Three-Parameter vs. Two-Parameter Weibull Distribution
Three-Parameter vs. Two-Parameter Weibull Distribution
www.elsevier.com/locate/compstruct
a
Department of Civil and Environmental Engineering, The University of Tennessee, Knoxville, TN 37996-2010, USA
b
School of Civil Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0355, USA
Abstract
The three-parameter and two-parameter Weibull distributions are compared using 26 mechanical property data sets of fiber-
reinforced polymeric (FRP) composite materials manufactured by the pultrusion process. Both strength and stiffness properties were
examined. The probability distributions were compared on the basis of goodness of fit, nominal design values, and allowable load to
achieve uniform reliability. It is recommended that the two-parameter Weibull distribution be used to characterize FRP composite
material properties. The primary basis for this recommendation is small differences in nominal design values and small differences in
allowable loads between the two-parameter and three-parameter Weibull distributions. Other supporting reasons for the recom-
mendation are similar observed significance levels in distribution fitting, computational efficiency, and the fact that the location
parameter of the three-parameter Weibull distribution is near the first order statistic.
Ó 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Composites; Probability; Statistical analysis; Weibull distribution; Reliabililty; Maximum likelihood; Method of moments
*
Corresponding author. Tel.: +1-865-974-7540; fax: +1-865-974- in which h is the scale parameter, b is the shape param-
2669. eter, and d is the location parameter. If d ¼ 0, the distri-
E-mail address: [email protected] (R.M. Bennett). bution becomes the two-parameter Weibull distribution.
0263-8223/02/$ - see front matter Ó 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 2 6 3 - 8 2 2 3 ( 0 2 ) 0 0 1 5 8 - 7
498 M. Alqam et al. / Composite Structures 58 (2002) 497–503
Zanakis [15] documents seventeen different methods E-glass/polyester pultruded box-shaped components in
for obtaining the parameters of the three-parameter another case [16]. Coupons from the former and latter
Weibull distribution. Two common parameter estima- study are labeled VG and PG, respectively. Coupon
tion methods used in engineering are the modified mo- data from a study [10] of E-glass/polyester plates is also
ment method and the maximum likelihood method, with included in this study, and labeled as PG-M. Both
those being the methods considered in this paper. The stiffness and strength parameters are examined, with a
two-parameter and three-parameter Weibull distribu- total of 26 data sets and over seven hundred data points
tions are compared on the basis of the goodness-of-fit of being considered. A summary of the data sets is given in
the distribution to FRP material property data and Table 1, which gives the sample size and coefficient of
lower tail behavior. The lower tail behavior is examined variation of each data set.
through the analysis of nominal design values and
probabilistic based allowable loads.
Data used for the comparisons are taken from two
2. Parameter estimation
different studies that examined the short-term axial
compressive strength of E-glass/vinylester pultruded
Two methods will be used for parameter estimation,
I- and box-shaped components in one case [17], and
the modified moment method and the maximum likeli-
the short-term eccentric axial compressive strength of
hood method. Moment methods are based on equating
Table 1
Description of data sets and location parameter for three-parameter Weibull distribution of d^=x1 and d^=x
Material property Sample size (2) Coefficient of Location parameter, d^
(1) variation (3) Modified moment method Maximum likelihood method
d^=x1 (4) d^=x (5) d^=x1 (6) d^=x (7)
Longitudinal tensile strength
VG 1–6 30 0.074 0.984 0.883 0.996 0.894
VG 7–12 30 0.130 0.949 0.760 0.925 0.741
VG 13–18 24 0.102 0.957 0.811 0.983 0.833
VG 19–24 24 0.077 0.913 0.786 0.979 0.843
PG 30 0.069 0.800 0.680 0.878 0.746
Longitudinal compressive strength
VG 1–6 30 0.073 0.985 0.886 0.994 0.894
VG 7–12 30 0.093 0.756 0.605 0.932 0.747
VG 13–18 24 0.114 0.969 0.819 0.995 0.841
VG 19–24 24 0.106 0.948 0.793 0.940 0.786
PG 30 0.122 0.935 0.750 0.938 0.752
PG-M 51 0.092 0.560 0.427 0.819 0.624
Shear strength
VG 19–24 24 0.072 0a 0a 0a 0a
PG 18 0.098 0.950 0.816 0.988 0.848
Transverse compressive strength
PG-M 52 0.060 0.974 0.871 0.982 0.877
Longitudinal tensile modulus
VG 1–6 30 0.068 0.949 0.836 0.938 0.826
VG 7–12 30 0.096 0a 0a 0a 0a
VG 13–18 24 0.116 0.917 0.742 0.962 0.778
VG 19–24 24 0.046 0.974 0.902 0.981 0.908
PG 30 0.063 0.947 0.839 0.947 0.838
Longitudinal compressive modulus
VG 1–6 30 0.054 0.940 0.843 0.973 0.873
VG 7–12 30 0.087 0.852 0.703 0a 0a
VG 13–18 24 0.132 0.940 0.753 0.984 0.788
VG 19–24 23 0.040 0.749 0.682 0a 0a
PG 30 0.040 0.941 0.864 0.962 0.884
Shear modulus
VG 19–24 24 0.113 0.841 0.666 0a 0a
PG 18 0.104 0.895 0.742 0.970 0.804
a
Indicates location parameter was set equal to zero.
