See discussions, stats, and author profiles for this publication at: [Link]

net/publication/238784663

Weibull Distributions and Their Applications

Chapter · January 2006

CITATIONS READS

17 4,337

2 authors, including:

Chin-Diew Lai
Massey University
217 PUBLICATIONS 4,756 CITATIONS

SEE PROFILE

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

Monitoring of high-yield processes View project

C D Lai View project

All content following this page was uploaded by Chin-Diew Lai on 26 August 2014.

The user has requested enhancement of the downloaded file.

 applications density Weibull probability plot
distribution Weibull distribution
function empirical Weibull derived
modeling WPP Weibull probability plot
estimation failure
Indexentriesonthispage

rate graphical
hazard plot
historical
hypothesis testing
modified moments
reliability

2. Weibull Distributions and Their

b
Weibull
Applications

Distri
Weibull models are used to describe various types 2.1.2 Relations to Other Distributions
of observed failures of components and .. 2
phenomena. They are widely used in
2.2 Properties........................................... 2
reliability and survival analysis. In addition to the
2.2.1 Basic Properties ........................
traditional two-parameter and threeparameter Weibull
2
distributions in the reliability or statistics literature,
2.2.2 Properties Related
many other Weibull-related distributions are available.
to Reliability............................. 3
The purpose of this chapter is to give a brief
introduction to those models, with the emphasis on 2.2.3 Simulation ...............................
models that have the potential for further applications. 4
After introducing the traditional Weibull distribution, 2.3 Modeling Failure Data.......................... 5
some historical development and basic properties 2.3.1 Probability Plots........................
arepresented. We also discuss estimation problems and 5
hypothesis-testing issues, with the emphasis on 2.3.2 Estimation and Hypothesis
graphical methods. Many extensions and Testing .............. 6
generalizations of the basic Weibull distributions are 2.3.3 Hypothesis Testing ....................
then summarized. Various applications in the reliability 7
context and some Weibull analysis software are also
provided. 2.4 Weibull-Derived Models ....................... 8
November200510:08

2.4.1 Taxonomy for Weibull Models

2.1 Three-Parameter Weibull ..... 8
Distribution... 2 2.4.2 Univariate Models
..................... 8 2.4.3 Type VI
2.1.1 Historical Models
CET
SpringerHandbook

Development.............. (Stochastic Point Process Models) 11

MSID:hb06-a2

2
4
 2.5 Empirical Modeling of Data .................. 2.6.2 Applications in Other
11 Areas......... 13
2.6.3 Weibull Analysis
2.6 Applications........................................
Software........... 13
12
2.6.1 Applications in References .................................................. 14
Reliability........... 12
The Weibull distribution is one of the best-known the Weibull distribution has
lifetime distributions. It adequately describes observed appearedinfivedifferentfor
failures of many different types of components and ms. Thetwocommon forms
phenomena. Over the last three decades, numerous of the distribution function
articles have been written on this distribution. Hallinan are as follows:
[2.1] gives an insightfulreview by presenting a number
of historical facts, the many forms of this distribution as
used by practitioners, and possible confusions and errors F (2.1)
that arise due to this non-uniqueness. Johnson et al. [2.2] and
devote a comprehensive chapter to a systematic study of F (2.2)
this distribution. More recently, Murthy et al. [2.3] The parameters of the distribution are given by the
presented a monograph that contains nearly every facet set θ = {α,β,τ} with α > 0, β > 0 and τ ≥ 0; where α is
relating to the Weibull distribution and its extensions.
a scale parameter, β is the shape parameter that
In Sect.2.1, we first define the three-parameter
determines the appearance or shape of the
Weibull distribution and then look at its historical
distribution and τ is the location parameter. The
development and relations to other distributions. Section
parameter λ combines both scale and shape features
2.2 studies the properties of the Weibull distribution, in
particular those relevant to reliability. A brief discussion as λ = α−β.
on simulation of Weibull variates is also included. Although one should use F(t,θ) instead of F(t),
We consider estimation problems and hypothesis where θ = (α,β,τ) denotes the vector of parameters,
testing in Sect.2.3. In particular, we emphasize the for notational convenience we suppress the
graphical methods as a tool for selection and parameter and use F(t) to denote F(t,θ) in this
parameter estimation of a Weibull model. We
chapter. Also, we do not intend to give an
devote Sect.2.4 to Weibullderived models, which
exhaustive review. Rather, we confine our
includes many extensions and generalizations.
discussion to aspects that are of relevance to the
Finally, in Sect.2.5, we outline various applications,
context of reliability theory.
especially those in the reliability context. Because
For τ = 0, (2.1) and (2.2) become the two-
of the vast literature, we are unable to refer to all
the source authors and we apologize in advance for parameter Weibull distribution with
any omissions in this regard. Symbols and
Abbreviations.
T Random variable F (2.3)
F(t) CDF, cumulative distribution
and
function
f(t) PDF, probability density function F (2.4)
h(t) Failure rate function (hazard rate) Murthy et al. [2.3] refer to this as the standard
α,β,τ Parameters WPP Weibull model, Johnson et al. [2.2] refer to a
Indexentriesonthispage

2 Part A Fundamental Statistics and its Applications

2.1 Three-Parameter Weibull Distribution

SpringerHandbook

Weibull probability plot standard Weibull when α =

MSID:hb06-a2

According to Hallinan [2.1] 1 (or λ = 1) in (2.3,4).

 2.1.1 Historical Development
TheWeibulldistributionisnamed after itsoriginator,the
Swedish physicist Waloddi Weibull, who in 1939 used
it to model the distribution of the breaking strength of
materials [2.4] and in 1951 for a wide range of other
applications[2.5].Thedistributionhasbeenwidelystudied
since its inception. It has been known that Weibull may
not be the first to propose this distribution. The name
Fréchet distribution is also sometimes used due to the
fact that it was Fréchet [2.6] who first identified this
distributiontobean extremaldistribution(latershowntobe
one of the three possible solutions by Fisher and Tippett
[2.7]). According to Hallinan [2.4], it was Weibull who
suggested a scale parameter and a location parameter
that made the distribution meaningful and useful. 2.1.2
Relations to Other Distributions
The Weibull distribution includes the exponential (β = 1)
and the Rayleigh distribution (β = 2) as special cases. If
X denotes the Weibull variable, then −X has a type 3
extreme-value distribution ([2.8], Chapt.06-c22TS0 ). A
PartA

simple log transformation will transform the Weibull

distribution into the Gumbel distribution (type 1
2.2

extreme-value distribution).
The Burr XII distribution, is given by t

−k

F,t ≥ 0 ;k ,α > 0 . (2.5)

Let k = α; as k → ∞, then the Burr distribution

approaches the Weibull (see, for example, [2.9]).
November200510:08CET

β
− −1 t− τ
(t) = βα β (t − τ) β exp −
α
(2.6) Moments
Let T denotetherandomvariablefromthethree-
and −1 parameterWeibulldistributiongivenby(2.1).Then
f(t) = βλ(t − τ) β exp − λ(t − τ) β , t ≥ τ. (2.7)
SpringerHandbook

thetransformedvariable T = (T − τ)/α,isthestan-
TS0 Pleaseprovidetitleofreferencedchapter
MSID:hb06-a2

Editor’sortypesetter’sannotations(willberemovedbeforethefinalTeXrun)
4
 Indexentriesonthispage

Weibull Distributions and Their Applications 2.2 Properties 3

2.2 Properties The distribution is positively skewed for small values of

β. The skewness index decreases and equals zero
2.2.1 Basic Properties for β = 3.6 (approximately). Thus, for values of β in the
vicinity of 3.6, the Weibull distribution is similar in
Density Function
shape to a normal distribution. The coefficient of
The probability density functions (PDF) (Fig.2.1) of
kurtosis β2 also decreases with β and then increases, β2
(2.1) and (2.2) are
has a minimum value of about 2.71 when β = 3.35
(approximately).

