares (WLSE) of the parameters α and
β are:
n n n n

  wi  wi xi yi   wi xi . w y i i
ˆ  i 1 i 1 i 1 i 1
2 (2)
n n  n 
 w 
i 1
i
i 1
wi xi2 



i 1
wi xi 



 n n 


i 1
wi yi  ˆ  w x 
i 1
i i
ˆ  exp   n  (3)



ˆ  i 1
wi 


The values of xi and yi are obtained in the same manner as for the LSE model. The values of the
weights w are modeled through a polynomial model that realizes a relation between Fi and wi to
carry out the proposed approximation. In [4] the approximations of the weights are given by:
wi  0.076  3.610Fˆi  6.867Fˆi 2  13.54Fˆi3  9.231Fˆi4 (4)
where the values of Fi are obtained by equation Y   X   ln  in i is the order number of failures
and n is the size of the sample.

http://www.pcf2018.ubi.pt PCF 2018

UBI, Covilhã, Portugal
XVI Portuguese Conference on Fracture 2018 23-24th April 2018

APPLICATION WITH EXPERIMENTAL LIFE-TO-FATIGUE DATA

This analysis of the performance of the estimation methods with the fatigue data is given by the
following steps:

1. Removal of random samples from the data set;

2. Estimates of the α and β parameters of the Weibull distribution;
3. Graphical analysis of the Weibull probability paper (WPP) and cumulated distribution (CDF).

Step 1: Through the Monte Carlos simulation, 10.000 random samples will be taken for each
sample from 50 life assays at fatigue of the HTA / 982A.
Step 2: It is generated 10.000 samples of size n=3, n=4 and n=5. The estimates of a and b values
of the Weibull distribution using the four estimation methods estimated by Equation (5):
10.000 10.000
 
E ˆ 
1
10.000  ˆi and E ˆ  
1
10.000  ˆ i (5)
i 1 i 1
Step 3: Through the Goodness-off-fit (GOF) tests it is possible to determine which estimated
distribution presents the best level of fit to the fatigue data. It is measured the distance between
the cumulative distribution function estimated, F, and the function of the empirical distribution,
Fn. Three goodness-of-fit statistics are performed: Kolmogorov-Smirnov; Cramer-von Mises;
and Anderson-Darling [9].

Table 1: Goodness-of-fit statistics as defined by D’Agostino and Stephens (1986)

RESULTS AND DISCUSSION

The results of the simulations showed the application of the WLSE method for small samples,
where is verified a better performance when compared to the others methods. When the WLSE
method is compared to the LSE, it is possible to verify that the results are close, it can still be
superior in most tests. But when compared to the MLE and MOM methods it is possible to
observe a much greater difference. This confirms that when performing fatigue tests at a constant
strain level it is important to use this method to have an estimate with the least precision error.
For the other methods to reach a good estimation of the distribution parameters it is necessary to
have larger samples when compared to the WLSE confirms that a smaller amount of
experimental data will be necessary, leading to a reduction of costs.
Analysing the results of Table 2 it is verified that the WLSE estimator has a better fit. The results
of the Goodness-of-fit statistics presents the WLSE method that had the best fit in all three test
statistics. The MLE and MOM methods presented a low adjustment, especially when we observe
the Anderson-Darling statistic that considers a greater weight in the adjustment of the tails of the
cumulative distribution.
http://www.pcf2018.ubi.pt PCF 2018
UBI, Covilhã, Portugal
XVI Portuguese Conference on Fracture 2018 23-24th April 2018

Table 2: Goodness-of-fit statistics

Method KS CvM AD
n 3 4 5 3 4 5 3 4 5
MLE 0.1583 0.1256 0.103 0.3834 0.2334 0.1545 6.4252 3.5592 2.2116
WLSE 0.0758 0.0633 0.0795 0.0682 0.0328 0.03 0.7626 0.3203 0.2451
MOM 0.16 0.1311 0.11 0.3924 0.2451 0.1675 6.3913 3.6748 2.3861
LSE 0.0851 0.0696 0.0652 0.101 0.0539 0.0429 0.9979 0.4925 0.3507

In this expanded abstract, it was possible to compare and suggest a method of estimation of the
parameters of the Weibull distribution most recommended to represent the fatigue life behaviour
of the composite materials for small samples. The WLSE method presented a good performance
in the estimation of Weibull parameters for small samples. The LSE and WLSE methods
generally present a good performance when the samples are smaller than 20. Samples smaller
than 5 the WLSE performance was higher than the other methods. This can be proven through
the Monte Carlo simulation that generated scenarios with small samples of the HTA / 982A
fatigue data. The analysis was performed through the graphs of Weibull's cumulated distribution
functions and through Goodness-of-fit statistics to prove this level of adherence. In all these
techniques of analysis of the estimation methods it was possible to realize that small samples,
can pass a distorted view of the behavior of the life to the fatigue of the materials, to minimize
the errors and to improve reliability in the estimation of the parameters of the distribution of
Weibull, it is recommended that high cycle fatigue test of composite materials use the weighted
least squares method and sample greater than five.

REFERENCES
[1] Wicaksono S, Chai GB. A review of advances in fatigue and life prediction of fiber-reinforced
composites. Proc Inst Mech Eng Part L J Mater Des Appl 2013;227:179–95.
doi:10.1177/1464420712458201.
[2] D’Amore A, Giorgio M, Grassia L. Modeling the residual strength of carbon fiber reinforced
composites subjected to cyclic loading. Int J Fatigue 2015;78:31–7.
doi:10.1016/j.ijfatigue.2015.03.012.
[3] Khalili A, Kromp K. Statistical properties of. J Mater Sci 1991;26:6741–52.
doi:10.1063/1.3050751.
[4] Hung W-L. Weighted least-squares estimation of the shape parameter of the Weibull distribution.
Qual Reliab Eng Int 2001;17:467–9.
[5] Yavuz AA. Estimation of the shape parameter of the Weibull distribution using linear regression
methods: Non-censored samples. Qual Reliab Eng Int 2013;29:1207–19. doi:10.1002/qre.1472.
[6] Kantar YM, Ibrahim A, Ilhan U, YEN\.ILMEZ I. Comparison of some estimation methods of the
two parameter Weibull distribution for unusual wind speed data cases. Int J Informatics Technol
2016;9:81.
[7] Zhang LF, Xie M, Tang LC. On Weighted Least Squares Estimation for the Parameters of Weibull
Distribution. Recent Adv. Reliab. Qual. Des., London: Springer London; 2008, p. 57–84.
doi:10.1007/978-1-84800-113-8_3.
[8] Lee J, Harris B, Almond DP, Hammett F. Fibre composite fatigue-life determination. Compos Part
A Appl Sci Manuf 1997;28:5–15. doi:10.1016/S1359-835X(96)00088-7.
[9] Stephens MA. Tests based on EDF statistics. Goodness-of-Fit Tech 1986;68:97–193.

http://www.pcf2018.ubi.pt PCF 2018

UBI, Covilhã, Portugal