Chap4 - 1 Time To Failure

Chapter 4
TIME TO FAILURE
DISTRIBUTIONS
4.1 Introduction
This chapter considers probability distributions which are most often used in
reliability as time to failure distributions. These distributions include the dis-
tributions which have well-known probability density functions for describing
time-to-failure (TTF) in many situations. The chapter also considers other
less used distributions of TTF, and presents characteristics and properties of
the distributions considered.
The probability distribution of TTF indicates beliefs about the likelihood
of failure times. These beliefs can be based on the frequency of data values
and have a frequency interpretation or they can be based on judgment and
have a frequency or a subjective interpretation. Basically, the probability dis-
tribution is a model for knowledge about the time to failure or the reliability
of a device.
In the rst sections of this chapter, probability distributions of TTF are
considered from the point of view of their corresponding hazard functions.
These sections also subscribe to the relatively simple point of view that a
single hazard function is sucient for describing the behavior of the TTF over
the entire range of the failure times. In the last section of this chapter, more
complex models are introduced where the use of several hazard functions or
several distributions is necessary to model the situation.
121
122 CHAPTER 4. TIME TO FAILURE DISTRIBUTIONS
4.2 Monotone hazard functions
The hazard function of a distribution was dened earlier but it is important
to consider such functions here since they describe the relationship between
the instantaneous probability of failure and time. Sometimes there is phys-
ical or data information about the hazard function and not the probability
distribution itself and then the hazard is useful in determining the distribu-
tion of TTF. There may also be more general information about the hazard
function, such as the fact that it is monotone increasing or decreasing. In this
case, the information is also useful because there are techniques of analysis
of failure times that are based solely on the monotonicity information.
In this chapter, however, the type of hazard function and its shape is
considered only from the point of view of determining the distribution of the
TTF. Models with monotone increasing hazards are often used because one
is often interested in the life of a device in the period of its life when an aging
process is in force. It is then useful to consider which distributions have this
type of hazard function. Models with monotone decreasing hazard functions
are used less often but can have application in the study of early lifetimes
of devices. Models with constant hazard functions are unique and are often
useful as baseline distributions to which other distributions are compared or
as simple models for failure modes resulting in random failures.
4.2.1 Constant hazard rate-the exponential distribu-
tion
In selecting a pdf to describe the TTF for a particular situation, it seems ap-
propriate to require that the properties of the distribution do not contradict
what is known about the failure behavior of the component of system under
study. For example, for most manufactured devices beyond the burn-in
period, the hazard rate function h(t) is monotone increasing, or constant. In
this case, it is important that the hazard rate function of the distribution
under consideration as a model for the failure times of the device incorporate
this property of monotone increasing or constant hazard.
In terms of constant hazard, the exponential distribution can be charac-
terized as the only distribution with a constant hazard function. Thus, if
a unit has failure times that follow the exponential distribution, it can be
said that the unit is not subject to burn-in nor is it subject to wear.
Conversely, if a unit has survived to time t and its probability of survival,
4.2. MONOTONE HAZARD FUNCTIONS 123
that is, its reliability, over an additional time period t
a
is the same regard-
less of the present age t, then the time to failure of the unit must follow
an exponential distribution. Thus, the exponential distribution is central in
reliability studies, in terms of its usage and in terms of this gurative way of
having a hazard function which is not increasing or decreasing.
For the exponential distribution, the density function f(t) is:
f(t) =
1
= e
t
, , > 0 (4.1)
where is the mean time to failure and is the hazard rate. It follows
that the cdf F(t), the reliability function R(t) and the hazard function h(t)
are respectively:
F(t) = 1 e
= 1 e
t
, R(t) = e
= e
t
and h(t) =
1
= (4.2)
If the failure time of a unit follows an exponential distribution,
P(T t) = R(t) = e
t
and
P(T t
a
|T t) =
R(t + t
a
)
R(t)
= e
ta
.
That is, the conditional probability of failure conditioned on an earlier
time period is independent of the earlier time period. This is usually referred
to as the memory-less property of the exponential distribution. It means
that devices whose failure times follow the exponential distribution model are
not subject to wear out. That is, a device that has survived to time t is as
good as new in terms of its remaining life. In contrast, with devices that
are subject to wear out, the conditional probability of failure conditioned on
an earlier time period is a decreasing function of the length of the earlier
time period.
Also, it is true that if a unit is functioning at time t and the probability
of survival over an additional time period is the same regardless of t, then
P(T t
a
|T t) =
R(t + t
a
)
R(t)
= R(t
a
) or R(t + t
a
) = R(t)R(t
a
)
and necessarily,
R(t) = e
t
and h(t) = .
That is, the distribution of the time to failure is exponential.
4.2.2 The exponential distribution and the geometric
distribution
Now, consider the time line being segmented into intervals of a length t
and let the probability of failure in any interval be p, where p is constant
for all intervals. Also, assume that the occurrence of a failure in an interval
is independent of that occurrence in any other interval. The occurrence
or non-occurrence of a failure in an interval can be called a trial and the
trials in this example make up a Bernoulli process. (These assumptions are
also equivalent to the assumptions or characteristics underlying the binomial
distribution.) Then the number of intervals, R, until the rst failure has a
geometric distribution and
P{R = r; p} = (1 p)
r1
p, r = 1, 2, 3, ...
Thus, the geometric distribution is the discrete analog of the exponential
distribution. Notice also that the assumptions on the constancy of p and the
independence of the occurrences in the intervals cause the process to operate
similarly to a continuous process with a constant failure rate. It can also
be shown that if the length of the interval t goes to zero, the geometric
distribution becomes the exponential distribution.
4.2.3 The exponential distribution and the Poisson pro-
cess
Consider, for example, that failures represent any time related occurrences
and the random variable of interest is the number of failures, X(t), that occur
in the interval [0, t]. Then, assume:
a) X(0)=0
b) The numbers of failures in non-overlapping intervals are independent
random variables
c) the distribution of the number of failures that occur in a given interval
depends only on the length of the interval and not on the location of
the failures
d) lim
h0
p{X(h) = 1}
h
=
4.3. INCREASING AND DECREASING HAZARD FUNCTIONS 125
e) lim
h0
p{X(h)2}
h
= 0
Under these assumptions, the random variable or counting process, X(t), the
number of failures in the interval [0, t] is said to be a Poisson process and
the number of failures in any interval of length t has a Poisson distribution
with parameter = t and the times between failures have an exponential
distribution with parameter . That is,
P{X(t) = r} =
e
t
(t)
r
r!
and
P{X
i+1
> t
i+1
|X
i
= t
i
} = P{0 events in (t
i
, t
i
+ t
i+1
)|X
i
= t
i
}
= P{0 events in (t
i
, t
i+1
)} = exp [(t
i+1
t
i
)]
4.3 Increasing and decreasing hazard func-
tions
If a unit has a conditional survival probability R(t
a
|t) that decreases as a
function of the time t, the time period previous to the present time period
t
a
, then the device is wearing out or aging and
R(t
a
|t) =
R(t + t
a
)
R(t)
is decreasing in t for each t
a
. It follows that
lim
ta0
_
1
t
a
_
1
R(t + t
a
)
R(t)
__
= lim
ta0
_
R(t) R(t + t
a
)
t
a
R(t)
_
= h(t)
is increasing in t and the distribution of t is called an increasing failure
rate (IFR) distribution. If R(t
a
|t) is increasing in t for each t
a
, h(t) is de-
creasing in t and R(t) is called a decreasing failure rate (DFR) distribution.
The exponential distribution has constant failure rate and is the only dis-
tribution that has. Thus much can be learned about a distribution of the