PartA
f ,t≥τ
Mode
Order Statistics
It follows from (2.6) that the mode is at α((β−1)/β)1/β +τ
Let T1,T2,··· ,Tn denote n independent and identically

2.2
for β > 1 and at τ for 0
distributed three-parameter Weibull random variables
with density function given in (2.6) and cumulative
Median distribution function (CDF) in (2.1). Further, let T(1) ≤
It follows from (2.3) that the median of the distribution T(2) ≤ ··· ≤ T(n) denote the order statistics from a sample of
is at α(log2)1/β +τ. n observations. The probability density function of T(1),
dard form (in the sense of Johnson et al. [2.2]) with the is given by
density function given by

f ,β > 0 . (2.8)
The moments of T (about zero) are easily obtained from f
the moments of T which are given below: (2.12)

It is obvious from (2.12) that T(1) is also distributed

µr , (2.9) as a Weibull random variable, except that α is
replaced by αn−1/β.
from which we get
The density function of T(r)(1 ≤ r ≤ n) is given by
E , (2.10) r−1

Skewness and Kurtosis

e

Density function ThesurvivalfunctionoftheWeibulldistributionis

1.8
β
1.6 t− τ
F̄( t) = 1 − F( t) = exp − , t ≥ τ.
1.4 α
α = 0.5, = 0.9
1.2 (2.15)
1.0
−1
0.8 Notethatat t = τ + α, F̄(τ + α) = 1 − e ≈ 0.3679.
0.6
0.4 α = 3, = 2 Failure Rate
α = 2, = 0.5 Thefailureratefunction(alsoknownasthehazardrate)
November200510:08CET

0.2
forthethree-parameterWeibullis
0.0
β− 1
0 1 2 3 4 5 t f(t) = β t − τ
h( t) = . (2.16)
SpringerHandbook

Fig.2.1 Two-parameterWeibulldensityplots F̄(t) α α

MSID:hb06-a2

CE1 Pleaseconfirmthepositionofthetimessymbolinthisequation(2.14)

Editor’sortypesetter’sannotations(willberemovedbeforethefinalTeXrun)
4
 It can be shown that Distributions
k Supposewe have two Weibullrandom variables X
i
and Y with distribution functions F(x) and G(y),
E , (2.14) respectively, given by
i=0
F ,
where G . (2.20)
We say that X ages faster than Y if the ratio of the
n
r− 1 failure rate of X over the failure rate of Y is an
k + k increasing function of t. This ratio is given by
ωr( ) =(r −1)!(!n − r) ! Γ 1 β i= 0
r− 1
( − 1) r i
α1β2 × α1ββ21−−11 tβ2−β1 (2.21) α2β1 α2 which is
× .
(n−r +i +1)1+(k/β) an increasing function of t if β2 > β1.

CE1
Effectiveness of Parallel Redundancy –
The Weibull Case
2.2.2 Properties Related to Reliability Denote the lifetime of one component by T and the
lifetime of a parallel system of two such independent
In this section, we consider only the first form of the components by Tp. The effectiveness of parallel
4 Part A Weibull distribution.
Fundamental Statistics and its Applications
Indexentriesonthispage

mean residual life (MRL)

Forthetwo-parametercase,itisgivenby
Failure rate function
β− 1 4
f ( t) t
h ( t) = = β . (2.17)
F̄ ( t) α α
3
Itisobviousthat h ( t) isadecreasingfunctionwhen
β< 1,constantwhen β = 1( theexponentialcase),and
anincreasingfunctionwhen β> [Link] 2
α = 0.5, = 0.9
PartA 2.2

behaviourofthefailureratefunction,theWeibulldis-
tributionoftenbecomessuitablewhentheconditions 1 α = 3, =2
for strictrandomness oftheexponentialdistribution
α = 2, = 0.5
arenotsatisfied,withtheshapeparameter β having 0
avaluedependinguponthefundamentalnaturebeing
0 1 2 3 4 5 t
considered.
Insomeway,havingafailureratefunctionofmono-
tonicshapehaslimitationsinreliabilityapplications.
Forthisreason,severalgeneralizedormodifiedWeibull Suppose E( X) = E(Y), i.e., α2 Γ( 1 + 1/β 2 ) =
distributionshavebeenproposed(seeSect.2.4formore α1 Γ( 1 + 1/β 1 ) . Lai and Xie [2.10] show that
details).Figure2.2showsplotsofthefailureratefunc-
tionsforsomeselectedparametersvalues.
Fig.2.2 Two-parameter Weibull failure rate functions Var(X)
≤ Var(Y).
Mean Residual Life redundancy of the component is defined by [2.11]
The mean residual life (MRL) of a lifetime random
variable T is defined as ep = [E(Tp)− E(T)]/E(T) . (2.22)
t
∞F dx F¯(t)
¯(x)
Suppose that T is a two-parameter Weibull distribution
µ(t) = E(T −t|T> t) =. (2.18)
with shape parameter β and scale parameter α. Xie and
For the Weibull distribution, the MRL is Lai [2.11] show that
complicated except for the two special cases β = 1
ep = 1−2−1/β , (2.23)
(exponential) and β = 2√ (Rayleigh distribution).
α = 2σ
√ location parameter1 τ = 0, the MRL of
Assuming the
= 2πσ [1 − Φ(t/σ) ] e 2σ 2 which decreases as β increases. Thus, a parallel
November200510:08

the Rayleigh distribution with scale parameter is redundancy is more effective for β < 1.
t2
µ(t), t > 0 , (2.19) where Φ(t/σ) denotes the 2.2.3 Simulation
distribution function of the standard normal The two-parameter Weibull distributionwith parameters
variable. β and λ has the probability density function
CET
SpringerHandbook
MSID:hb06-a2

Relative Ageing of Two Two-Parameter Weibull

4
 Indexentriesonthispage

Weibull Distributions and Their Applications 2.3 Modeling Failure Data 5

β 1
f(x) = λβx − e− .
λxβ
(2.24) The simple inverse-probability integral-transform
method applied to the standard Weibull distribution (λ =
1) is quite efficient. The formula is simply x = (−logu)1/β ,
whereu [Link], an
acceptance/rejection method could also be used to avoid
the evaluation of the logarithm in the distribution
function. The standard Weibull variate is

PartA
then scaled by λ−(1/β) to obtain a two-
parameter Weibull observation with density
(2.24).

2.3
2.3 Modeling Failure Data The process of black box modeling involves the

Failure data can be classified into two types: complete •

andincomplete(censored).Forcompletedata,theactual following steps: •• Step 1 – Model selection, Step 2 –
realized values are known for each observation in the
data set, whereas for censored data, the actual realized Parameter estimation, Step 3 – Model validation.
values are not known for some or all of the observations. To select a model out of many possible models, one
Further, there are at least two types of censoring, details requires a good understanding of their different
ofwhichmaybefoundLawless[2.12]andNelson[2.13]. properties. This often is a trial-and-error process. Liao
Intheirintroduction,Murthyetal.[2.3]pointoutthat the and Shimokawa [2.14] consider goodness-of-fit testing
two different approaches to building mathematical as a key procedure for selecting the statistical
models are as follows. distribution that best fits the observed data. To execute
the remaining two steps, we need various tools and
1. Theory-based modeling:here, the modeling is based
techniques.
on the established theories for component failures.
A large number of Weibull-relatedmodels have been
This kind of model is also called a physics based
derived in the literature. In Sect.2.4, various such
model or white-box model.
models are grouped into several categories. Selecting an
2. Empirical modeling: here, the data available forms
appropriate model from the family of Weibull-related
the basis for the model building. This kind of model
distributions can often be based on some of the
is also called a data-dependent model or black-box
probability plots to be discussed below. Moreover, they
model.
can provide crude estimates of model parameters. There
In empirical modeling, the type of mathematical are many different statisticaltests for validatinga model
formulations needed for modeling is dictated by a and we postpone this discussion until the next section.
preliminary analysis of data available. If the analysis
indicates that there is a high degree of variability, then 2.3.1 Probability Plots
November200510:08

one needs to use models that can capture this variability.

This requires probabilistic and stochastic models to Weibull Probability Plot
model a given data set. Weibull probability plots (WPP) can be constructed
in several ways [2.15]. In the early 1970s a special
paper was developed for plotting the data in the
CET
SpringerHandbook

form of F(t) versus t on a graph paper with a log–

MSID:hb06-a2

log scale on the vertical axis and a log scale on the

4
 horizontal axis. A WPP plotting of data involves 2.3.2 Estimation and Hypothesis Testing
computing the empirical distribution function,
which can be estimated in different ways; the two
standard methods are:

•
• Fˆˆ(ti) == i/(−n+1), the “mean rank” estimator, and

F(ti) (i 0.5)/n, the “median rank” estimator.

