9/1/2021

Failure Distributions
(Time-dependent Failure Rates)
Nagi Gebraeel
Professor of Industrial & Systems Engineering
Georgia Tech

Failure Models
Time-Dependent Failure Models
Weibull Distribution
Normal Distribution
Lognormal Distribution
Gamma Distribution

1
 9/1/2021

Time-Dependent Failure Models

 Time-dependent failures models include probability distributions that
model failure processes that have hazard rate functions that are not
constant over time.

 Probability distributions that fall under this category include the

Weibull, Normal, Lognormal, and Gamma distributions.

Weibull Distribution
 The Weibull distribution is used to model both increasing and decreasing
failure rates.
 Its hazard rate function is characterized by the following form:
𝛽 𝑡
𝜆 𝑡 =
𝜃 𝜃
𝛽 𝑡′
𝑅 𝑡 = exp − d𝑡 =𝑒
𝜃 𝜃
𝛽 𝑡
𝑓 𝑡 = 𝑒 𝜃 > 0, 𝛽 > 0, 𝑡 ≥ 0
𝜃 𝜃

2
 9/1/2021

Weibull Distribution
 β is referred to as the shape parameter.
 It affects the shape of the distribution in the sense that for 𝛽 < 1 the
distribution looks similar to an exponential. In fact, for 𝛽 = 1 the
distribution is indeed Exponential with 𝜆 = 1⁄𝜃
 Whereas for 𝛽 > 3, the distribution is close to symmetrical, and for
1 < 𝛽 < 3 it is skewed.
𝛽 𝑡
𝜆 𝑡 =
𝜃 𝜃

Weibull Distribution
 𝜃 is the scale parameter and it influences both the mean and the spread
of the distribution.
 𝜃 is also called the characteristic life, and usually has a unit of time.
 Reliability at the characteristic life, 𝑅 𝜃 , is always 0.368. How?

𝜃
𝑅 𝜃 = exp − = exp −1 = 0.368
𝜃
 In other words, we expect that 63.2% of all the Weibull failures will occur
by time 𝑡 = 𝜃.

3
 9/1/2021

Weibull Distribution
 Given a desired reliability R,

𝑅 𝑡 =𝑒 =𝑅
the design life can be estimated as follows;
⁄
𝑡 = 𝜃 − ln 𝑅
 When 𝑅 = 0.99, 𝑡 . is referred to as the B1 life. The time at which 1% of
the population will have failed.
 When 𝑅 = 0.5, then
⁄
𝑡 . =𝑡 = 𝜃 − ln 0.5
= 𝜃 0.69315 ⁄
is the median time to failure.

Beta values
Weibull plots: PDF and Reliability Function
𝑓(𝑡) 𝑅(𝑡)

4
 9/1/2021

Weibull plots: Hazard Function

6

= f(t) / R(t) Beta

0.5
5
1.5

2.0
4
4.0
Hazard Rate

Theta = 2
2

0 t
0.0 1.0 2.0 3.0 4.0 5.0 6.0

Weibull plots: Characteristic Life 𝜽

Smaller theta = Earlier Failure
𝑓(𝑡) 1.6

1.4
Theta
1.2 0.5

1.0
1
2.0
0.8

Beta = 1.5
0.6

0.4

0.2

0
0.0 1.0 2.0 3.0 4.0 5.0 6.0
t

10

5
 9/1/2021

Weibull Distribution
 The MTTF and the variance of the Weibull distribution are given as follows:

where;
Γ 𝑥 = 𝑦 𝑒 d𝑦

Γ 𝑥 = 𝑥−1 Γ 𝑥−1

11

Weibull Distribution: Shape Parameter 𝜷

Value Property
0<𝛽<1 Decreasing Failure Rate (DFR)
𝛽=1 Exponential Distribution (CFR)
1<𝛽<2 IFR-concave
𝛽=2 Rayleigh Distribution (LFR)
𝛽>2 IFR - Convex
3≤𝛽≤4 IFR - Approaches Normal
Distribution – Symmetrical
• 𝛽 =   Most closely approximates the normal
• 𝛽 =   Mean = Median