time to failure if it is either IFR or DFR by comparing it to the exponen-
tial. For example, the Weibull distribution is IFR when its shape parameter
> 1 and DFR for 0 < < 1. The gamma distribution is IFR when its
shape parameter r > 1 and DFR for 0 < r < 1. Both the Weibull and the
gamma are exponential for r = = 1. There are also distributions with
non-monotone failure rates. For example, the log normal and the log-logistic
are distributions used in reliability which have non-monotone failure rates.
4.4 Time to failure distributions with mono-
tone failure rate
The popular choices for time to failure (TTF) distributions with monotone
hazard or failure rates are: the exponential, the Weibull, the gamma and the
normal. The density functions of these distributions are frequently chosen
as models for the frequency of occurrence of TTF values. Often, this choice
is made because either their theoretical properties are consistent with the
conditions of use and the physics of failure of the device, or because the
density adequately describes the failure history of the device. The above
distributions are outlined rst in this chapter. The Type I Extreme Value
distribution, a distribution with a monotonically increasing hazard function
and less popular in its use than the above distributions, is outlined later in
the chapter.
Other distributions of time to failure with non-monotone hazard func-
tions, including the log-normal distribution, will also be presented later in
the chapter. A relatively recently proposed choice for the distribution of time
to failure, the Burr distribution, is outlined also later in this chapter. The
Burr and the extreme value distributions are less known and used, especially
the Burr, but are nevertheless important in the list of available choices for
models and ll important gaps in the span of possible hazard or failure rate
considerations.
In an earlier chapter, the bathtub curve and its use in reliability was
recognized. Note that the burn-in portion of the curve represents distri-
butions with DFR and that the wear-out portion represents distributions
with IFR. The constant hazard portion is often called the useful life portion
of the hazard curve. For electronic components, the useful life portion is
usually longer than for mechanical components. See, for example, Figures
3.9 and 3.10
4.5. EXPONENTIAL DISTRIBUTION 127
4.5 Exponential distribution
In Section 4.2.3, the exponential distribution was characterized by its rela-
tionship with the Poisson process. But the gamma distribution is also related
to the Poisson process. The Poisson process is a model for the number of
events that occur in a time interval with mild assumptions that failure events
often meet. The model then species that the distribution of the number of
events is Poisson and that the distribution of the time between events is ex-
ponential. The time to the rst failure is also exponentially distributed but
the time to the rth failure has a gamma distribution with shape parameter
r.
The exponential distribution with rate parameter or mean has prop-
erties:
R(t) = e
t
, f(t) = e
t
, h(t) = , H(t) = t
E(T) = = =
1
, V (T) =
2
=
2
=
2
,
3
= 2
3
,
4
= 9
4
,
coecient of skewness =
1
= 2, coecient of excess =
2
3 = 6,
coecient of variation CV =
V (T)
E(T)
= 1,
and t
R
=
1
lnR, where t
R
is the reliable life for the value R.
The exponential density (pdf) and cdf are illustrated in Figures 4.1 and
4.2 for means = 2, 10, 20, 50 or rates = 0.5, 0.1, 0.05, 0.02.
The exponential distribution is widely used in its own right as a model
of the times to failure in some situations and is also widely used as the stan-
dard to which other distributions are compared. It can be chosen for its
constant hazard property or because it models the observed times. Note,
however, that the coecients of skewness and excess are xed values for all
exponential distributions. This fact indicates that the shape of the exponen-
tial distribution remains the same for all values of its parameter. This fact
also indicates a lack of exibility in tting dierent models of failure time
data. There is agreement among practitioners that there must exist strong
information about a constant hazard in order to justify the assumption and
subsequent use of the exponential distribution as a failure time model. Listed
below in Table 4.1 are 96 simulated values from an exponential distribution
with mean (MTBF) of 1000 cycles.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
0 5 10 15 20 25 30 35 40 45 50
t
f
(
t
)
f(t),mean=2,skewness=2,excess=9
f(t),mean=10,skewness=2,excess=9
f(t).mean=20,skewness=2,excess=9
f(t).mean=50,skewness=2,excess=9
Figure 4.1: Exponential Density Function
Table 4.1
2100 2232 625 1852 299 559 54 660 562 69 91 541
427 50 571 306 412 9 1129 749 208 490 792 485
140 351 398 610 1463 1237 884 1985 2083 588 1640 363
315 631 353 853 767 418 855 726 4707 329 196 160
224 29 300 2892 2051 62 312 495 2622 832 555 963
2607 1136 405 263 723 967 1516 1731 1785 2278 163 1476
712 44 789 666 2894 116 1513 1112 338 1072 2168 2707
788 67 1096 572 1013 225 269 1406 644 181 19 1041
By examining these data, one notices that they are all over the place.
They range from 9 to 4707. Remember that the standard deviation of the
exponential distribution is equal to its mean; in this case, 1000 cycles. One
can also compute that the proportion of values larger than the mean 1000 is
30/96 = 0.3125. More than 60% of the values are below the mean; this is
one of the characteristics of the exponential distribution. In fact, note that
4.5. EXPONENTIAL DISTRIBUTION 129
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 5 10 15 20 25 30 35 40 45 50
t
F
(
t
)
F(t),mean=2,skewness=2,excess=9
Figure 4.2: Exponential Cumulative Function
for any device subject to exponential failures, there is a 0.632 probability of
failure before the mean time to failure, or, equivalently, 63.2% of such devices
will fail before the mean. It is rather common in reliability to set goals and
develop metrics in terms of the mean time to failure. However, in light of
the above, it seems questionable to set a reliability goal in terms of an event
in which most (63.2%) of what may happen before the event is unfavorable
(failure).
EXAMPLE: 4.5 Suppose that the TTF density of a machine is given
by:
f(t) =
1
1000
e
(t/1000)
, t > 0,
that is, the TTF has an exponential distribution with mean 1000. The prob-
ability of surviving the interval (0,1000) is given by: R(1000) = exp(1) =
0.368 or 1 0.632.
There are several viable alternatives to focusing on the mean in this way.
Among them are the use of distributional properties, such as percentiles and
reliable life. The reliable life, t
r
is dened as the time at which the reliability
is r. For example, t
0.9
= 300 cycles means that after 300 cycles of operation,
the reliability is 0.9.
EXAMPLE: 4.6 For the exponential distribution, any value of the
mean life, determines a life for a specied value of R; i.e., the reliable
life may be established for any mean and desired reliability. Since R(t) =
exp(t/), t
R
= ln R. For example, if = 1000 cycles, then:
R = 0.1 0.3 0.5 0.7 0.9 0.95 0.99 0.999
t
R
= 2303 1204 693 357 105 51 10 1
It becomes clear in examining the table that if the mean number of cycles
to failure is 1000, high reliabilities are achieved only for relatively short num-
bers of cycles. Thus, it seems more meaningful to set a reliability goal, not in
terms of , but in terms of a statement that a 0.95 reliability for 1000 cycles
of operation is required. If the exponential distribution is appropriate, the
value equivalent to this goal is easily obtained using = t/ ln R(t). In this
case, = 1000/ ln(.95) = 19496 cycles. The concept of reliable life will be
referred to frequently throughout this text.
EXAMPLE: 4.7 A semiconductor is characterized by an exponential
TTF distribution with = 1 10
6
failures per hour. The probability of no
failure in the interval (10
6
, 10
7
) given no failure in (0, 10
6
) is:
R(10
7
)
R(10
6
)
=
e
10
6
(10
7
)
e
10
6
(10
6
)
=
e
10
e
1
= 0.0001234
The conditional reliability in the interval (67 10
6
, 76 10
6
) is:
R(76 10
6
)
R(67 10
6
)
=
e
10
6
(7610
6
)
e
10
6
(6710
6
)
=
e
76
e
67
= e
9
= 0.0001234
that is, with the exponential distribution, the reliability in an interval of
equal length is the same regardless of where in the time zone that interval
occurs.
4.6. WEIBULL DISTRIBUTION 131
4.6 Weibull distribution
The Weibull distribution is perhaps the most widely used of all the failure
time distributions and it is noted for its exibility as a model. The Weibull
distribution with scale parameter and shape parameter has properties:
R(t) = e
(
t
,
f(t) =
t
1
e
(
t
, h(t) =
t
1
, H(t) =
_
t
E(T) = =
_
1 +
1
_
, V (T) =
2
=
2
=
2
2
=
2
_
_
1 +
2
2
_
1 +
1
__
,
where () =
_
0
x
1
e
x
dx,
3
=
3
_
1 +
3
_
3
2
+ 2
3
,
where
i
=
i
_
1 +
i
_
,
4
=
4
_
1 +
4
_
4
3
+6
2
3
4
, mode =
_
( 1)
_1
, > 1
(4.3)
1
=

