Вы находитесь на странице: 1из 17


discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/37628953

Weibull Distributions and Their Applications

Article February 2006

DOI: 10.1007/978-1-84628-288-1_3 Source: OAI


6 2,551

3 authors, including:

Chin-Diew Lai
Massey University


Some of the authors of this publication are also working on these related projects:

C D Lai View project

All content following this page was uploaded by Chin-Diew Lai on 11 August 2017.

The user has requested enhancement of the downloaded file.

applications Weibull probability plot
density Weibull distribution
distribution function Weibull derived
empirical modeling WPP Weibull probability plot
failure rate
Index entries on this page

hazard plot
hypothesis testing

Weibull Distrib
2. Weibull Distributions and Their Applications

2.1.2 Relations to Other Distributions .. 2

Weibull models are used to describe various

Part A 2
types of observed failures of components 2.2 Properties ........................................... 2
and phenomena. They are widely used in 2.2.1 Basic Properties ........................ 2
reliability and survival analysis. In addition 2.2.2 Properties Related
to the traditional two-parameter and three- to Reliability............................. 3
parameter Weibull distributions in the reliability 2.2.3 Simulation ............................... 4
or statistics literature, many other Weibull-related 2.3 Modeling Failure Data.......................... 5
distributions are available. The purpose of this 2.3.1 Probability Plots........................ 5
chapter is to give a brief introduction to those 2.3.2 Estimation
models, with the emphasis on models that and Hypothesis Testing .............. 6
have the potential for further applications. After 2.3.3 Hypothesis Testing .................... 7
introducing the traditional Weibull distribution,
2.4 Weibull-Derived Models ....................... 8
some historical development and basic properties
2.4.1 Taxonomy for Weibull Models ..... 8
are presented. We also discuss estimation problems
2.4.2 Univariate Models ..................... 8
and hypothesis-testing issues, with the emphasis
2.4.3 Type VI Models
on graphical methods. Many extensions and (Stochastic Point Process Models) 11
generalizations of the basic Weibull distributions
are then summarized. Various applications in 2.5 Empirical Modeling of Data .................. 11
the reliability context and some Weibull analysis 2.6 Applications ........................................ 12
software are also provided. 2.6.1 Applications in Reliability........... 12
2.6.2 Applications in Other Areas......... 13
2.6.3 Weibull Analysis Software........... 13
2.1 Three-Parameter Weibull Distribution ... 2
2.1.1 Historical Development .............. 2 References .................................................. 14

The Weibull distribution is one of the best-known life- We consider estimation problems and hypothesis testing
time distributions. It adequately describes observed in Sect. 2.3. In particular, we emphasize the graphical
failures of many different types of components and phe- methods as a tool for selection and parameter estimation
nomena. Over the last three decades, numerous articles of a Weibull model. We devote Sect. 2.4 to Weibull-
have been written on this distribution. Hallinan [2.1] derived models, which includes many extensions and
gives an insightful review by presenting a number of his- generalizations. Finally, in Sect. 2.5, we outline various
torical facts, the many forms of this distribution as used applications, especially those in the reliability context.
by practitioners, and possible confusions and errors that Because of the vast literature, we are unable to refer to
arise due to this non-uniqueness. Johnson et al. [2.2] all the source authors and we apologize in advance for
devote a comprehensive chapter to a systematic study any omissions in this regard.
of this distribution. More recently, Murthy et al. [2.3]
presented a monograph that contains nearly every facet Symbols and Abbreviations.
relating to the Weibull distribution and its extensions.
T Random variable
In Sect. 2.1, we first define the three-parameter
F(t) CDF, cumulative distribution function
4 November 2005 10:08 CET

Weibull distribution and then look at its historical devel-

f(t) PDF, probability density function
opment and relations to other distributions. Section 2.2
h(t) Failure rate function (hazard rate)
studies the properties of the Weibull distribution, in
, , Parameters
particular those relevant to reliability. A brief discus-
Springer Handbook

WPP Weibull probability plot

sion on simulation of Weibull variates is also included.
MS ID: hb06-a2
Index entries on this page

2 Part A Fundamental Statistics and its Applications

2.1 Three-Parameter Weibull Distribution

According to Hallinan [2.1] the Weibull distribution has 2.1.1 Historical Development
appeared in five different forms. The two common forms
of the distribution function are as follows: The Weibull distribution is named after its originator, the
t Swedish physicist Waloddi Weibull, who in 1939 used
F(t) = 1 exp , t (2.1) it to model the distribution of the breaking strength of

Part A 2.2

materials [2.4] and in 1951 for a wide range of other ap-

and plications [2.5]. The distribution has been widely studied
  since its inception. It has been known that Weibull may
F(t) = 1 exp (t ) , t . (2.2)
not be the first to propose this distribution. The name
The parameters of the distribution are given by the set Frchet distribution is also sometimes used due to the
= {, , } with > 0, > 0 and 0; where is fact that it was Frchet [2.6] who first identified this dis-
a scale parameter, is the shape parameter that deter- tribution to be an extremal distribution (later shown to be
mines the appearance or shape of the distribution and one of the three possible solutions by Fisher and Tippett
is the location parameter. The parameter combines [2.7]). According to Hallinan [2.4], it was Weibull who
both scale and shape features as = . suggested a scale parameter and a location parameter
Although one should use F(t, ) instead of F(t), that made the distribution meaningful and useful.
where = (, , ) denotes the vector of parameters,
for notational convenience we suppress the parameter 2.1.2 Relations to Other Distributions
and use F(t) to denote F(t, ) in this chapter. Also, we
do not intend to give an exhaustive review. Rather, we The Weibull distribution includes the exponential
confine our discussion to aspects that are of relevance to ( = 1) and the Rayleigh distribution ( = 2) as
the context of reliability theory. special cases. If X denotes the Weibull variable,
For = 0, (2.1) and (2.2) become the two-parameter then X has a type 3 extreme-value distribu-
Weibull distribution with tion ([2.8], Chapt. 06-c22 TS0 ). A simple log trans-
    formation will transform the Weibull distribution
F(t) = 1 exp ,t 0 (2.3) into the Gumbel distribution (type 1 extreme-value
and The Burr XII distribution, is given by
F(t) = 1 exp t , t . (2.4) t
F(t) = 1 1 + , t 0 ; k , > 0 . (2.5)
Murthy et al. [2.3] refer to this as the standard Weibull
model, Johnson et al. [2.2] refer to a standard Weibull Let k = ; as k , then the Burr distribution ap-
when = 1 (or = 1) in (2.3, 4). proaches the Weibull (see, for example, [2.9]).