12

6
 9/1/2021

Example: Weibull Distribution

 A compressor experiences wearout that can be characterized by an linearly
increasing hazard rate function:
2 𝑡
𝜆 𝑡 =
1000 1000
 Calculate the MTTF and the Design Life that ensures a reliability 0.99.
Before we start we need to identify the parameters of the Weibull distribution.
Let’s take a closer look at the expression of the hazard function.
2 𝑡 𝛽 𝑡
𝜆 𝑡 = −−→ 𝜆 𝑡 =
1000 1000 𝜃 𝜃
It is clear that 𝜷 = 𝟐 𝐚𝐧𝐝 𝜽 = 𝟏𝟎𝟎𝟎 𝐡𝐫

13

Example: Weibull Distribution

 A compressor experiences wearout that can be characterized by an linearly
increasing hazard rate function:
2 𝑡
𝜆 𝑡 =
1000 1000
 Calculate the MTTF and the Design Life that ensures a reliability 0.99.
1 1
𝑀𝑇𝑇𝐹 = 𝜃 × Γ 1 + = 1000 × Γ 1 + = 886.23
𝛽 2
Note that Γ 1 + is evaluated as “=GAMMA(1.5)” in Excel

14

7
 9/1/2021

Example: Weibull Distribution

 A compressor experiences wearout that can be characterized by an linearly
increasing hazard rate function:
2 𝑡
𝜆 𝑡 =
1000 1000
 Calculate the MTTF and the Design Life that ensures a reliability 0.99.
For a reliability of 0.99, we have the following:

𝑅 𝑡 =𝑒 =𝑒 = 0.99
Thus design life for 0.99 reliability is given by,

𝑡 = 1000 − ln 0.99 = 100.25 hr

15

Knowledge Check
A compressor experiences wearout that can be characterized by an linearly
increasing hazard rate function:
2 𝑡
𝜆 𝑡 =
1000 1000

What are the values of 𝜃 and 𝛽

a) 2 and 1000
b) 3 and 1000
c) Unknown

16

8
 9/1/2021

Failure Modes
 Complex systems may experience failures in multiple ways, i.e., resulting from
different physical characteristics.
 These failure categories are referred to as failure modes. Independent
 If we define 𝑅 (𝑡) as the reliability of the 𝑖 failure mode, i.e., the probability
that the 𝑖 failure mode does not occur before time 𝑡, and assume that these
failure modes are independent, then the system reliability, which we denote by
𝑅(𝑡) can be expressed as follows:

Note: This is the probability that none of the 𝑛 failure modes occurs before time 𝑡.

18

Failure Modes
 Consider a system with n independent failure
modes, (or with n components in series)
each having an independent Weibull failure
distribution with shape parameter 𝛽 and
scale parameter 𝜃 , then the system hazard
rate function can be determined as follows:

𝛽 𝑡 1
𝜆 𝑡 = = 𝛽𝑡
𝜃 𝜃 𝜃

 This property is true only when each

component has the same shape parameter.

19

9
 9/1/2021

Identical Weibull Components When they are i.i.d:

 If a systems of n serially related and independent component have identical

hazard rate functions.
𝛽 𝑡
𝜆 𝑡 =
𝜃 𝜃
 Then the hazard rate of the system can be written as follows:

𝛽 𝑡
𝜆 𝑡 =
𝜃 𝜃

 Therefore,
𝑡
𝑅 𝑡 = exp −𝑛
𝜃

20

Burn-In Screening for the Weibull

 Many companies undergo what is known as Burn-in screening to reduce
unexpected failures of new products.
 Recall that burn-in typically represents the first part of the bathtub curve
with a decreasing hazard rate function.
 If a product survives burn-in screening, it is expected to have a higher
reliability.
 We can use conditional reliability to estimate reliability after burn-in
Do you recall
screening period, 𝑇 . the Exponential
𝑡+𝑇 Case?
exp − 𝜃 𝑡+𝑇 𝑇
𝑅 𝑡|𝑇 = = exp − + The memoryless
𝑇 𝜃 𝜃 Property. Doesn’t
exp − apply here!
𝜃

21

10
 9/1/2021

Example: Burn-in Screening

 Givn a Weibull failure distribution with a shape parameter of and a scale
parameter of 16,000.
 Define the reliability function and calculate the design life if a 90% reliability is
desired.
𝑡 /
𝑅 𝑡 = exp − =0.9
16,000
𝑡 = 16,000 − ln 0.9 = 18.71 hr

