Reliability Engineering and System Safety 92 (2007) 719726

A exible Weibull extension

Mark Bebbington
a,
, Chin-Diew Lai
a
, Ric ardas Zitikis
b
a
Institute of Information Sciences and Technology, Massey University, Private Bag 11222, Palmerston North, New Zealand
b
Department of Statistical and Actuarial Sciences, University of Western Ontario, London, Ont., Canada N6A 5B7
Received 1 December 2005; received in revised form 27 March 2006; accepted 28 March 2006
Available online 12 May 2006
Abstract
We propose a new two-parameter ageing distribution which is a generalization of the Weibull and study its properties. It has a simple
failure rate (hazard rate) function. With appropriate choice of parameter values, it is able to model various ageing classes of life
distributions including IFR, IFRA and modied bathtub (MBT). The ranges of the two parameters are clearly demarcated to separate
these classes. It thus provides an alternative to many existing life distributions. Details of parameter estimation are provided through a
Weibull-type probability plot and maximum likelihood. We also derive explicit formulas for the turning points of the failure rate function
in terms of its parameters. This, combined with the parameter estimation procedures, will allow empirical estimation of the turning
points for real data sets, which provides useful information for reliability policies.
r 2006 Elsevier Ltd. All rights reserved.
Keywords: Reliability; IFR; IFRA; Bathtub shape; Weibull-type probability plot; Maximum likelihood estimates
1. Introduction
Motivated by engineering applications, Weibull [1]
suggested a distribution that has proved to be of seminal
importance in reliability. The survival function is given by
the equation

Ft expt=b
a
; t40, (1)
with parameters a; b40. The corresponding failure rate
function is
ht
a
b
t
b
_ _
a1
. (2)
The Weibull distribution has been widely used to model the
failures of many materials, and in numerous other
applications. A very small sample of the vast literature
includes applications to the yield strength and fatigue life
of steel [1], fracture strength of glass [2], pitting corrosion
in pipes [3], adhesive wear in metals [4], and failure of
carbon bre composites [5], coatings [6], brittle materials
[7], composite materials [8], and concrete components [9].
Recently, it has been applied in mixture models, particu-
larly for automobile warranty data [1012].
A recent paper by Murthy et al. [13] discusses additional
applications, and gives a methodological review of the
Weibull area. It also suggests further study of various
Weibull-type distributions, their properties, related plots,
and model selection. In this paper we introduce yet another
member of the Weibull family, which we call the exible
Weibull distribution. We will rst dene it mathematically,
then examine its properties, and nally details of its
application.
We see from (2) that the Weibull distribution has a
monotonic failure rate function, although this may be
increasing or decreasing. In more complex systems, such as
electronic ones, the failure rate is often non-monotonic.
This usually takes the form of increased failure rate
early (wear-in) and late (wear-out) in the component
lifetime. This is usually termed a bathtub-shaped failure
rate, or if the limit of the failure rate at time zero is zero, a
modied bathtub (MBT) (or roller coaster)-shaped
failure rate. Given the utility of the bathtub-shaped failure
rate functions in reliability engineering, many of the
variations on the Weibull distribution have been motivated
by the desire to produce a bathtub-shaped failure rate
function.
ARTICLE IN PRESS
www.elsevier.com/locate/ress
0951-8320/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ress.2006.03.004

Corresponding author. Tel.: +64 6 3569099x2452; fax: +64 6 3502261.