2.2 Properties
2.2.1 Basic Properties Mode
It follows from (2.6) that the mode is at t =
Density Function (( 1)/)1/ + for > 1 and at for 0 < 1.
The probability density functions (PDF) (Fig. 2.1) of
(2.1) and (2.2) are Median
    It follows from (2.3) that the median of the distribution
1 t is at (log 2)1/ + .
f(t) = (t ) exp , t

4 November 2005 10:08 CET

(2.6) Moments
Let T denote the random variable from the three-
  parameter Weibull distribution given by (2.1). Then
f(t) = (t )1 exp (t ) , t . (2.7)
Springer Handbook

the transformed variable T  = (T )/, is the stan-

MS ID: hb06-a2

0 Please provide title of referenced chapter

Editors or typesetters annotations (will be removed before the final TeX run)
Index entries on this page

Weibull Distributions and Their Applications 2.2 Properties 3

dard form (in the sense of Johnson et al. [2.2]) with the a sample of n observations. The probability density
density function given by function of T(1) , is given by
f(t) = t 1 exp t , x > 0 , > 0 . (2.8) f (1) (t) = n[1 F(t)]n1 f(x)
The moments of T (about zero) are easily obtained from n t 1 n[(t)/]
= e ,t .
the moments of T  which are given below:

Part A 2.2
r = E(T  ) =
+1 , (2.9)
It is obvious from (2.12) that T(1) is also distributed as
a Weibull random variable, except that is replaced by
from which we get n 1/ .
  The density function of T(r) (1 r n) is given by
E(T  ) = +1 , (2.10)
n! r1
f (r) (t) = 1 e[(t)/]
(r 1)!(n r)!
2 (nr+1)
2 1 e[(t)/] (t )1 , t 0 .
Var(T ) = +1 +1 . (2.11)

It can be shown that

Skewness and Kurtosis
The distribution is positively
skewed for small values  k
of . The skewness index 1 decreases and equals E (T(r) )k = i ki ki
(r) , (2.14)
zero for = 3.6 (approximately). Thus, for values of i=0
in the vicinity of 3.6, the Weibull distribution is simi-
lar in shape to a normal distribution. The coefficient of
kurtosis 2 also decreases with and then increases,   r1
n! k 
2 has a minimum value of about 2.71 when = 3.35 k(r) = 1+
(r 1)!(n r)!
(approximately). i=0
(1)r r1 i
Order Statistics .
(n r + i + 1)1+(k/)
Let T1 , T2 , , Tn denote n independent and identically 1
distributed three-parameter Weibull random variables
with density function given in (2.6) and cumula-
tive distribution function (CDF) in (2.1). Further, let 2.2.2 Properties Related to Reliability
T(1) T(2) T(n) denote the order statistics from
In this section, we consider only the first form of the
Weibull distribution.
Density function The survival function of the Weibull distribution is
1.6 t
F(t) = 1 F(t) = exp , t .
= 0.5, = 0.9
1.2 (2.15)
0.8 Note that at t = + , F( + ) = 1 e1 0.3679.
Failure Rate
0.4 = 3, = 2 = 2, = 0.5 The failure rate function (also known as the hazard rate)
4 November 2005 10:08 CET

for the three-parameter Weibull is
0 1 2 3 4 5 t f(t) t 1
h(t) = = . (2.16)
Springer Handbook

Fig. 2.1 Two-parameter Weibull density plots F(t)

MS ID: hb06-a2

1 Please confirm the position of the times symbol in this equation (2.14)

Editors or typesetters annotations (will be removed before the final TeX run)
mean residual life (MRL)
Index entries on this page

4 Part A Fundamental Statistics and its Applications

For the two-parameter case, it is given by

  Failure rate function
f(t) t 1 4
h(t) = = . (2.17)
It is obvious that h(t) is a decreasing function when
< 1, constant when = 1 (the exponential case), and
an increasing function when > 1. Because of the 2
= 0.5, = 0.9
Part A 2.2

behaviour of the failure rate function, the Weibull dis-

tribution often becomes suitable when the conditions 1 = 3, = 2
for strict randomness of the exponential distribution
= 2, = 0.5
are not satisfied, with the shape parameter having
a value depending upon the fundamental nature being
0 1 2 3 4 5 t
In some way, having a failure rate function of mono- Fig. 2.2 Two-parameter Weibull failure rate functions
tonic shape has limitations in reliability applications.
For this reason, several generalized or modified Weibull Suppose E(X) = E(Y), i. e., 2 (1 + 1/2 ) =
distributions have been proposed (see Sect. 2.4 for more 1 (1 + 1/1 ). Lai and Xie [2.10] show that
details). Figure 2.2 shows plots of the failure rate func- Var(X) Var(Y).
tions for some selected parameters values.
Effectiveness of Parallel Redundancy
Mean Residual Life The Weibull Case
The mean residual life (MRL) of a lifetime random Denote the lifetime of one component by T and
variable T is defined as the lifetime of a parallel system of two such in-
 dependent components by Tp . The effectiveness of
F(x) dx
(t) = E(T t|T > t) = t . (2.18) parallel redundancy of the component is defined
F(t) by [2.11]
For the Weibull distribution, the MRL is complicated
except for the two special cases = 1 (exponential) and ep = [E(Tp ) E(T )]/E(T ) . (2.22)
= 2 (Rayleigh distribution). Assuming the location
parameter = 0, the MRLof the Rayleigh distribution Suppose that T is a two-parameter Weibull distribution
with scale parameter = 2 is with shape parameter and scale parameter . Xie and
1 2
t Lai [2.11] show that
(t) = 2 [1 (t/)] e 2 2 , t > 0 , (2.19)
where (t/) denotes the distribution function of the ep = 1 21/ , (2.23)
standard normal variable.
which decreases as increases. Thus, a parallel redun-
Relative Ageing of Two Two-Parameter Weibull dancy is more effective for < 1.
Suppose we have two Weibull random variables X and Y 2.2.3 Simulation
with distribution functions F(x) and G(y), respectively,
given by The two-parameter Weibull distribution with parameters
F(x) = 1 exp (x/2 )2 , and has the probability density function
G(y) = 1 exp (y/1 )1 . (2.20)
f(x) = x 1 ex . (2.24)
We say that X ages faster than Y if the ratio of the failure
rate of X over the failure rate of Y is an increasing The simple inverse-probability integral-transform
function of t. This ratio is given by
4 November 2005 10:08 CET

method applied to the standard Weibull distribu-

1 2 1 2
1 tion ( = 1) is quite efficient. The formula is
t 2 1 (2.21) simply
2 1 1 1
Springer Handbook

which is an increasing function of t if 2 > 1 . x = ( log u)1/ ,

MS ID: hb06-a2
Index entries on this page

Weibull Distributions and Their Applications 2.3 Modeling Failure Data 5

where u is a uniform number between 0 and 1. Of course, function. The standard Weibull variate is then scaled by
an acceptance/rejection method could also be used to (1/) to obtain a two-parameter Weibull observation
avoid the evaluation of the logarithm in the distribution with density (2.24).

