40 Probability Distributions Used in Reliability Engineering

Life Distribution

2. Common Life Distributions

 Exponential Continuous Distribution 41

2.1. Exponential Continuous

Distribution
Probability Density Function - f(t)

Exponential
0.90 λ=0.5
0.80 λ=1
0.70
λ=2
0.60
0.50 λ=3
0.40
0.30
0.20
0.10
0.00
0 0.5 1 1.5 2 2.5 3 3.5 4

Cumulative Density Function - F(t)

1.00

0.80

0.60 λ=0.5
λ=1
0.40
λ=2
0.20 λ=3

0.00
0 0.5 1 1.5 2 2.5 3 3.5 4

Hazard Rate - h(t)

3.00

2.50

2.00

1.50

1.00

0.50

0.00
0 0.5 1 1.5 2 2.5 3 3.5 4
 42 Common Life Distributions

Parameters & Description

Parameters Scale Parameter: Equal to the hazard
𝜆 𝜆>0
rate.
Limits 𝑡 ≥ 0
Exponential

Function Time Domain Laplace Domain

𝜆
PDF 𝑝(𝑡) = 𝜆e−λt 𝑝(𝑠) = , 𝑠 > −𝜆
𝜆+𝑠
𝜆
CDF 𝐹(𝑡) = 1 − e−λt 𝐹(𝑠) =
𝑠(𝜆 + 𝑠)
1
Reliability R(t) = e−λt 𝑅(𝑠) =
𝜆+𝑠
1
𝑚(𝑚) = e−λx 𝑚(𝑠) =
𝜆+𝑠
Conditional
Survivor Function Where
𝑃(𝑇 > 𝑚 + 𝑡|𝑇 > 𝑡) 𝑡 is the given time we know the component has survived to.
𝑚 is a random variable defined as the time after 𝑡. Note: 𝑚 = 0 at 𝑡.

Mean Residual 1 1
Life 𝑠(𝑡) = 𝑠(𝑠) =
𝜆 𝜆𝑠
𝜆
Hazard Rate ℎ(𝑡) = 𝜆 ℎ(𝑠) =
𝑠
Cumulative 𝜆
𝐻(𝑡) = 𝜆𝑡 𝐻(𝑠) =
Hazard Rate 𝑠2
Properties and Moments
𝑠𝑓(2)
Median
𝜆
Mode 0
1
Mean - 1st Raw Moment
𝜆
1
Variance - 2nd Central Moment
𝜆2
Skewness - 3rd Central Moment 2
th
Excess kurtosis - 4 Central Moment 6
𝑠𝜆
Characteristic Function
𝑡 + 𝑠𝜆
1
100α% Percentile Function 𝑡𝛼 = − ln(1 − 𝛼)
𝜆
 Exponential Continuous Distribution 43

Parameter Estimation
Plotting Method
Least Mean X-Axis Y-Axis
Square - 𝜆̂ = −𝑚
𝑦 = 𝑚𝑚 + 𝑐 𝑡𝑖 𝑠𝑓[1 − 𝐹(𝑡𝑖 )]

Exponential
Likelihood Function
nF F
nS S
nI LI UI
𝐿(𝐸|𝜆) = λnF � e−λ.ti . � e−ti . � �e−λti − e−λti �
�������
i=1���� �����
i=1 �� ���������������
i=1
failures survivors interval failures
Likelihood
Functions when there is no interval data this reduces to:

𝐿(𝐸|𝜆) = 𝜆nF 𝑅 −𝜆𝑡𝑇 𝑤ℎ𝑅𝑠𝑅 𝑡𝑇 = � t Fi + � t Si = 𝑡𝑓𝑡𝑠𝑠 𝑡𝑠𝑚𝑅 𝑠𝑓 𝑡𝑅𝑠𝑡

nF nS nI
LI RI
Λ(𝐸|𝜆) = r. ln(λ) − � λt Fi − � λt s + � ln �e−λti − e−λti �
�����������i=1 ���
i=1 ���������������
i=1
failures survivors interval failures
Log-Likelihood
Functions when there is no interval data this reduces to:

Λ(𝐸|𝜆) = nF . ln(𝜆) − 𝜆𝑡𝑇 where 𝑡𝑇 = � t Fi + � t Si

solve for 𝜆 to get 𝜆̂:

nF nS nI LI RI
∂Λ nF t LI
i e
λti
− t RI
i e
λti
=0 − � t Fi − � t Si − � � �=0
∂λ λ 𝜆𝑡 𝐿𝐼
𝑅 𝑖 −𝑅 𝑖 𝜆𝑡 𝑅𝐼
����� ��
i=1 ���
i=1 ���������������
i=1
failures survivors interval failures

Point When there is only complete and right-censored data the point
Estimates estimate is:
nF
𝜆̂ = 𝑤ℎ𝑅𝑠𝑅 𝑡𝑇 = � t Fi + � t Si = 𝑡𝑓𝑡𝑠𝑠 𝑡𝑠𝑚𝑅 𝑠𝑓 𝑡𝑅𝑠𝑡
𝑡𝑇
Fisher 1
𝐼(𝜆) =
Information 𝜆
λ lower - λ upper - λ upper -
2 Sided 2 Sided 1 Sided
100𝛾%
Confidence 𝜒 21−γ (2nF ) 𝜒 21+γ (2nF + 2) 2
𝜒(γ) (2nF + 2)
Type I (Time � � � �
Interval 2 2
Terminated) 2𝑡𝑇
2𝑡𝑇 2𝑡𝑇
(excluding
interval data) Type II (Failure 𝜒 21−γ (2nF ) 𝜒 21+γ (2nF ) 2
𝜒(γ) (2nF )
� � � �
2 2
Terminated) 2𝑡𝑇
2𝑡𝑇 2𝑡𝑇
 44 Common Life Distributions

2
𝜒(𝛼) is the 𝛼 percentile of the Chi-squared distribution. (Modarres et
al. 1999, pp.151-152) Note: These confidence intervals are only valid
for complete and right-censored data or when approximations of
interval data are used (such as the median). They are exact
confidence bounds and therefore approximate methods such as use of
the Fisher information matrix need not be used.
Exponential

Bayesian
Non-informative Priors 𝝅(𝝀)
(Yang and Berger 1998, p.6)
Type Prior Posterior
Uniform Proper Prior 1 Truncated Gamma Distribution
with limits 𝜆 ∈ [𝑠, 𝑏] 𝑏−𝑠 For a ≤ λ ≤ b
𝑐. 𝐺𝑠𝑚𝑚𝑠(𝜆; 1 + nF , t T )

Otherwise 𝜋(𝜆) = 0
Uniform Improper Prior 1 ∝ 𝐺𝑠𝑚𝑚𝑠(1,0) 𝐺𝑠𝑚𝑚𝑠(𝜆; 1 + nF , t T )
with limits 𝜆 ∈ [0, ∞)
Jeffrey’s Prior 1 1 𝐺𝑠𝑚𝑚𝑠(𝜆; 12 + nF , t T )
∝ 𝐺𝑠𝑚𝑚𝑠(2, 0)
√𝜆 when 𝜆 ∈ [0, ∞)
Novick and Hall 1 𝐺𝑠𝑚𝑚𝑠(𝜆; nF , t T )
∝ 𝐺𝑠𝑚𝑚𝑠(0,0)
𝜆 when 𝜆 ∈ [0, ∞)
where 𝑡𝑇 = ∑ t Fi + ∑ t Si = total time in test
Conjugate Priors
UOI Likelihood Evidence Dist of Prior Posterior
Model UOI Para Parameters
𝜆 𝑓𝐹 failures
𝑘 = 𝑘𝑜 + 𝑓𝐹
from Exponential in 𝑡𝑇 unit of Gamma 𝑘0 , Λ 0
Λ = Λ 𝑜 + 𝑡𝑇
𝐸𝑚𝑝(𝑡; λ) time
Description , Limitations and Uses
Example Three vehicle tires were run on a test area for 1000km have
punctures at the following distances:
Tire 1: No punctures
Tire 2: 400km, 900km
Tire 3: 200km

Punctures are a random failure with constant failure rate therefore

an exponential distribution would be appropriate. Due to an
exponential distribution being homogeneous in time, the renewal
process of the second tire failing twice with a repair can be
considered as two separate tires on test with single failures. See
example in section 1.1.6.