M. Alqam et al. / Composite Structures 58 (2002) 497–503 499
sample moments to the corresponding distribution mo- mum likelihood method. This method requires solving
ments. The mean value or the first moment, l, of the the following simultaneous Eqs. [4].
Weibull distribution is: ! !1
X n
^ X
n
^
1 ðxi d^Þ lnðxi d^Þ
b
ðxi d^Þ b
shape parameter is 6–25. In these five cases, the location The modified moment method and the maximum
parameter was taken as zero, or the two-parameter likelihood method were used to determine parameters of
Weibull distribution was used. the three-parameter Weibull distribution for each of the
The maximum likelihood estimator of the location 26 data sets being considered in this paper. The maxi-
parameter is generally just slightly less than the mini- mum likelihood method was also used to determine
mum data point. For the 19 cases with a positive loca- parameters for the two-parameter Weibull distribution
tion parameter, the average values of d^=x1 and d^=x are by setting d ¼ 0 and only using Eqs. (8) and (9). Al-
0.96 and 0.82, respectively. The location parameter from though the parameters for the two-parameter Weibull
the maximum likelihood method is given in Table 1. distribution could be determined using moment meth-
ods, the maximum likelihood method is the generally
used method for the two-parameter Weibull distribution
3. Goodness of fit [12]. An OSL was obtained for each of the distributions
for each data set. The results are shown in Table 2.
Many methods, such as Chi-square, the Kolmogo-
rov–Smirnov, exist for determining the goodness of fit of
a probability distribution to a set of data. The Ander-
son–Darling test was chosen for this study as it is more Table 2
OSL using the Anderson–Darling test for different Weibull distribu-
sensitive to the tail behavior [9], and has been recom- tions
mended for statistical analysis of composites [12]. The
Material OSL
sensitivity to the tail behavior is particularly useful in
property (1) Two-parameter Three-parameter Weibull
structural engineering applications, where the tail is
Weibull (2)
important in computing the structural reliability. The Modified Maximum
Anderson–Darling statistic is obtained as: moment likelihood
method (3) method (4)
Z 1 e 2
½ F n ðxÞ F0 ðxÞ Longitudinal tensile strength
A2n ¼ n dF0 ðxÞ ð11Þ VG 1–6 0.009 0.823 0.817
1 F0 ðxÞ½1 F0 ðxÞ
VG 7–12 0.302 0.045 0.072
in which Fen ðxÞ is a step function that jumps at the order VG 13–18 0.138 0.224 0.146
statistics of x, and F0 ðxÞ is the hypothesized continuous VG 19–24 0.008 0.208 0.420
PG 0.699 0.871 0.860
cumulative distribution function. The Anderson–Dar-
ling statistic is a measure of the square of the error be- Longitudinal compressive strength
tween the data and the hypothesized distribution VG 1–6 0.001 0.536 0.320
VG 7–12 0.212 0.475 0.727
weighted so that the tails of the data are more important VG 13–18 0.088 0.183 0.117
than the central portion. For computation purposes, the VG 19–24 0.469 0.272 0.303
Anderson–Darling statistic can be obtained as: PG 0.214 0.131 0.096
n PG-M 0.122 0.255 0.157
X 1 2i
A2n ¼ ln½F0 ðxðiÞ Þ þ ln 1 F0 ðxðnþ1iÞ Þ n Shear strength
i¼1
n VG 19–24 0.517 0.610a 0.517a
ð12Þ PG 0.073 0.544 0.486
Transverse compressive strength
in which xðiÞ is the ith order statistic of the data set. For PG-M 0.020 0.546 0.424
the Weibull distribution, an observed significance level
Longitudinal tensile modulus
(OSL) is obtained as follows [12]:
VG 1–6 0.179 0.505 0.522
1 VG 7–12 0.022 0.030a 0.022a
OSL ¼ ð13Þ VG 13–18 0.010 0.023 0.012
1 þ exp½0:10 þ 1:24lnAD þ 4:48AD VG 19–24 0.177 0.622 0.508
PG 0.180 0.504 0.414
in which
Longitudinal compressive modulus
0:2 2 VG 1–6 0.153 0.518 0.483
AD ¼ 1 þ pffiffiffi An ð14Þ
n VG 7–12 0.052 0.012 0.052a
VG 13–18 0.010 0.042 0.036
The OSL is the probability of obtaining a value of the VG 19–24 0.299 0.386 0.299a
test statistic at least as large as that obtained from the PG 0.504 0.753 0.678
data if the hypothesis that the data are actually from Shear modulus
the distribution being tested is true. Typically, a 5% VG 19–24 0.065 0.312 0.065a
significance level is used, so that the null hypothesis is PG 0.116 0.598 0.701
a
only rejected if the OSL is less than 0.05. Indicates a location parameter of 0.0.