2.3 Modeling Failure Data

Part A 2.3
Failure data can be classified into two types: complete bility plots to be discussed below. Moreover, they can
and incomplete (censored). For complete data, the actual provide crude estimates of model parameters. There are
realized values are known for each observation in the many different statistical tests for validating a model and
data set, whereas for censored data, the actual realized we postpone this discussion until the next section.
values are not known for some or all of the observations.
Further, there are at least two types of censoring, details 2.3.1 Probability Plots
of which may be found Lawless [2.12] and Nelson [2.13].
In their introduction, Murthy et al. [2.3] point out that Weibull Probability Plot
the two different approaches to building mathematical Weibull probability plots (WPP) can be constructed in
models are as follows. several ways [2.15]. In the early 1970s a special paper
was developed for plotting the data in the form of F(t)
1. Theory-based modeling:here, the modeling is based
versus t on a graph paper with a loglog scale on the ver-
on the established theories for component failures.
tical axis and a log scale on the horizontal axis. A WPP
This kind of model is also called a physics based
plotting of data involves computing the empirical dis-
model or white-box model.
tribution function, which can be estimated in different
2. Empirical modeling: here, the data available forms
ways; the two standard methods are:
the basis for the model building. This kind of model
is also called a data-dependent model or black-box F(ti ) = i/(n + 1), the mean rank estimator, and
model. F(ti ) = (i 0.5)/n, the median rank estimator.
In empirical modeling, the type of mathematical Here, the data consists of successive failure times ti ,
formulations needed for modeling is dictated by a pre- t1 < t2 < . . . tn . For censored data (right-censored or
liminary analysis of data available. If the analysis interval), the approach to obtain the empirical distri-
indicates that there is a high degree of variability, then bution functions needs to be modified; see, for example,
one needs to use models that can capture this variabil- Nelson [2.13].
ity. This requires probabilistic and stochastic models to These days, most computer reliability software
model a given data set. packages contain programs to produce these plots au-
The process of black box modeling involves the tomatically given a data set. A well-known statistical
following steps: package Minitab provides a Weibull probability plot
under the Graph menu  Probability Plot.
Step 1 Model selection,
We may use an ordinary graph paper or spreadsheet
Step 2 Parameter estimation,
software with unit scale for plotting. Taking logarithms
Step 3 Model validation.
twice of both sides of each of the CDF in (2.3) yields
To select a model out of many possible models, one
requires a good understanding of their different prop- log[ log F(t)] = log(t ) log . (2.25)
erties. This often is a trial-and-error process. Liao and 
Shimokawa [2.14] consider goodness-of-fit testing as Let y = log log F(t) and x = log(t ). Then we
a key procedure for selecting the statistical distribution have
that best fits the observed data. To execute the remaining y = x log . (2.26)
two steps, we need various tools and techniques.
4 November 2005 10:08 CET

A large number of Weibull-related models have been The plot is now on a linear scale.
derived in the literature. In Sect. 2.4, various such mod-
els are grouped into several categories. Selecting an Weibull Hazard Plot
appropriate model from the family of Weibull-related The hazard plot is analogous to the probability plot,
Springer Handbook

distributions can often be based on some of the proba- the principal difference being that the observations are
MS ID: hb06-a2
Index entries on this page

6 Part A Fundamental Statistics and its Applications

plotted against the cumulated hazard (failure) rate rather suggested to set the initial value of = T(1) and then
than the cumulated probability value. Moreover, this is adjust this estimate to achieve a straight line. We also
designed for censored data. note that it is common for to be 0.
Let H(t) denote the cumulative hazard rate, then
F(t) = exp[H(t)], so Estimation of the Shape Parameter. It follows from
  (2.26) that the shape parameter can be estimated from the
t slope of the WPP. However, care needs need to be given
H(t) = log F(t) = , (2.27)
Part A 2.3

because possible confusion may arise due to several

H(t)1/ / = (t ) . (2.28) possible forms of the Weibull distribution.

Let y = log(t ) and x = log H(t), then we have Estimation of the Scale Parameter. To estimate the
1 scale parameter, we first observe from the standard
y = log + x . (2.29) Weibull that, when t is at the 63.2 percentile, F(t) = 1

e1 , i. e., F(t) = e1 . Or equivalently, for t = + ,
Rank the n survival times (including the censoreddata) F( + ) = e1 (so + is the 63.2 percentile). It
in ascending order and let K denote the reverse ranking now follows from (2.26) that log(t ) = log so that
order of the survival time, i. e., K = n for the smallest = t , where t is the 63.2 percentile. Hence the scale
survival time and K = 1 for the largest survival time. parameter is simply estimated from the x-intercept of
The hazard is estimated from 100/K (a missing value the plot (2.26): = e(x-intercept) .
symbol is entered at a censored failure time). The cu- Under the probability plot option, Minitab not only
mulative hazard is obtained by cumulating the hazards. gives a WPP, but also provides an estimate for the scale
The Weibull hazard plot is simply the plot arising from and shape parameters and .
(2.29). See Nelson [2.16] for further details. Figure 2.3 illustrates how a WPP plot appears in
a Minitab graph with = 9.062 and = 1.661.
2.3.2 Estimation and Hypothesis Testing
Statistical Methods of Estimation
Parameter estimation is the second step of our model- There are several statistical methods for estimating the
ing process. We consider both graphical and statistical model parameters. These include the method of moment,
methods for the three-parameter Weibull distribution. the method of percentile, the method of maximum like-
lihood, and the Bayesian method. The estimates can be
Graphical Methods
Graphical plots are designed not just to assess if the
Probability (%)
data follows a Weibull population, they are also useful 99
for estimating Weibull parameters (at least for initial Shape 1.661
estimates). These can be obtained from the smooth fit 90 Scale 9.062
80 N 25
to a WPP plot of data and involve exploiting properties AD 0.300
such as asymptotes, intersection points, slope, or points 60 P-Value > 0.250
of inflection. 50
Estimation of Location Parameter. The determination
of a suitable location parameter is not a simple task. 20
If a data set graphs as a straight line on a WPP, then the
data set is indeed adequately described by a Weibull dis- 10
tribution. However, an incorrect selection of will yield
a curved plot of data that is in fact Weibull-distributed.
If the graph of a set of data appears concave upward, 3
4 November 2005 10:08 CET

then the plot can be straightened by decreasing the value 2

of . Conversely, if the data are concave downward, then
the plot can be straightened by increasing the value of . 0.1 1.0 10.0 100.0
The value of cannot be chosen larger than the smallest Data set 1
Springer Handbook

failure-time value, say T(1) , in the sample. It is generally Fig. 2.3 A typical Weibull probability plot CE2
MS ID: hb06-a2

2 Please define abbreviation AD used in this figure

Editors or typesetters annotations (will be removed before the final TeX run)
Index entries on this page

Weibull Distributions and Their Applications 2.3 Modeling Failure Data 7

either point estimates or interval estimates. The details of isfy the following equations
these methods can be found, for example, in Kalbfleisch  1/
and Prentice [2.17]. n

We note that these methods may not be appropriate = (X i ) (2.32)
for small data sets. Hence, the estimates obtained from i=1
the WPP plot may be taken as the final estimates when 

the data size is small. = (X i ) log(X i )

Part A 2.3
 n 1 1
Moment Estimation. By equating the first three mo-  1

ments of the Weibull distribution to the first three sample (X ) log(X i )
moments and solving, it is possible to find the moment i=1 i=1
of , and . Now, the first moment ratio (2.33)
1 depends
only on the shape parameter , and hence
once 1 has been estimated from the sample coef-
ficient of skewness, can be determined numerically. 

