3

3 Probability and Life Distributions

for Reliability Analysis

In reliability engineering, data are often collected from analysis of incoming parts and
materials, tests during and after manufacturing, elded products, and warranty
returns. If the collected data can be modeled, then properties of the model can be
used to make decisions for product design, manufacture, reliability assessment, and
logistics support (e.g., maintainability and operational availability).
In this chapter, discrete and continuous probability models (distributions) are intro-
duced, along with their key properties. Two discrete distributions (binomial and
Poisson) and ve continuous distributions (exponential, normal, lognormal, Weibull,
and gamma) that are commonly used in reliability modeling and hazard rate assess-
ments are presented.

3.1 Discrete Distributions

A discrete random variable is a random variable with a nite (or countably in nite)
set of values. If a discrete random variable (X ) has a set of discrete possible values
(x1, x2, . . . xn), a probability mass function (pmf), f(xi) , is a function such that

f ( xi ) ≥ 0, for all i
n

∑ f (x ) = 1
i =1
i (3.1)

f ( xi ) = P {X = xi }.

The cumulative distribution function (cdf) is written as:

F (xi ) = P {X ≤ xi }. (3.2)

Reliability Engineering, First Edition. Kailash C. Kapur and Michael Pecht.

© 2014 John Wiley & Sons, Inc. Published 2014 by John Wiley & Sons, Inc.

45
3 Probability and Life Distributions for Reliability Analysis

The mean, µ, and variance, σ2, of a discrete random variable are de ned using the
pmf as (see also Chapter 2):

µ = E [X ] = ∑ x f (x )
i
i i (3.3)

σ 2 = V [X ] = ∑ (x − µ)
i
i
2
f (xi ) = ∑x
i
i
2
f ( xi ) − µ 2 . (3.4)

3.1.1 Binomial Distribution

The binomial distribution is a discrete probability distribution applicable in situations
where there are only two mutually exclusive outcomes for each trial or test. For
example, for a roll of a die, the probability is one to six that a speci ed number will
occur (success) and ve to six that it will not occur (failure). This example, known as
a “Bernoulli trial,” is a random experiment with only two possible outcomes, denoted
as “success” or “failure.” Of course, success or failure is de ned by the experiment.
In some experiments, the probability of the result not being a certain number may be
de ned as a success.
The pmf, f(x), for the binomial distribution gives the probability of exactly k suc-
cesses in m attempts:

 m
f (k ) =   p k q m−k , 0 ≤ p ≤ 1, q = 1− p, k = 0, 1, 2, … , m, (3.5)
 k 

where p is the probability of the de ned success, q (or 1 – p) is the probability of

failure, m is the number of independent trials, k is the number of successes in m trials,
and the combinational formula is de ned by

 m m!
  ≡ Ckm = , (3.6)

k  k !( m − k )!

where ! is the symbol for factorial. Since (p + q) equals 1, raising both sides to a
power j gives
j
( p + q ) =1. (3.7)

The general equation is

∑ f (k ) = F (m) = ( p + q )
k =0
m
= 1. (3.8)

The binomial expansion of the term on the left in Equation 3.7 gives the probabilities
of j or less number of successes in j trials, as represented by the binomial distribution.
For example, for three components or trials each with equal probability of success (p)
or failure (q), Equation 3.7 becomes:
3
( p + q ) = p3 + 3 p2 q + 3 pq 2 + q3 = 1. (3.9)

46
 3.1 Discrete Distributions

The four terms in the expansion of (p + q)3 give the values of the probabilities for
getting 3, 2, 1, and no successes, respectively. That is, for m = 3 and the probability
of success = p, f(3) = p3, f(2) = 3p2q, f(1) = 3pq2, and f(0) = q3.
The binomial expansion is also useful when there are products with different success
and failure probabilities. The formula for the binomial expansion in this case is
m

∏( p + q ) = 1,
i =1
i i (3.10)

where i pertains to the ith component in a system consisting of m components. For

example, for a system of three different components, the expansion takes the form

( p1 + q1 )( p2 + q2 )( p3 + q3 ) = p1 p2 p3 + ( p1 p2 q3 + p1q2 p3 + q1 p2 q3 )
(3.11)
+ ( p1q2q3 + q1 p2 q3 + q1q2 p3 ) + q1q2 q3 = 1,

where the rst term on the right side of the equation gives the probability of success
of all three components, the second term (in parentheses) gives the probability of
success of any two components, the third term (in parentheses) gives the probability
of success of any one component, and the last term gives the probability of failure
for all components.
The cdf for the binomial distribution, F(k), gives the probability of k or fewer suc-
cesses in m trials. It is de ned by using the pmf for the binomial distribution,
k
F (k ) = ∑(
i =0
m
i ) pi q(m−i ). (3.12)

For a binomial distribution, the mean, µ is given by

µ = mp (3.13)

and the variance is given by

σ 2 = mp(1− p). (3.14)

Example 3.1

An engineer wants to select four capacitors from a large lot of capacitors in which 10
percent are defective. What is the probability of selecting four capacitors with:

(a) Zero defective capacitors?

(b) Exactly one defective capacitor?
(c) Exactly two defective capacitors?
(d) Two or fewer defective capacitors?

Solution:
Let success be de ned as “getting a good capacitor.” Therefore, p = 0.9, q = 0.1, and
m = 4. Using Equation 3.5 and Equation 3.6, f (4) is the probability of all four being

47
3 Probability and Life Distributions for Reliability Analysis

good (no defectives)—that is, based on four components (trials), the values of p and
q are equal for all the capacitors.

 4
f (4 ) =  (0.9) (0.1) = 0.6561.
4 0
4

Another way to solve this problem is by de ning success as “getting a certain

number of defective capacitors” with p = 0.1 and thus q = 0.9. In this case, f(0) gives
the probability that there will be no defectives in the four selected samples. That is,

4
(a) f (0) =  (0.1) (0.9) = 0.6561
0 4
0

Continuing with the latter approach, the solution to problems (b), (c), and (d), respec-
tively, are:

 4
(b) f (1) =  (0.1) (0.9) = 0.2916
1 3
1
4
(c) f (2) =  (0.1) (0.9) = 0.0486
2 2
2
(d) F (2) = f (0) + f (1) + f (2) = 0.9963.

Example 3.2

Consider a product with a probability of failure in a given test of 0.1. Assume 10 of

these products are tested.

(a) What is the expected number of failures that will occur in the test?
(b) What is the variance in the number of failures?
(c) What is the probability that no product will fail?
(d) What is the probability that two or more products will fail?

Solution:
Here m = 10, and p = 0.1.

(a) The expected number of failures is the mean,

µ = mp = (10 × 0.1) = 1.

(b) The variance is:

σ 2 = mp (1− p) = [10 ×0.1×(1− 0.1)] = 0.9.

(c) The probability of having no failures is the pmf with k = 0. That is,

10
f (0) =  × 0.10 ×(1− 0.1) = 0.349.
10
 0 

48
 3.1 Discrete Distributions

(d) The probability of having two or more failures is the same as 1 minus the prob-
ability of having zero or one failures. It is given by:

Pr (two or more failures ) = [1− { f (0) + f (1)}]

= 1− 0.349 − {10 × 0.1×(1− 0.1) }
9

= 0.264.

Example 3.3

An electronic automotive control module consists of three identical microprocessors

in parallel. The microprocessors are independent of each other and fail independently.
For successful operation of the module, at least two microprocessors must operate
normally. The probability of success of each microprocessor for the duration of the
warranty is 0.95. Determine the failure probability of the control module during
warranty.

Solution:
The module fails when two or more microprocessors fail. In other words, the module
fails when only one or none of the microprocessors is working. So the probability of
failure of the module during warranty will be given by:

Pr (module fails during warranty ) = [ f (0) + f (1)],

where m = 3 components, k = 0 or 1 is the total number of working components,

p = 0.95, and q = 0.05. Therefore:

Pr (module fails during warranty ) = (0.05) + {3× 0.95×(0.05) }

3 2

= 0.00725.

Example 3.4

The probability of a Black Hawk helicopter surviving a mission is 0.91. If seven

helicopters are sent on a mission and ve must succeed for mission success, what is
the probability of mission success?

Solution:
This is also called a 5-out-of-7 system in reliability (see Chapter 17). If the number
of successes is ve or more, the mission will be a success. Hence, the probability of
mission success or mission reliability is

m 7
 m i 7
RS = ∑
i =k
  R (1− R )m−i =
 i  ∑ i R (1− R )
i =5
i 7−i

7 7 7

=   0.915 (0.09) +   0.916 (0.09) +   0.917 (0.09)
2 1 0
5 6 7
= 0.1061 + 0.3577 + 0.5168 = 0.9806.