E-mail address: M.Bebbington@massey.ac.nz (M. Bebbington).
Gurvich et al. [14] introduced a class of distributions
characterized by the cumulative distribution function
Ft 1 expfaGtg; t40, (3)
with parameter a40, where Gt is a monotonically
increasing function of t. Equivalently, the survival function
is

Ft expfaGtg.
Nadarajah and Kotz [15] have since shown that several
existing life distributions, such as the modied Weibull
extension Ft 1 explat=a
b
1, see [16], may be
expressed in the form (3). Clearly, (3) is a generalized
Weibull distribution. Xie et al. [17] reviewed several
families of extended Weibull distributions, while a com-
prehensive taxonomy of Weibull models can be found
in [18].
In this paper, we propose a new life distribution of a
similar form to (3), with the difference that Gt is not a
monotonic function of t. We will completely characterize
the failure rate function, and consider parameter estima-
tion. The new distribution is shown to be quite exible,
being able to model both IFR and IFRA ageing classes.
Also, it can yield a MBT-shaped failure rate distribution,
and in particular, allows considerable exibility in model-
ling the pre-useful (i.e., infancy) period.
2. The new exible Weibull distribution
We will dene our model in terms of the survival
function. Let T be a lifetime random variable, and

Ft expe
atb=t
; t40, (4)
with parameters a; b40: Clearly, this has the form of
distribution (3) with a 1 and
Gt e
atb=t
. (5)
However, function (5) is not monotonic in t and so the
distribution under consideration differs from that in (3)
considered by Gurvich et al. [14]. Note that when b 0, if
we set a logl, distribution (4) becomes exponential and
thus the proposed life distribution may be regarded as a
generalization of the Weibull.
The density function corresponding to (4) is
f t a b=t
2
expat b=t expe
atb=t
.
Though formulas for the mean and the variance are
difcult to obtain explicitly, the quantiles are easy to
evaluate. Let t
p
be the pth quantile of T. By considering the
loglog transformation of p

Ft, we have loglog p
at
2
b=t and so t
p
is a solution of the quadratic equation
in t
at
2
loglog p t b 0.
Since the solutions have to be non-negative, the only one is
t
p

1
2a
loglog p

floglog pg
2
4ab
_
_ _
.
The failure rate function has a reasonably simple form
ht
f t

Ft
a b=t
2
expat b=t. (6)
The shape of the density and failure rate function are
illustrated for selected values of a and b in Figs. 1 and 2.
For reference, the solid curve is the same in both gures.
We see that as b decreases, the failure rate function
becomes more bathtub-like. While, as a increases, the
bathtub becomes shallower. In particular, we see that the
model has a great deal of exibility in the existence and
weight of the failure mode corresponding to wear-in.
Various properties of f t and ht that follow from their
denitions and/or can be seen in Figs. 1 and 2 are discussed
in detail in the next section.
3. Ageing behaviour
We note that lim
t!1
ht 1, thus the failure rate
function is ultimately increasing. Also, with the notation
s 1=t for computational convenience,
lim
t!0
ht lim
s!1
a bs
2
expbs a=s
lim
s!1
2bsb a=s
2

1
expbs a=s
lim
s!1
2bb 2a=s
2
2a=s
3
a
2
=s
4

1
expbs a=s
0,
and thus a pure bath-tub form is impossible.
The rest of this section is subdivided into three
subsections in which we consider more subtle properties
of the failure rate function ht. First we determine the
values of a and b for which ht is increasing, that is,
belongs to the IFR class, and then we determine the values
of a and b for which ht belongs to the IFRA class.
Interestingly, we shall show that there are values of a and b
for which ht is IFRA but not IFR. Finally, we will
consider in detail the case when ht is not IFR, and show
that ht takes on the MBT shape. Such failure rate
functions are particularly useful models in many practical
situations (cf., e.g., [19]).
3.1. IFR
We need to determine the parameter values for which the
failure rate function ht is an increasing function, that is,
belongs to the IFR class. Naturally, we solve the problem
by specifying those a and b for which h
0
t is strictly
positive for all tX0.
Consider the derivative of the failure rate function ht. It
follows from (6) that
h
0
t
2b
t
3
e
atb=t

at
2
b
t
2
a b=t
2
e
atb=t

at
2
b
2
2bt
t
4
e
atb=t
.
ARTICLE IN PRESS
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 720
Clearly the sign of h
0
t is the same as the sign of
at
2
b
2
2bt. Let a gb. Then
at
2
b
2
2bt b
2
gt
2
1
2
2bt40
if and only if h
0
t40. This requires b
2
gt
2
1
2
42bt or
gt
2
142t=b
1=2
(note that t; b40), or
gt gt
2