3
2
3/2
, coecient of excess =
2
3 =

4
2
2
3,
coecient of variation =CV =
(1+
2
)
2
(1+
1
)
(1+
1
)
, t
R
= (lnR)
1
.
The examples of Weibull pdf, cdf and hazard function are shown in Fig-
ures 3.4-3.6.
For < 1, the Weibull distribution is DFR, for > 1, the Weibull is
IFR and for = 1, the Weibull becomes the exponential. That is, if < 1,
h(t) is monotone decreasing, if > 1, h(t) is monotone increasing, and if
= 1, h(t) is constant. Also, note that if is approximately 3.5, f(t) is
approximately symmetrical, if 1 < < 3.5, f(t) is positively skewed, and
if > 3.5, f(t) is negatively skewed. Note that the coecient of skewness
and the coecient of excess

2
3 are both functions only of the shape
parameter since the scale parameter gets cancelled in the ratio.
Since the Weibull distribution allows both increasing and decreasing fail-
ure rates, it can be used to represent the times to failure for both the burn-
in portion of the hazard curve as well as the wear-out portion. It has
been successfully used in both situations.
The Weibull distribution also has a justication related to the extreme
value of a group of failure times. This aspect will be discussed as part of the
discussion of the extreme value distribution.
EXAMPLE: 4.8 Failures associated with the grasping function of a
robot arm follow a Weibull distribution with parameters = 2.1 and =
2200 hours. The probability of the robot arm grasping function surviving
past 1000 hours is
R(1000) = e
(
1000
2200
)
2.1
= 0.826
EXAMPLE: 4.9 The times to failures for a pick and place machine
used in surface mount technology appear to follow a Weibull distribution,
with = 2 and = 30 weeks. The probability of surviving 40 weeks is
R(40) = e
(
40
30
)
2
= e
1.778
= 0.169
This Weibull is IFR, since h(t) =
t
450
. The mean (MTTF) is 30(1+
1
2
) = 26.6
weeks.
The probability that the pick and place machine will fail in the interval
(40, 50) given survival in (0, 40) is:
F(50) F(40)
R(40)
=
1 e
(50/40)
2
(1 e
(40/40)
2
)
e
(40/40)
2
=
e
1
e
1.5625
e
1
= 1 e
0.5625
= 1 0.5625 = 0.4302
EXAMPLE: 4.10 Suppose that one knew that a device had a reliability
of 0.60 at 1000 hours of operation and a reliability of 0.10 at 4000 hours of
operation and that the failure times followed a Weibull distribution. In this
case it follows that:
e
(
1000
= 0.60 ande
(
4000
= 0.10
Using logarithms, one can solve the two equations and can nd that the
appropriate distribution of time to failure is a Weibull with = 1.68 and
= 2437.5.
4.7. GAMMA DISTRIBUTION 133
4.7 Gamma distribution
Another distribution that is closely related to the exponential distribution
is the gamma distribution. The gamma distribution has a scale parameter,
denoted here as and a shape parameter, denoted here as r. The gamma
distribution can also arise from a Poisson process as the distribution of the
time to the rth occurrence. In terms of the Bernoulli process, the number of
trials to the rth occurrence is distributed as negative binomial and the nega-
tive binomial is the discrete analog of the gamma. It follows that the time to
the rth occurrence is gamma distributed and, in the same consideration, the
sum of r independent exponential random variables is gamma distributed.
The gamma distribution with scale parameter and shape parameter r has
properties:
f(t) =

r
t
r1
e
t
(r)
, t 0 R(t) = e
t
r1
i=0
(t)
i
i!
, if r is an integer,
R(t) =
1
(r)
_
t

r
u
r1
e
u
du =
1
(r)
(r, t), otherwise.
h(t) =

r
t
r1
e
t
(t, r)
E(T) = =
r
, V (T) =
2
=
2
=
r
2
,
4
=
r
2
+6r
4
, where,
i
=
(r+i)
(r)
,
mode =
r 1
, if r > 1, coecient of skewness =