49
3 Probability and Life Distributions for Reliability Analysis

3.1.2 Poisson Distribution

In situations where the probability of success (p) is very low and the number (m) of
samples tested (i.e., the number of Bernoulli trials conducted) is large, it is cumber-
some to evaluate the binomial coef cients. A Poisson distribution is useful in such
cases.
The pmf of the Poisson distribution is given as:

µ k −µ
f (k ) = e ; k = 0, 1, 2, … , (3.15)
k!

where µ is the mean and also the variance of the Poisson random variable.
For a Poisson distribution for m Bernoulli trials, with the probability of success in
each trial equal to p, the mean and the variance are given by:

µ = mp, σ 2 = mp. (3.16)

The Poisson distribution is widely used in industrial and quality engineering

applications. It is also the foundation of some of the attribute control charts.
For example, it is used in applications such as determination of particles of contami-
nation in a manufacturing environment, number of power outages, and aws in rolls
of polymers.

Example 3.5

Solve Example 3.2 using the Poisson distribution approximation.

Solution:
The expected number of failures is the same as the mean,

µ = (10 )(0.1) = 1.

The variance is also equal to 1.

The probability of obtaining no failures is the same as the pmf with k = 0,

f (0) = e−µ = e−1 = 0.3678.

The probability of getting two or more failures is the same as 1 minus the probabil-
ity of obtaining zero or one failures. It is given by:

Pr (two or more failures ) = [1 − { f (0) + f (1)}]

= [1 − {0.3678 + e−1 }] = 0.2642.

Note the differences from Example 3.2, because m is not very large.

3.1.3 Other Discrete Distributions

Other discrete distributions that are used in reliability analysis include the geometric
distribution, the negative binomial distribution, and the hypergeometric distribution.

50
 3.2 Continuous Distributions

With the geometric distribution, the Bernoulli trials are conducted until the rst
success is obtained. The geometric distribution has the “lack of memory” property,
implying that the count of the number of trials can be started at any trial without
affecting the underlying distribution. In this regard, this distribution is similar to the
continuous exponential distribution, which will be described later.
With the negative binomial distribution (a generalization of the geometric distribu-
tion), the Bernoulli trials are conducted until a certain number of successes are
obtained. Negative binomial distribution is, however, conceptually different from the
binomial distribution, since the number of successes is predetermined, and the number
of trials is a random variable.
With the hypergeometric distribution, testing or sampling is conducted without
replacement from a population that has a certain number of defective products. The
hypergeometric distribution differs from the binomial distribution in that the popula-
tion is nite and the sampling from the population is made without replacement.

3.2 Continuous Distributions

If the range of a random variable, X, extends over an interval (either nite or in nite)
of real numbers, then X is a continuous random variable. The cdf is given by:

F (x ) = P {X ≤ x}. (3.17)

The probability density function (pdf) is analogous to pmf for discrete variables,
and is denoted by f(x) , where f(x) is given by (if F(x) is differentiable):

d
f (x) = F ( x ), (3.18)
dx
which yields
x

F (x ) = ∫
−∞
f (u )du. (3.19)

The mean, µ, and variance, σ2, of a continuous random variable are de ned over
the interval from –∞ to +∞ in terms of the probability density function as (see
Chapter 2):
+∞

µ= ∫ xf (x)dx
−∞
(3.20)

+∞ +∞

∫ (x − µ ) ∫x
2
σ =
2
f (x )dx = 2
f ( x)dx − µ 2. (3.21)
−∞ −∞

Reliability is concerned with the time to failure random variable T and thus X is
replaced by T. Thus, Equation 3.19 corresponds to Equation 2.5 and Equation 3.20
corresponds to Equation 2.46.

51
3 Probability and Life Distributions for Reliability Analysis

Example 3.6

The pdf for the time to failure of an appliance is given by:

1
f (t) = t ⋅ e−t 4,
16

where t is in years, and t > 0.

(a) What is the probability of failure in the rst year?

(b) What is the probability of the appliance lasting at least 5 years?
(c) If no more than 5% of the appliances will require warranty service, what is the
maximum number of months for which the appliance should be warranted?

Solution:

(a) For the given pdf, the cdf is

t
1 t 
F (t) =
16 ∫ τ ⋅e
0
−τ 4
⋅ d τ = 1−  + 1 e−t 4.
4 

The probability of failure during the rst year = F(1) = 0.0265.

(b) The probability of lasting more than 5 years is = [1 – F(5)] = [1 – 0.3554] =

0.6446.
(c) For this case, F(t0) has to be less than or equal to 0.05, where t0 is the
warranty period. From the above results, we nd that the time has to be more
than 1 year. Also, F(2) is equal to 0.09, hence the warranty period should be
between 1 and 2 years. We can nd that for no more than 5% warranty service,
t0 = 1.42 years. Therefore, the warranty should be set at no greater than 17
months.

Example 3.7

The time-to-failure random variable, T, of a product follows the following probability

density function:

t
F (t) = , 0 ≤ t ≤ 400
80, 000
= 0, otherwise.

We give solutions to the following four parts.

(a) Find the standard deviation for the time-to-failure random variable.

52
 3.2 Continuous Distributions

Solution:

400
400
t t3
E [T ] = ∫ 0
t⋅
80, 000
dt =
240, 000 0
= 266.67

400
400
t t4
E [T 2 ] = ∫ 0
t2 ⋅
80, 000
dt =
320, 000 0
= 80, 000,

then variance V[T] and the standard deviation are given by (see Eq. 2.35 in
Chapter 2)

V [T ] = E [T 2 ]− (E [T ]) = 8888.89
2

Standard deviation = V [T ] = 94.28.

(b) Find the coef cient of skewness of the distribution for the time-to-failure
random variable.

Solution:

µ1 = µ = µ1′ = E [T ]
µ2′ = E [T 2 ] µ2 = V [T ]
400
400
t t5
µ3′ = ∫ 0
t 3
80, 000
dt =
400, 000 0
= 25.6 ×106.

Using Equation 2.52 (Chapter 2), we have

µ3 = µ3′ − 3µµ2′ + 3µ 2 µ1′ − µ3

= µ3′ − 3µµ2′ + 2 µ3 = −474, 074.

Hence,

µ3 −474, 074
α3 = = = −0.5657.
µ232
(8888.89)1.5

The above triangular distribution is negatively skewed, which is good in terms of

reliability because the time to failure is a “larger the better” characteristic for the
product.

(c) Find the B5 and B50 life of the product based on the above probability density
function.

Solution:
Using Equation 2.5 and Equation 2.40, give

53
 3 Probability and Life Distributions for Reliability Analysis

t
u t2
F (t) = ∫0
80, 000
du =
160, 000
B52
F (B5 ) = 0.05 =
160, 000
B5 = 89.44
B502
F (B50 ) = 0.5 =
160, 000
B50 = 282.843.

(d) Draw the failure rate (or hazard rate) curve for the above product by evaluating
it at t = 0, 50, 100, 300, 400.

Solution:
Using Equation 2.5, the following table can be developed and the hazard rate function
h(t) is drawn as shown in Figure 3.1.

Value of h(t) vs. t

t f(t) R(t) h(t)

0 0 1 0
50 0.000625 0.984375 0.000635
100 0.00125 0.9375 0.001333
300 0.00375 0.4375 0.008571
400 0.005 0 Infinity

The values of the hazard function are given in Figure 3.1. It is clear that for the
above triangular distribution, the failure rate is increasing; such distributions have the
property of an increasing failure rate (IFR). Many products that wear or deteriorate
with time will exhibit IFR behavior.

h(t )
0.02

0.015
Hazard rate h(t )

0.01
h(t )

0.005

5E-17
0 50 100 150 200 250 300 350
Time (t ) Figure 3.1 Hazard rate function, h(t).

54
 3.2 Continuous Distributions

3.2.1 Weibull Distribution

The Weibull distribution is a continuous distribution developed in 1939 by Waloddi
Weibull (1939), and who presented it in detail in 1951 (Weibull 1951). The Weibull
distribution is widely used for reliability analyses because a wide diversity of hazard
rate curves can be modeled with it. The distribution can also be approximated to other
distributions under special or limiting conditions. The Weibull distribution has been
applied to life distributions for many engineered products, and has also been used for
reliability testing, material strength, and warranty analysis.
The probability density function for a three-parameter Weibull probability distribu-
tion function is

 t−γ β
− 
−β β −1  η 
f (t) = βη (t − γ ) e , (3.22)

where β > 0 is the shape parameter, η > 0 is the scale parameter, which is also denoted
by θ in many references and books, and γ is the location or time delay parameter. The
reliability function is given by

∞  t−γ β
− 
 η 
R (t) = ∫t
f (τ ) d τ = e . (3.23)

It can be shown that Equation 3.23 gives, for a duration t = γ + η, starting at time
t = 0, a reliability value of R(t) = 36.8%, regardless of the value of β. Thus, for any
Weibull failure probability density function, 36.8% of the products survive for
t = γ + η.
The time to failure of a product with a speci ed reliability, R, is given by
1β
t = γ + η [− ln R (t)] . (3.24)

The hazard rate function for the Weibull distribution is given by

β−1
f (t) β  t − γ 
h (t) = =   . (3.25)
R (t) η  η 

The conditional reliability function is (see Eq. 2.39):

R (t + t1 )
R (t, t1 ) =
R (t1 )
  t + t1 − γ  β  t1 − γ  β  (3.26)
= exp −   +  .
  η   η  
 

Equation 3.26 gives the reliability for a new mission of duration t for which t1 hours
of operation were previously accumulated up to the beginning of this new mission.
It is seen that the Weibull distribution is generally dependent on both the age at the
beginning of the mission and the mission duration (unless β = 1). In fact, this is true
for most distributions, except for the exponential distribution (discussed later).