Total distance on test is 3 × 1000 = 3000km. Total number of

 Exponential Continuous Distribution 45

failures is 3. Therefore using MLE the estimate of 𝜆:

nF 3
𝜆̂ = = = 1E-3
𝑡𝑇 3000

With 90% confidence interval (distance terminated test):

Exponential
2 2
𝜒(0.05) (6) 𝜒(0.95) (8)
� = 0.272𝐸-3, = 2.584𝐸-3�
6000 6000

A Bayesian point estimate using the Jeffery non-informative

improper prior 𝐺𝑠𝑚𝑚𝑠(12, 0), with posterior 𝐺𝑠𝑚𝑚𝑠(𝜆; 3.5, 3000) has
a point estimate:
3.5
𝜆̂ = E[𝐺𝑠𝑚𝑚𝑠(𝜆; 3.5, 3000)] = = 1.16̇ E − 3
3000

With 90% confidence interval using inverse Gamma cdf:

[𝐹𝐺−1 (0.05) = 0.361𝐸-3, 𝐹𝐺−1 (0.95) = 2.344𝐸-3]

Characteristics Constant Failure Rate. The exponential distribution is defined by a

constant failure rate, 𝜆. This means the component is not subject to
wear or accumulation of damage as time increases.

𝒇(𝟎) = 𝝀. As can be seen, 𝜆 is the initial value of the distribution.

Increases in 𝜆 increase the probability density at 𝑝(0).

HPP. The exponential distribution is the time to failure distribution

of a single event in the Homogeneous Poisson Process (HPP).

𝑇~𝐸𝑚𝑝(𝑡; 𝜆)
Scaling property
𝜆
𝑠𝑇~𝐸𝑚𝑝 �𝑡; �
𝑠
Minimum property
𝑛

min{𝑇1 , 𝑇2 , … , 𝑇𝑛 }~𝐸𝑚𝑝 �𝑡; � 𝜆𝑖 �

𝑖=1
Variate Generation property
ln(1 − 𝑠)
𝐹 −1 (𝑠) = , 0<𝑠<1
−𝜆
Memoryless property.
Pr(𝑇 > 𝑡 + 𝑚|𝑇 > 𝑡) = Pr(𝑇 > 𝑚)

Properties from (Leemis & McQueston 2008).

Applications No Wearout. The exponential distribution is used to model
occasions when there is no wearout or cumulative damage. It can
be used to approximate the failure rate in a component’s useful life
period (after burn in and before wear out).

Homogeneous Poisson Process (HPP). The exponential

 46 Common Life Distributions

distribution is used to model the inter arrival times in a repairable

system or the arrival times in queuing models. See Poisson and
Gamma distribution for more detail.

Electronic Components. Some electronic components such as

capacitors or integrated circuits have been found to follow an
Exponential

exponential distribution. Early efforts at collecting reliability data

assumed a constant failure rate and therefore many reliability
handbooks only provide a failure rate estimates for components.

Random Shocks. It is common for the exponential distribution to

model the occurrence of random shocks An example is the failure
of a vehicle tire due to puncture from a nail (random shock). The
probability of failure in the next mile is independent of how many
miles the tire has travelled (memoryless). The probability of failure
when the tire is new is the same as when the tire is old (constant
failure rate).

In general component life distributions do not have a constant

failure rate, for example due to wear or early failures. Therefore the
exponential distribution is often inappropriate to model most life
distributions, particularly mechanical components.

Online:
http://www.weibull.com/LifeDataWeb/the_exponential_distribution.h
tm
http://mathworld.wolfram.com/ExponentialDistribution.html
http://en.wikipedia.org/wiki/Exponential_distribution
http://socr.ucla.edu/htmls/SOCR_Distributions.html (web calc)
Resources
Books:
Balakrishnan, N. & Basu, A.P., 1996. Exponential Distribution:
Theory, Methods and Applications 1st ed., CRC.

Nelson, W.B., 1982. Applied Life Data Analysis, Wiley-

Interscience.
Relationship to Other Distributions
2-Para Exp Special Case:
Distribution 1
𝐸𝑚𝑝(𝑡; 𝜆) = 𝐸𝑚𝑝(𝑡; µ = 0, β = )
𝐸𝑚𝑝(𝑡; 𝜇, 𝛽) 𝜆
Let
𝑇1 … 𝑇𝐸 ~𝐸𝑚𝑝(𝜆) 𝑠𝑓𝑝 𝑇𝑡 = 𝑇1 + 𝑇2 + ⋯ + 𝑇𝐸
Then
Gamma
𝑇𝑡 ~𝐺𝑠𝑚𝑚𝑠(𝑘, 𝜆)
Distribution
The gamma distribution is the probability density function of the
sum of k exponentially distributed time random variables sharing
𝐺𝑠𝑚𝑚𝑠(𝑡; 𝑘, 𝜆)
the same constant rate of occurrence, 𝜆. This is a Homogeneous
Poisson Process.
 Exponential Continuous Distribution 47