Using , can be determined from the standard devia- ( 1) (X i )1 = (X i )1 .

tion, and lastly is determined from the sample mean. i=1 i=1
If is known, then can be estimated from the ratio (2.34)
(standard deviation/(mean-). If the value satisfying (2.32)(2.34) is larger than X 1 =
For modified moment estimation, see Cohen min X i , then it is the maximum-likelihood estimate of .
et al. [2.18]. Otherwise, we set that = X 1 and then (2.32) and (2.33)
must be solved for and .
Maximum-Likelihood Method. Maximum-likelihood The method is suitable for large data sets because
methods for the three-parameter Weibull distribution of its asymptotic properties. Modifications of these es-
have been reviewed by Zanakis and Kyparisis [2.19]; timates can be found, for example, in Johnson et al.
see also Johnson et al. [2.2, Chapt. 21] for other details. [2.2].
The most common situation is when the location The reader may consult Chapt. 21 of Johnson et al.
parameter is known. Without loss of generality, we [2.2] and Chapt. 4 of Murthy et al. [2.3] for the following
may assume that = 0 so that the model becomes the methods of estimation,
two-parameter Weibull distribution.
The maximum-likelihood estimators, and of Best linear unbiased estimation,
and , respectively, satisfy the equations Percentile estimator,
 n 1/ Bayesian estimator,
1  Interval estimation,
= Xi (2.30) Minimum quantile distance estimation.
and For the two-parameter Weibull distribution (i. e., = 0),
  n 1 methods of inference can be obtained via the type 1

 extreme-value distribution.
= X log X i
i X i
i=1 i=1 2.3.3 Hypothesis Testing
Goodness-of-Fit Tests for the Weibull
log X i , (2.31)
There are other general goodness-of-fit tests based on the
where X i , i = 1, 2, . . . n are n independent observations empirical distribution functions derived. For example,
from the two-parameter Weibull distribution. If = 0, the chi-square test, KolmogorovSmirnov test, Cramer
then each X i is replaced by X i in the above equa-
4 November 2005 10:08 CET

von Mises test and AndersonDarling test, and so on,

tions. The value needs to be solved from (2.31) and could also be used to test the goodness of fit of the
then (2.30) is used to obtain . Weibull model.
Suppose the location parameter is also unknown, Generally, a goodness-of-fit test for Weibull distri-
Springer Handbook

then the maximum-likelihood estimates , and sat- bution can be described as:H0 : the population follows
MS ID: hb06-a2
Index entries on this page

8 Part A Fundamental Statistics and its Applications

a Weibull model, versus H1 : the Weibull model is not goodness-of-fit test for the two-parameter Weibull and
suitable. its power is compared with other traditional goodness-
Minitab under probability plot gives the Anderson of-fit tests.
Darling statistic as well as the confidence bands.
Lawless [2.12] summarizes the goodness-of-fit tests Hypothesis Testing
for the Weibull or extreme-value distribution. Four tests Hypothesis testing of Weibull parameters may be per-
(the likelihood ratio test as a sub-model of the gamma formed by the likelihood ratio, score and Wald tests.
Part A 2.4

distribution, MannScheuerFertig test, Tiku test and It is well known that all these tests are asymptotically
Cramervon Mises test) are discussed in detail. Dodson optimal.
[2.20] also includes and compares some goodness-of- For testing Weibull versus the exponential distribu-
fit tests for Weibull distributions in Chapt. 4 of his tion, we test = 1 (or = 1 of the second form) versus
book. Liao and Shimokawa [2.14] construct a new = 0.

2.4 Weibull-Derived Models

There are many extensions, generalizations and modi- rived Weibull model. T is a random variable from F(t)
fications to the Weibull distribution. They arise out of and Z is a random variable from G(t).
the need to model features of empirical data sets that
cannot be adequately described by a three-parameter Type I Models
Weibull model: for example, the monotonic property of (Transformation of Weibull Variable)
the Weibull, which is unable to capture the behaviour Here Z and T are related by a transformation. The
of a data set that has a bathtub-shape failure rate. Xie transformation can be either linear or nonlinear. These
et al. [2.21] review several Weibull-related distributions include the following:
that exhibit bathtub-shaped failure rates. Plots of mean
residual life from several of these Weibull-derived mod- Reflected Weibull Distribution.
els are given in Lai et al. [2.22]. For simplicity, we    
simply refer to these Weibull-related models as Weibull G(t) = exp , < t < . (2.35)

This is also known as a type 3 extreme-value distribution

2.4.1 Taxonomy for Weibull Models
(see Johnson et al. [2.8], Chapt. 22).
According to Murthy et al. [2.3], the taxonomy for
Double Weibull Distribution.
Weibull models involves seven types, each of which
divided into several sub-types. These models can be g(t) = (1/2) |t|(1) exp( |t| ) , < t < .
grouped into three groups: (2.36)
1. Univariate models (types IV), This is an obvious extension to include the negative real
2. Multivariate models (type VI), line as its support.
3. Stochastic process models (type VII).
In this chapter we confine our discussion to Log Weibull Distribution. The distribution is de-
univariate and stochastic process models, while rived from the logarithmic transformation of the
Murthy et al. [2.23] looks at bivariate models (a special two-parameter Weibull distribution. The distribution
case of multivariate models). In this chapter, only some function is

selective models will be included. We refer our readers t

to the book by Murthy et al. [2.3] for more details. G(t) = 1 exp exp , < t < .

4 November 2005 10:08 CET

2.4.2 Univariate Models
This is also known as a type 1 extreme-value distribution
The starting point is the two-parameter Weibull model or Gumbel distribution. In fact, it is the most commonly
Springer Handbook

with distribution function F(t). Let G(t) denote the de- referred to in discussions of extreme-value distributions
MS ID: hb06-a2
Index entries on this page

Weibull Distributions and Their Applications 2.4 Weibull-Derived Models 9

(see [2.8], Chapt. 22). The density function is The density function is

1 t t 1 (t/)  1
g(t) = exp exp exp , g(t) = (t/)

t e 1 e
and the failure rate function is quite simple and given by
  For an appropriate choice of the parameter set, it will
1 t give rise to a bathtub-shaped failure function. Jiang and
h(t) = exp .

Part A 2.4
Murthy [2.27] use a graphical approach to study this
distribution. The failure rate function is given by
Inverse (or Reverse) Weibull Model. Let Z = 2 /T ,
where T has a two-parameter Weibull distribution, then 1 (t/) 1 e(t/)
the distribution function of T is h(t) = t e  .

1 (1 e(t/) )
G(t) = exp[(t/) ] , t 0 . (2.38)

This is also known as a type 2 extreme-value distribution

Four-Parameter Weibull Distribution [2.28].
or Frchet distribution (see [2.8], Chapt. 22). The failure
rate function is as follows:    
t a
t 1 e(/t) G(t) = 1 exp ,
h(t) = . bt
1 e (/t)
0at b<. (2.43)

Type II Models (Modification/Generalization

of the Weibull Distribution) Doubly Truncated Weibull Distribution.
Here G(t) is related to F(t) by some relationship.
F(t) F(a)
G(t) = ,
Extended Weibull Distribution [2.24]. F(b) F(a)

0<at b<. (2.44)
G(t) = 1 . (2.39)
1 (1 ) e (t)

The failure rate function is Modified Weibull Extension [2.29].

h(t) =  
1 (1 ) e(t) G(t) = 1 exp e(t/) 1 ,
The failure rate function is increasing if 1, 1 and t 0, , , > 0 . (2.45)
decreasing if 1, 1.
This is known as the generalized exponential power
Modified Weibull Distribution [2.25]. model originally studied by Smith and Bain [2.30]. The
case of = 1 is also studied in Chen [2.31]. The failure
G(t) = 1 exp(t et ) , t 0 (2.40)
rate function is
h(t) = (t/)1 exp (t/) . (2.46)
h(t) = ( + t)t exp(t) . (2.41)

For = 0, this reduces to a Weibull distribution. For It approaches to a two-parameter Weibull distribution
0 < < 1, h(t) is initially decreasing and then increasing when with in such a manner that 1 / is
in t, implying that the failure rate function has a bathtub held constant.
shape. When > 1, h(t) is increasing in t.
4 November 2005 10:08 CET

Type III Models (Models Involving Two or More

Exponentiated Weibull Distribution [2.26].
These are univariate models derived from two or more
G(t) = [F(t)] = {1 exp[(t/) ]} , t 0 . distributions with at least one being either the standard
Springer Handbook

(2.42) Weibull model or a distribution derived from it.