_
1
=
2
r
(4.4)
coecient of excess =
2
3 =
6
r
, coecient of variation = CV =
1
r
, t
R
is the solution t
R
of R(t
R
) = R, or the solution to (r, t
r
) = 1 R.
The gamma pdf is illustrated in Figure 4.3, the gamma cdf is illustrated
in Figure 4.4 and the gamma hazard function is illustrated in Figure 4.5 for
pairs of values of the scale parameter and the shape parameter r equal to
(0.2, 0.75), (0.333, 1), (0.25, 2) and (1, 3).
The chi-square distribution with degrees of freedom is a special case of
the gamma distribution, when = 1/2 and r = /2. The gamma distribu-
tion is DFR for r < 1, decreasing to a constant value . The gamma is an
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 2 4 6 8 10 12 14 16 18 20
t
f
(
t
)
f(t),r=0.75,lam=0.2,mean=3.75,st.dev=4.33,skewness=2.31,excess=8
f(t),r=1,lam=0.333,mean=3,st.dev=3,skewness=2,excess=6
f(t),r=2,lam=0.25,mean=8,st.dev=5.66,skewness=1.41,excess=3
f(t),r=3,lam=1,mean=3,st.dev=1.73,skewness=1.15,excess=2
Figure 4.3: Gamma Density Function
exponential when r = 1. When r > 1, the gamma is IFR, increasing from
zero to a constant value . The primary use of the gamma in reliability is
as the distribution of the sum of r exponential random variables when such
is needed in practice. The use of the gamma distribution as the sum of r
independent exponential random variables is also of use in the theoretical
development of inferences related to the exponential.
EXAMPLE: 4.11 The time to failure of a particular device follows the
exponential distribution with mean time to failure 100 hours. A sample of
n=15 such devices are tested to failure and the sample mean is observed. The
sum of the sample times to failure, T, is distributed gamma with = .01
and r=15. This sum can be transformed into a chi-squared variable by mul-
tiplying by 2; that is, 2T is chi-squared distributed with 30 degrees of
freedom. (If X is distributed exponential with hazard rate , then X is ex-
ponentially distributed with hazard rate 1. 2X is exponential with hazard
4.7. GAMMA DISTRIBUTION 135
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12 14 16 18 20
t
F
(
t
)
F(t),r=0.75,lam=0.2,mean=3.75,st.dev=4.33,skewness=2.31,excess=
8
F(t),r=1,lam=0.333,mean=3,st.dev=3,skewness=2,excess=6
F(t),r=2,lam=0.25,mean=8,st.dev=5.66,skewness=1.41,excess=3
F(t),r=3,lam=1,mean=3,st.dev=1.73,skewness=1.15,excess=2
Figure 4.4: Gamma Cumulative Function
rate 2 and 2T is gamma with parameters = 2 and r=n.) In this example,
P{2T < 43.773} = 0.95 and thus P{T < 2188.65} = 0.95.
EXAMPLE: 4.12 A sensitive microswitch with an exponential failure
rate of 2 per year has been replicated 4 times to increase system reliability.
For example, all 4 switches must fail for unreliability. What is the reliability
of this switch over 3 years?
R(3) = 1 F(3) = 1
k=4
e
2(3)
6
k
k!
=
3
k=0
e
6
6
k
k!
= 0.285
Also, note that the MTBF= E(t) =
r
=
4
2
= 2 years.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12 14 16 18 20
t
h
(
t
)
h(t),r=0.75,lam=0.2,mean=3.75,st.dev=4.33,skewness=2.31,excess=8
h(t),r=1,lam=0.333,mean=3,st.dev=3,skewness=2,excess=6
h(t),r=2,lam=0.25,mean=8,st.dev=5.66,skewness=1.41,excess=3
h(t),r=3,lam=1,mean=3,st.dev=1.73,skewness=1.15,excess=2
Figure 4.5: Gamma Hazard Function
4.8 Normal distribution
Strictly speaking, since the normal distribution allows negative values and
since the life of an item is always positive, some caution should be observed in
using the normal distribution to represent the time to failure distribution. In
some places, it is simply stated that the normal distribution should be used
in reliability only if it can be assumed that (/) > 3. This would essentially
ensure that only positive values occur. The normal distribution with mean
and standard deviation has properties
f(t) =
1
2
e
1
2
(
x
)
2
, R(t) = 1
_
t
_
,
where
_
t
_
=
_ t
2
e
z
2
2
dz, h(t) =
f(t)
R(t)
,
4.8. NORMAL DISTRIBUTION 137
H(t) = ln
_
1
_
t
__
, E(T) = , V (T) =
2
=
2
,
3
= 0,
4
= 3
4
1
= 0, coecient of excess =
2
3 = 0,
coecient of variation CV =

, t
R
=
1
(1
R
) + .
The pdf, cdf and hazard function of the normal distribution are presented
in Figures 4.6, 4.7 and 4.8 respectively, and for mean and standard deviation
pairs (2, 0.5), (5, 1), (10, 2) and (12, 0.75).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 2 4 6 8 10 12 14 16
t
f
(
t
)
f(t),mean=2,st dev=0.5,CV=.25
f(t),mean=5,st dev=1,CV=.20
f(t),mean=10,st dev=2,CV=.20
f(t),mean=12,st dev=0.75,CV=.0625
Figure 4.6: Normal Distribution Density Function
The normal distribution is an increasing failure rate (IFR) distribution,
and for values of t larger than the mean, the hazard rate approaches an
asymptote
(t)
, which increases linearly in t.