55
 3 Probability and Life Distributions for Reliability Analysis

Table 3.1 Weibull distribution parameters

Location γ
Shape parameter β
Scale parameter η
Mean (arithmetic average) γ + η + Γ(1/β + 1)
Median (B50, or time at 50% failure) γ + η(ln2)1/β
Mode (highest value of f(t)) for β > 1 γ + η(l − 1/β)1/β
for β = 1 γ
Standard deviation 2  1 
η Γ  + 1 − Γ2  + 1
 β   β 

β=3
Probability density function f(t)

β=2

β=1

Figure 3.2 Effects of shape parameter β on prob-

Operating time ability density function, where η = 1 and γ = 0.

Table 3.1 lists the key parameters for a Weibull distribution and values for mean,
median, mode, and standard deviation. The function Γ is the gamma function,
for which the values are available from statistical tables and also are provided in
Appendix B.
The shape parameter of a Weibull distribution determines the shape of the hazard
rate function. With 0 < β < 1, the hazard rate decreases as a function of time, and
can represent early life failures (i.e., infant mortality). A β = 1 indicates that the
hazard rate is constant and is representative of the “useful life” period in the “ideal-
ized” bathtub curve (see Figure 2.6). A β > 1 indicates that the hazard rate is increas-
ing and can represent wearout. Figure 3.2 shows the effects of β on the probability
density function curve with η = 1 and γ = 0. Figure 3.3 shows the effect of β on the
hazard rate curve with η = 1 and γ = 0.
The scale parameter η has the effect of scaling the time axis. Thus, for a xed γ and
β, an increase in η will stretch the distribution to the right while maintaining its start-
ing location and shape (although there will be a decrease in the amplitude, since the
total area under the probability density function curve must be equal to unity). Figure
3.4 shows the effect of η on the probability density function for β = 2 and γ = 0.
The location parameter locates the distribution along the time axis and thus esti-
mates the earliest time to failure. For γ = 0, the distribution starts at t = 0. With
γ > 0, this implies that the product has a failure-free operating period equal to γ.
Figure 3.5 shows the effects of γ on the probability density function curve for β = 2
and η = 1. Note that if γ is positive, the distribution starts to the right of the t = 0

56
 3.2 Continuous Distributions

β=3
β = 0.5
Hazard rate h(t)

β=1

Figure 3.3 Dependence of hazard rate on shape

Operating time parameter, where η = 1 and γ = 0.

0.8
Probability density function f(t)

η=1
0.6

η=2
0.4
η=3
0.2

0
0 1 2 Figure 3.4 Effects of scale parameter η on the pdf
Operating time of a Weibull distribution, where β = 2 and γ = 0.

line, or the origin. If γ is negative, the distribution starts to the left of the origin, and
could imply that failures had occurred prior to the time t = 0, such as during trans-
portation or storage. Thus, there is a probability mass F(0) at t = 0, and the rest of
the distribution for t > 0 is given in Figure 3.5. The Weibull distribution can also be
formulated as a two-parameter distribution with γ = 0.
The reliability function for the two-parameter Weibull distribution is
∞  t β
− 
R (t ) = ∫t
f (τ ) d τ = e  η 
(3.27)

and the hazard rate function is

β −1
f (t) β  t 
h (t) = =   . (3.28)
R (t ) η  η 

The two-parameter Weibull distribution can be used to model skewed data. When
β < 1, the failure rate for the Weibull distribution is decreasing and hence can be used
to model infant mortality or a debugging period, situations when the reliability in
terms of failure rate is improving, or reliability growth. When β = 1, the Weibull
distribution is the same as the exponential distribution. When β > 1, the failure rate
is increasing, and hence can model wearout and the end of useful life. Some examples
of this are corrosion life, fatigue life, or the life of antifriction bearings, transmission
gears, and electronic tubes.

57
 3 Probability and Life Distributions for Reliability Analysis

1
Probability density function f(t) γ=0 γ>0

γ<0

0
0 1 2 3 4 5
Figure 3.5 Effects of location parameter γ,
Operating time
where β = 2 and η = 1.

The three-parameter Weibull distribution is a model when there is a minimum life

or when the odds of the component failing before the minimum life are close to zero.
Many strength characteristics of systems do have a minimum value signi cantly
greater than zero. Some examples are electrical resistance, capacitance, and fatigue
strength.

Example 3.8

Assume that the time to failure of a product can be described by the Weibull
distribution, with estimated parameter values of η = 1000 hours, γ = 0, and β = 2.
Estimate the reliability of the product after 100 hours of operation. Also determine
the MTTF.

Solution:
From Equation 3.27, we have:
2
R (100 ) = e−(100 /1000) = 0.990.

And from Table 3.1 we have

MTTF = 1000Γ (1 2 + 1) = 1000Γ (1.50) = 886 hours.

where the value of Γ (1.50) can be found from the table in Appendix B.

Example 3.9

Suppose that the life distribution for miles to failure for a give failure mode for the
transmission of a GM Cadillac model follows the two-parameter Weibull distribution
with η = 150,000 mi, β = 4.5.

(a) Find the mean miles between failures or the expected life in miles of these
transmissions.

58
 3.2 Continuous Distributions

Solution:
Using the table in Appendix B, we have

 1  1 
E [T ] = ηΓ1 +  = 150, 000Γ1 +
 β   4.5 
= 150, 000 ×Γ(1.22222) = 150, 000 × 0.912573
= 136, 886 mi.

(b) Find the standard deviation for the miles to failure random variable.

Solution:
Again, using the table in Appendix B,

  2  1 
V [T ] = η 2 Γ1 +  − Γ2 1 + 
  β   β 
= 150, 0002  Γ (1.4444) − Γ (1.2222) 
2

= 150, 0002 ×(0.8858 − 0.9125732 ) = 1.19×109

Standard deviation = V (T ) = 1.19×109 = 34, 513 mi.

(c) If GM gives a warranty for 70,000 mi on these transmissions, what percent of

these transmissions will fail during the warranty period?

Solution:

 70 ,000 4.5
− 
150 ,000 
1− R (70, 000) = 1− e = 0.03188.

Thus, 3.188% of the transmissions will fail during the warranty period for the given
failure mode.

Example 3.10

Suppose that the life distribution (life in years of continuous use) of hard disk drives
for a computer system follows a two-parameter Weibull distribution with the following
parameters: β = 3.10 and η = 5 years.

(a) The manufacturer gives a warranty for 1 year. What is the probability that a
disk drive will fail during the warranty period?

Solution:

  1 3.10 
F (1) = 1− R (1) = 1− exp −  
 5  
 
= 1− 0.993212 = 0.006788.

(b) Find the mean life and the median life (B50) for the disk drive.

59
3 Probability and Life Distributions for Reliability Analysis

Solution:

 1 
mean = 5Γ 1 + = 5Γ (1.32258)
 3.10 
= 5 ×0.89431 = 4.47155 years.

And to nd the median life, we have

  B 3.10 
R (B50 ) = exp − 50   = 0.5
  5  
 
3.10
 B50 
 = − ln (0.5) = 0.693147
 5 
B50 = 0.888492 ×5 = 4.44246 years.

(c) By what time will 95% of the disk drives fail? (Find the B95 life).

Solution:

  B 3 10 
.
R (B95 ) = exp − 95   = 0.05
  5  
 
3.110
 B95 
 = − ln (0.05) = 2.99573
 5 
B95 = 1.42466×5 = 7.12329 years.

Example 3.11

The failure rate of a component, in failures per year, is given by:

h (t) = 0.003t 2 , t ≥ 0.

(a) Find an expression for the reliability function and the probability density func-
tion for the time to failure of the component.

Solution:
Using Equation. 2.35 and Equation 2.36 in Chapter 2, we have

h (t) = 0.003t 2
t

H (t) = ∫ 0.003x dx = 0.001t

0
2 3

R (t) = exp[−0.001t3 ]
f (t ) = 0.003t 2 exp[−0.001t3 ].

60
 3.2 Continuous Distributions

This is easily recognizable as a Weibull distribution with the following values of the
parameters:

Weibull: β = 3, η = 10.

(b) Find the expected life (MTTF) for the component.