MS ID: hb06-a2
Index entries on this page

10 Part A Fundamental Statistics and its Applications

n-Fold Mixture Model. Arrhenius Weibull Model.

G(t) = pi Fi (t) , pi 0 , pi = 1 . (2.47)
(S) = exp(0 + 1 S) . (2.51)
i=1 i=1
Jiang and Murthy [2.32, 33] categorize the possible Power Weibull Model.
shapes of the failure rate function for a mixture of any e0
two Weibull distributions in terms of five parameters. (S) = . (2.52)
Part A 2.4

Gurland and Sethurama [2.34] also consider a mix-

ture of the Weibull distribution with failure rate t 1 These types of models have been used extensively in
and the exponential distribution with failure rate 1 . For accelerated life testing [2.39] in reliability theory. As
> 1, the Weibull distribution is IFR CE3 . They found a result, they are referred to as accelerated failure
that the resulting mixture distribution is ultimately DFR. models.
In fact, the mixture has a failure rate with an upside-down
Weibull Proportional Hazard Models.
bathtub shape.
h(t) = (S)h 0 (t) , (2.53)
n-Fold Competing Risk Model.
where h 0 (t) is called the baseline hazard for a two-

G(t) = 1 [1 Fi (t)] . (2.48) parameter Weibull distribution. The only restriction on
the scalar function (S) is that it be positive. Many
different forms for (S) have been proposed. One such
Jiang and Murthy [2.35] also give a parametric study of is the following:
a competing risk model involving two Weibull distribu-  
tions.  k
(S) = exp b0 + bi si . (2.54)
n-Fold Multiplicative Model (Complementary Risk i=1
Model). For more on such models, see Cox and Oakes [2.39] and

n Kalbfleisch and Prentice [2.17].
G(t) = Fi (t) . (2.49)
i=1 Type V Models (Discrete Weibull Models)
The multiplicative model involving two Weibull distri- Here T can only assume non-negative integer values and
bution is considered in Jiang and Murthy [2.36]. this defines the support for F(t).

n-Fold Sectional Model. Model 1 [2.40].

1 q t
k1 F1 (t) ,

0 t t1 ,
F(t) =
t = 0, 1, 2, 3 ,

1 k F (t) , t < t t , (2.55)
2 2 1 2
0 t<0.
G(t) = (2.50)

. . .

1 kn Fn (t) , t > tn1 , Model 2 [2.41]. The cumulative hazard function is given
where the sub-populations Fi (t) are two- or three-
parameter Weibull distributions and the ti s (called ct 1 t = 1, 2, , m
partition points) are an increasing sequence. Jiang and H(t) = (2.56)
Murthy [2.37] and Jiang et al. [2.38] consider the case 0 t<0,
n = 2 in details.
where m is given by
Type IV Models
int c[1/(1)]  if > 1 ,
4 November 2005 10:08 CET

(Weibull Models with Varying Parameters)

m= (2.57)
For models belonging to this group, the parameters of the if 1 ,
model are either functions of the independent variable (t)
or some other variables (such as the stress level s, etc.), and int() represents the integer part of the quantity
Springer Handbook

or are random variables. inside the brackets.

MS ID: hb06-a2

3 Please ensure that the abbreviations IFR and DFR are defined on first use in the text

Editors or typesetters annotations (will be removed before the final TeX run)
Index entries on this page

Weibull Distributions and Their Applications 2.5 Empirical Modeling of Data 11

Model 3 [2.42]. Weibull process, see for example, Bain and Engelhart
  [2.44, Chapt. 9].

F(t) = 1 exp r(i)
Proportional Intensity Model
  (Cox and Oaks [2.45])
The intensity function is given by
= 1 exp ci 1 , t = 0, 1, 2, .
i=1 (t; S) = 0 (t)(S) , t 0 , (2.60)

Part A 2.5
where 0 (t) is of the form given above and (S) is
2.4.3 Type VI Models a function of the explanatory variables S. The only
(Stochastic Point Process Models) restriction on (S) is that (S) > 0.

These are stochastic point process models with links to Ordinary Renewal Process (Yannaros [2.46])
the standard Weibull model. Here the point process is a renewal process with the
time between events being independent and identically
Power Law Process (Bassin [2.43]) distributed with the distribution function given by the
This is a point process model with the intensity function standard Weibull distribution.
given by
 1  Modified Renewal Process
(t) = . (2.59) Here the distribution function for the time to first
event, F0 (t), is different from that for the subsequent
This model has been called by many different names: inter-event times, which are identical and independent
power law process; RaschWeibull process; Weibull in- random variables with distribution function F(). Note
tensity function); WeibullPoisson process and Weibull that F0 (t) and/or F(t) are standard Weibull distributions.
process. We note that the inter-event times do not have When F0 (t) = F(t), this reduces to the ordinary renewal
Weibull distributions. For further discussion on the process.

2.5 Empirical Modeling of Data

We have seen a large number of Weibull-related models
Probability (%)
which we simply refer as Weibull models. They ex- 99.9
hibit a wide range of shapes for the density and failure 99 Shape 1.015
rate functions, which make them suitable for modeling Scale 30.33
90 N 50
complex failure-data sets. AD 0.891
Recall, we mention that empirical modeling usu- 70 P-Value 0.021
ally involves three steps: model selection, estimation of 60
model parameters and model validation. In the context 40
of Weibull models, a selection procedure may be based 30
on WPP plots. This is possible due to the availability of 20
WPP or generalized WPP plots for all the Weibull mod-
els of types IIII. Of course, the shape of the density and 10
failure rate functions will also be valuable in the selec-
tion step. An added advantage of the WPP plots is that 5
they provide crude estimates of model parameters; these 3
4 November 2005 10:08 CET

serve as a starting point for steps 2 and 3. 2

It has been suggested that an alternative method to
estimate model parameters is through a least-squares 0.1 1.0 10.0 100.0 1000.0
fit. Basically speaking, this involves selecting the pa- Data set 2
Springer Handbook

rameters the parameters to minimize a function given Fig. 2.4 Probability plot of data set from Table 2.1
MS ID: hb06-a2
Index entries on this page

12 Part A Fundamental Statistics and its Applications

Table 2.1 Data set of failure test (data set 2)