The use of normal distribution can also be justied as a limit of the
sums (based on Central Limit Theorem); for example, when approximating
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12 14 16 18
t
F
(
t
)
F(t),mean=2,st
dev=0.5,CV=.25
F(t),mean=5,st dev=1,CV=.20
F(t),mean=10,st dev=2,CV=.20
F(t),mean=12,st
dev=0.75,CV=.0625
Figure 4.7: Normal Distribution Cumulative Function
Gamma distribution with large r.
EXAMPLE: 4.13 The refractory lining of an oven used as part of a
curing process wears out (must be replaced) periodically. The wear-out dis-
tribution is believed to be normal with a mean of 2700 hours and a standard
deviation of 250 hours. After how many hours of operation should the oven
be relined so the wear out is virtually impossible?
Schedule relining so that it occurs at a point in the extreme left tail of
the distribution, say at 3. Reline every 2700 3(250) = 1950 hours.
EXAMPLE: 4.14 Suppose one knew that the failure process of a device
was such that essentially all of the devices will fail before 10000 hours, but
that 95% of the devices will operate beyond 2000 hours. Further, it is known
that the failure times follow a normal distribution. In this case, it is known
4.9. EXTREME VALUE DISTRIBUTIONS 139
0
1
2
3
4
5
6
7
0 2 4 6 8 10 12 14 16
t
h
(
t
)
h(t),mean=2,st dev=0.5,CV=.25
h(t),mean=5,st dev=1,CV=.20
h(t),mean=10,st dev=2,CV=.20
h(t),mean=12,st dev=0.75,CV=.0625
Figure 4.8: Normal distribution Hazard Function
that:
+ 3 = 10000 and 1.645 = 2000
These equations result in the determination of the mean = 4833 hours and
the standard deviation = 1722 hours for the normal distribution of failure
times.
4.9 Extreme value distributions
The distribution of the rst order statistic T
(1)
, from a sample of n, is, in
terms of the reliability function or in terms of the cumulative distribution
function,
R
T
(1)
(t) = [R
T
(t)]
n
or F
T
(1)
(t) = 1 [1 f
T
(t)]
n
.
This will be discussed further in a later section on order statistics. It can
also be shown that when n gets large, the random variable nF
T
(T
(1)
) = Y
n
goes to the random variable Y which has distribution F
Y
(y) = 1 e
y
and
T
(1)
goes to the random variable F
1
T
(Y/n).
Then, if the random variable T is bounded from below and with a weak
conditions on F
T
, the limiting distribution for T
(1)
is a Weibull distribution.
This case is called the Type III asymptotic distribution of the smallest ex-
treme and indicates another physical scenario for the usage of the Weibull
distribution as a distribution of the time to failure.
Suppose a complex device consists of parts and material with many modes
of failure and each mode for each part or material has a corresponding time to
failure and the failure of any part or material will cause failure of the device.
Then the rst order statistic from a large sample of n modes of failure is the
determining factor for the time to failure of the device. This is the weakest
link theory and many of the devices of interest in reliability seem to follow
this theory for the failure process. Thus, under the above conditions, the
Weibull distribution is the distribution of the time to failure.
The term extreme value distribution is, however, more often associated
with the Type I asymptotic distribution of the smallest extreme. This type
of asymptotic distribution relates to the situation where the underlying dis-
tribution has no lower bound on its values. Then the limiting distribution of
the rst order statistic T
(1)
is of the form given below with the corresponding
properties.
For the Type I Extreme Value Distribution of random variable Y with
location parameter and scale parameter :
f(y) =
e
(y)
e
e
(y)
, < y < , R(y) = e

e
(y)
,
h(y) =
e
(y)
, H(y) = e
(y)
, mode = , E(Y ) = = 0.5772,

where 0.5772 . . . is Eulers constant, V (Y ) =
2
=
2
=

2
6

2
, coecient of
skewness =
1
= 1.3, coecient of excess =
2
3 = 2.4,
coecient of variation = CV =