2t
b
_ _
1=2
140. (7)
ARTICLE IN PRESS
0 1 2 3 4 5 6 7 8 9 10
0
0.2
0.4
0.6
0.8
1
t
f
(
t
)
0 1 2 3 4 5 6 7 8 9 10
0
0.5
1
1.5
2
t
h
(
t
)
Fig. 1. Density (top) and failure rate (bottom) as a function of b: b
3
8
(dashed), b
9
8
(solid), b
27
8
(dash-dot), b
81
8
(dotted); a
1
8
everywhere.
0 1 2 3 4 5 6 7 8 9 10
0
0.2
0.4
0.6
0.8
1
t
f
(
t
)
0 1 2 3 4 5 6 7 8 9 10
0
0.5
1
1.5
2
t
h
(
t
)
Fig. 2. Density (top) and failure rate (bottom) as a function of a: a
1
8
(solid), a
2
8
(dot-dash), a
3
8
(dashed), a
4
8
(dotted); b
9
8
everywhere.
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 721
Now in order for ht to be IFR, this condition must hold
for all t40. It is easy to see that gt has a minimum when
2gt 2bt
1=2
0, or at t

8g
2
b
1=3
. Thus, the dis-
tribution is IFR if and only if
0ogt

g8g
2
b
2=3
2=b
1=2
8g
2
b
1=6
1
1 0:75g
1=3
b
2=3
,
or, in other words, if and only if gb
2
4
27
64
. Recalling that
g a=b, we nd that the distribution is IFR if and only if
ab4
27
64
. Since lim
t!1
ht 1, it is not possible to give rise
to a DFR distribution.
3.2. IFRA
A distribution F is said to be is IFRA if and only if
_
t
0
hs ds=t log

Ft=t is increasing in t. The IFRA
class features prominently in the classic text [20]. An
important reason for its popularity is due to the closure
property which states that a coherent system has an IFRA
life distribution if each of the independent components has
an IFRA distribution. It is well known that an IFR implies
IFRA, that is, an IFR class is contained in an IFRA class.
There are few known life distributions that are IFRA but
not IFR (i.e. ht is not strictly increasing for some interval
of t). Our new distribution has this property with suitable
choice of a and b. To see this we rst note that
log

Ft
t

e
atb=t
t
.
Let s 1=t and us se
bsa=s
. Then
u
0
s 1 sb a=se
bsa=s
,
and thus log

Ft=t is increasing in t if and only if
u
0
sp0, i.e., 1 sb a=sp0. The latter is equivalent to
s
2
s=b a=bX0: Now
s
2
s=b a=b s
1
2b
_ _
2