M. Alqam et al. / Composite Structures 58 (2002) 497–503 501
where R is the resistance, L is a live load, D is a dead Modified moment Maximum likelihood
method (2) method (3)
load, and g < 0 is failure. The load is considered to be
half dead load and half live load. The dead load is as- Longitudinal tensile strength
VG 1–6 1.07 1.07
sumed to follow a normal distribution with a coefficient VG 7–12 1.18 1.18
of variation of 0.10, and the live load is assumed to VG 13–18 1.09 1.09
follow an Extreme Type I distribution with a coefficient VG 19–24 1.08 1.08
of variation of 0.25. These are approximately the sta- PG 1.01 1.01
tistics and load ratio applicable to both building and Longitudinal compressive strength
bridge loads [6,11]. VG 1–6 1.08 1.07
Each data set was used for the resistance, R, and a VG 7–12 1.07 1.08
dead and live load, D and L, were determined using first- VG 13–18 1.15 1.15
VG 19–24 1.12 1.12
order reliability methods [2] such that the reliability PG 1.16 1.16
index would be 3.00. Both the three-parameter Weibull PG-M 1.05 1.06
distribution determined using the modified moment
Shear strength
method and the maximum likelihood method were used. VG 19–24 0.99 1.00
The ratios of the allowable load from the three-pa- PG 1.09 1.09
rameter Weibull distribution to the allowable load from
Transverse compressive strength
the two-parameter Weibull distribution are shown in PG-M 1.03 1.03
Table 4.
The three-parameter Weibull distributions resulted Longitudinal tensile modulus
VG 1–6 1.02 1.02
in allowable loads that averaged 7% greater than al- VG 7–12 0.95 1.00
lowable loads from the two-parameter Weibull distri- VG 13–18 1.10 1.11
bution, with the modified moment method giving on VG 19–24 1.00 1.00
the average just slightly higher allowable loads. Given PG 1.01 1.01
the fact that different chosen probability distributions Longitudinal compressive modulus
can result in allowable loads that vary by as much as VG 1–6 1.00 1.00
50% [7], the 7% variation in allowable load is not very VG 7–12 0.99 1.00
significant. VG 13–18 1.20 1.21
VG 19–24 1.00 1.00
The allowable load was also examined for a live load PG 1.00 1.00
to dead load ratio of four instead of one. This was based
on composite structures being lighter, resulting in higher Shear modulus
VG 19–24 1.16 1.00
live to dead load ratios. For a live load to dead load PG 1.12 1.13
ratio of 4, the allowable load from the three-parameter
Weibull distributions averaged 3% greater than the al-
lowable load using the two-parameter Weibull distri- used. This is because of the relative ease in estimating
bution. The decrease in the ratio of allowable loads was the distribution parameters, the less likelihood in ob-
expected since the resistance statistics become less im- taining a negative location parameter, and the similarity
portant as the uncertainty in the load increases with in the results to those obtained using the maximum
increasing live to dead load ratios. The 3% difference is likelihood method.
virtually insignificant. The two-parameter Weibull distribution could not be
rejected at the 5% level for 19 out of the 26 data sets.
Although the two-parameter Weibull resulted in lower
6. Evaluation of results nominal design values and lower allowable loads than
either of the three-parameter Weibull distributions, the
Not much difference was observed between the differences were well less than ten percent. This differ-
modified moment method and the maximum likelihood ence is insignificant in most structural engineering ap-
method for the three-parameter Weibull distribution. plications.
Both resulted in similar OSL, nominal design values, Even though the three-parameter Weibull distribu-
and allowable loads. If a three-parameter Weibull dis- tion is more tobust than the two-parameter Weibull
tribution is to be used to characterize the data, it is distribution, there are some problems with the three-
recommended that the modified moment method be parameter Weibull distribution. The basic conceptual
M. Alqam et al. / Composite Structures 58 (2002) 497–503 503