Here, the data consists of successive failure times ti,

t1 < t2 < ...tn. For censored data (right-censored or
interval), the approach to obtain the empirical
distribution functions needs to be modified; see, for
example, Nelson [2.13].
These days, most computer reliability software
packages contain programs to produce these plots
automatically given a data set. A well-known
statistical package Minitab provides a Weibull
probability plot under the Graph menu Probability
Plot.
We may use an ordinary graph paper or spreadsheet
software with unit scale for plotting. Taking logarithms
twice of both sides of each of the CDF in (2.3) yields
log[−log F¯(t)] = β log(t −τ)−β logα . (2.25)

Let y and x = log(t −τ). Then we

have

y = βx −β logα . (2.26)

The plot is now on a linear scale.

Weibull Hazard Plot

The hazard plot is analogous to the probability plot, the
principal difference being that the observations are
plotted against the cumulated hazard (failure) rate
rather than the cumulated probability value. Moreover,
this is designed for censored data.
Let H(t) denote the cumulative hazard rate, then
F¯(t) = exp[−H(t)], so

H , (2.27)
H(t)1/β/α = (t −τ) . (2.28)

Let y = log(t −τ) and x = log H(t), then we have

y . (2.29)

Rank the n survival times (including the

censoreddata) in ascending order and let K denote
the reverse ranking order of the survival time, i.e.,
K = n for the smallest survival time and K = 1 for
the largest survival time. The hazard is estimated
from 100/K (a missing value symbol is entered at a
censored failure time). The cumulative hazard is
obtained by cumulating the hazards. The Weibull
SpringerHandbook

hazard plot is simply the plot arising from (2.29).

MSID:hb06-a2

See Nelson [2.16] for further details.

 Indexentriesonthispage

6 Part A Fundamental Statistics and its Applications

Parameter estimation is the second step of our
modeling process. We consider both graphical and
statistical methods for the three-parameter Weibull
distribution. suggested to set the initial value of τ =
T(1) and then adjust this estimate to achieve a
straight line. We also note that it is common for τ to
be 0.
Estimation of the Shape Parameter. It follows from
PartA

(2.26)thattheshapeparametercanbeestimatedfromthe
slope of the WPP. However, care needs need to be given
because possible confusion may arise due to several
2.3

possible forms of the Weibull distribution.

Estimation of the Scale Parameter. To estimate the scale

parameter, we first observe from the standard Weibull
that, when t is at the 63.2 percentile, F(t) = 1− e−1, i.e.,
F¯(t) = e−1. Or equivalently, for t = τ +α, F¯(τ +α) = e−1 (so
τ +α is the 63.2 percentile). It now follows from (2.26)
that log(t −τ) = logα so that αˆ = t−τˆ, where t is the 63.2
percentile. Hence the scale parameter is simply
estimated from the x-intercept of the plot (2.26): αˆ = e(x-
intercept).

Under the probability plot option, Minitab not only

gives a WPP, but also provides an estimate for the scale
and shape parameters α and β.
SpringerHandbook
MSID:hb06-a2
 Indexentriesonthispage

Weibull Distributions and Their Applications 2.3 Modeling Failure Data 7

Figure 2.3 illustrates how a WPP plot appears in a and
Minitab graph with αˆ = 9.062 and βˆ = 1.661.

Statistical Methods of Estimation βˆ

There are several statistical methods for estimating the
[Link], the
method of percentile, the method of maximum , (2.31)
likelihood, and the Bayesian method. The estimates can where Xi,i = 1,2,...n are n independent

PartA
be [Link]
these methods can be found, for example, in Kalbfleisch observations from the two-parameter
and Prentice [2.17].

2.3
We note that these methods may not be appropriate Weibull distribution. If τ = 0, then each Xi
for small data sets. Hence, the estimates obtained from
is replaced by Xi −τ in the above equations.
the WPP plot may be taken as the final estimates when
the data size is small. The value βˆ needs to be solved from (2.31) and then

Moment Estimation. By equating the first three moments (2.30) is used to obtain αˆ.
of theWeibulldistributionto thefirst threesample
moments and solving, it is possible to find the moment Suppose the location parameter τ is also unknown,
estimators of α,β and τ. Now, the first moment ratio then the maximum-likelihood estimates α,ˆ βˆ and τˆ
depends only on the shape parameter β, and hence satisfy the following equations
has been estimated from the sample
coefficient of skewness, βˆ can be determined
numerically. Using βˆ, αˆ can be determined from the αˆ =(2.32)
standard deviation, and lastly τˆ is determined from the
sample mean. If τ is known, then β can be estimated
from the ratio
(standard deviation/(mean-τ).
For modified moment estimation, see Cohen et al.
[2.18].

Maximum-Likelihood Method.
Maximum-likelihood methods for
the three-parameter Weibull =
distribution have been reviewed by Zanakis and
Kyparisis [2.19]; see also Johnson et al. [2.2, Chapt. 21] and
for other details. n n
The most common situation is when the location
parameter τ is known. Without loss of generality, we (βˆ τ)ˆ βˆ−1 .
may assume that τ = 0 so that the model becomes the i=1 i=1

two-parameter Weibull distribution. (2.34)

November200510:08

The maximum-likelihood estimators, βˆ and αˆ of β

Ifthevalueτˆ satisfying(2.32)–(2.34)islargerthan X1 =
and α, respectively, satisfy the equations min Xi, then it is the maximum-likelihood estimate
1/βˆ
of τ. Otherwise, we set that τˆ = X1 and then (2.32)
CET
SpringerHandbook

αˆ =(2.30) and (2.33) must be solved for αˆ and βˆ.

MSID:hb06-a2
4
 The method is suitable for large data sets Generally, a goodness-of-fit test for Weibull
because of its asymptotic properties. Modifications distribution can be described as:H0 : the population
of these estimates can be found, for example, in follows a Weibull model, versus H1 : the Weibull model
Johnson et al. is not suitable.
[2.2]. Minitab under probability plot gives the
The reader may consult Chapt. 21 of Johnson et Anderson– Darling statistic as well as the
al. [2.2] and Chapt. 4 of Murthy et al. [2.3] for the confidence bands.
following methods of estimation, Lawless [2.12] summarizes the goodness-of-
fit tests for the Weibull or extreme-value
distribution. Four tests (the likelihood ratio test
as a sub-model of the gamma distribution, Mann–
••••• Best linear unbiased Scheuer–Fertig test, Tiku test and Cramer–von
Mises test) are discussed in detail. Dodson [2.20]
estimation, Percentile also includes and compares some goodness-offit
tests for Weibull distributions in Chapt. 4 of his
estimator, Bayesian book. Liao and Shimokawa [2.14] construct a new
goodness-of-fit test for the two-parameter Weibull
and its power is compared with other traditional
estimator,
goodnessof-fit tests.
Interval estimation,
Hypothesis Testing
Minimum quantile distance estimation.
Hypothesis testing of Weibull parameters may be
For the two-parameter Weibull distribution (i.e., τ = performed by the likelihood ratio, score and Wald tests.
It is well known that all these tests are asymptotically
0), methods of inference can be obtained via the
optimal.
type 1 extreme-value distribution. For testing Weibull versus the exponential
distribution, we test α = 1 (or λ = 1 of the second form)
2.3.3 Hypothesis Testing
versus α = 0.
Goodness-of-Fit Tests for the Weibull
Distribution
Thereareothergeneralgoodness-of-fittestsbasedonthe
empirical distribution functions derived. For
example, the chi-square test, Kolmogorov–Smirnov
test, Cramer– von Mises test and Anderson–Darling
test, and so on, could also be used to test the
goodness of fit of the Weibull model.
SpringerHandbook
MSID:hb06-a2
 Indexentriesonthispage

8 Part A Fundamental Statistics and its Applications

2.4 Weibull-Derived Models

There are many extensions, generalizations and G
modifications to the Weibull distribution. They This is also known as a type 3 extreme-value
arise out of the need to model features of empirical distribution (see Johnson et al. [2.8], Chapt. 22).
data sets that cannot be adequately described by a
three-parameter Weibull model: for example, the Double Weibull Distribution. g
monotonic property of the Weibull, which is unable
to capture the behaviour of a data set that has a
PartA

bathtub-shape failure rate. Xie et al. [2.21] review

several Weibull-related distributions that exhibit
This is an obvious extension to include the negative real
bathtub-shaped failure rates. Plots of mean residual
2.4

line as its support.