a
b

1
4b
2
which is greater than or equal to zero for all s if and only if
a=b 1=4b
2
X0. That is, the distribution is IFRA if and
only if abX
1
4
, which includes the IFR case, as required.
We see that if
1
4
oabo
27
64
, then the distribution is IFRA, but
not IFR.
3.3. Modied bathtub shape
It is widely believed that many products, particularly
electronic items, exhibit a bathtub-shaped failure rate
function. This belief is supported by much experience and
extensive data collection in many industries. Essentially, a
bathtub-shaped failure rate function is decreasing during
the infant mortality phase, approximately constant during
the useful life phase, and nally, increasing during the
wearout phase. Several Weibull-related distributions that
can model a bathtub phenomenon are given in [17].
A comprehensive review of this important ageing class
can be found in [21].
Some researchers (see, e.g., [22]), upon collecting failure
data found that component failure rates often follow a
more complex pattern than those described by bathtub
curves. The modications of a bathtub curve are known
as generalized/modied bathtub curves. The major
difference between the traditional and the modied bath-
tub curves is the behaviour during their infant mortality
phase. Jensen and Petersen [19] suggested a two-stage
model of infant mortality: the rst stage has an increasing
failure rate, indicating failures that rise from comparatively
coarse defects, such as those from imperfect manufactur-
ing, improper handling, or defective control processes.
Such a failure rate in this rst stage peaks quickly and is
followed by a period of decreasing failure rate and then
increasing. In another words, the failure rate function
increases at the beginning and then follows a bathtub
shape. We use the abbreviation MBT to denote a modied
bathtub shape and refer to a distribution F as a MBT
distribution if its failure rate has a MBT shape. We note
that MBT distributions are often found in mixture
distributions such as mixtures of Gammas and mixtures
of Weibulls (cf., e.g., [2326]).
Coming now back to our distribution in (4), we rst
see that in view of the fact lim
t!0
ht 0, as proved
earlier, a bath-tub shape is not possible. However, we
have also proved above that ht is IFR if and only if
ab4
27
64
. It is obvious from the same proof that if abo
27
64
then, being a parabola, gt 0 as dened in (7) will have
two solutions, and thus ht must be initially increasing,
and then follow a bathtub shape; decreasing again and
ultimately increasing. Hence, when abo
27
64
, distribution (4)
is MBT.
Some further detail about the distribution (4) can
be obtained by performing some tedious but elementary
computations. Namely, when abo
27
64
, then the failure
rate function ht is strictly increasing on the interval
0; t
1
, strictly decreasing on t
1
; t
2
, and strictly increas-
ing on t
2
; 1. The points t
1
and t
2
are given by the
formulas:
t
1

1
2

4b
3a
A B
_

1
2

8b
3a
A B 4b
_
a
2

4b
3a
A B
_
_ _

_
,
8
t
2

1
2

4b
3a
A B
_

1
2

8b
3a
A B 4b
_
a
2

4b
3a
A B
_
_ _

_
,
9
ARTICLE IN PRESS
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 722
where
A:
2
11=3
b
2
3 27a
2
b
2
32a
3
b
3
3

3
p

27a
4
b
4
64a
5
b
5
_
_ _
1=3
,
B:
2
1=3
27a
2
b
2
32a
3
b
3
3

3
p

27a
4
b
4
64a
5
b
5
_
_ _
1=3
3a
2
.
We show in the Appendix that when a gets larger, then
both t
1
and t
2
get smaller, provided that the product ab
remains xed. In this case, the rate of convergence to 0 of t
1
and t
2
is of the order O1=a. At the threshold ab
27
64
, we
nd that t
1
t
2
3=8a, and thus ht is strictly increasing
everywhere except at t 3=8a, where it attens momen-
tarily. This supplements Figs. 1 and 2, where either a or b is
xed and the other parameter is allowed to vary.
4. Parameter estimation
We will briey outline two simple methods of estimating
the parameters a and b, and illustrate them using a real
data set. We will suppose that we observe failure times
t
1
; . . . ; t
n
.
4.1. Weibull-type probability plot
In many practical applications, a graphical approach is
often utilized to provide initial estimates of the unknown
parameters. For most Weibull-related distributions, a
generalized Weibull probability plot can be constructed.
It follows from (4) that a loglog transformation of the
survival function

Ft gives
y loglog

Ft at b=t.
Letting x
1
t and x
2
1=t, we have a linear regression
y ax
1
bx
2
. Letting t take on the observed values
t
1
; :::; t
n
, we obtain the corresponding values of x
1
, x
2
,
and y, which we denote by x
i1
, x
i2
, and y
i
, for 1pipn. Let
x
1
, x
2
, and y be the column vectors with the coordinates
x
i1
, x
i2
, and y
i
, respectively. Let
x
1
x
2 denote the n 2
matrix with the two columns x
1
and x
2
. Then we have the
equation
y x
1
x
2