 Calculate the design life for the same reliability level (90%) after a 10 hour burn-in
screening period was conducted.
/ /
𝑅 𝑡|10 = exp − + = 0.9
, ,
/
𝑡 = 16,000 − ln 0.9 + − 10 = 101.24 hr
,

22

Three-parameter Weibull
 Wherever there is a minimum life 𝑡 such that 𝑇 > 𝑡 , the three-parameter
Weibull may be appropriate.
 The distribution assumes that no failures will take place prior to time 𝑡 .
Reliability is expressed as :
𝑡−𝑡
𝑅 𝑡 = exp − 𝑡≥𝑡
𝜃
𝛽 𝑡−𝑡
𝜆 𝑡 = 𝑡≥𝑡
𝜃 𝜃
 The parameter 𝑡 is known as the location parameter. The following are the parameters
of this distribution:
1
𝑀𝑇𝑇𝐹 = 𝑡 + 𝜃Γ 1 +
𝛽
/
𝑇 = 𝑡 + 𝜃 0.69315
/
𝑡 = 𝑡 + 𝜃 − ln 𝑅

23

11
 9/1/2021

Minimum Extreme Value Distribution

 An extreme value distribution is the asymptotic distribution resulting
from finding the minimum or maximum of a large number of
observations from an underlying unbounded parent distribution.
 It is related to the Weibull distribution.
 If the underlying distribution is the failure distribution of many
competing failure modes, then we are interested in finding the
minimum that corresponds to the component’s failure time.

24

Minimum Extreme Value Distribution

 To derive the distribution, we begin by considering a 2-parameter
exponential with a scale parameter 𝛼 and a location parameter 𝜇.
1
𝜆 𝑡 = 𝑒 𝛼 > 0, −∞ < 𝑡 < ∞
𝛼
1 _
𝑅 𝑡 = exp − 𝑒 𝑑𝑡′ = exp −𝑒 = exp −𝑒
𝛼
 The pdf
_
𝑑 exp −𝑒 _
1 _
𝑓 𝑡 = = 𝑒 𝑒
𝑑𝑡 𝛼
MTTF
 The mean of the distribution is 𝜇 − 𝛼𝛾, where 𝛾 = 0.577216 is called the Euler
constant, and a standard deviation 𝛼𝜋/ 6.
 The location parameter 𝜇 is the mode of the distribution.

25

12
 9/1/2021

Minimum Extreme Value Distribution

 If 𝑥 has a Weibull distribution with parameters 𝛽 and 𝜃, then 𝑡 = ln 𝑥
has a minimum extreme value distribution with parameters 𝛼 = and
𝜇 = ln 𝜃 where 𝑥 > 0, and 𝑡 has the reliability function.
_
𝑅 𝑡 = exp −𝑒 = −𝑒 / =𝑒

 = 3 and  = 12

26

Example:
 A manufactured part is subjected to a large number of (cyclical) stress points at
which a fracture line may initiate. The weakest stress point will show a fracture
line the earliest resulting in the part failing. Laboratory tests have determined
that the failure time is a minimum extreme value distribution with a location
parameter of 5 years and scale parameter of 0.4.

𝑀𝑇𝑇𝐹 = 𝜇 − 𝛼𝛾 = 5 − 0.4 0.577216 = 4.77 yrs

𝛼𝜋 0.4𝜋
𝜎= =
= 0.513 yr
6 6
𝑡 = 𝜇 − 0.366513 𝛼 = 5 − 0.366513 0.4 = 4.85 yrs.

27

13
 9/1/2021

Normal Distribution
 One of the most popular distributions for modeling various types of
phenomena .
 The normal distribution has been widely used to model fatigue and
wear-out phenomena.
1 1 𝑡−𝜇
𝑓 𝑡 = exp − −∞ <𝑡 < ∞
2𝜋𝜎 2 𝜎

where 𝜇 and 𝜎 are the mean & variance,

 The CDF of the normal distribution is,
1 1 𝑡′ − 𝜇
𝐹 𝑡 = exp − d𝑡′
2𝜋𝜎 2 𝜎

30

Normal Distribution
 The CDF of the Normal is evaluated numerically.
 The first step to evaluate cdf/pdf is to transform the normal random variable T
as shown below,
𝑇−𝜇
𝑧=
𝜎
to a Standard Normal r.v. with mean 0 and variance 1 defined by the following
PDF,
1
𝜙 𝑧 = 𝑒 /
2𝜋
and the following CDF,

Pr Z ≤ z = Φ 𝑧 = 𝜙 𝑧′ d𝑧′

31

14
 9/1/2021

Example: Normal Distribution

 Wear-out in a machine follows a normal distribution with mean 10 months
and standard deviation 2. Calculate the probability that a failure will occur
between 9 and 11 months.