Solution:

 1
E [T ] = θΓ1 + 
 β 
 1
= 10Γ 1 +  = 10 × 0.89298 = 8.9298 yeears.
 3

(c) Find the B10 (the 10th percentile) for the life of the component.

Solution:
We need to nd the value of t, such that the item has a 10% chance of failing. This
is equivalent to nding the point at which R(t) = 0.9. Solving for t:
3
0.9 = e−0.001(B10 )
3
ln 0.9 = −0.001(B10 )
− ln (0.9)
B10 = 3 = 4.723 years.
0.001

3.2.2 Exponential Distribution

The exponential distribution is a single-parameter distribution that can be viewed as
a special case of a Weibull distribution, where β = 1. The probability density function
has the form

f (t) = λ0 e−λ0t, t ≥ 0, (3.29)

where λ0 is a positive real number, often called the constant failure rate. The parameter
λ0 is typically an unknown that must be calculated or estimated based on statistical
methods discussed later in this section. Figure 3.6 gives a graph for an exponential
distribution, with λ0 = 0.10. Table 3.2 summarizes the key parameters for the expo-
nential distribution.
Once λ0 is known, the reliability can be determined from the probability density
function as

∞ ∞
R (t) = ∫ t
f (τ ) d τ = ∫ t
λ0 e−λ0t d τ = e−λ0t. (3.30)

The cdf or unreliability is given by

F (t ) = Q (t ) = 1− exp[−λ0t ]. (3.31)

61
3 Probability and Life Distributions for Reliability Analysis

0.01

0.008

f(t) 0.006

0.004

0.002

100 200 300 400 500 Figure 3.6 An example of an exponential

t distribution.

Table 3.2 Exponential distribution parameter

Scale parameter 1/λ0

Median (B50) 0.693/λ0
Mode (highest value of f(t)) 0
Standard deviation 1/λ0
Mean 1/λ0

As mentioned, the hazard rate for the exponential distribution is constant:

f (t) 1
h (t) = = −λ0t (λ0 e−λ0t ) = λ0. (3.32)
R (t) e

The conditional reliability is

R (t + t1 )
R (t, t1 ) = = e−λ0 (t+t1) e −λ0t1 = e−λ0t. (3.33)
R (t1 )

Equation 3.33 shows that previous usage (e.g., tests or missions) do not affect future
reliability. This “as good as new” result stems from the fact that the hazard rate is a
constant and the probability of a product failing is independent of the past history
or use of the product.
The mean time to failure (MTTF) for an exponential distribution, also denoted
by θ, is determined from the general equation for the mean of a continuous
distribution:
∞ ∞
1
MTTF = ∫ R (t)dt = ∫ e
0 0
−λ0t
dt =
λ0
. (3.34)

Thus, the MTTF or the MTBF is inversely proportional to the constant failure rate,
and thus the reliability can be expressed as

R (t ) = e−t MTBF . (3.35)

The MTBF is sometimes misunderstood to be the life of the product or the time
by which 50% of products will fail. For a mission time of t = MTBF, the reliability

62
 3.2 Continuous Distributions

calculated from Equation 3.30 gives R(MTBF) = 0.368. Thus, only 36.8% of the
products survive a mission time equal to the MTBF.

Example 3.12

Show that the exponential distribution is a special case of the Weibull distribution.

Solution:
From Equation 3.22, set β = 1 and γ = 0
t
1 −
f (t) = e η .
η

Thus, in this case, the Weibull distribution reduces to the single-parameter exponential
distribution with λ0 = 1/η. The reliability and the hazard rate functions simplify to:
t
−
R (t) = e η

1
h (t) = ,
η

where

1
η= .
λ0

If β = 1 and γ > 0, then the Weibull distribution is the same as the exponential dis-
tribution with minimum life, γ, or is also a two-parameter exponential distribution.

Example 3.13

Consider an electronic product that exhibits a constant hazard rate. If the MTBF is
5 years, at what time will 10% of the products fail?

Solution:
Using Equation 3.35 with R = 0.90 and MTBF ≈ 43,800 hours (5 years), we solve
for t, where t is in hours. Thus, t = –[(MTBF) × ln(R)] ≈ 4600 hours, or nearly half
a year.

Example 3.14

Here we consider a mixture of exponential distributions. The pdf for the life of a
device is given by the following probability density function, which is a mixture of
two exponential distributions.

1 3
f (t) = e−t + e−2t, t ≥ 0.
4 2
(a) Prove that the above function is a valid pdf.

63
3 Probability and Life Distributions for Reliability Analysis

Solution:
∞ ∞
 1 −t 3 −2t  1 3 1
∫
0
 e + e  dt = − e−t − × e−2t
4 2 
 4 2 2 0

1 3
= − (0 − 1) − (0 −1) = 1.
4 4
Therefore, the above function is a valid pdf.

(b) Find the probability that a device will last at least 3 hours.

Solution:
1 3
f (t) = e−t + e−2t, t ≥ 0.
4 2
∞ ∞
1 3 
R (t) = ∫ f (τ )d τ = ∫  4 e
t t
−τ
+ e−2τ  d τ
2 
∞
 1 3  1 3
= − e−τ − e−2 τ  = e−t + e−2t
 4 4  t 4 4
1 3
R (3) = e−3 + e−6 = 0.01431.
4 4
Alternatively: R(t) = 1 − F(t).
t t
 1 −τ 3 −2 τ   1 3 
F (t) = ∫
0
 e + e  d τ = − e−τ − e−2 τ 
4 2   4 4 
0

1 3
= − e−t − e−2t + 1
4 4
1 3
F (3) = − e−3 − e−6 + 1 = 0.98569
4 4
R (3) = 1− F (3) = 1 = 1 − 0.98569 = 0.01431.
(c) Find the expected life or the MTBF of the device.

Solution:
∞ ∞
1 3 
MTBF = ∫ R (t)dt = ∫  4 e
0 0
−t
+ e−2 t  dt
4 
∞
 1 3  1 3 5
= − e−t − e−2t  = + = hours.
 4 8 0 4 8 8

3.2.3 Estimation of Reliability for Exponential Distribution

For reliability tests in which the hazard rate is assumed to be constant, and the time
to failure can be assumed to follow an exponential distribution, the constant failure

64
 3.2 Continuous Distributions

1 t1
2
. t2
No. of units

.
r tr
.
.
Suspension
n Suspension Figure 3.7 Failure-truncated test. Failures are
Time denoted by .

1
2
No. of units

. ...
. ...
. ...
n
t0
Time Figure 3.8 Time-truncated test. , Failures.

rate can be estimated by life testing. There are various ways to test the items. Figure
3.7 gives an example of a failure-truncated test, in which n items on individual test
stands are monitored to failure. The test ends as soon as there are r failures (without
replacement r ≤ n), as shown in Figure 3.7.
The total time on test, TT, considering both failed and unfailed (or suspended) units,
is calculated by the following equation:

r
TT = ∑ t + (n − r)t .
i =1
i r (3.36)

Another test situation is called time-truncated testing. In Figure 3.8, there are n test
stands (or n items on test in a test chamber). The units are monitored and replaced
as soon as they fail. Testing for these units continues until some predetermined time,
t0. In this case, the total time on test is

TT = nt0. (3.37)

Then the point estimator (minimum variance unbiased estimator) for θ, the
MTBF, is

T
θˆ = T . (3.38)
r

Further details are given in Chapter 13. Also, the point estimator for λ is

1
λˆ = . (3.39)
θˆ

Chapter 13 will present the methodology for the point estimation and con dence
interval for several test situations and underlying life distributions.

65
3 Probability and Life Distributions for Reliability Analysis

Example 3.15

Seven prototypes are monitored for some failure during development testing or as
elded products. The failures on the products are xed (it is assumed that we can
renew the product) and testing continues. Then testing is stopped at the times given
below for each product:

Product no. Hours when failures are recorded Hours when testing is stopped

01 2467; 3128; 3283; 7988 8012

02 None 6147
03 1870; 6121; 6175 9002
04 3721; 4393; 5848; 6425; 6353 11,000
05 498 4651
06 184; 216; 561; 2804 5012
07 2342; 4213 12,718

Estimate the MTBF for this product.

Solution:
In this case, TT, the total time on test, is obtained by adding all the hours when testing
was stopped:

TT = 8012 + + 12, 718 = 56,542 hours.

During this total test period, there were 19 failures. Thus, the point estimator for the
MTBF, under the assumption that the time between failures follows the exponential
distribution, is

θˆ = 56, 542 19 = 2975 hours.