a
b
_ _
X
a
b
_ _
,
whose least-squares solution is of course
^ a
^
b
_ _
X
T
X
1
X
T
y.
4.2. Maximum likelihood estimation
The likelihood function is
L

n
i1
f t
i

n
i1
a
b
t
2
i
_ _
exp a

n
i
t
i

n
i1
b
t
i
_ _
exp

n
i1
e
at
i
b=t
i

_ _
.
Let L denote the log-likelihood function,
L log L

n
i1
log a
b
t
2
i
_ _
a

n
i1
t
i

n
i1
b
t
i
_ _

n
i1
e
at
i
b=t
i

. 10
Then
qL
qa

n
i1
1
a b=t
2
i

n
i1
t
i

n
i1
t
i
e
at
i
b=t
i

n
i1
1
a b=t
2
i
t
i
1 e
at
i
b=t
i

_ _
and
qL
qb

n
i1
1
at
2
i
b

n
i1
1
t
i

n
i1
1
t
i
e
at
i
b=t
i

n
i1
1
at
2
i
b

1
t
i
1 e
at
i
b=t
i

_ _
.
The maximum likelihood estimates ^ a and
^
b can then be
obtained by solving the simultaneous equations
qL
qa
0;
qL
qb
0,
or by direct maximization of (10).
5. An example
The data in Table 1, for the time between failures of
secondary reactor pumps [27] produces estimates of ^ a
0:0207;
^
b 2:5875 by the MLE method, and ^ a
0:0191;
^
b 2:0384 by the Weibull-type probability plot
(regression) method. The tted survival and failure rate
functions are shown in Fig. 3. We see the clear shape of a
modied bathtub distribution.
For comparison purposes, we will consider some
alternative models. First, of course, is the standard Weibull
distribution (1). The modied Weibull distribution [28]
has survival function

Ft expat
b
e
lt
; t40,
with parameters a40, bX0 and l40. (Note that the
distribution reduces to the Weibull when l 0.) This
distribution has a failure rate function
ht ab ltt
b1
e
lt
.
ARTICLE IN PRESS
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 723
ARTICLE IN PRESS
Table 1
Time between failures (thousands of hours) of secondary reactor pumps [27]
2.160 0.746 0.402 0.954 0.491 6.560 4.992 0.347
0.150 0.358 0.101 1.359 3.465 1.060 0.614 1.921
4.082 0.199 0.605 0.273 0.070 0.062 5.320
0 10 20 30 40 50 60
0
0.1
0.2
0.3
0.4
t
h
(
t
)
0 10 20 30 40 50 60
0
0.2
0.4
0.6
0.8
1
t
S
(
t
)
Fig. 3. Fitted failure rate for secondary reactor pumps: MLE estimates (dashed) and regression estimates (dotted). The empirical survival function is given
as the solid line in the lower plot.
0 10 20 30 40 50 60
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t
S
(
t
)
Fig. 4. Fitted survival functions for secondary reactor pumps: exible
Weibull (solid curve), Weibull (dashed) reduced additive Weibull (dot-
dash) and extended Weibull (dotted). The empirical survival function is
the solid step-function.
0 10 20 30 40 50 60
0
0.05
0.1
0.15
0.2
0.25
t
h
(
t
)
Fig. 5. Fitted failure rate functions for secondary reactor pumps: exible
Weibull (solid curve), Weibull (dashed) reduced additive Weibull (dot-
dash) and extended Weibull (dotted).
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 724
The reduced additive Weibull model [29,30] has survival
function

Ft exp at
b
at
1=b
lt
_ _
for all t40, with parameters aX0, b41, and lX0. The
corresponding failure rate function is
ht abat
b1
a=bat
1=b1
l.
Finally, the extended Weibull distribution [31] has
survival function