0.12 0.43 0.92 1.14 1.24 1.61 1.93 2.38 4.51 5.09
6.79 7.64 8.45 11.90 11.94 13.01 13.25 14.32 17.47 18.10
18.66 19.23 24.39 25.01 26.41 26.80 27.75 29.69 29.84 31.65
32.64 35.00 40.70 42.34 43.05 43.40 44.36 45.40 48.14 49.10
49.44 51.17 58.62 60.29 72.13 72.22 72.25 72.29 85.20 89.52
Part A 2.6

by tests for validating a model. These generally require

n data that is different from the data used for model selec-
J() = [y(ti ; ) yi ]2 , (2.61) tion and parameter estimation. A smaller data set may
i=1 pose a problem, as there will be no separate data left af-
where y(ti ; ) uses the vector parameter and yi is ter model selection and parameter estimation. Various
the corresponding value obtained from the data. The solutions have been proposed in the case of a small data
optimization can be carried out using any standard op- set. We refer the readers to the book by Meeker and
timization package. The least-squares method not only Escobar [2.48] for further details.
furnishes us with parameter estimates, but also helps to Example 1:
select a Weibull model. If one of the potential candi- 50 items are tested to failure. The failure times are auto-
dates has a value for J() which is considerably smaller matically recorded and given in Table 2.1. The Weibull
than that for the other models, then this can undoubtedly probability plot indicates that the two-parameter Weibull
be accepted as the most appropriate model for model- model is not appropriate.
ing the given data set. If two or more Weibull models Based on the plot, we observe that the plot shows
give rise to roughly the same value of J(), one would a decreasing slope at the beginning, and an increasing
need to examine additional properties of the WPP plots slope at the end. This is an indication of a distribution
to decide on the final model. Other approaches such as with a bathtub-shaped failure-rate function. Some of the
bootstrap and jackknife may be employed for the final models of type II or type III can be used. One possibility
selection. For more on this, see Murthy et al. [2.3, 47]. is to fit the early part and last part of the plot with separate
A couple of comments on step 3 of our empirical lines (Fig. 2.4). This will result in a two-fold competing
modeling may be in order. There are many statistical risk model (2.48).

2.6 Applications
2.6.1 Applications in Reliability carried out during the development phase. Here the pro-
totype is tested until a failure occurs and the causes of the
Product reliability depends on the design, development failure are analyzed. Based on this, design changes are
and manufacturing decisions made prior to the launch made to overcome the identified failure causes. This pro-
(pre-launch stage) of the product and it in turn affects cess is continued until the reliability target is achieved.
the failures when the product is put into operation after The reliability of the items produced during manufac-
launch (post-launch stage). turing tends to vary from the design target value due
The pre-launch stage involves several phases. In the to variations resulting from the manufacturing process.
feasibility phase, study is carried out using the specified Through proper process and quality control during the
target value for product reliability. During the design manufacturing phase, these variations are controlled.
phase, product reliability is assessed in terms of part and In the post-launch stage, the reliability of an item
component reliabilities. Product reliability increases as decreases due to deterioration resulting from age and/or
4 November 2005 10:08 CET

the design is improved. However, this improvement has usage. This deterioration is affected by several factors,
an upper limit. If the target value is below this limit, including the environment, operating conditions and
then the design using available parts and components maintenance. The rate of deterioration can be controlled
achieves the desired target value. If not, then a program through effective preventive maintenance actions. Poor
Springer Handbook

to improve the reliability through testfixtest cycles is reliability results in higher maintenance cost for the
MS ID: hb06-a2
Index entries on this page

Weibull Distributions and Their Applications 2.6 Applications 13

Table 2.2 A sample of reliability applications

Author(s) Topic
Weibull [2.5] Yield strength of steel, fatigue life of steel
Keshevan et al. [2.49] Fracture strength of glass
Sheikh et al. [2.50] Pitting corrosion in pipes

Part A 2.6
Quereshi and Sheikh [2.51] Adhesive wear in metals
Durham and Padget [2.52] Failure of carbon-fiber composites
Almeida [2.53] Failure of coatings
Fok et al. [2.54] Failure of brittle materials
Newell et al. [2.55] Failure of composite materials
Li et al. [2.56] Concrete components

buyer. It also leads to higher warranty cost for the manu- used to model the in-control duration in the de-
facturer resulting from the cost of rectifying all failures sign of control charts to detect the change from
within the warranty period subsequent to the sale of the in-control to out-of-control. For more on this, see
product. Rahim [2.59], Costa and Rahim [2.60], and Chen and
Most products are composed of several components Yang [2.61]. Nelson [2.62] deals with control charts
and the failure of the product is due to failure of its com- for items with Weibull failure distributions where
ponents. Weibull models (the two- and three-parameter conforming and nonconforming items differ in the
models as well as a variety of types IIII Models) have scale parameter but have the same shape parameter.
been used to model the failures of many components Murthy et al. [2.63] and Djamaludin et al. [2.64]
and the literature is vast. A very small sample of the deal with lot production and look at optimal lot
literature follows. size to control the occurrence of nonconforming
These models allow the determination of product items.
reliability in terms of component reliabilities during the When the failure rate has a bathtub shape, it is prone
design phase. to early failure. For products modeled by Weibull models
During the development phase, it is often necessary exhibiting bathtub failure rates, burn-in is a technique
to use accelerated testing to hasten the process leading that can be used to weed out such failures and improve
to component failures. A variety of type IV models have product reliability before it is released for sale. For more
been used for the design of experiments to carry out this on burn-in, see, e.g., Kececioglu and Sun [2.65].
testing. For more on such models, see Nelson [2.39] and Warranty cost analysis for products with a Weibull
Meeker and Escobar [2.48]. failure distribution has received a lot of attention. For
The improvement in reliability (also referred to as more on this, see, e.g., Blischke and Murthy [2.66, 67]
reliability growth) during the development phase has and Murthy and Djamaludin [2.68].
been modeled in many different ways. Duane [2.57] Preventive maintenance of products with a Weibull
used a Weibull intensity model formulation to model failure distribution has received considerable attention.
the improvement in failure rate as a function of the In the age policy, an item is replaced preventively when it
development time. The breakthroughs leading to im- reaches some specified age and Tadikmalla [2.69] deals
provements can be viewed as random points along with this in the context of the Weibull failure distribution.
the time axis. Crow [2.58] modeled this by a Weibull In the block policy, items are replaced preventively at set
power-law process (type VI Weibull model). For mod- clock times and Blischke and Murthy [2.70] deals with
els that are modification of this model, see Murthy this in the context of the Weibull failure distribution.
4 November 2005 10:08 CET

et al. [2.3].
In the manufacturing phase, the fraction of noncon- 2.6.2 Applications in Other Areas
forming items is small when the process is in-control
while it increases significantly when the process Weibull distribution has been used as a model in diverse
Springer Handbook

goes out-of-control. the Weibull distribution has been disciplines to study many different issues. There are
MS ID: hb06-a2
Index entries on this page

14 Part A Fundamental Statistics and its Applications

Table 2.3 A sample of other applications

Discipline Topic Author(s)
Geophysics Wind-speed data analysis Al-Hasan 2.71
Earthquake magnitude Huillet and Raynaud 2.72
Volcanic occurrence data Bebbington and Lai 2.73
Low-flow analysis Durrans 2.74
Part A 2

Regional flood frequency Heo et al. 2.75

Food science Sterility in thermal preservation method Mafart et al. 2.76
Social science Unemployment duration Roed and Zhang 2.77
Environment Environment radioactivity Dahm et al. 2.78
Nature Ecological application Fleming 2.79
Medical science Survival data Carroll 2.80

several thousand papers and we give a very small sample Weibull++ 6 by ReliaSoft Corp., Tucson, AZ;
of this vast literature http://weibull.reliasoft.com/
Relex Weibull by Relex Software Corp., Greens-
2.6.3 Weibull Analysis Software burg, PA; http://www.relexsoftware.com/products
In any analysis of statistical data, computer software is
required. In addition to standard statistical software such WeibullPro by Isograph Inc., Newport Beach, CA;
as Minitab, SPSS, SAS, etc., or spreadsheet software http://www.isograph-software.com/avsoverwbl.htm
such as Excel, some specialized Weibull analysis soft- WinSMITH Weibull by Barringer & Associates,
ware are also available. Below is a list of some common Inc., Humble, TX; http://www.barringer1.com
ones. /wins.htm