Example 3.16

Estimate the MTBF (point estimator) or (θ̂) for the following reliability test
situations:

(a) Failure terminated, with no replacement. Twelve items were tested until the
fourth failure occurred, with failures at 200, 500, 625, and 800 hours.
(b) Time terminated, with no replacement. Twelve items were tested up to 1000
hours, with failures at 200, 500, 625, and 800 hours.
(c) Failure terminated, with replacement. Eight items were tested until the third
failure occurred, with failures at 150, 400, and 650 hours.
(d) Time terminated, with replacement. Eight items were tested up to 1000 hours,
with failures at 150, 400, and 650 hours.
(e) Mixed replacement/nonreplacement. Six items were tested through 1000 hours
on six different test stands. The rst failure on the test stand occurred at 300
hours, and its replacement failed after an additional 400 hours. On the second
test stand, failure occurred at 350 hours, and its replacement failed after an
additional 500 hours. On the third test stand, failure occurred at 600 hours, and

66
 3.2 Continuous Distributions

its replacement did not fail up to the completion of the test. The items on the
other three test stands did not fail for the duration of the test.

Solution:

(a) MTBF(e ) = θˆ = (200 + 500 + 625 + 800 + 8(800)) 4 = 2,131 hours

(b) MTBF(e) = θˆ = (200 + 500 + 625 + 800 + 8(1000 )) 4 = 2, 531 hours
(c) MTBF(e) = θˆ = (8)(650) 3 = 1, 733 hours
(d) MTBF(e) = θˆ = (8)(1000) 3 = 2, 667 hours
(e) MTBF(e ) = θˆ =(700 + 850 +1000 + (3)(1000 )) 5 = 1,110 hours .

Example 3.17

Forty modules were placed on life test for 20 days (24 hours per day). Failed boards
were replaced on the test stands with new ones. The test produced two failures. Esti-
mate the MTBF or the failure rate for the modules.

Solution:
In this case, the total time on test is

TT = 20× 24 × 40 = 19, 200 hours

θˆ = TT r = 19, 200 2 = 9600 hours

λˆ = r TT = 2 19, 200 = 1.04×10 − 4 failures per hour.

3.2.4 The Normal (Gaussian) Distribution

The normal distribution occurs whenever a random variable is affected by a sum of
random effects, such that no single factor dominates. This motivation is based on
central limit theorem, which states that under mild conditions, the sum of a large
number of random variables is approximately normally distributed. It has been used
to represent dimensional variability in manufactured goods, material properties, and
measurement errors. It has also been used to assess product reliability.
The normal distribution has been used to model various physical, mechanical,
electrical, or chemical properties of systems. Some examples are gas molecule velocity,
wear, noise, the chamber pressure from ring ammunition, the tensile strength of
aluminum alloy steel, the capacity variation of electrical condensers, electrical power
consumption in a given area, generator output voltage, and electrical resistance.
The probability density function for the normal distribution is based on the follow-
ing Gaussian function:

1  1  t − µ 
2 
f (t) = exp −    , −∞ ≤ t ≤ +∞, (3.40)
σ 2π  2  σ  
 
where the parameter µ is the mean or the MTTF, and σ is the standard deviation of
the distribution. The parameters for a normal distribution are listed in Table 3.3.

67
 3 Probability and Life Distributions for Reliability Analysis

Table 3.3 Normal distribution parameters

Mean (arithmetic average) µ

Median (B50 or 50th percentile) µ
Mode (highest value of f(t)) µ
Location parameter µ
Shape parameter/standard deviation σ
s (an estimate of σ) B50 − B16
Failure probability distribution f(t)

Figure 3.9 Probability density function for

Mean normal distribution.

Figure 3.9 shows the shape of the probability density function for the normal
distribution.
The cdf, or unreliability, for the normal distribution is:

t
1  1  x − µ  2 
F (t) =
σ 2π −∞∫exp − 
 2  σ 

  dx.


(3.41)

A normal random variable with mean equal to zero and variance of 1 is called a
standard normal variable (Z), and its pdf is given by

1 −z2 2
φ (z ) = e , (3.42)
2π

where z ≡ (t – µ)/σ.
The properties of the standard normal variable, in particular the cumulative prob-
ability distribution function, are tabulated in statistical tables (provided in Appendix
C). Table 3.4 provides the percentage values of the areas under the normal curve at
different distances from the mean in terms of multiples of σ. For example,

P [ X ≤ µ − 3σ ] = 0.00135 (3.43)

and

68
 3.2 Continuous Distributions

Table 3.4 Areas under the normal curve

µ – 1σ = 15.87% µ + 1σ = 84.130%
µ – 2σ = 2.28% µ + 2σ = 97.720%
µ – 3σ = 0.135% µ + 3σ = 99.865%
µ – 4σ = 0.003% µ + 4σ = 99.997%

P [ X ≤ µ + 3σ ] = 0.99865. (3.44)

There is no closed-form solution to the integral of Equation 3.41, and, therefore,

the values for the area under the normal distribution curve are obtained from the
standard normal tables by converting the random variable, t, to a random variable,
z, using the transformation:

t −µ
z= , (3.45)
σ

given by Equation 3.42. We have

 t − µ 
F (t) = Φ ( z ) = Φ (3.46)
 σ 

or

 t − µ 
R (t) = 1− Φ  (3.47)
 σ 

and

φ [(t − µ) σ ]
h (t) = , (3.48)
σ R (t)

where φ(.) is the pdf for the standard normal distribution and Φ(z) is the cdf for the
standard normal random variable Z.
From Equation 3.48, we can prove that the normal distribution has an increasing
hazard rate (IHR). The normal distribution has been used to describe the failure
distribution for products that show wearout and that degrade with time. The life of
tire tread and the cutting edges of machine tools t this description. In these situa-
tions, life is given by a mean value of µ, and the variability about the mean value is
de ned through standard deviation. When the normal distribution is used, the prob-
abilities of a failure occurring before or after this mean time are equal because the
mean is the same as the median.

Example 3.18

A machinist estimates that there is a 90% probability that the washer in an air com-
pressor will fail between 25,000 and 35,000 cycles of use. Assuming a normal distribu-
tion for washer degradation, nd the mean life and standard deviation of the life of
the washers.

69
3 Probability and Life Distributions for Reliability Analysis

Solution:
Assuming that 5% of the failures are at fewer than 25,000 cycles and 5% are at more
than 35,000 cycles, the mean of the distribution will be centered at 30,000 cycles of
use, that is, µ = 30,000.
In this condition:

25, 000 − µ
Φ ( z1 ) = 0.05, z1 =
σ
35, 000 − µ
Φ ( z2 ) = 0.95, z2 = .
σ

From the normal distribution table, z1 = –1.65, and z2 = 1.65. Hence, –1.65σ = 25,000 – µ
and 1.65σ = 35,000 – µ.
Solving the above two equations with the mean value of 30,000 cycles results in a
σ of 3030 cycles.

Example 3.19

The time for failure due to fungi growth is normally distributed with mean µ = 2.8
hours and standard deviation σ = 0.6 hours.

(a) What is the probability that the failure due to fungi growth will occur in 1.5
hrs?
(b) If we accept a probability of failure due to fungi growth of only 10%, after
what time from the start should the fungi be analyzed?

Solution:

(a) The probability that the fungi will grow in less than 1.5 hours is given by:

P {T < 1.5} = Q (1.5) = Φ ( z ),

z = (t − µ) σ = (1.5 − 2.8) 0.6 = −2.1667.

From the standard normal table, Φ(–2.1667) = 0.0151.

(b) For this condition, F(t) = 0.1 = Φ(z), then from the standard normal table, z
is approximately –1.28. Therefore, –t + µ = 1.28σ, hence t = 2.03 hours.

Example 3.20

A component has the normal distribution for time to failure, with µ = 20,000 hours
and σ = 3000 hours.

(a) Find the probability that the component will fail between 14,000 hours and
15,000 hours.

70
 3.2 Continuous Distributions

Solution:

P [14, 000 ≤ T ≤ 15, 000 ]

14, 000 − 20, 000 T − µ 15, 000 − 20, 000 
= P ≤ ≤ 
 3000 σ 3000 
= P [−2 ≤ Z ≤ −1.667 ]
= Φ (−1.667 )− Φ (−2 )
= (1− 0.95225) − (1− 0.97725)
= 0.04775 − 0.02275 = 0.025.
(b) Find the failure rate of a component that has been working for 14,000 hours.

Solution:

14, 000 − 20, 000 

f (14, 000) = φ   3000
 3000
1.7997 ×10−5
= = 5.999×10−9
3000
14, 000 − 20, 000 
pdf φ (14, 000) = φ   = 1.7997 ×10
−5
 3000
based on MS Excel evaluation
 14, 000 − 20, 000 
R (t) = P Z >  = P [ Z > −2 ] = Φ (2) = 0.977249
 3000 
f (t)
h (t) = = 6.13865×10−9 failures cycle.
R (t)

Example 3.21

The time to failure random variable for a light bulb made by Company X follows a
normal distribution, with µ = 1600 hours and σ = 250 hours.
(a) Find the B10 life of these light bulbs.

Solution:
Setting the value of reliability at B10 equal to 0.90, we have

R (B10 ) = 0.9
B10 −1600
z = −1.28 =
250
B10 = (−1.28)(250) + 1600 = 1280 hourrs.