Ft
le
t=b
a
1 1 le
t=b
a
with failure rate function
ht
a=bt=b
a1
1 1 le
t=b
a
.
The tted survival functions are shown in Fig. 4 and the
hazard functions in Fig. 5. As the modied Weibull curves
are very close to the Weibull in this case, they have been
omitted to aid clarity. The tted parameters and log-
likelihoods are given in Table 2, which also gives the
KolmogorovSmirnov distance (D). However, none of the
latter values are signicant for a data set of this size.
Murthy et al. [13] select models by examining graphically
the t to the survival curve in the WPP plot, which gives a
similar result to examining the KolmogorovSmirnov
distances. Here we choose to select models based on their
likelihood. We see that the exible Weibull model provides
by far the best t to the data, without even allowing for it
having only two parameters.
Acknowledgements
Ric ardas Zitikis was partially supported by a research
grant at Massey University, as well as by a Discovery
Research Grant from the Natural Sciences and Engineering
Research Council (NSERC) of Canada. Two anonymous
referees suggested improvements in layout and content.
Appendix
Proof that the turning points of the failure rate function
t
1
and t
2
tend to 0 at the rate O1=a when a ! 1 but the
product ab remains a xed constant.
We will rst rewrite the formulas (8) and (9) for t
1
and t
2
.
Let go
27
64
be a constant, and let a and b satisfy the
relationship ab g. Given g, we dene another two
constants
d
1
:
2
11=3
g
2
3 27g
2
32g
3
3