6
_
0.5772
_, y
R
= + [ln(ln R)].
(4.5)
The extreme value distribution density, cumulative and hazard functions
are illustrated in Figures 4.9, 4.10 and 4.11, respectively, for the scale pa-
rameter and location parameter pairs (2, 0), (1, 0), (0.5, 0) and (0.2857,
0).
0
0.2
0.4
0.6
0.8
1
1.2
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
f
(
y
)
f(y),location=0,scale=2,mean=-1.15,st
dev=2.57,skewness=1.3,excess=2.4
f(y),location=0,scale=1,mean=-.58,st
f(y),location=0,scale=0.5,mean=-.29,st
dev=.64,skewness=1.3,excess=2.4
f(y),location=0,scale=0.2857,mean=-.16,st
Figure 4.9: Extreme Value Density Function
The form of the Type I extreme value distribution that is used here is
slightly dierent from the form used in some texts on distribution theory (e.g.,
Johnson and Kotz (1970)). The form above that is used here is typical of the
form used in reliability texts, theory and applications. Most of the uses of the
Type I extreme value distribution in reliability is based on its relationship to
the Weibull distribution, and the above form of the distribution makes that
relationship straight-forward and easily accessible. Note that the form here is
negatively skewed with the same value of skewness and excess for all versions
of the distribution. Thus, one should note that the parameters of the Type I
extreme value distribution, and , are purely location and scale parameters
respectively, which means that all of the extreme value distributions have
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-5 -4 -3 -2 -1 0 1 2 3
y
F
(
y
)
F(y),location=0,scale=2,mean=-1.15,st
F(y),location=0,scale=1,mean=-.58,st
F(y),location=0,scale=0.5,mean=-.29,st
F(y),location=0,scale=0.2857,mean=-
.16,st dev=.37,skewness=1.3,excess=2.4
Figure 4.10: Extreme Value Cumulative Function
the same shape. As was mentioned earlier with the exponential distribution,
this fact diminishes the practical usefulness of the distribution for tting
data. The extreme value distribution is an IFR distribution with the hazard
increasing from zero exponentially.
The Type I extreme value distribution is also related to the Weibull dis-
tribution in the same way that the normal and the lognormal are related,
that is, the log base e of the Weibull times are distributed as extreme value.
If the Weibull random variable T has scale and shape , the distribution
of Y = ln T is extreme value with = ln , and =
1
.
EXAMPLE: 4.15 In an earlier example, the times to failure for a pick
and place machine used in surface mount technology were assumed to follow a
Weibull distribution with shape parameter = 2 and scale parameter = 30
0
1
2
3
4
5
6
7
8
9
10
-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1
y
h
(
y
)
h(y),location=0,scale=2,mean=-1.15,st
h(y),location=0,scale=1,mean=-.58,st
h(y),location=0,scale=0.5,mean=-.29,st
h(y),location=0,scale=0.2857,mean=-.16,st
Figure 4.11: Extreme Value Hazard Function
weeks. The reliability at 40 weeks was
R(40) = e
(
40
30
)
= 0.169.
The distribution of the logarithm of the failure time, ln T, has the extreme
value distribution with = ln30 = 3.401 and = 0.5. The reliability of
ln(40) = 3.689 is
R(3.689) = e
e
(3.6893.401)
.5
= e
e
0.576
= e
1.77
= 0.169.
The relationship of the extreme value distribution to the Weibull distri-
bution is also useful in the theoretical development of analytical procedures
for the Weibull distribution. But the extreme value distribution often has
useful practical properties as a model for the failure times in much the same
way as the lognormal distribution has. A major dierence in considering the
usage of a model based on the extreme value distribution and the lognormal
distribution is the fact that the extreme value distribution is IFR and the
lognormal distribution has a non-monotone failure rate.
4.10 Non-monotone hazard rate distributions
There are three failure time distributions that are considered in this chap-
ter for which the failure rate is non-monotone. These are similar in their
application, since in their most interesting cases they all have failure rates
that are initially increasing to a maximum and then decrease. They are the
lognormal, the log-logistic and the inverse Gaussian; and all have advantages
and disadvantages. The log-logistic is considered in its more general form,
usually referred to as the Burr Type XII distribution.
Although the hazard rate of these distributions is non-monotone, the
distributions are all increasing in popularity of use as models of failure time.
The families of these distributions are exible in their use as descriptive
models, that is, some member of the family can be found to t or model a
distribution which represents the failure time distribution for failure times
from a wide variety of failure time models.
The lognormal distribution is the most popular of these non-monotone
hazard rate distributions but the Burr distribution is as exible for tting
models of failure time data and has the added advantage of having a cdf that
is a simple closed form function. This advantage makes computations with
the Burr distribution easy and also has a theoretical advantage when one
works with censored data.
The usual hazard function of these distributions increases over time to
a maximum and then decreases. Since this implies that the failure risk de-
creases after some time and then continues to decrease as the item of interest
ages, it is a concern of some practitioners and prospective users of these dis-
tributions as models of failure times as they note that for many items the
risks of failure do not decrease with increasing time. In spite of this concern,
there are at least three reasons that these distributions with non-monotone
hazard functions might be useful as models of time-to-an-event data:
1) The failure process might be such that it actually has an non-monotone
hazard rate. For example, with modern components with high reliabilities,
there may actually be a hardening process at work so that the hazard rate is
4.10. NON-MONOTONE HAZARD RATE DISTRIBUTIONS 145
non-monotone and decreasing with larger ages. In addition, some analysts of
time to death of cancer patients have noted that the observed hazard func-
tion was such that mortality reached a peak after a nite period and then
slowly declined (Farewell and Prentice (1979) and Bennett (1983)).
2) The time to an event process might be other than a failure process (such
as the duration of a strike (Lawrence (1984)) with a non-monotone hazard
function.
3) The large values of time may not be of interest or the interest in the time
to failure is only over a nite interval of time in which the physical hazard
rate is matched by that of one of the non-monotone hazard distributions.
4.10.1 Lognormal distribution
The lognormal distribution is one of the most widely used distributions of
time to failure. The lognormal is denoted by that name since, if T is the
random variable representing the lognormal time to failure, the random vari-
able, Y = ln T, is normally distributed with mean and standard deviation
.
The failure rate of the lognormal distribution increases from zero at t = 0
to a maximum and then decreases to zero as t increases. The amount of
increase or decrease of the failure rate depends mostly on the value of and
for a certain range of values the failure rate is essentially constant for an
important range of failure time values. For very small values of the hazard
function of the lognormal essentially increases over most of the time values
and behaves similarly to the normal hazard function. For large values of
the hazard function essentially decreases over most of the time values. The
lognormal distribution appears to be able to represent the distribution of the
time to failure for many failure processes. It is becoming more popular in its
usage and has become perhaps the second most used distribution of failure
times, second in use to the Weibull.
In attempting to model the distribution of a value associated with fail-
ure in terms of the physics of failure, it seems that the use of the lognormal
distribution can be justied whenever the accumulated causes of failure act
in a multiplicative manner. If the cause of failure is due to a build-up of
accumulated value from many causes and failure occurs when the accumu-
lated build-up passes a threshold value, and, in addition, the accumulated
build-up occurs multiplicatively, then the value at failure is approximately
lognormal. In this case, failure depends on the value of a product so that
the failure value, denoted by T, is such that T =
n
i=1
T
c
i
, where T
c
i
is the
accumulated value associated with the ith cause. Then the distribution of
T is of interest and, if n is large, the distribution of log T is normal by the
Central Limit Theorem.
For the log normal distribution with parameters and :
Y = ln T N(, ), f(t) =
1
t
2
e
(ln t)
2
2
2
, t > 0
R(t) = 1
_
ln t
_
, h(t) =
f(t)
R(t)
, H(t) = ln
_
1
_
ln t
__
,
E(T) =
T
= e
2
2
, V (T) =
2
T
=
_
e
(2+
2
)
__
e
2
1
_
,
2T
= e
2
e
2
2
_
e
2
1
_
3T
=
_
e
2
_3
2
_
e
2
1
_
2
_
e
2
+ 2
_
e
3
,
4T
=
_
e
2
_
2
_
e
2
1
_
2
_
_
e
2
_
4
+ 2
_
e
2
_
3
+ 3
_
e
2
_
2
3
_
e
4
1
=
_
e
2
1
_1
2
_
e
2
+ 2
_
> 0,
coecient of excess =
2
3 =
_
e
2
_
4
+ 2
_
e
2
_
3
+ 3
_
e
2
_
2
6,
mode= e
2
, median = e
, E(T) =
T
> median{T} > mode {T},
coecient of variation = CV =
V (T)
E(T)
, t
R
= e
[
1
(1R)+]
.
The lognormal density, cumulative and hazard function are illustrated in
Figures 4.12, 4.13 and 4.14, respectively, for mean and standard deviation
pairs: (0, 0.35), (1, 0.5), (2, 0.3) and (2.5, 1).
EXAMPLE: 4.16 The time (in minutes) to repair a tracking malfunc-
tion on a robot arm used in circuit card assembly tends to follow a log normal
distribution with parameters = 4.2 and = 1. The probability that a repair
action will be started and completed in 30 minutes is:
_
ln 30 4.2
1
_
=
_
3.4 4.2
1
_
= (0.8) = 0.2119.
EXAMPLE: 4.17 The failure of a certain metal under stress and tem-
perature due to fatigue cracks is known to follow a lognormal distribution
4.11. BURR DISTRIBUTION WITH SPECIAL CASE LOG-LOGISTIC147
0
0.2
0.4
0.6
0.8
1
1.2
0 2 4 6 8 10 12 14 16
t
f
(
t
)
f(t),mu=0=0,sigma=0.35,mean=1.06,st.dev=0.38,skewness=1.13,excess=2.35,mode=0.88
f(t),mu=1,sigma=0.5,mean=3.08,st.dev=1.64,skewness=1.75,excess=5.90,mode=2.12
f(t),mu=2,sigma=0.3,mean=7.73,st.dev=2.37,skewness=0.95,excess=1.64,mod
e=6.75
f(t),mu=2.5,sigma=1,mean=20.09,st.dev=26.33,skewness=6.18,excess=110.94,mode=4.48
Figure 4.12: Lognormal Density Function
with parameter = 2. What would the value of the parameter have to be
in order for the reliable life with 0.95 reliability to be 10000 hours?
t
0.95
= 10000 = e
(2
1
(0.05)+)
, so that ln(10000) = 2
1
(0.05) +
9.21034 (1.645(2)) = 12.5 = .
That is, one needs a lognormal distribution with mean: e
(12.5+2)
= 1982759.26.
4.11 Burr distribution with special case log-
logistic
A distribution which is not well known in reliability practice at this point,
but which has the possibility of eective usage is the Burr distribution. The
Burr distribution has properties which are useful in reliability studies and it
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12 14 16
t
F
(
t
)
F(t),mu=0,sigma=0.35,mean=1.06,st.dev=0.38,skewness=1.
13,excess=2.35,mode=0.88
F(t),mu=1,sigma=0.5,mean=3.08,st.dev=1.64,skewness=1.7
F(t),mu=2,sigma=0.3,mean=7.73,st.dev=2.37,skewness=0.9
F(t),mu=2.5,sigma=1,mean=20.09,st.dev=26.33,skewness=6
.18,excess=110.94,mode=4.48
Figure 4.13: Lognormal Cumulative Function
is easy to use since the Burr reliability function can be written in closed form.
The Burr distribution is similar to the lognormal distribution in the sense
that it can be thought of as representing a random variable which is the log of
a random variable having distribution similar to a normal distribution. That
is, the Burr random variable can be thought of as the log of an extension of
the logistic random variable. The Burr distribution also has a non-monotone
failure rate as does the lognormal distribution. But an advantage of the
Burr family of distributions over the lognormal family is that the reliability
function of the Burr family can be written in a simple explicit form and can
easily be computed without the use of tables. In addition, the log-logistic
distribution is becoming more used in reliability studies, especially in the
biological and medical sciences studies, and the log-logistic is a special case
of the general Burr distribution.
The Burr distribution is easier to handle with censored data than the
lognormal because the cdf of the Burr can be written in closed form and
0
0.5
1
1.5
2
2.5
0 2 4 6 8 10 12 14 16
t
h
(
t
)
h(t),mu=0,sigma=0.35,mean=1.06,st
dev=0.38,skewness=1.13,excess=2.35,mode=0.88
h(t),mu=2.5,sigma=1,mean=20.09,st
Figure 4.14: Lognormal Hazard Function
thus the probability statements and likelihoods for the Burr are more easily
viewed and manipulated.
The potential for the use of the Burr distribution in reliability is great
(see Zimmer, Keats and Wang (1998) and Gupta, Gupta and Lvin (1996))
and lies in several directions. The Burr distribution has been shown to
approximate many important distributions, including the normal, lognormal,
gamma, logistic and extreme-value. See Rodriguez (1977). As the log-logistic
becomes more used, it can be seen that the Burr, being more general, can
become an important alternative to the lognormal.
The original Burr distribution, proposed by Burr (1942) was a two-
parameter distribution, although when Burr used the distribution for approx-
imating other distributions, he, by necessity, used four parameters. However,
most references to the Burr distribution uses only the two-parameter version.
In this text, the 4-parameter Burr distribution will presented in the chap-
ter where the Burr distribution is discussed in more detail, that is, Chapter
6. In this section the 3-parameter Burr distribution will be outlined with
parameters c, k and S and has properties:
f(t) =
kc
_
t
S
_
c1
dt
S
_
1 +
_
t
S
_
c
_
k+1
t > 0
R(T) =
_
1 +
_
t
S
_
c
_
k
, S, c, k > 0
h(t) =
kc
S
_
t
S
_
c1
1 +
_
t
S
_
c
, H(t) = k ln
_
1 +
_
t
S
_
c
_
, T
R
= S(R
1
k
1)
1
c
,
E(T) = =
S
_
1
c
_
_
k
1
c
_
c(k)
,
V (T) =
2
=
2S
2
_
2
c
_
_
k
2
c
_
c(k)