(b) Find the reliability of the light bulbs for 1100 hours.

Solution:
 1100 −1600 
R (1100) = Pr Z ≥  = Pr (Z ≥−2.0) = 0.9773.
 250

71
3 Probability and Life Distributions for Reliability Analysis

(c) What is the failure rate or hazard rate of a light bulb that has not failed for
1100 hours?

Solution:

f (1100) φ (−2.0) 250 0.05399 250

h (1100) = = =
R (1100) Φ (−2.0) 0.9773
= 0.0002210 failures hour .

where:

1100 −1600 
φ   = φ (−2 ) = 0.05399.
 250

based on MS Excel evaluation, where:

(−2)2 

1 − 2 
φ (−2.0) = e = (0.39894)(0.13533) = 0.05399.
2π

Alternatively,

 (−2)2 
− 
1 
 2 
f (1100) = e 
2πσ
0.05399
= = 0.0002160
250
f (1100)
h (1100) =
R (1100 )
0.0002160
= = 0.0002210.
0.9773

(d) Company X has 5 million light bulbs in the eld that have been in use for 1100
hours and have not failed so far. How many light bulbs will fail in the next day
(or 24 hours)? Assume that light bulbs are used on average for 10 hours per
day.

Solution:
Using the concepts covered in Section 2.2.1 of Chapter 2, we have Ns(1100) = 5 × 106
and ∆t = 10 hours.

h (1100) = 0.0002210 =
N S (1100)− N S (1100 + 10) N S (1100)− N S (1100 + 10)
= .
N S (1100)×∆t 5×106 ×10

Then the number of failures between 1100 and 1110 hours is given by:

(0.0002210)(5×106 )(10 ) = 11, 050 light bulbs.

72
 3.2 Continuous Distributions

3.2.5 The Lognormal Distribution

For a continuous random variable, there may be a situation in which the random
variable is a product of a series of random variables. The lognormal distribution is a
positively skewed distribution and has been used to model situations where large
occurrences are concentrated at the tail (left) end of the range. Some examples are
the amount of electricity used by different customers, the downtime of systems, the
time to repair, the light intensities of light bulbs, the concentration of chemical process
residues, and automotive mileage accumulation by different customers. For example,
the wear on a system may be proportional to the product of the magnitudes of the
loads acting on it. Thus, a random variable may be modeled as a lognormal random
variable if it can be thought of as the multiplicative product of many independent
random variables each of which is positive. If a random variable has lognormal dis-
tribution, then the logarithm of the random variable is normally distributed. If X is
a random variable with a normal distribution, then Y = eX has a lognormal distribu-
tion; or if Y has lognormal distribution, then X = log Y has normal distribution.
Suppose Y is the product of n independent random variables given by

Y = YY
1 2Y3 ……Yn. (3.49)

Taking the natural logarithm of Equation 3.49 gives

lnY = lnY1 + lnY2 + lnY3 + + lnYn. (3.50)

Then ln Y may have approximately normal distribution based on the central limit
theorem.
The lognormal distribution has been shown to apply to many engineering situa-
tions, such as the strengths of metals and the dimensions of structural elements, and
to biological parameters, such as loads on bone joints. Lognormal distributions have
been applied in reliability engineering to describe failures caused by fatigue and to
model time to repair for maintainability analysis. The probability density function for
the lognormal distribution is:

1  1  ln t − µ  2 
f (t) = exp −   , (3.51)
σt 2π  2  σ  
 

where σ is the standard deviation of the logarithms of all times to failure, and µ is
the mean of the logarithms of all times to failure. If random variable T follows a
lognormal distribution with parameters µ and σ, then ln T follows a normal distribu-
tion so that

E [ ln T ] = µ and V [ ln T ] = σ 2. (3.52)

The cdf (unreliability) for the lognormal distribution is:

t
1 1  1  ln x − µ 2 
Q (t) =
σ 2π ∫ x
exp − 
 2  σ  

  dx

0 (3.53)
 ln t − µ 
= Φ  .
 σ 

73
 3 Probability and Life Distributions for Reliability Analysis

Failure probability distribution f(t)

σ = 0.1

σ = 0.5
Figure 3.10 Lognormal probability density
Time function where σ = 0.1 and σ = 0.5.

Table 3.5 Lognormal distribution parameters

Mean exp[µ + 0.5σ2]

Variance (eσ2
− 1)e2µ +σ
2

Median (B50 or time at 50% failures) B50 = eµ

Mode (highest value of f(t)) t = exp[µ – σ2]
Location parameter eµ
Shape parameter σ
s (estimate of σ) ln(B50/B16)

The probability density function for two values of σ are as shown in Figure
3.10. The key parameters for the lognormal distribution are provided in Table 3.5.
The MTTF for a population for which the time to failure follows a lognormal
distribution is given by

 σ2 
MTTF = exp µ +  (3.54)
 2 

and the failure rate is given by

 ln t − µ 
h (t) = φ  tσ R (t). (3.55)
 σ 

The hazard rate for the lognormal distribution is neither always increasing nor always
decreasing. It takes different shapes, depending on the parameters µ and σ. We can
prove that the hazard rate of a lognormal distribution is increasing on average (called
IHRA).
From the basic properties of the logarithm operator, it can be shown that if vari-
ables X and Y are distributed lognormally, then the product random variable Z = XY
is also lognormally distributed.

74
 3.2 Continuous Distributions

Example 3.22

A population of industrial circuit breakers was found to have a lognormal failure

distribution with parameters µ = 3 and σ = 1.8. What is the MTTF of the popula-
tion? What is the estimate of reliability of these circuit breakers for continuous opera-
tion over 30 years?

Solution:
From Equation 3.54 for the MTTF,

MTTF = exp (3 + 0.5 ×(1.8) 2) = 101.5 years.

For a 30-year operation (from Eq. 3.53),

ln(30) − 3 3.41− 3
z= = = 0.223.
1.8 1. 8
Hence, from the table of standard normal distribution, the estimate of reliability for
a 30-year operation is given by:

R (30) = [1 − Φ( z )] = [1− Φ(0.223)] = [1− 0.588] = 0.412 .

Example 3.23

The time to repair a copy machine follows the lognormal distribution with µ = 2.50
and σ = 0.40. Time is in minutes.

(a) Find the probability that the copy machine will be repaired in 20 minutes.

Solution:

 ln 20 − 2.5 
P [T ≤ 20 ] = P [ lnT ≤ ln 20 ] = P  Z ≤ 
 0.40 
= P [ Z ≤ 1.23933] = 0.89239.

(b) Find the median value, or B50 life, for the time to repair a random variable.

P [T ≤ B50 ] = 0.5 = P [ Z ≤ 0 ]
ln T − 2.5
0=
0.40
T = 12.185.

3.2.6 Gamma Distribution

The probability density function for the gamma distribution is given by

λ η η−1 −λt
f (t) = t e , t ≥ 0, (3.56)
Γ (η )

75
3 Probability and Life Distributions for Reliability Analysis

where Γ(η) is the gamma function (values for this function are given in Appendix B).
The gamma distribution has two parameters, η and λ, where η is called the shape
parameter and λ is called the scale parameter. The gamma distribution reduces to
the exponential distribution if η = 1. Adding η exponential distributions, η ≥ 1,
with the same parameter λ, provides the gamma distribution. Thus, the gamma dis-
tribution can be used to model time to the ηth failure of a system if the underlying
system/component failure distribution is exponential with parameter λ. We can also
state that if Ti is exponentially distributed with parameter λ, i = 1, 2, . . . , η, then
T = T1 + T2 + . . . + Tη has a gamma distribution with parameters λ and η.
This distribution could be used if we wanted to determine the system reliability for
redundancy with identical components all having a constant failure rate.
From Equation 3.56, the cumulative distribution or the unreliability function is

t
λ η η−1 −λt
F (t ) = ∫
0
Γ (η )
τ e d τ , t ≥ 0. (3.57)

If η is an integer, then the gamma distribution is also called the Erlang distribution,
and it can be shown by successive integration by parts that

∞ k
(λt) exp[−λt]
F (t) = ∑k =η
k!
. (3.58)

Then,

η−1 k
(λt) exp[−λt ]
R (t) = ∑
k =0
k!
(3.59)

and

f (t)
h (t) = . (3.60)
R (t)

Also,

η
E (T ) = (3.61)
λ

and

η
V (T ) = . (3.62)
λ2

The failure rate for the gamma distribution is decreasing when η < 1, is constant when
η = 1(because it is an exponential distribution), and is increasing when η > 1.

76
 3.3 Probability Plots

Example 3.24

The time to a major failure in hours for a copy machine follows a gamma distribution
with parameters η = 3 and λ = 0.002.

(a) What is the expected life, or mean time between failures (MTBF), for the copy
machine?