life from several of these Weibull-derived models
are given in Lai et al. [2.22]. For simplicity, we
Log Weibull Distribution. The distribution is derived from
simply refer to these Weibull-related models as
the logarithmic transformation of the two-parameter
Weibull models.
Weibull distribution. The distribution function is
2.4.1 Taxonomy for Weibull Models
G .
According to Murthy et al. [2.3], the taxonomy for (2.37)
Weibull models involves seven types, each of This is also known as a type 1 extreme-value
which divided into several sub-types. These models distribution or Gumbel distribution. In fact, it is the
can be grouped into three groups: most commonly referred to in discussions of extreme-
1. Univariate models (types I–V), value distributions (see [2.8], Chapt. 22). The density
2. Multivariate models (type VI),3. function is
Stochastic process models (type VII).
In this chapter we confine our discussion to g ,
univariate and stochastic process models, while and the failure rate function is quite simple and given by
Murthy et al. [2.23] looks at bivariate models (a
special case of multivariate models). In this chapter, h .
only some selective models will be included. We
refer our readers to the book by Murthy et al. [2.3]
for more details. Inverse (or Reverse) Weibull Model. Let Z = α2/T, where T
has a two-parameter Weibull distribution, then the
2.4.2 Univariate Models distribution function of T is
The starting point is the two-parameter Weibull G(t) = exp[−(t/α)−β] ,t ≥ 0 . (2.38)
model with distribution function F(t). Let G(t)
denote the derived Weibull model. T is a random This is also known as a type 2 extreme-value
variable from F(t) and Z is a random variable from distribution or Fréchet distribution (see [2.8], Chapt.
G(t). 22). The failure rate function is as follows:
Type I Models βαβt−β−1 e−(α/t)β h(t) =
November200510:08

(Transformation of Weibull Variable) β

Here Z and T are related by a transformation. The .

transformation can be either linear or nonlinear. These
1− e−(α/t)
include the following:
CET
SpringerHandbook

Reflected Weibull Distribution.

MSID:hb06-a2
4
 Indexentriesonthispage

Weibull Distributions and Their Applications 2.4 Weibull-Derived Models 9

Type II Models (Modification/Generalization of G,
the Weibull Distribution) Here G(t) is related to F(t)
by some relationship. (2.43)

Extended Weibull Distribution [2.24].

(λt)β Doubly Truncated Weibull Distribution.
)
G(t β . (2.39) F(t)− F(a)
=
G( t),

PartA
−(λt)
F(b)− F(a)
The failure rate function is
0<a≤t≤b<∞.

2.4
(2.44)

h .
Thefailureratefunctionisincreasingif ν ≥ 1,β ≥ 1 and
Modified Weibull Extension [2.29].
decreasing if ν ≤ 1,β ≤ 1.
G ,
Modified Weibull Distribution [2.25]. t ≥ 0,α,β,λ > 0 . (2.45)
G(t) = 1−exp(−λt e ) ,t ≥ 0
β νt (2.40)

This is known as the generalized exponential power

with
model originally studied by Smith and Bain [2.30].
h(t) = λ(β+νt)tβ−1 exp(νt) . (2.41)
The case of α = 1 is also studied in Chen [2.31]. The
For ν = 0, this reduces to a Weibull distribution. For 0 < failure rate function is
β < 1,h(t)isinitiallydecreasingandthenincreasing in t,
implying that the failure rate function has a bathtub h . (2.46)

shape. When β > 1, h(t) is increasing in t. It approaches to a two-parameter Weibull

distribution when λ → ∞ with α in such a manner
Exponentiated Weibull Distribution [2.26].
that αβ−1/λ is held constant.
G(t) = [F(t)]ν = {1−exp[−(t/α)β]}ν ,t ≥ 0 .
Type III Models (Models Involving Two or More
(2.42)
Distributions)
The density function is
These are univariate models derived from two or
more distributions with at least one being either the
g . standard Weibull model or a distribution derived
from it.
For an appropriate choice of the parameter set, it n-Fold Mixture Model.
will give rise to a bathtub-shaped failure function. n n
Jiang and Murthy [2.27] use a graphical approach
G pi = 1 .
to study this distribution. The failure rate function is
given by (2.47) i=1 i=1
November200510:08

Jiang and Murthy [2.32, 33] categorize the possible

shapes of the failure rate function for a mixture of
h (t/α) . any two Weibull distributions in terms of five
parameters. Gurland and Sethurama [2.34] also
consider a mixture of the Weibull distribution with
CET
SpringerHandbook

Four-Parameter Weibull Distribution [2.28]. failure rate λβtβ−1 and the exponential distribution
with failure rate λ1. For β > 1, the Weibull
MSID:hb06-a2
4
 Indexentriesonthispage

10 Part A Fundamental Statistics and its Applications

distribution is IFRCE3 . They found that the resulting Weibull Proportional Hazard Models.
mixture distribution is ultimately DFR.
Infact,themixturehasafailureratewithanupside-down h(t) = ψ(S)h0(t), (2.53)
bathtub shape. n-Fold Competing Risk Model.
n where h0(t) is called the baseline hazard for a
G . (2.48) i=1 twoparameter Weibull distribution. The only restriction
on the scalar function ψ(S) is that it be positive. Many
Jiang and Murthy [2.35] also give a parametric different forms for ψ(S) have been proposed. One such
study of a
PartA

is the following:
competing
risk model
involving
2.4

. (2.54)
two For more on such models, see Cox and Oakes [2.39] and
Weibull distributions.
Kalbfleisch and Prentice [2.17].
n-Fold Multiplicative Model (Complementary Risk
Type V Models (Discrete Weibull Models) Here T
Model).
n
can only assume non-negative integer values and this
defines the support for F(t).
G (2.49) i=1
Model 1 [2.40].
The multiplicative model involving two Weibull


distribution is considered in Jiang and Murthy
[2.36].  1 qt t
0,1,2,
n-Fold Sectional Model. 3,0t<0.
F(t) =− β = ··· (2.55)
), 0 ≤ t ≤ t1 ,
,
G(2.50)
Model 2 [2.41]. The cumulative hazard function is given
by

1−kn F¯n(t) , t > tn−1 ,

where the sub-populations Fi(t) are two- or

threeparameter Weibull distributions and the tis
,m
(called partition points) are an increasing sequence.
Jiang and Murthy [2.37] and Jiang et al. [2.38] H(2.56)
consider the case n = 2 in details.
Arrhenius Weibull Model.
α(S) = exp(γ0 +γ1S) . (2.51)

Power Weibull Model.

eγ0 α(S)
(2.52) S
= .
γ1

These types of models have been used extensively in

SpringerHandbook

accelerated life testing [2.39] in reliability theory. As a

MSID:hb06-a2

result, they are referred to as accelerated failure models.

 November200510:08CET
 −[ 1/(β − 1)] if
(Weibull Models with Varying Parameters)
Formodelsbelongingtothisgroup,theparametersofthe m = int c β> 1 , (2.57)
 ∞ if β ≤ 1 ,
modelareeitherfunctionsoftheindependentvariable(t)
orsomeothervariables(suchasthestresslevel s,etc.), andint ( •) representstheintegerpartofthequantity
Indexentriesonthispage

orarerandomvariables. insidethebrackets.
CE3 PleaseensurethattheabbreviationsIFRandDFRaredefinedonfirstuseinthetext

4 Editor’sortypesetter’sannotations(willberemovedbeforethefinalTeXrun)

Weibull Distributions and Their Applications 2.5 Empirical Modeling of Data 1

Type IV Models 1
where m is given by
Model 3 [2.42].

, t = 0,1,2,··· .

PartA
(2.58)

2.4.3 Type VI Models (Stochastic Point

2.5
Process Models)
These are stochastic point process models
with links to the standard Weibull model.

Power Law Process (Bassin [2.43]) This is a point

process model with the intensity function given by

. (2.59)
This model has been called by many different names:
power law process; Rasch–Weibull process; Weibull
intensity function); Weibull–Poisson process and
Weibull process. We note that the inter-event times do
not have Weibull distributions. For further discussion on
the Weibull process, see for example, Bain and
Engelhart [2.44, Chapt. 9].

Proportional Intensity Model

(Cox and Oaks [2.45])
The intensity function is given by

λ(t; S) = λ0(t)ψ(S) ,t ≥ 0 , (2.60)

where λ0(t) is of the form given above and ψ(S) is a

function of the explanatory variables S. The only
restriction on ψ(S) is that ψ(S) > 0.