3
p
27g
4
64g
5
_
_ _
1=3
,
d
2
:
2
1=3
27g
2
32g
3
3

3
p
27g
4
64g
5
_
_ _
1=3
3
.
In turn, we dene another two constants
D
1

1
2

4g
3
d
1
d
2
_

1
2

8g
3
d
1
d
2
4g
_

4g
3
d
1
d
2
_
_ _

_
,
D
2

1
2

4g
3
d
1
d
2
_

1
2

8g
3
d
1
d
2
4g
_

4g
3
d
1
d
2
_
_ _

_
.
Obviously, D
1
pD
2
. Now we write the following equations:
t
1

D
1
a
and t
2

D
2
a
.
Clearly, when a gets larger, both t
1
and t
2
get smaller.
References
[1] Weibull W. A statistical distribution of wide applicability. J Appl
Mech 1951;18:2937.
[2] Keshevan K, Sargent G, Conrad H. Statistical analysis of the
Hertzian fracture of pyrex glass using the Weibull distribution
function. J Mater Sci 1980;15:83944.
[3] Sheikh A, Boah JK, Hansen DA. Statistical modelling of pitting
corrosion and pipeline reliability. Corrosion 1990;46:1906.
[4] Queeshi FS, Sheikh AK. Probabilistic characterization of adhesive
wear in metals. IEEE Trans Reliab 1997;46:3844.
ARTICLE IN PRESS
Table 2
Fitted parameters (MLEs) for secondary reactor pumps
Model ^ a ^
b
^
l
log L D
Flexible Weibull 0.0207 2.5875 NA 83.3424 0.1342
Weibull 0.8077 13.9148 NA 85.4734 0.1173
Modied Weibull 0.1213 0.7924 0.0009 85.4677 0.1188
Reduced additive Weibull 0.0070 1.7292 0.0452 86.0728 0.1619
Extended Weibull 0.4189 1.0212 10.2778 86.6343 0.1057
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 725
[5] Durham SD, Padgett WJ. Cumulative damage model for system
failure with application to carbon bers and composites. Techno-
metrics 1997;39:3444.
[6] Almeida JB. Application of Weibull statistics to the failure of
coatings. J Mater Process Technol 1999;93:25763.
[7] Fok SL, Mitchell BC, Smart J, Marsden BJ. A numerical study on the
application of the Weibull theory to brittle materials. Eng Fract Mech
2001;68:11719.
[8] Newell JA, Kurzeja T, Spence M, Lynch M. Analysis of recoil
compressive failure in high performance polymers using two- and
four-parameter Weibull models. High Perform Polym 2002;14:
42534.
[9] Li QS, Fang JQ, Liu DK, Tang J. Failure probability prediction of
concrete components. Cem Concr Res 2003;33:16316.
[10] Majeske KD. A mixture model for automobile warranty data. Reliab
Eng Syst Saf 2003;81:717.
[11] Bucar T, Nagode M, Fajdiga M. Reliability approximation using
nite Weibull mixture distributions. Reliab Eng Syst Saf 2004;84:
24151.
[12] Attardi L, Guida M, Pulcini G. A mixed-Weibull regression model
for the analysis of automotive warranty data. Reliab Eng Syst Saf
2005;87:26573.
[13] Murthy DNP, Bulmer M, Eccleston JA. Weibull model selection for
reliability modelling. Reliab Eng Syst Saf 2004;86:25767.
[14] Gurvich MR, Dibenedetto AT, Rande SV. A new statistical
distribution for characterizing the random length of brittle materials.
J Mater Sci 1997;32:255964.
[15] Nadarajah S, Kotz S. On the recent papers on modied Weibull
distributions. IEEE Trans Reliab 2005;54:612.
[16] Xie M, Tang Y, Goh TN. A modied Weibull extension with bathtub
shaped failure rate function. Reliab Eng Syst Saf 2002;76:27985.
[17] Xie M, Lai CD, Murthy DNP. Weibull-related distributions for
modelling bathtub-shaped failure rate functions. In: Lindqvist BH,
Doksum K, editors. Mathematical and statistical methods in
reliability. Singapore: World Scientic Publishing; 2003. p. 28397.
[18] Murthy DNP, Xie M, Jiang R. Weibull models. New York: Wiley;
2003.
[19] Jensen F, Petersen NE. Burn-in: an engineering approach to the
design and analysis of burn-in procedures. New York: Wiley; 1982.
[20] Barlow RE, Proschan F. Statistical theory of reliability and life
testing. Silver Spring, MD; 1981.
[21] Lai CD, Xie M, Murthy DNP. Bathtub-shaped failure rate
distributions. Hand Stat 2001;10:69104.
[22] Kuo W, Kuo Y. Facing the headaches of early failures: a state of the
art review of burn-in decisions. Proc IEEE 1983;71:125766.
[23] Gupta RC, Warren R. Determination of change points of non-
monotonic failure rates. Commun StatTheory Methods 2001;30:
190320.
[24] Jiang R, Murthy DNP. Mixture of Weibull distributionsparametric
characterization of failure rate function. Appl Stochastic Models
Data Anal 1998;14:4765.
[25] Block HW, Li Y, Savits TH. Initial and nal behavior of failure rate
functions for mixtures and systems. J Appl Probab 2003;40:72140.
[26] Wondmagegnehu ET. On the behavior and shape of mixture failure
rates from a family of IFR Weibull distributions. Naval Res Logistics
2004;51:491500.
[27] Salman Suprawhardana M, Prayoto, Sangadji. Total time on test plot
analysis for mechanical components of the RSG-GAS reactor. Atom
Indones 1999;25(2).
[28] Lai CD, Xie M, Murthy DP. A modied Weibull distribution. IEEE
Trans Reliab 2003;52:337.
[29] Xie M, Lai CD. Reliability analysis using additive Weibull model
with bathtub-shaped failure rate function. Reliab Eng Syst Saf
1995;52:8793.
[30] Lai CD, Zhang LY, Xie M. Mean residual life and other properties of
Weibull related bathtub-shaped failure rate distributions. Int J Reliab
Qual Saf Eng 2004;11:11332.
[31] Marshall AW, Olkin I. A new method for adding a parameter to a
family of distributions with application to the exponential and
Weibull families. Biometrika 1997;84:64152.
ARTICLE IN PRESS
M. Bebbington et al. / Reliability Engineering and System Safety 92 (2007) 719726 726