Special Case:
𝐸𝑚𝑝(𝑡; 𝜆) = 𝐺𝑠𝑚𝑚𝑠(𝑡; 𝑘 = 1, 𝜆)
Let
𝑇1 , 𝑇2 … ~𝐸𝑚𝑝(𝑡; 𝜆)
Given

Exponential
𝑡𝑠𝑚𝑅 = 𝑇1 + 𝑇2 + ⋯ + 𝑇𝐾 + 𝑇𝐾+1 …
Then
Poisson 𝐾~𝑃𝑓𝑠𝑠(k; µ = 𝜆𝑡)
Distribution
The Poisson distribution is the probability of observing exactly k
𝑃𝑓𝑠𝑠(𝑘; 𝜇) occurrences within a time interval [0, t] where the inter-arrival times
of each occurrence is exponentially distributed. This is a
Homogeneous Poisson Process.

Special Cases:
𝑃𝑓𝑠𝑠(k = 1; µ = 𝜆𝑡) = 𝐸𝑚𝑝(𝑡; 𝜆)
Let
𝑋~𝐸𝑚𝑝(𝜆) 𝑠𝑓𝑝 𝑌 = X1/β
Weibull Then
Distribution −1
𝑌~𝑊𝑅𝑠𝑏𝑠𝑠𝑠(𝛼 = 𝜆 𝛽 , 𝛽)
𝑊𝑅𝑠𝑏𝑠𝑠𝑠(𝑡; 𝛼, 𝛽) Special Case:
1
𝐸𝑚𝑝(𝑡; 𝜆) = 𝑊𝑅𝑠𝑏𝑠𝑠𝑠 �𝑡; 𝛼 = , 𝛽 = 1�
𝜆
Let
𝑋~𝐸𝑚𝑝(𝜆) 𝑠𝑓𝑝 𝑌 = [𝑋], 𝑌 𝑠𝑠 𝑡ℎ𝑅 𝑠𝑓𝑡𝑅𝑅𝑅𝑠 𝑓𝑝 𝑋
Geometric Then
Distribution 𝑌~𝐺𝑅𝑓𝑚𝑅𝑡𝑠𝑠𝑐(𝛼, 𝛽)

𝐺𝑅𝑓𝑚𝑅𝑡𝑠𝑠𝑐(𝑘; 𝑝) The geometric distribution is the discrete equivalent of the

continuous exponential distribution. The geometric distribution is
also memoryless.
Let
Rayleigh
𝑋~𝐸𝑚𝑝(𝜆) 𝑠𝑓𝑝 𝑌 = √X
Distribution
Then
1
𝑅𝑠𝑦𝑠𝑅𝑠𝑅ℎ(𝑡; 𝛼) 𝑌~𝑅𝑠𝑦𝑠𝑅𝑠𝑅ℎ(𝛼 = )
√𝜆
Special Case:
Chi-square 1
𝜒 2 (𝑚; 𝑠) 𝜒 2 (𝑚; 𝑠 = 2) = 𝐸𝑚𝑝 �𝑚; λ = �
2
Let
Pareto
𝑌~𝑃𝑠𝑠𝑅𝑡𝑓(𝜃, 𝛼) 𝑠𝑓𝑝 𝑋 = ln(𝑌⁄𝜃 )
Distribution
Then
𝑃𝑠𝑠𝑅𝑡𝑓(𝑡; 𝜃, 𝛼)
𝑋~𝐸𝑚𝑝(𝜆 = 𝛼)
 48 Common Life Distributions

Let
𝑅 −𝑋
Logistic 𝑋~𝐸𝑚𝑝(𝜆 = 1) 𝑠𝑓𝑝 𝑌 = ln � �
Distribution 1 + 𝑅 −𝑋
𝐿𝑓𝑅𝑠𝑠𝑡𝑠𝑐(µ , 𝑠) Then
𝑌~𝐿𝑓𝑅𝑠𝑠𝑡𝑠𝑐(0,1)
Exponential

(Hastings et al. 2000, p.127):