Ordinary Renewal Process (Yannaros [2.46])

Here the point process is a renewal process with the
time between events being independent and
identically distributed with the distribution function
given by the standard Weibull distribution.
November200510:08

Modified Renewal Process

CET
SpringerHandbook
MSID:hb06-a2
4
 Here the distribution function for the time to first
event, F0(t), is different from that for the subsequent
inter-event times, which are identical and
independent random variables with distribution
function F(·). Note that F0(t) and/or F(t) are standard
Weibull distributions. When F0(t) = F(t), this
reduces to the ordinary renewal process.
2.5 Empirical Modeling of Data
We have seen a large number of Weibull-related models which we simply refer as Weibull models. They exhibit a

wide range of shapes for the density and failure rate functions, which make them suitable for modeling complex
failure-data sets.
Recall, we mention that empirical modeling usually involves three steps: model selection, estimation of model
parameters and model validation. In the context of Weibull models, a selection procedure may be based on WPP
plots. This is possible due to the availability of WPP or generalized WPP plots for all the Weibull models of types I–
III. Of course, the shape of the density and failure rate functions will also be valuable in the selection step. An added
advantage of the WPP plots is that they provide crude estimates of model parameters; these serve as a starting point
for steps 2 and 3.
It has been suggested that an alternative method to estimate model parameters is through a least-squares fit.
Basically speaking, this involves selecting the parameters the parameters to minimize a function given
Fig.2.4 Probability plot of data set from Table 2.1
SpringerHandbook
MSID:hb06-a2
 PartA

Indexentriesonthispage
2.6

12 Part A Fundamental Statistics and its Applications

Table2.1 Data set of failure test (data set 2)

by tests for validating a model. These generally require data

that is different from the data used for model selection
nJ , (2.61) i=1
and parameter estimation. A smaller data set may pose a
where y(ti;θ) uses the vector parameter θ and yi is the problem, as there will be no separate data left after
corresponding value obtained from the data. The model selection and parameter estimation. Various
optimization can be carried out using any standard solutions have been proposed in the case of a small data
optimization package. The least-squares method not set. We refer the readers to the book by Meeker and
only furnishes us with parameter estimates, but also Escobar [2.48] for further details.
helps to select a Weibull model. If one of the potential Example 1:
candidates has a value for J(θ) which is considerably 50 items are tested to failure. The failure times are
smaller than that for the other models, then this can automatically recorded and given in Table 2.1. The
undoubtedly be accepted as the most appropriate model Weibull probabilityplotindicatesthatthetwo-
for modeling the given data set. If two or more Weibull parameterWeibull model is not appropriate.
models give rise to roughly the same value of J(θ), one Based on the plot, we observe that the plot shows a
would need to examine additional properties of the WPP decreasing slope at the beginning, and an increasing
plots to decide on the final model. Other approaches slope at the end. This is an indication of a distribution
such as bootstrap and jackknife may be employed for the with a bathtub-shaped failure-rate function. Some of the
final selection. For more on this, see Murthy et al. models of type II or type III can be used. One possibility
November200510:08

[2.3,47]. istofittheearlypartandlastpartoftheplotwithseparate lines

A couple of comments on step 3 of our empirical (Fig.2.4). This will result in a two-fold competing risk
modeling may be in order. There are many statistical model (2.48).
CET
SpringerHandbook
MSID:hb06-a2
4
 2.6 Applications through effective preventive maintenance actions. Poor
reliability results in higher maintenance cost for the
2.6.1 Applications in Reliability buyer. It also leads to higher warranty cost for the
manufacturer resulting from the cost of rectifying all
Product reliability depends on the design, failures within the warranty period subsequent to the
development and manufacturing decisions made sale of the product.
prior to the launch (pre-launch stage) of the product Most products are composed of several components
and it in turn affects the failures when the product and the failure of the product is due to failure of its
is put into operation after launch (post-launch components. Weibull models (the two- and three-
stage). parameter models as well as a variety of types I–III
The pre-launch stage involves several phases. In Models) have been used to model the failures of
the feasibility phase, study is carried out using the many components and the literature is vast. A very
specified target value for product reliability. During small sample of the literature follows.
the design phase, product reliability is assessed in These models allow the determination of product
terms of part and component reliabilities. Product reliability in terms of component reliabilities during the
reliability increases as the design is improved. design phase.
However, this improvement has an upper limit. If During the development phase, it is often necessary
the target value is below this limit, then the design to use accelerated testing to hasten the process leading
using available parts and components achieves the to component failures. A variety of type IV models have
desired target value. If not, then a program to been used for the design of experiments to carry out this
improve the reliability through test–fix–test cycles testing. For more on such models, see Nelson [2.39] and
is carried out during the development phase. Here Meeker and Escobar [2.48].
the The improvement in reliability (also referred to as
prototypeistesteduntilafailureoccursandthecausesoft reliability growth) during the development phase has
he failure are analyzed. Based on this, design been modeled in many different ways. Duane [2.57]
changes are madetoovercome used a Weibull intensity model formulation to model the
[Link] is continued improvement in failure rate as a function of the
until the reliability target is achieved. The reliability development time. The breakthroughs leading to
of the items produced during manufacturing tends improvements can be viewed as random points along the
to vary from the design target value due to time axis. Crow [2.58] modeled this by a Weibull
variations resulting from the manufacturing power-law process (type VI Weibull model). For models
process. Through proper process and quality control that are modification of this model, see Murthy et al.
during the manufacturing phase, these variations are [2.3].
controlled. In the manufacturing phase, the fraction of
In the post-launch stage, the reliability of an item nonconforming items is small when the process is in-
decreases due to deterioration resulting from age and/or control while it increases significantly when the process
usage. This deterioration is affected by several factors, goes out-of-control. the Weibull distribution has been
including the environment, operating conditions and used to model the in-control duration in the design of
maintenance. The rate of deterioration can be controlled control charts to detect the change from in-control to
Indexentriesonthispage

Weibull Distributions and Their Applications 2.6 Applications 13

Table2.2 A sample of reliability applications

November200510:08

PartA

Author(s)
Topic
Weibull [2.5]
Yieldstrengthofsteel,fatiguelifeofsteel
2.6

Keshevan et al. [2.49]

Fracturestrengthofglass
CET
SpringerHandbook

Sheikh et al. [2.50]

Pittingcorrosioninpipes
MSID:hb06-a2

Quereshi and Sheikh [2.51]

Adhesivewearinmetals
Durham and Padget [2.52]
Failureofcarbon-fibercomposites
4

Almeida [2.53]
Failureofcoatings
Fok et al. [2.54]
Failureofbrittlematerials
Newell et al. [2.55]
Failureofcompositematerials
Li et al. [2.56]
Concretecomponents
 out-of-control. For more on this, see Rahim [2.59],
Costa and Rahim [2.60], and Chen and Yang [2.61].
Nelson [2.62] deals with control charts for items with
Weibull failure distributions where conforming and
nonconforming items differ in the scale parameter but
have the same shape parameter. Murthy et al. [2.63] and
Djamaludin et al. [2.64] deal with lot production and
look at optimal lot size to control the occurrence of
nonconforming items.
When the failure rate has a bathtub shape, it is
prone
[Link]
exhibiting bathtub failure rates, burn-in is a
technique that can be used to weed out such failures
and improve product reliability before it is released
for sale. For more on burn-in, see, e.g., Kececioglu
and Sun [2.65].
Warranty cost analysis for products with a
Weibull failure distribution has received a lot of
attention. For more on this, see, e.g., Blischke and
Murthy [2.66,67] and Murthy and Djamaludin
[2.68].
Preventive maintenance of products with a
Weibull failure distribution has received
considerable attention.
Intheagepolicy,anitemisreplacedpreventivelywhenit
reaches some specified age and Tadikmalla [2.69]
deals
withthisinthecontextoftheWeibullfailuredistribution.
In the block policy, items are replaced preventively
at set clock times and Blischke and Murthy [2.70]
deals with this in the context of the Weibull failure
distribution. 2.6.2 Applications in Other Areas

Weibull distribution has been used as a model in diverse

disciplines to study many different issues. There are
SpringerHandbook
MSID:hb06-a2
 Indexentriesonthispage

14 Part A Fundamental Statistics and its Applications

Table2.3 A sample of other applications

Discipline Topic Author(s)
Geophysics Wind-speeddataanalysis Al-Hasan 2.71
Earthquakemagnitude Huillet and
Raynaud 2.72
Volcanicoccurrencedata Bebbington and
Lai 2.73
Low-flowanalysis Durrans 2.74
Regionalfloodfrequency Heo et al. 2.75
Food science Sterilityinthermalpreservationmethod Mafart et al. 2.76
Social science Unemploymentduration Roed and Zhang
2.77
Environment Environmentradioactivity Dahm et al. 2.78
Nature Ecologicalapplication Fleming 2.79
Medical science Survivaldata Carroll 2.80

several thousand References

papers and we • Weibull++ 6
give a very small
sample of this by ReliaSoft
vast literature
Corp.,
2.6.3 Weibull
Tucson, AZ;
Analysis
Software [Link]