_
S
_
1
c
_
_
k
1
c
_
c(k)
_
2
,
if k is an ineger,
E(T) =
_
k 1
1
c
_
. . .
_
1
1
c
_
mc(k 1)! sin

c
, V (T) =
2
_
k 1
2
c
_
. . .
_
1
2
c
m
2
c(k 1)! sin
2
c
[E(T)]
2
,
The Burr density, cumulative and hazard function are illustrated in Fig-
ures 4.15, 4.16 and 4.17, respectively and for c, k and S triples:
(3, 2, 0.5), (4.8737, 6.15784, 1), (2.105, 5.0, 1.5267), (10, 5, 2).
Note that the Burr distribution with parameters (4.8737, 6.15784, 1) has
properties that are essentially the same as the normal distribution. This
will be discussed further in Chapter 6 where the Burr approximations to the
normal, lognormal and Weibull will be illustrated.
The Burr density is unimodal if c > 1 and L-shaped if c 1. If c > 1, the
hazard rate increases to a single maximum at mt = (c 1)
1
c
, then decreases.
The hazard function is decreasing when c 1. For values of c > 1 but near
1, the hazard is almost constant after the maximum. When comparing the
Burr hazard to the lognormal hazard, values of c > 1 are more interesting for
the comparison. Also, note that the rth moment for the Burr distribution
only exists if ck > r.
0
0.5
1
1.5
2
2.5
0 1 2 3 4 5 6
t
f
(
t
)
f(t), c=3, k=2, S=0.5, mean=0.40, st
dev=0.20,skewness=1.59, excess=7.81, mode=0.33
f(t), c=4.8737, k=6.15784, S=1, mean=0.64, st
dev=0.16, skewness=0.00, excess=0.00, mode=0.65
f(t), c=2.105, k=5.0, S=1.5267, mean=0.68, st
dev=.38, skewness=1.11, excess=2.36, mode=0.50
f(t), c=10, k=5, S=2, mean=1.64, st dev=0.21,
skewness=-0.39, excess=0.40, mode=1.68
Figure 4.15: Burr Probability Density Function
EXAMPLE: 4.18 Suppose that a device will almost always fail before
200 hours and the time (in hours) to failure of the device follows the Burr
distribution with parameters: c=3.33, k=4 and S=100. From a Burr table
(See Zimmer and Burr (1963)), the mean and standard deviation for t/S are
the values: = 0.62384 and = 0.23406, thus the mean time to failure is
62.384 hours with a standard deviation of 23.406 hours. It is of interest to
compute the probability that the device survives past 100 hours:
P{t > 100} = R(100) =
1
_
1 + ([100/100])
3.33
4
=
1
2
4
=
1
16
= 0.0625.
It is also of interest to the practitioners to compute the probability that the
device fails before the time + 3 :
P{t < +3} = P{t < 132.602} = 1
1
_
1 + ([132.602]/100)
3.33
4
= 10.00623 = 0.9938
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 1 2 3 4 5 6
t
F
(
t
)
F(t), c=3, k=2, S=0.5, mean=0.40, st dev=0.20,
skewness=1.59, excess=7.81, mode=0.33
F(t), c=4.8737, k=6.15784, S=1, mean=0.64, st
dev=0.16, skewness=0.00, excess=0.00,
mode=0.65
F(t), c=2.105, k=5.0, S=1.5267, mean=0.68, st
dev=0.38, skewness=1.11, excess=2.36,
mode=0.50
F(t), c=10, k=5, S=2, mean=1.64, st dev=0.21,
skewness=-0.39, excess=0.40, mode=1.68
Figure 4.16: Burr Cumulative Function
EXAMPLE: 4.19 Suppose that one has the above information and that
it is of interest to nd the probability that the device will survive past 100
hours, given that it has survived to 50 hours. This probability is easy to
compute under the Burr distribution assumption and is given by:
R(100|50) = P(T > 100|T > 50) =
_
1 +
_
50
100
_
3.33
_
4
_
1 +
_
100
100
_
3.33
_
4
=
1.461
16
= 0.091
4.12 Inverse Gaussian distribution
Another distribution that is used in reliability in more and more applications
is the inverse Gaussian (or inverse Normal). The inverse Gaussian is usually
presented as a two-parameter distribution with parameters which repre-
4.12. INVERSE GAUSSIAN DISTRIBUTION 153
0
2
4
6
8
10
12
14
16
18
20
0 1 2 3 4 5 6
t
h
(
t
)
h(t), c=3, k=2, S=0.5,
mean=0.40, st dev=0.20,
skewness=1.59, excess=7.81,
mode=0.33
h(t), c=4.8737, k=6.15784, S=1,
mean=0.64, st dev=0.16,
mode=0.65
h(t), c=2.105, k=5.0, S=1.5267,
mean=0.68, st dev=0.38,
mode=0.50
h(t), c=10, k=5, S=2, mean=1.64,
st dev=0.21, skewness=-0.39,
excess=0.40, mode=1.68
Figure 4.17: Burr Hazard Function
sents the mean, and which represents the shape parameter. It has a non-
monotonic hazard but has an advantage that the hazard does not necessarily
decrease to zero as the time increases. The inverse Gaussian distribution is a
unimodal distribution. Also, the inverse Gaussian is asymptotically normal
as . For the inverse Gaussian distribution with parameters and :
f(t) =
_