Solution:
Using Equation 3.61, we have

η 3
MTBF = = = 1500 hours.
λ 0.002
(b) What is the reliability of the copy machine for 500 hours of continuous
operation?

Solution:
Using Equation 3.59,
η−1 k
(λt) e−λt
R (t) = ∑k =0
k!
3−1
(0.002 ×500)k e−0.002×500
R (500) = ∑k =0
k!
= 0.919698.

(c) What is the failure rate of a copy machine that has been working for 500 hours?

Solution:
Using Equation 3.56 and Equation 3.60, we have

λ η η−1 −λt
f (t) = t e
Γ (η )
0.0023
f (500) = 5003−1 e−(0.002*500) = 0.000368
Γ (3)
h (500 ) = f (500 ) R (500)
= 0.0004001 failures per hour.

Thus, the failure rate is 0.0004 failures per hour, or 4 failures per 10,000 hours of
total use.

3.3 Probability Plots

Probability plotting is a method for determining whether data (observations) conform

to a hypothesized distribution. Typically, computer software is used to assess the

77
3 Probability and Life Distributions for Reliability Analysis

Table 3.6 Examples of cdf estimates for N = 20

Estimate of cumulative distribution function or unreliability

Rank Midpoint plotting Expected plotting Median plotting

order (i) position position position Median rank

1 2.5 4.8 3.4 3.406

2 7.5 9.5 8.3 8.251
3 12.5 14.3 13.2 13.147
4 17.5 19.0 18.1 18.055
5 22.5 23.8 23.0 22.967
6 27.5 28.6 27.9 27.880
7 32.5 33.3 32.8 32.795
8 37.5 38.1 37.7 37.710
9 42.5 42.8 42.6 42.626
10 47.5 47.6 47.5 47.542
11 52.5 52.4 52.5 52.458
12 57.5 57.1 57.4 57.374
13 62.5 61.9 62.3 62.289
14 67.5 66.7 67.2 67.205
15 72.5 71.4 72.1 72.119
16 77.5 76.4 77.0 77.033
17 82.5 80.1 81.9 81.945
18 87.5 85.7 86.8 86.853
19 92.5 90.5 91.7 91.749
20 97.5 95.2 96.6 96.594

hypothesized distribution and determine the parameters of the underlying distribu-

tion. The method used by the software tools is analogous to using constructed prob-
ability plotting paper to plot data. The time-to-failure data is ordered from the
smallest to the largest in value in an appropriate metric (e.g., time to failure and cycles
to failure). An estimate of the percent of unreliability is selected. The data are plotted
against a theoretical distribution in such a way that the points should form a straight
line if the data come from the hypothesized distribution. The data are plotted on
probability plotting papers (these are distribution speci c), with ordered times to
failure in the x-axis and the estimate of percent unreliability as the y-axis. A best- t
straight line is drawn through the plotted data points.
The time to failure data used for the x-axis is obtained from the eld or testing.
The estimate of unreliability against which to plot this time-to-failure data is not that
obvious. Several different techniques, such as “midpoint plotting position,” “expected
plotting position,” “median plotting position,” “median rank,” and Kaplan–Meier
ranks (in software) are used for this estimate. Table 3.6 provides estimates for unreli-
ability based on different estimation schemes for a sample size of 20.
The median rank value for the ith failure, Qi, is given by the solution to the follow-
ing equation:
N
N!
∑ k !(N − k)! (1−Q )
k =i
i
N −k
Qik = 0.5, (3.63)

where N is the sample size, i is the failure number, and Qi is the median rank (or
estimate of unreliability at the failure time of the ith failure). Equation 3.64, which

78
 3.3 Probability Plots

estimates the median plotting positions, can be used in place of the median rank as
an approximation:

100×(i − 0.3)
Qi = . (3.64)
N + 0.4

The axes used for the plots are not linear. The axes are different for each probability
distribution and are created by linearizing the cmf or unreliability function, typically
by taking the logarithm of both sides repeatedly. For example, mathematical manipu-
lation based on Equation 3.27 for a two-parameter Weibull distribution will result in
an ordinate (y-axis) as log log reciprocal of R(t) = 1 – Q(t) scale and the abscissa as
a log scale of time to failure, and is derived below:

 t β
− 
 η 
Q (t) = 1− e
 − t β 
  
ln(1− Q (t)) = lne  η  
  (3.65)
 
t
ln (− ln(1− Q (t))) = β ln  
 η 
y = β x − β ln (η ),

where x = ln(t) and y = ln(−ln(1 − Q(t))).

Once the probability plots are prepared for different distributions, the goodness of
t of the plots is one factor in determining which distribution is the right t for the
data. Probability distributions for data analysis should be selected based on their
ability to t the data and for physics-based reasons. There should be a physics-based
argument for selection of a distribution that draws from the failure model for the
mechanism(s) that caused the failures. These decisions are not always clear-cut. For
example, the lognormal and the Weibull distribution both model fatigue failure data
well, and hence it is often possible for both to t the failure data; thus, experience-
based engineering judgments need to be made.
There is no reason to assume that all the time-to-failure data taken together need
to t only one failure distribution. Since the failures in a product can be caused by
more than one mechanism, it is possible that some of the failures are caused by one
mechanism and the others by a different mechanism. In that case, no single probability
distribution will t the data well. Even if it appears that one distribution ts all the
data, that distribution may not have good predictive ability. That is why it may be
necessary to separate the failures by mechanisms into sets and then t separate dis-
tributions for each set.
Table 3.7 shows times to failure separated into two groups by failure mechanism.
Figure 3.11 shows the Weibull probability plots for the competing failure mecha-
nism data. Note that the shape and scale factors for the two sets are distinct, with
one set having a decreasing hazard rate (β = 0.67) and the other set having
an increasing hazard rate (β = 4.33). If the data are plotted together, the result
shows an almost constant hazard rate. However, spare part and support decisions
made based on results from a combined data analysis can be misleading and
counterproductive.

79
 3 Probability and Life Distributions for Reliability Analysis

Table 3.7 Time to failure data separated by failure mechanism

Ordered Data State F or S Time to F or S Failure Mechanism Group

1 F 2 V
2 F 10 V
3 F 13 V
4 F 23 V
5 F 23 V
6 F 28 V
7 F 30 V
8 F 65 V
9 F 80 V
10 F 88 V
11 F 106 V
12 F 143 V
13 F 147 W
14 F 173 V
15 F 181 W
16 F 212 W
17 F 245 W
18 F 247 V
19 F 261 V
20 F 266 W
21 F 275 W
22 F 293 W
23 S 300
24 S 300
25 S 300
26 S 300
27 S 300
28 S 300
29 S 300
30 S 300

F, failure; S, suspension; V, failure mechanism 1; W, failure mechanism 2.

99.000

95.000

CFM 1 Points
50.000 CFM 2 Points
CFM 1 Line
Unreliability Q(t)

CFM 2 Line
Probability Line

10.000

5.000

Figure 3.11 Weibull probability plot

1.000 for competing failure mechanism
1.000 10.000 100.000 1000.000 data shown in Table 3.7. β1 = 0.67,
Time (t) η1 = 450; β2 = 4.33, η2 = 340.

80
 3.3 Probability Plots

99.000
90.000

50.000
Unreliability Q(t)

Data Points
Probability Line

10.000
5.000

1.000 Figure 3.12 Two-parameter Weibull

10.000 100.000 1000.000 probability plot for time-to-failure data
Time (t) shown in Table 3.8. β1 = 0.65, η1 = 825.

Table 3.8 Test Data for Example 3.25

Sample number Time to failure (hours) Sample number Time to failure (hours)

1 14 6 563
2 58 7 –
3 130 8 –
4 245 9 –
5 382 10 –

Example 3.25

Figure 3.12 shows reliability test data for 10 identical products out of which six
products failed within the test duration of 600 hors. The time to failure is plotted
on two-parameter Weibull probability plotting paper. Using the plot, estimate the
following:

(a) The unreliability and reliability at the end of 50 hours.

(b) The reliability for a new period of 50 hours, starting after the end of the previ-
ous 50-hour period.
(c) The longest duration that will provide a reliability of 95% assuming the opera-
tion starts at 50 hours.

Solution:

(a) For this example, we nd that β = 0.65, and η is estimated to be 825 hours. It
is now possible to write the equation for the reliability and use it for analysis.
The plotted straight line can also be used to determine the reliability values
directly.
From Figure 3.12, the unreliability estimate for a mission time of 50 hours can
be read directly from the straight line. The value is Q(50) = 15%. Thus, the
reliability for this duration is R(50) = 1 – Q(50) = 85%.

81
3 Probability and Life Distributions for Reliability Analysis

(b) The reliability for a new 50-hour period starting with an age of 50 hours is
given by the conditional reliability equation as

R (50 + 50) R (100) 0.78

R (50, 50) = = = = 91.7%,
R (50) R (50) 0.85

where R(100) = 1 – Q(100) can be taken directly from the curve.