In any analysis of [Link]/

statistical data,
computer • Relex
software is Weibull by
required. In
additionto Relex
standard Software
statisticalsoftwar Corp.,
esuch as Minitab,
SPSS, SAS, etc., Greensburg,
or spreadsheet PA;
software such as [Link]
Excel, some
specialized lexsoftware.c
Weibull analysis om/products /
software are also
[Link]
available. Below
is a list of some • WeibullPro
common ones.
SpringerHandbook

by Isograph
 Inc., Newport Weibull by
Beach, CA; Barringer &
[Link] Associates,
ograph- Inc., Humble,
[Link]/ TX;
[Link] [Link]
[Link]
m •
/[Link]
WinSMITH
2.1 A 2.2 N
.J .L
.J .
r. J
H o
al h
li n
n s
a o
n: n,
A S.
r K
e o
vi tz
e ,
w N
o .
f B
t al
h a
e kr
W is
ei h
b n
ul a
l n:
di C
st o
ri n
b ti
u n
ti u
o o
n, u
J. s
Q U
u n
al iv
. a
T ri
e a
c t
h e
n D
ol is
. tr
2 i
5, b
8 u
5 ti
– o
9 n
3 s,
(1 V
9 ol
SpringerHandbook

9 .1
3) ,
 2 A
n st
d a
e ti
d s
n. ti
( c
W al
il t
e h
y, e
N o
e ry
w o
Y f
o t
rk h
1 e
9 st
9 r
4) e
2.3 D n
. gt
N h
.P o
PartA

. f
M m
u at
2

rt e
h ri
y, al
M ,
. In
Xi g.
e, V
R. e
Ji t
a e
n n
g: s
W k
ei a
b p
u a
ll A
M c
o a
d d.
el H
s a
( n
W dl
il in
e g
y, a
N r
e 1
w 5
Y 1,
o 1
rk –
2 4
0 5
0 (1
November200510:08

3) 9
2.4 W 3
. 9)
W 2.5 W
hb06-

ei .
CET a2

b W
SpringerHandbook
MSI
D:

ul ei
l: b
4
 ul la
l: lo
A i
st d
a e
ti p
s r
ti o
c b
al
a
di
bi
st
lit
ri
é
b
u d
ti e
o l’
n é
f c
u a
n rt
c m
ti a
o xi
n m
o u
f m
w ,
id A
e n
a
n.
p
S
pl
o
ic
c.
a
bi P
lit ol
y, o
J. n
A ai
p s
pl e
. M
M at
e h.
c C
h. r
1 a
8, c
2 o
9 vi
3
e
–
6,
2
9
9
3
6
–
(1
9 1
5 1
1) 6
2.6 M (1927)
. 2.7 R.
Fr A
é .
Fi
c
s
h
h
e
e
t:
r,
S
L.
SpringerHandbook

u M
r .C
 . m
Ti b
p ri
p d
e g
tt e
: P
Li hi
m lo
iti s.
n S
g o
f c.
o 2
r 4,
m 1
s 8
o 0
f –
fr 1
e 9
q 0
u (1
e 9
n 2
c 8)
y 2.8 N
di .L
st .
ri J
b o
u h
ti n
o s
n o
o n,
f S.
t K
h o
e tz
la ,
r N
g .
e B
st al
o a
r kr
s is
m h
al n
le a
st n:
m C
e o
m n
b ti
e n
r u
o o
f u
a s
s U
a n
m iv
pl a
e, ri
P a
r t
o e
c. D
SpringerHandbook

C is
a tr
 i n
b al
u y
ti si
o s,
n J.
s, Q
V u
ol al
.2 .
, T
2 e
n c
d h
e n
d ol
n. .
( 3
W 0(
il 4)
e ,
y, 3
N 8
e 6
w –
York 3
1995) 9
2.9 W 4
.J (1
. 9
Zi 9
m 8)
m 2.10 C.
e D
r, .
J. L
B. ai
K ,
e M
a .
ts Xi
, e:
F. R
K. el
W a
a ti
n v
g: e
T a
h g
e ei
B n
u g
rr f
XI o
I r
di t
st w
ri o
b p
u a
ti r
o al
n le
in l
r s
el y
ia st
bi e
lit m
SpringerHandbook

y s
a a
 n e
d ct
r e
el d
a lif
t e
e ti
d m
p e
r b
o y
bl p
e a
m r
s, al
M le
a l
t r
h. e
C d
o u
m n
p. d
M a
o n
d c
el y,
. A
3 si
8, a
1 P
3 a
3 ci
9 fi
– c
1 J.
3 O
4 p
5 e
(2 r.
0 R
0 e
3) s.
2.11 M 1
. 3,
Xi 1
e, 7
C. 1
D –
. 1
L 7
ai 9
: (1
O 9
n 9
t 6)
h 2.12 J.
e F.
in L
cr a
e w
a le
s ss
e :
o S
f
t
t
a
h
ti
e
s
e
SpringerHandbook

x ti
p c
 a (
l W
M il
o e
d y,
el N
s e
a w
Y
n
o
d
rk
M
1
e
9
t
8
h 2)
o 2.14 M
d .
sf Li
o a
r o,
Li T.
f S
e hi
ti m
m o
e k
D a
a w
t a:
a A
n
(
e
W
w
il
g
e
o
y,
o
N d
e n
w e
Y ss
o -
rk o
1 f-
9 fi
8 t
2) t
2.13 W e
. st
N f
el o
s r
o T
n: y
A p
p e-
p I
li e
e xt
d r
Li e
f m
e e-
D v
a al
t u
a e,
A 2-
n p
a a
r
SpringerHandbook

ly
si a
s m
 e .
t T
e h
r o
W m
ei p
b s
ul o
l n:
di W
st ei
ri b
b ul
u l
ti p
o r
n o
s b
w a
it bi
h lit
e y
s p
ti a
m p
a e
t rs
e ,
d J.
p Q
a u
r al
a .
m T
e e
t c
e h
rs n
, ol
J. .
St 3,
a 1
t. 4
C 0
o –
m 1
p 4
u 6
t. (1
Si 9
m 7
. 1)
6 2.16 W
4, .
2 N
3 el
– s
4 o
8 n:
(1 T
9 h
9 e
9) o
2.15 W ry
. a
N n
el d
s a
o p
n, pl
SpringerHandbook

V ic
.C a
 ti h,
o R.
n L.
o P
f r
h e
a n
z ti
a c
r e:
d T
pl h
o e
t S
ti t
n a
g ti
f s
o ti
r c
c a
e l
n A
s n
o a
r ly
e si
d s
fa o
il f
u F
r a
e il
d u
a r
t e
a, D
J. a
Q t
u a
al (
. W
T il
e e
c y,
h N
n e
ol w
. Y
1 o
4, rk
9 1
3 9
5 8
– 0)
9 2.18 A
6 .C
6 .
(1 C
9 o
7 h
2) e
2.17 J. n,
D B.
. J.
K W
al hi
b tt
fl e
ei n,
SpringerHandbook

s Y.
c D
 in –
g: 1
M 6
o 7
di (1
fi 9
e 8
d 4)
m 2.19 S.
o H
m .
e Z
n a
t n
e a
s ki
ti s,
m J.
a K
ti y
o p
n a
f ri
o si
r s:
t A
h r
e e
t vi
h e
r w
e o
e- f
p m
a a
r xi
a m
m u
e m
t li
e k
r el
W ih
ei o
b o
ul d
l e
di s
st ti
ri m
b a
u ti
ti o
o n
n, m
J. e
Q t
u h
al o
. d
T s
e f
c o
h r
n t
ol h
. e
1 t
6, h
1 r
SpringerHandbook