2 t
3
_1
2
e

2
2
t
(t)
2
, t , > 0
R(t) =
_
_
t
_1
2
_
1
t
_
_
e
2
t
_1
2
_
1 +
t
_
_
,
where represents the standard normal cdf, h(t) =
f(t)
R(t)
, E(T) = , V (T) =
2
=

3
,
3
=
3
5
2
,
4
= 15
_
3
_
+3
_
2
_
, coecient of skewness =
1
= 3
_
,
coecient of excess = 3 = 15
, coecient of variation=CV=
_
,
mode=
3
2
2
+
_
1 +
9
2
4
2
_1
2
.
The inverse Gaussian density, cumulative and hazard function are illus-
trated in Figure 4.23, Figure 4.24 and Figure 4.25, respectively and for and
pairs: (1, 1), (2, 0.60), (1, 5), (3, 125).
0
0.2
0.4
0.6
0.8
1
1.2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t
f
(
t
)
f(t), mu=1, lam=1, mean=1, st
dev=1,skewness=3, excess=15, mode=0.30
f(t), mu=2, lam=0.6, mean=2, st dev=3.65,
skewness=5.48, excess=50, mode=0.20
f(t), mu=1, lam=5, mean=1, st dev=0.45,
f(t), mu=3, lam=125, mean=3, st dev=0.46,
skewness=0.46, excess=0.36, mode-2.89
Figure 4.18: Inverse Gaussian Density Function
The mode of the hazard is between t
mode
and
2
3
, and the asympotote
of the hazard as time increases is

2
2
. Since the hazard function does not
decrease asymptotically to zero as does the lognormal, this is seen by some
to be an advantage of the inverse Gaussian to the lognormal. The distribution
of the sample mean from a sample of n independent inverse Gaussian random
variables is also inverse Gaussian with parameters and n.
4.12. INVERSE GAUSSIAN DISTRIBUTION 155
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t
F
(
t
)
F(t), mu=1, lam=1, mean=1, st
dev=1,skewness=3, excess=15, mode=0.30
F(t), mu=2, lam=0.6, mean=2, st dev=3.65,
F(t), mu=1, lam=5, mean=1, st dev=0.45,
F(t), mu=3, lam=125, mean=3, st dev=0.46,
Figure 4.19: Inverse Gaussian Distribution Function
EXAMPLE: 4.20 It is assumed that the time to failure of a particular
device follows an inverse Gaussian distribution with parameters = 4 hours
and = 16. In this case, the mean is 4 hours and the standard deviation is
2 hours. The sample mean

T from a sample of 10 devices also has an inverse
Gaussian distribution with mean
T
= 4 hours,
T
= 160 and standard
deviation
0.4 = 0.63. For this example, the probability that the mean of a
sample of 10 devices is greater than 3 hours is:
P{
T > 3} = R(3) = (1.826) e

8
(12.78) = 0.966 0 = 0.966.
EXAMPLE: 4.21 Suppose it is known that the time to failure of a device
has an inverse Gaussian distribution with mean 50 hours and a standard
deviation of 20 hours. In this case, one knows that = 50 and one can
compute that:
0
0.5
1
1.5
2
2.5
3
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t
h
(
t
)
h(t), mu=1, lam=1, mean=1, st dev=1,skewness=3,
excess=15, mode=0.30
h(t), mu=2, lam=0.6, mean=2, st dev=3.65,
h(t), mu=1, lam=5, mean=1, st dev=0.45,
h(t), mu=3, lam=125, mean=3, st dev=0.46,
Figure 4.20: Inverse Gaussian Hazard Function
400 =
125000
or = 312.5
If one is interested in the probability that the device will survive until +
2.5 = 100 hours, one nds that:
R(100) =
_
_
312.5
100
_1
2
_
1
100
50
_
_
e
2(312.5)
50
_
_
312.5
100
_1
2
_
1
100
50
_
_
R(100) = 0.0386 0.0153 = 0.0233

Chap4 - 1 Time To Failure

Uploaded by

Copyright:

Available Formats

Chap4 - 1 Time To Failure

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chap4 - 1 Time To Failure

Uploaded by

Copyright:

Available Formats

Chapter 4

and the coecient of excess

, if r > 1, coecient of skewness =

, which increases linearly in t.

, < y < , R(y) = e

, mode = , E(Y ) = = 0.5772,

T > 3} = R(3) = (1.826) e

You might also like