 9  10 x  10 11  10 
P  9  X  11  P    
 2 2 2 
 P 0.5 z 0.5 

  
 P  z  0.5   P  z  0.5 
 0.69146  0.30854  0.38292

Using Excel
0.38292 = NORMDIST(11,10,2,TRUE) - NORMDIST(9,10,2,TRUE)
NORMDIST( value, mean, sd, T)

32

Standard Normal Table

33

15
 9/1/2021

Normal Reliability Function

 The Normal reliability function is given as,
𝑇−𝜇 𝑡−𝜇
𝑅 𝑡 =𝑃 𝑇≥𝑡 =𝑃 ≥
𝜎 𝜎
𝑡−𝜇 𝑡−𝜇
=𝑃 𝑧≥ = 1−Φ
𝜎 𝜎
 The hazard function of the normal distribution can be expressed as
follows:
𝑓(𝑡) 𝑓(𝑡)
𝜆 𝑡 = =
𝑅(𝑡) 1 − Φ 𝑡 − 𝜇 /𝜎

 The hazard function for the normal distribution can be shown to be an

increasing function, and thus commonly used for IFR processes.

34

Example: Normal Distribution

 A new fan belt is developed from a higher grade of material. It has a time to
failure distribution which is normal with a mean of 35,000 vehicle miles and a
standard deviation of 7,000 vehicle miles. Find its designed life if a .97
reliability is desired.
𝑡−𝜇 𝑡 − 35,000
𝑅 𝑡 = 1−Φ = 1−Φ = 0.97, find 𝑡?
𝜎 7,000
From the tables we find that 1 − Φ −1.88 = 0.96995
𝑡 − 35,000
Therefore = −1.88 and 𝑡 . = 21,480 miles
7,000

35

16
 9/1/2021

Knowledge Check
Tire wear follows a normal distribution with mean 30,000 miles and standard
deviation 3039.5 miles. Determine the tire reliability at 24,000 miles.
a) 0.0244
b) 0.8
c) 0.9756

36

Central Limit Theorem

 The central limit theorem (CLT) provides a strong motivation for many of the
statistical applications of the normal distribution.
 The CLT states that under fairly general conditions the sum of 𝑛 random variables
approaches a normal distribution as 𝑛 approaches infinity.
 More formally, if 𝑌 = 𝑋 + 𝑋 + ⋯ + 𝑋 where 𝑋 , 𝑋 , … , 𝑋 are 𝑛 independent
random variables with finite mean 𝐸(𝑋 ) and finite variance 𝑉 𝑋 .
 For a sufficiently large 𝑛, 𝑌 has an approximately normal distribution with the
following mean and variance:

𝐸 𝑌 = 𝐸 𝑋 𝑉 𝑌 = 𝑉 𝑋

 This result holds regardless of the distribution of the random variables.

37

17
 9/1/2021

Central Limit Theorem

 The central limit theorem (CLT) provides a strong motivation for many of the
statistical applications of the normal distribution.
 The CLT states that under fairly general conditions the sum of 𝑛 random variables
approaches a normal distribution as 𝑛 approaches infinity.
 More formally, if 𝑌 = 𝑋 + 𝑋 + ⋯ + 𝑋 where 𝑋 , 𝑋 , … , 𝑋 are 𝑛 independent
random variables with finite mean 𝐸(𝑋 ) and finite variance 𝑉 𝑋 .
 For a sufficiently large 𝑛, 𝑌 has an approximately normal distribution with the
following mean and variance:

𝐸 𝑌 = 𝐸 𝑋 𝑉 𝑌 = 𝑉 𝑋

 This result holds regardless of the distribution of the random variables.