2.1 A. J. Jr. Hallinan: A review of the Weibull distribu- 2.9 W. J. Zimmer, J. B. Keats, F. K. Wang: The Burr XII
tion, J. Qual. Technol. 25, 8593 (1993) distribution in reliability analysis, J. Qual. Technol.
2.2 N. L. Johnson, S. Kotz, N. Balakrishnan: Continuous 30(4), 386394 (1998)
Univariate Distributions, Vol. 1, 2nd edn. (Wiley, New 2.10 C. D. Lai, M. Xie: Relative ageing for two parallel sys-
York 1994) tems and related problems, Math. Comp. Model. 38,
2.3 D. N. P. Murthy, M. Xie, R. Jiang: Weibull Models 13391345 (2003)
(Wiley, New York 2003) 2.11 M. Xie, C. D. Lai: On the increase of the expected
2.4 W. Weibull: A statistical theory of the strength of lifetime by parallel redundancy, Asia Pacific J. Oper.
material, Ing. Vetenskapa Acad. Handlingar 151, 145 Res. 13, 171179 (1996)
(1939) 2.12 J. F. Lawless: Statistical Models and Methods for Life-
2.5 W. Weibull: A statistical distribution function of wide time Data (Wiley, New York 1982)
applicability, J. Appl. Mech. 18, 293296 (1951) 2.13 W. Nelson: Applied Life Data Analysis (Wiley, New
2.6 M. Frchet: Sur la loi de probabilit de lcart max- York 1982)
imum, Ann. Soc. Polonaise Math. Cracovie 6, 93116 2.14 M. Liao, T. Shimokawa: A new goodness-of-fit
(1927) test for Type-I extreme-value, 2-parameter Weibull
2.7 R. A. Fisher, L. M. C. Tippett: Limiting forms of fre- distributions with estimated parameters, J. Stat.
quency distribution of the largest or smallest Comput. Sim. 64, 2348 (1999)
member of a sample, Proc. Cambridge Philos. Soc. 2.15 W. Nelson, V. C. Thompson: Weibull probability pa-
24, 180190 (1928) pers, J. Qual. Technol. 3, 140146 (1971)
2.8 N. L. Johnson, S. Kotz, N. Balakrishnan: Continuous 2.16 W. Nelson: Theory and application of hazard plotting
4 November 2005 10:08 CET

Univariate Distributions, Vol. 2, 2nd edn. (Wiley, New for censored failure data, J. Qual. Technol. 14, 935
York 1995) 966 (1972)
2.17 J. D. Kalbfleisch, R. L. Prentice: The Statistical Anal-
ysis of Failure Data (Wiley, New York 1980)
Springer Handbook
MS ID: hb06-a2
Index entries on this page

Weibull Distributions and Their Applications References 15

2.18 A. C. Cohen, B. J. Whitten, Y. Ding: Modified mo- tributions, Int. J. Reliab. Quality Saf. Eng. 4, 1734
ment estimation for the three-parameter Weibull (1997)
distribution, J. Qual. Technol. 16, 159167 (1984) 2.36 R. Jiang, D. N. P. Murthy: Parametric study of multi-
2.19 S. H. Zanakis, J. Kyparisis: A review of maximum plicative model involving two Weibull distributions,
likelihood estimation methods for the three- Reliab. Eng. Sys. Saf. 55, 217226 (1997)
parameter Weibull distribution, J. Stat. Comp. Simul. 2.37 R. Jiang, D. N. P. Murthy: Reliability modeling in-
25, 5373 (1986) volving two Weibull distributions, Reliab. Eng. Sys.
2.20 B. Dodson: Weibull Analysis (ASQC Quality, Milwau- Saf. 47, 187198 (1995)

Part A 2
kee 1994) see also 2.81 2.38 R. Jiang, M. Zuo, D. N. P. Murthy: Two sectional mod-
2.21 M. Xie, C. D. Lai, D. N. P. Murthy: Weibull-related els involving two Weibull distributions, Int. J. Reliab.
distributions for modelling of bathtub shaped fail- Qual. Saf. Eng. 6, 103122 (1999)
ure rate functions. In: Mathematical and Statistical 2.39 W. Nelson: Accelerated Testing (Wiley, New York
Methods in Reliability, ed. by B. H. Lindqvist, 1990)
K. A. Doksum (World Scientific, Singapore 2003) 2.40 T. Nakagawa, S. Osaki: The discrete Weibull distri-
pp. 283297 bution, IEEE Trans. Reliab. 24, 300301 (1975)
2.22 C. D. Lai, L. Y. Zhang, M. Xie: Mean residual life and 2.41 W. E. Stein, R. Dattero: A new discrete Weibull dis-
other properties of Weibull related bathtub shape tribution, IEEE Trans. Reliab. 33, 196197 (1984)
failure rate distributions, Int. J. Reliab. Qual. Saf. 2.42 W. J. Padgett, J. D. Spurrier: Discrete failure models,
Eng. 11, TS4 (2004) IEEE Trans. Reliab. 34, 253256 (1985)
2.23 D. N. P. Murthy, J. Baik, M. Bulmer, R. J. Wilson: 2.43 W. M. Bassin: A Bayesian optimal overhaul model for
Two-dimensional modelling of failures. In: Springer the Weibull restoration process, J. Am. Stat. Ass. 68,
Handbook Of Engineering Statistics, ed. by H. Pham 575578 (1973)
(Springer, Heidelberg, Berlin 2004) 2.44 L. J. Bain, M. Engelhart: Statistical Analysis of
2.24 A. W. Marshall, I. Olkin: A new method for adding Reliability and Life-testing Models, 2nd edn. (Mar-
aparameter to afamily of distributions with ap- cel Dekker, New York 1991)
plication to the exponential and Weibull families, 2.45 D. R. Cox, D. D. Oakes: Analysis of Survival Data
Biometrika 84, 641652 (1997) (Chapman Hall, New York 1984)
2.25 C. D. Lai, M. Xie, M. D. N. P. Murthy: Modified Weibull 2.46 N. Yannaros: Weibull renewal processes, Ann. Inst.
model, IEEE Trans. Reliab. 52, 3337 (2003) Stat. Math. 46, 641648 (1994)
2.26 G. S. Mudholkar, D. K. Srivastava: Exponentiated 2.47 D. N. P. Murthy, M. Bulmer, J. E. Eccleston: Weibull
Weibull family for analyzing bathtub failure-rate modelling, Reliab. Eng. Sys. Saf. TS5 , (2004) in print
data, IEEE Trans. Reliab. 42, 299302 (1993) 2.48 W. Q. Meeker, L. A. Escobar: Statistical Methods for
2.27 R. Jiang, D. N. P. Murthy: The exponentiated Weibull Reliability Data (Wiley, New York 1998)
family: A graphical approach, IEEE Trans. Reliab. 48, 2.49 K. Keshevan, G. Sargent, H. Conrad: Statistical anal-
6872 (1999) ysis of the Hertzian fracture of pyrex glass using
2.28 J. A. Kies: The Strength of Glass, Report, Vol. 5093 the Weibull distribution function, J. Mater. Sci. 15,
(Naval Research Lab., Washington 1958) 839844 (1980)
2.29 M. Xie, T. N. Goh, Y. Tang: A modified Weibull ex- 2.50 A. K. Sheikh, J. K. Boah, D. A. Hansen: Statistical
tension with bathtub-shaped failure rate function, modelling of pitting corrosion and pipeline relia-
Reliab. Eng. Syst. Saf. 76, 279285 (2002) bility, Corrosion 46, 190196 (1990)
2.30 R. M. Smith, L. J. Bain: An exponential power life- 2.51 F. S. Queeshi, A. K. Sheikh: Probabilistic characteri-
testing distribution, Commun. Stat. 4, 469481 zation of adhesive wear in metals, IEEE Trans. Reliab.
(1975) 46, 3844 (1997)
2.31 Z. Chen: A new two-parameter distribution with 2.52 S. D. Durham, W. J. Padgett: Cumulative damage
bathtub shape or increasing failure rate function, model for system failure with application to car-
Stat. Probab. Lett. 49, 155161 (2000) bon fibers and composites, Technometrics 39, 3444
2.32 R. Jiang, D. N. P. Murthy: Modeling failure data by (1997)
mixture of two Weibull distributions: A graphical 2.53 J. B. Almeida: Application of Weilbull statistics to the
approach, IEEE Trans. Reliab. 44, 477488 (1995) failure of coatings, J. Mater. Proces. Technol. 93,
2.33 R. Jiang, D. N. P. Murthy: Mixture of Weibull 257263 (1999)
distributions-parametric characterization of failure 2.54 S. L. Fok, B. C. Mitchell, J. Smart, B. J. Marsden: A nu-
rate function, Stochastic Models Data Anal. 14, 4765 merical study on the application of the Weibull
(1998) theory to brittle materials, Eng. Fract. Mech. 68,
4 November 2005 10:08 CET