5 e
9 e
 p y,
a M
r il
a w
m a
e u
t k
e e
r e
W 1
ei 9
b 9
ul 4)
l s
di e
st e
ri al
b s
u o
ti 2.
o 8
n, 1
J. 2.21 M
St .
a Xi
t. e,
C C.
o D
m .
p. L
Si ai
m ,
ul D
. .
2 N
5, .P
5 .
3 M
– u
7 rt
3 h
(1 y:
9 W
8 ei
6) b
2.20 B. ul
D l-
o r
d el
s at
o e
n: d
W di
ei st
b ri
u b
ll u
A ti
n o
a n
ly s
si f
s o
( r
A m
S o
Q d
C el
Q li
u n
SpringerHandbook

al g
it o
 f y,
b e
a d.
t b
h y
t B.
u H
b .
s Li
h n
a d
p q
e vi
d st
fa ,
il K.
u A
r .
e D
r o
a k
t s
e u
f m
u (
n W
c o
ti rl
o d
n S
s. ci
In e
: n
M ti
a fi
t c,
h Si
e n
m g
a a
ti p
c o
a r
l e
a 2
n 0
d 0
S 3)
t p
a p.
ti 2
s 8
ti 3
c –
a 2
l 9
M 7
e 2.22 C.
t D
h .
o L
d ai
s ,
i L.
n Y.
R Z
el h
i a
a n
b g,
SpringerHandbook

ili M
t .
 Xi n
e: s,
M In
e t.
a J.
n R
r el
e ia
si b.
d Q
u u
al al
lif .
e S
a af
n .
d E
o n
t g.
h 1
e 1,
r TS
4
p
r (2
o 0
p 0
e 4)
r 2.23 D
ti .
e N
s .P
o .
f M
W u
ei rt
b h
ul y,
l J.
r B
el ai
a k,
t M
e .
d B
b ul
a m
t e
h r,
t R.
u J.
b W
s il
h s
a o
p n:
e T
fa w
il o
u -
r di
e m
r e
a n
t si
e o
di n
st al
ri m
b o
u d
SpringerHandbook

ti el
o li
 n B
g e
o rli
f n
fa 2
il 0
u 0
r 4)
e 2.24 A
s. .
In W
: .
S M
p a
ri rs
n h
g al
e l,
r I.
H O
a lk
n in
d :
b A
o n
o e
k w
O m
f e
E t
n h
g o
i d
n f
e o
e r
ri a
n d
g di
S n
t g
a a
ti p
s a
ti r
c a
s, m
e e
d. t
b e
y r
H t
. o
P af
h a
a m
m il
(S y
p o
ri f
n di
g st
e ri
r, b
H u
ei ti
d o
el n
b s
e w
SpringerHandbook

r it
g, h
 a M
p u
pl rt
ic h
a y:
ti M
o o
n di
t fi
o e
t d
h W
e ei
e b
x ul
p l
o m
n o
e d
n el
ti ,
al IE
a E
n E
d T
W r
ei a
b n
ul s.
l R
fa el
m ia
ili b.
e 5
s, 2,
Bi 3
o 3
m –
e 3
tr 7
ik (2
a 0
8 0
4, 3)
6 2.26 G
4 .S
1 .
– M
6 u
5 d
2 h
(1 ol
9 k
9 a
7) r,
2.25 C. D
D .K
. .
L Sr
ai iv
, a
M st
. a
Xi v
e, a:
M E
. x
D p
. o
N n
SpringerHandbook

.P e
. n
 ti n
a g,
t D
e .
d N
W .P
ei .
b M
ul u
l rt
fa h
m y:
il T
y h
f e
o e
r x
a p
n o
al n
y e
zi n
n ti
g at
b e
a d
t W
h ei
t b
u ul
b l
fa fa
il m
u il
r y:
e- A
r g
a r
t a
e p
d hi
a c
t al
a, a
IE p
E p
E r
T o
r a
a c
n h,
s. IE
R E
el E
ia T
b. r
4 a
2, n
2 s.
9 R
9 el
– ia
3 b.
0 4
2 8,
(1 6
9 8
9 –
3) 7
2.27 R. 2
SpringerHandbook

Ji (1
a 9
 9 G
9) o
2.28 J. h,
A Y.
. T
Ki a
e n
s: g:
T A
h m
e o
S di
tr fi
e e
n d
g W
t ei
h b
o ul
f l
G e
l xt
a e
s n
s, si
R o
e n
p w
o it
rt h
, b
V at
ol h
.5 t
0 u
9 b
3 -
( s
N h
a a
v p
al e
R d
e fa
s il
e u
a r
rc e
h r
L at
a e
b. f
, u
W n
a c
s ti
hi o
n n,
gt R
o el
n ia
1 b.
9 E
5 n
8) g.
2.29 M S
. y
Xi st
e, .
T. S
SpringerHandbook

N af
. .
 7 8
6, 1
2 (1975)
7 2.31 Z.
9 C
– h
2 e
8 n:
5 A
(2 n
0 e
0 w
2) t
2.30 R. w
M o
. -
S p
m a
it r
h, a
L. m
J. e
B t
ai e
n: r
A di
n st
e ri
x b
p u
o ti
n o
e n
n w
ti it
al h
p b
o at
w h
e t
r u
lif b
e s
t h
e a
s p
ti e
n o
g r
di in
st cr
ri e
b a
u si
ti n
o g
n, fa
C il
o u
m r
m e
u r
n. at
St e
a f
t. u
4, n
4 c
6 ti
9 o
SpringerHandbook

– n,
4 St
 a l
t. di
P st
r ri
o b
b u
a ti
b. o
L n
e s:
tt A
. g
4 r
9, a
1 p
5 hi
5 c
– al
1 a
6 p
1 p
(2 r
0 o
0 a
0) c
2.32 R. h,
Ji IE
a E
n E
g, T
D r
. a
N n
.P s.
. R
M el
u ia
rt b.
h 4
y: 4,
M 4
o 7
d 7
el –
in 4
g 8
fa 8
il (1
u 9
r 9
e 5)
d 2.33 R.
a Ji
t a
a n
b g,
y D
m .
ix N
t .P
u .
r M
e u
o rt
f h
t y:
w M
o ix
W t
ei u
SpringerHandbook

b r
ul e
 Indexentriesonthispage

Weibull Distributions and Their Applications References 15

o o
f c
W h
ei a
b s
ul ti
l c
di M
st o

PartA
ri d
b el
u s
ti D

2
o at
n a
s- A
p n
a al
r .
a 1
m 4,
e 4
tr 7
ic –
c 6
h 5
a (1
r 9
a 9
ct 8)
e J. Gurland, J.
ri Sethuraman:
z How pooling
a failure data
ti may reverse
o increasing
n failure rates,
o J. Am. Stat.
f Ass. 90,
fa 1416–1423
il (1995)
u R. Jiang,
r D.N.P.
e Murthy:
r Parametric
a study of
t competing
e risk model
f involving
u two Weibull
n distributions
c , Int. J.
ti Reliab.
o Quality Saf.
SpringerHandbook

n, Eng. 4, 17–
St 34 (1997)
 2.36 R. 2.38 R.
Jian Jian
g, g,
D.N. M.
P. Zuo,
Mur D.N.
thy: P.
Para Mur
met thy:
ric Two
stud secti
y of onal
mul mod
tipli els
cati invo
ve lvin
mod gtw
el o
invo Wei
lvin bull
g distr
two ibuti
Wei ons,
bull Int.J
distr .Reli
ibuti ab.
ons, Qua
Reli l.
ab. Saf.
Eng. Eng.
Sys. 6,
Saf. 103
55, –
217 122
– (199
226 9)
(199 2.39 W.
7) Nels
2.37 R. on:
Jian Acc
g, eler
D.N. ate
P. d
Mur Tes
thy: ting
Reli (Wil
abili ey,
ty New
mod York
elin 1990)
g 2.40 T.
invo Nak
lvin aga
g wa,
two S.
Wei Osa
bull ki:
distr The
ibuti disc
ons, rete
Reli Wei
ab. bull
Eng. distr
Sys. ibuti
Saf. on,
47, IEEE
187 Tran
– s.
198 Reli
(199
SpringerHandbook

ab.
5) 24,
300
 – tion
301 proc
(197 ess,
5) J.
2.41 W.E Am.
. Stat.
Stei Ass.
n, R. 68,
Datt 575
ero: –
A 578
new (197
disc 3)
rete 2.44 L.J.
Wei Bain
bull , M.
distr Eng
ibuti elha
on, rt:
IEEE Sta
Tran tisti
s. cal
Reli Ana
ab. lysis
33, of
196 Reli
– abili
197 ty
(198 and
4) Life-
2.42 W.J. testi
Pad ng
gett, Mo
J.D. dels,
Spur 2nd
rier: edn.
Disc (Ma
rete rcel
failu Dek
re ker,
mod New
els, York
IEEE 199
Tran 1)
s. 2.45 D.R.
Reli Cox,
ab. D.D.
34, Oak
253 es:
– Ana
256 lysis
(198 of
5) Surv
2.43 W. ival
M. Dat
Bass a
in: A (Cha
Bay pma
esia n
n Hall,
opti New
mal York
over 198
haul 4)
mod 2.46 N.
el Yan
for naro
the s:
Wei Wei
SpringerHandbook