38

Lognormal Distribution
 Lognormal model is a very popular reliability model.
 If 𝑇 is lognormal then ln 𝑇 is normally distributed with mean 𝜇 and variance 𝜎 .
 The PDF of the lognormal is
1 1 𝑡
𝑓 𝑡 = exp − ln ,𝑡 ≥ 0
2𝜋 𝑠𝑡 2𝑠 𝑡
where 𝑠 is the shape parameter and 𝑡 is the location parameter

39

18
 9/1/2021

Lognormal/Normal Relationship
Distribution Lognormal Normal
Random Variable 𝑇 ln 𝑇  (t ) 
/ 1   ( ln )
Mean 𝑡 𝑒 ln 𝑡

Variance 𝑡 𝑒 𝑒 −1 𝑠 =𝜎

1 𝑡 𝑡−𝜇
Reliability 𝑅 𝑡 = 1−Φ ln 𝑅 𝑡 = 1−Φ
𝑠 𝑡 𝜎
𝑓(𝑡) 𝑓(𝑡)
Hazard 𝜆 𝑡 = 𝜆 𝑡 =
1 𝑡
1 − Φ 𝑠 ln 𝑡 1 − Φ 𝑡 − 𝜇 /𝜎

40

Lognormal Distribution
 The PDF of the lognormal can be expressed in an alternative manner that uses
the mean and standard deviation of ln 𝑇
1 1 ln 𝑡 − 𝜇
𝑓 𝑡 = exp −
2𝜋 𝑡𝜎 2 𝜎
Where 𝜇 and 𝜎 are mean and standard deviation of ln 𝑡 .
 In this form, the reliability function is,
ln 𝑡 − 𝜇
𝑅 𝑡 =𝑃 >𝑡 =𝑃 𝑧>
𝜎

41

19
 9/1/2021

Example: Lognormal Distribution

 Fatigue wearout of a component has a lognormal distribution with 𝑡 =
5000 hrs and 𝑠 = 0.2.
 Calculate the MTTF and the corresponding standard deviation.
. /
𝑀𝑇𝑇𝐹 = 5000 𝑒 = 5101 hr
. .
𝜎 = 5000 𝑒 𝑒 − 1 = 1.0619 × 10

 Calculate the reliability at 3000 hours

ln 𝑡 − 𝜇 1 3000
𝑅 𝑡 =𝑃 𝑧> =1−Φ ln = 0.995
𝜎 0.2 5000

42

Gamma Distribution
 The gamma model covers a wide range of hazard-rate functions, decreasing,
constant and increasing. It is characterized by two parameters; shape parameter 𝜈
and a scale parameter 𝛼.
𝑡 𝑒 /
𝑓 𝑡 = 𝜈, 𝛼 > 0 and 𝑡 ≥ 0
𝛼 Γ(𝜈)
 When 0 < 𝜈 < 1, then failure rate decreases from infinity to 1/𝛼 as time goes to
infinity.
 When 𝜈 > 1 , the failure rate increases from 1 to infinity. Furthermore, there is a
single peak of the pdf which occurs at time 𝑡 = 𝛼(𝜈 − 1)
 When 𝜈 = 1 , the failure rate is constant and equals 1/𝛼 , i.e., the gamma
distribution becomes an exponential.

43

20
 9/1/2021

Hazard Rate Function

 The mean of the gamma distribution is
given as 𝑀𝑇𝑇𝐹 = 𝜈𝛼

 Standard deviation is given as 𝜎 = 𝜈𝛼

𝛼 𝛾 − 1 for 𝛾 > 1
 The mode is 𝑡 =
0 Otherwise

Hazard Rate
Shape Parameter
Function
0<𝜈<1 DFR
𝜈=1 CFR
𝜈>1 IFR

44

Gamma Distribution
 The CDF can be obtained as follows;
𝑡
𝑡′ 𝑒 / 𝐼 𝛼 ,𝜈
𝐹 𝑡 = d𝑡 =
𝛼 Γ(𝜈) Γ(𝜈)

where 𝐼 , 𝜈 is called an incomplete gamma function which is evaluated

numerically.
 The reliability function is therefore given as follows:
𝑡
𝐼 𝛼 ,𝜈
𝑅 𝑡 =1−
Γ(𝜈)

45

21
 9/1/2021

Example: Gamma Distribution

 Failure of a critical machine part due to cyclical vibration can be characterized
by gamma distribution with shape parameter 2.3 and scale parameter 2000
operating hours. Calculate the MTTF and its variance.
𝑀𝑇𝑇𝐹 = 𝛾 × 𝛼 = 2.3 × 2000 = 4600 hr

𝜎= 𝛾𝛼 = 2.3 2000 = 3033.15 hr

46

22