2.34 J. Gurland, J. Sethuraman: How pooling failure data 11711179 (2001)

may reverse increasing failure rates, J. Am. Stat. Ass. 2.55 J. A. Newell, T. Kurzeja, M. Spence, M. Lynch: Analy-
90, 14161423 (1995) sis of recoil compressive failure in high performance
2.35 R. Jiang, D. N. P. Murthy: Parametric study of polymers using two-, four-parameter Weibull mod-
Springer Handbook

competing risk model involving two Weibull dis- els, High Perform. Polym. 14, 425434 (2002)
MS ID: hb06-a2

TS4 Please give pages

TS5 Please give any update if possible.

Editors or typesetters annotations (will be removed before the final TeX run)
Index entries on this page

16 Part A Fundamental Statistics and its Applications

2.56 Q. S. Li, J. Q. Fang, D. K. Liu, J. Tang: Failure prob- 2.70 W. R. Blischke, D. N. P. Murthy: Reliability (Wiley,
ability prediction of concrete components, Cement New York 2000)
Concrete Res. 33, 16311636 (2003) 2.71 M. Al-Hasan, R. R. Nigmatullin: Identification of the
2.57 J. T. Duane: Learning curve approach to reliability generalized Weibull distribution in wind speed data
monitoring, IEEE Trans. Aerosp. 40, 563566 (1964) by the Eigen-coordinates method, Renewable En-
2.58 L. H. Crow: Reliability analysis for complex systems. erg. 28(1), 93110 (2003)
In: Reliability and Biometry, ed. by F. Proschan, 2.72 T. Huillet, H. F. Raynaud: Rare events in a log-
R. J. Serfling (SIAM, Philadelphia 1974) pp. 379410 Weibull scenarioApplication to earthquake mag-
Part A 2

2.59 M. A. Rahim: Economic design of X charts assuming nitude data, Eur. Phys. J. B 12, 457469 (1999)
Weibull in-control times, J. Qual. Technol. 25, 296 2.73 M. S. Bebbington, C. D. Lai: On nonhomogeneous
305 (1993) models for volcanic eruptions, Math. Geol. 28, 585
2.60 A. F. B. Costa, M. A. Rahim: Economic design of X and 599 (1996)
R charts under Weibull shock models, Qual. Reliab. 2.74 S. R. Durrans: Low-flow analysis with aconditional
Eng. Int. 16, 143156 (2000) Weibull tail model, Water Resour. Res. 32, 17491760
2.61 Y. S. Chen, Y. M. Yang: Economic design of x-control (1996)
charts with Weibull in-control times, when there 2.75 J. H. Heo, D. C. Boes, J. D. Salas: Regional flood fre-
are multiple assignable causes, Int. J. Prod. Econ. quency analysis based on a Weibull model: Part
77, 1723 (2002) 1. Estimation, asymptotic variances, J. Hydrol. 242,
2.62 P. R. Nelson: Control chart for Weibull process, IEEE 157170 (2001)
Trans. Reliab. 28, 283288 (1979) 2.76 P. Mafart, O. Couvert, S. Gaillard, I. Leguerinel:
2.63 D. N. P. Murthy, I. Djamaludin, R. J. Wilson: Product On calculating sterility in thermal preservation
warranty and quality control, Qual. Reliab. Eng. 9, methods: application of the Weibull frequency dis-
431443 (1993) tribution model, Int. J. Food Microbiol. 72, 107113
2.64 I. Djamaludin, D. N. P. Murthy, R. J. Wilson: Lot- (2002)
sizing and testing for items with uncertain quality, 2.77 K. Roed, T. Zhang: A note on the Weibull distribution
Math. Comp. Model. 22, 3544 (1995) and time aggregation bias, Appl. Econ. Lett. 9, 469
2.65 D. Kececioglu, F. B. Sun: Burn-in Testing: Its Quan- 472 (2002)
tification and Optimization (Prentice Hall, Upper 2.78 H. Dahm, J. Niemeyer, D. Schroder: Application of
Saddle River 1997) the Weibull distribution to describe the vertical dis-
2.66 W. R. Blischke, D. N. P. Murthy: Product Warranty tribution of cesium-137 on a slope under permanent
Handbook (Marcel Dekker, New York 1994) pasture in Luxembourg, J. Environ. Radioac. 63,
2.67 W. R. Blischke, D. N. P. Murthy: Warranty Cost Anal- 207219 (2002)
ysis (Marcel Dekker, New York 1996) 2.79 R. A. Fleming: The Weibull model and an ecologi-
2.68 D. N. P. Murthy, I. Djamaludin: Product warranty: cal application: describing the dynamics of foliage
A review, Int. J. Prod. Econ. 79, 231260 (2002) biomass on Scots pine, Ecol. Model. 138(1-3), 309
2.69 P. R. Tadikmalla: Age replacement policies for 319 (2001)
Weibull failure rates, IEEE Trans. Reliab. 29, 8890 2.80 K. J. Carroll: On the use and utility of the Weibull
(1980) model in the analysis of survival data, Control. Clin.
Trials 24, 682701 (2003)
2.81 N. L. Johnson, S. Kotz, N. Balakrishnan: Continuos
Univariate Distributions, Vol. 1, 2nd edn. (Wiley, New
York 1994) pp. 657661
4 November 2005 10:08 CET
Springer Handbook
MS ID: hb06-a2

View publication stats