bull bull
rest rene
ora wal
 proc Con
esse rad:
s, Stati
Ann. stica
Inst. l
Stat. anal
Mat ysis
h. of
46, the
641 Hert
– zian
648 frac
(199 ture
4) of
2.47 D.N. pyre
P. x
Mur glas
thy, s
M. usin
Bul g
mer the
, J.E. Wei
Eccl bull
esto distr
n: ibuti
Wei on
bull func
mod tion,
ellin J.
g, Mat
Reli er.
ab. Sci.
Eng. 15,
Sys. 839
Saf. –
5
TS , 844
(200 (198
4) in 0)
prin 2.50 A.K.
t Shei
2.48 W.Q kh,
. J.K.
Mee Boa
ker, h,
L.A. D.A.
Esco Han
bar: sen:
Sta Stati
tisti stica
cal l
Met mod
hod ellin
s for g of
Reli pitti
abili ng
ty corr
Dat osio
a n
(Wil and
ey, pipe
New line
York relia
199 bilit
8) y,
2.49 K. Corr
Kes osio
hev n
an, 46,
G. 190
–
SpringerHandbook

Sarg
ent, 196
H.
 (199 2.53 J.B.
0) Alm
2.51 F.S. eida
Que :
eshi Appl
, icati
A.K. on
Shei of
kh: Weil
Pro bull
babi stati
listic stics
char to
acte the
riza failu
tion re
ofad of
hesi coa
vew ting
ear s, J.
inm Mat
etal er.
s, Proc
IEEE es.
Tran Tec
[Link] hnol
liab. . 93,
46, 257
38– –
44 263
(199 (199
7) 9)
2.52 S.D. 2.54 S.L.
Dur Fok,
ham B.C.
, Mitc
W.J. hell,
Pad J.
gett: Sma
Cum rt,
ulati B.J.
ve Mar
dam sde
age n: A
mod num
el eric
for al
syst stud
em y on
failu the
re appl
with icati
appl on
icati of
on the
to Wei
carb bull
on theo
fibe ry
rs to
and britt
com le
posi mat
tes, erial
Tec s,
hno Eng.
met Frac
rics t.
39, Mec
34– h.
SpringerHandbook

44 68,
(1997) 117
 1– approa
117 ch to
9 reliabili
(200 ty
1) monito
J.A. ring,
IEEE

PartA
Trans.
Aerosp.
40,

2
563–
566
(1964)
Newell, 2.58 L.H.
T. Crow:
Kurzeja Reliabili
, M. ty
Spence, analysis
M. for
Lynch: comple
Analysi x
s of system
recoil s. In:
compre Reliabil
ssive ity and
failure Biomet
in high ry, ed.
perfor by F.
mance Prosch
polyme an, R.J.
rs using Serfling
two-, (SIAM,
otations (will be removed before the final TeX run)
four- Philade
parame lphia
ter 1974)
Weibull pp.379
models –410
, High 2.59 M.A.
Perfor Rahim:
m. Econo
Polym. mic
14, design
425– of X
434 charts
(2002) assumi
2.56 Q.S. Li, ng
J.Q. Weibull
Fang, in-
D.K. control
Liu, J. times,
Tang: J. Qual.
Failure Techno
probabi l. 25,
lity 296–
predicti 305
on of (1993)
concret 2.60 A.F.B.
e Costa,
compo M.A.
nents, Rahim:
Cemen Econo
t mic
Concre design
te Res. of X
33, and R
1631– charts
1636 under
(2003) Weibull
2.57 J.T. shock
SpringerHandbook

Duane: models
Learnin , Qual.
g curve Reliab.
 Eng. , R.J.
Int. 16, Wilson:
143– Lotsizin
156 g and
(2000) testing
2.61 Y.S. for
Chen, items
Y.M. with
Yang: uncerta
Econo in
mic quality,
design Math.
of x- Comp.
control Model.
charts 22, 35–
with 44
Weibull (1995)
in- 2.65 D.
control Kececio
times, glu,
when F.B.
there Sun:
are Burn-in
multipl Testing
e : Its
assigna Quanti
ble fication
causes, and
Int. J. Optimi
Prod. zation
Econ. (Prentic
77, 17– e Hall,
23 Upper
(2002) Saddle
2.62 P.R. River
Nelson: 1997)
Control 2.66 W.R.
chart Blischk
for e,
Weibull D.N.P.
process Murthy
, IEEE :
Trans. Produc
Reliab. t
28, Warra
283– nty
288 Handb
(1979) ook
2.63 D.N.P. (Marcel
Murthy Dekker,
, I. New
Djamal York
udin, 1994)
R.J. 2.67 W.R.
Wilson: Blischk
Produc e,
t D.N.P.
warran Murthy
ty and :
quality Warra
control, nty
Qual. Cost
Reliab. Analysi
Eng. 9, s
431– (Marcel
443 Dekker,
(1993) New
2.64 I. York
Djamal 1996)
SpringerHandbook

udin, 2.68 D.N.P.

D.N.P. Murthy
Murthy , I.
 Indexentriesonthispage

16 Part A Fundamental Statistics and its Applications

Djamal R.R.
udin: Nigmat
Produc ullin:
t Identifi
warran cation
ty: A of the
review, general
Int. J. ized
Prod. Weibull
Econ. distribu
79, tion in
231– wind
260 speed
(2002) data by
2.69 P.R. the
Tadikm Eigen-
alla: coordin
Age ates
replace metho
ment d,
policies Renew
for able
Weibull Energ.
failure 28(1),
93–110
November200510:08

(2003)
2.72 T.
Huillet,
H.F.
hb06-

Raynau
CETa2

d: Rare
MSI

events
D:

in a
logWei
bull
4

rates, scenari
IEEE o—
Trans. Applica
Reliab. tion to
29, 88– earthq
90 uake
mag-
(1980)
nitude data, Eur.
2.70 W.R.
Blischk Phys. J. B 12, 457–
e, 469 (1999)
D.N.P. 2.73 M.S.
Murthy Bebbin
: gton,
Reliabil C.D.
ity Lai: On
(Wiley, nonho
New mogen
York eous
2000) models
SpringerHandbook

2.71 M. Al- for

Hasan, volcani
 c ds:
eruptio applica
ns, tion of
Math. the
Geol. Weibull
28, freque
585– ncy
599 distribu
(1996) tion
2.74 S.R. model,
Durran Int. J.
s: Low- Food
flow Microbi
analysis ol. 72,
with 107–
acondi 113
tional (2002)
Weibull 2.77 K.
tail Roed,
model, T.
Water Zhang:
Resour. A note
Res. 32, on the
1749– Weibull
1760 distribu
(1996) tion
2.75 J.H. and
Heo, time
D.C. aggreg
Boes, ation
J.D. bias,
Salas: Appl.
Region Econ.
al flood Lett. 9,
freque 469–
ncy 472
analysis (2002)
based 2.78 H.
on a Dahm,
Weibull J.
model: Niemey
Part 1. er, D.
Estima Schrod
tion, er:
asympt Applica
otic tion of
varianc the
es, J. Weibull
Hydrol. distribu
242, tion to
157– describ
170 e the
(2001) vertical
2.76 P. distribu
Mafart, tion of
O. cesium-
Couver 137 on
t, S. a slope
Gaillard under
, perman
[Link] ent
rinel: pasture
On in
calcula Luxem
ting bourg,
sterility J.
in Environ
therma .
l Radioa
c. 63,
SpringerHandbook

preserv
ation 207–
metho
 219 Model.
(2002) 138(1-
2.79 R.A. 3),
Flemin 309–
g: The 319
Weibull (2001)
model 2.80 K.J.
and an Carroll:
ecologi On the
cal use and
applica utility
tion: of the
describi Weibull
ng the model
dynami in the
cs of analysis
foliage of
biomas survival
s on data,
Scots Control
pine, . Clin.
Ecol.
View publication stats

Trials 24, 682–701 (2003)

2.81 N.L. Johnson, S. Kotz, N.
Balakrishnan: Continuos
Univariate Distributions, Vol.1,
2nd edn. (Wiley, New
York 1994) pp.657–661
SpringerHandbook