(c) For a mission time, t, that starts after a 50-hour period and must have a reli-
ability of 95%,

R (t + 50) R (t + 50 )
R (t, 50 ) = = = 0.95
R (50) 0.85

or

R (t + 50) = 0.95× 0.85 = 0.808.

To obtain this reliability, the unreliability is 0.192 or 19.2%. From the curve, the time
to obtain this unreliability is about 75 hours. Thus, 50 + t = 75 gives a maximum new
mission time of 25 hours in order to have a reliability of 95%.

When the life data contains two or more life segments—such as infant mortality,
useful life, and wearout—a mixed Weibull distribution can be used to t parts of the
data with different distribution parameters. A curved or S-shaped Weibull probability
plot (in either two or three parameters) is an indication that a mixed Weibull distribu-
tion may be present.
Statistical analysis provides no magical way of projecting into the future. The results
from an analysis are only as good as the assumed model and assumptions, including
how failure is de ned, the validity of the data, how the model is used, and taking into
consideration the tail of the distribution and the limits of extrapolations and inter-
polations. The following example demonstrates the absurdity of extrapolating times
to failure beyond their reasonable limits.

Example 3.26

A Weibull probability plot was made for a population collected over the rst 10 years
of its life containing failures (see Figure 3.13).

(a) Estimate the percentage of this population expected to fail by 300 years.
(b) Does the answer make sense if the time-to-failure data is for human mortality?
Explain.

Solution:

(a) The results show that the probability of failure at 300 years is approximately
2%.
(b) The mortality data for over a billion people for a 10-year period from the time
of birth ts a Weibull distribution very well. This looks impressive, but is nev-
ertheless all wrong. It is clear that this data should not be used for making any

82
 3.4 Summary

10.00

5.00

Estimated failure probability (%)

1.00

0.50

0.10

0.05

0.01
1.00 10.00 100.00 1000.00
Years
Figure 3.13 Weibull probability plot of time-to-failure data for Example 3.26.

judgment on human longevity, even though all the calculations are correct. The
mortality pattern of humans in the rst 10 years of life cannot be extrapolated,
because the mortality pattern changes with age. This is also often true for engi-
neered goods. Failures that occur in postmanufacturing tests are often caused
by defects introduced in manufacturing. The rst 10 years of time-to-failure
data will result in a shape factor (β) of less than one. However, during early
childhood through a large part of adulthood, the shape factor will be close to
one, where most deaths can be considered random (e.g., caused by many causes
such as accidents). Then the population will enter a wearout stage during which
people die from old age. Complete human mortality data should be modeled
using a mixed Weibull distribution.

3.4 Summary

The reliability function is used to describe the probability of successful system opera-
tion during a system’s life. A natural question is then, “What is the shape of a reli-
ability function for a particular system?” There are basically three ways in which this
can be determined:

1. Test many systems to failure using a mission pro le identical to use conditions.
This would provide an empirical curve based on the histogram that can give
some idea about the nature of the underlying life distribution.

83
3 Probability and Life Distributions for Reliability Analysis

2. Test many subsystems and components to failure under use conditions recreated
in the test environment. This empirically provides the component reliability
functions. Then derive analytically or numerically or through simulation the
system reliability function. (Chapter 17 covers topics related to system
reliability.)
3. Based on past experience with similar systems, hypothesize the underlying failure
distribution. Fewer systems can be tested to determine the parameters needed
to adapt the failure distribution to a particular situation. However, this will not
account for new failure mechanisms or new use conditions.

In some cases, the failure physics involved in a particular situation may lead to the
hypothesis of a particular distribution. For example, fatigue of certain metals tends
to follow either a lognormal or Weibull distribution. Once a distribution is selected,
the parameters for a particular application can be ascertained using statistical or
graphical procedures.
In this chapter, various distributions were presented. However, the most appropriate
distribution(s) for a particular failure mechanism or product that exhibits certain
failure mechanisms must be determined by the actual data, and not guessed. The
distribution(s) that best t the data and that also make sense in terms of the failure
processes should be used.

Problems

3.1 Prove that for a binomial distribution in which the number of trials is m and the
probability of success in each trial is p, the mean and the variance are equal to mp
and mp(1 – p), respectively.

3.2 Prove that for a Poisson distribution, the mean and the variance are equal to the
Poisson parameter µ.

3.3 Compare the results of Examples 3.2 and 3.5. What is the reason for the
differences?

3.4 Consider a system that has seven components; the system will work if any ve of
the seven components work. Each component has a reliability of 0.930 for a given
period. Find the reliability of the system.

3.5 For an exponential distribution, show that the time to 50% failure is given by
0.693/λ0.

3.6 For an exponential distribution, show that the standard deviation is equal to 1/λ0.

3.7 Show that for a two-parameter Weibull distribution, for t = η, the reliability
R(t) = 0.368, irrespective of β.

3.8 The front wheel roller bearing life for a car is modeled by a two-parameter Weibull
distribution with the following two parameters: β = 3.7, θ = 145,000 mi. What is the
100,000-mi reliability for a bearing?

84
 Problems

3.9 The life distribution (life in years of continuous use) of hard disk drives for a
computer system follows the Weibull distribution with the following parameters:

β = 2.7 and θ = 5.5 years.

(a) The manufacturer gives a warranty for 1 year. What is the probability that a disk
drive will fail during the warranty period?
(b) Find the mean life and the median life (B50) for the disk drive.
(c) By what time will 99% of the disk drives fail? (That is, nd the B99 life.)

3.10 The life distribution for miles to failure for the engine of a Lexus car follows the
Weibull distribution with

β = 3.8 and θ = 185, 000 mi.

(a) Find the mean miles between failures, or the expected life for the engine.
(b) Find the standard deviation for miles to failure.
(c) What percent of these engines will fail by 100,000 mi?
(d) What is the failure rate of an engine that has a life of 100,000 mi?

If a certain model has 200,000 engines in the eld with a life of 100,000 mi, how many
engines on average will fail in the next 100 mi of use out of the 200,000 engines?

3.11 A component has the normal distribution for time to failure, with µ = 26,000
hours and σ = 3500 hours.

(a) Find the probability that the component will fail between 22,000 hours and 23,000
hours.
(b) Find the failure rate of a component that has been working for 22,000 hours.

3.12 The time to failure random variable for a battery follows a normal distribution,
with µ = 800 hours and σ = 65 hours.

(a) Find the B10 life of these batteries.

(b) Find the probability that a battery will fail between 700 and 710 hours, given that
it has not failed by 700 hours.
(c) What is the failure rate or hazard rate of a battery that has a life of
(i) 700 hours
(ii) 710 hours.

3.13 The time to failure for the hard disk drives for a computer system follows a
normal distribution with

µ = mean life = 14, 000 hours

σ = standard deviation = 1500 hours.

85
3 Probability and Life Distributions for Reliability Analysis

(a) A manufacturer gives a warranty for 1 year of continuous use, or 365 × 24

hours of use. What percentage of hard disk drives will fail during this warranty
period?
(b) What is the failure or hazard rate of a drive that has been working successfully
for 1 year of continuous use?
(c) An IT manager of a large company, based on eld surveys and inventory
management, nds that the company has 250,000 of these drives on which the
warranty has just expired—that is, they are working today after one year of con-
tinuous use. What is the expected number of these drives that will fail in the next
24 hours?

3.14 The time to repair a communication network system follows a lognormal distri-
bution with µ = 3.50 and σ = 0.75. The time is in minutes.

(a) What is the probability that the communication network will be repaired by 60
minutes?
(b) Find the B20 value (the 20th percentile) for the time-to-repair random variable.
(c) Find the mean time to repair (MTTR) for the communication network.

3.15 The time to repair a copy machine follows the lognormal distribution with
µ = 2.70 and σ = 0.65. Time is in minutes.

(a) Find the probability that the copy machine will be repaired in 30 minutes.
(b) Find the median value or B50 life for the time-to-repair random variable.

3.16 The time to failure for a copy machine follows a gamma distribution with param-
eters η = 2 and λ = 0.004.

(a) What is the expected or mean time between failures (MTBF) for the copy
machine?
(b) What is the reliability of the copy machine for 200 hours of continuous
operation?
(c) What is the failure rate of a copy machine that has been working for 200
hours?

3.17 Describe two examples of systems that require a failure-free operating period,
without any maintenance. What are the timeframes involved?

3.18 Describe two examples of systems that require a failure-free operating period,
but may allow a maintenance period. Discuss the timeframes.

3.19 Show that the mode of the three parameter Weibull distribution is for

1β
t = γ + η (1−1 β )

for β > 1.

86
 Problems

3.20 A company knows that approximately 3 out of every 1000 processors that it
manufactures are defective. What is the probability that out of the next 20 processors
selected (at random):

(a) All 20 are working processors?

(b) Exactly 2 defective processors?
(c) At most 2 defective processors?
(d) At least 18 are defective?

87