Technical Support Document

Distribution Models for Reliability Data

Introduction
When performing reliability analysis, you must choose a distribution to model your data. The more closely the
distribution fits your data, the more likely the reliability statistics will accurately describe the performance of
your product. In addition, a well-fitting model can be used to make reasonable projections when extrapolating
beyond the range of data.
To choose a distribution, rely on your practical knowledge and experience with the product performance. Ask
yourself the following questions:

Do the data follow a symmetric distribution? Are they skewed in one direction?
Is the failure rate constant? Increasing? Decreasing?
What distribution has worked in the past for this situation?

You can also evaluate the fit of your data using Minitabs Distribution ID plot (Stat > Reliability/Survival >
Distribution Analysis (Right-Censoring or Arbitrary Censoring). The output displays probability plots for
many types of distributions commonly used in reliability analysis, including:

Weibull distribution
Exponential distribution
Lognormal distribution
Smallest Extreme Value distribution
Normal distribution

Frequently, you can model a given set of data with more than one distribution, or with a distribution that is
defined by one, two, or three parameters. For example, distributions that can model each type of data are
summarized below:

Left-skewed data In many cases, you can fit the Weibull or smallest extreme value distribution.

Symmetric data You can generally obtain a good fit with a Weibull or lognormal distribution. In some
cases, you can fit the normal distribution (depending on the heaviness of the tails) and obtain similar
results.

Right-skewed data You can often fit either the Weibull or the lognormal distribution and obtain a good
fit to the data.

A given set of data can sometimes be modeled using either 2 or 3 parameters. A 3-parameter model may
provide a better fit for some data, but may also result in overfitting the model. In general, experts generally
advise choosing the simplest model that works.
In this paper, well discuss the characteristics of each distribution and provide some common applications.

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 1

Technical Support Document

Distribution Models for Reliability Data

The Weibull Distribution
The Weibull distribution is the most commonly used distributions for modeling reliability data. This distribution
is easy to interpret and extremely versatile. By adjusting the value of its shape parameter, , you can model the
characteristics of many different life distributions, as shown in the table below.
Shape
Probability
Hazard function
parameter Distribution Function (failure rate)
(PDF)

Type of Product Failure

0<<1
Early failures, also known as
Infant mortality, because
they occur in initial period of
product life

Exponentially decreasing
from infinity

Initially high failure rate

that decreases over time
(first part of bathtub
shaped hazard function).

These failures may

necessitate a product burnin period to reduce risk of
initial failure.

=1
Random failures, multiplecause failures
Models useful life of
product

Exponentially decreasing
from 1/
scale parameter)

Constant failure rate over

life of product

= 1.5
Early wear-out failure

Rises to peak then

decreases

Increasing failure rate, with

most rapid increase initially

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 2

Technical Support Document

Distribution Models for Reliability Data

=2
Risk of wear-out failure
increases steadily over the
life of the product

Rayleigh distribution

Linearly increasing failure

rate

3 4
Rapid wear-out failures
Models the final period of
product life, when most
failures occur
Bell-shaped

Rapidly increasing

> 10
Very rapid wear-out failures
Models the final period of
product life, when most
failures occur
Very rapidly increasing
Similar to extreme value
distribution

Applications of the Weibull Distribution

The Weibull distribution can model data that are right-skewed, left-skewed, or symmetric. For this reason, the
distribution is used to evaluate reliability across diverse applications, including vacuum tubes, capacitors, ball
bearings, relays, and material strengths. The Weibull distribution can also model a hazard function that is
decreasing, increasing or constant, allowing it to describe any phase of an items lifetime.
Increasing hazard function
An increasing hazard function is probably the most common scenario, where items are more likely to fail with
time. For example, many mechanical items that are prone to stress or fatigue will have an increased risk of
failure over the lifetime of the product. Engineers might use a test to simulate wear-out stress, such as using a
machine to simulate extended usage of a light bulb over time and then recording the time until a failure occurs.
A Weibull distribution is often used to model this type of wear-out failure.
Decreasing hazard function
A decreasing hazard function indicates failures that are more likely to occur early in the life of a product. One
example is products or parts composed of metals that harden with use and thus grow stronger as time passes.
Another example is bugs in a computer program, which may be likely to appear initially but then decrease as
time passes. Often, this type of data can be modeled using a Weibull distribution with a shape parameter less
than 1.
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 3

Technical Support Document

Distribution Models for Reliability Data

Constant hazard function
A constant hazard function indicates failures that are equally likely to occur at any time in the products life.
This relatively constant period of low failure risk characterizes the middle portion of the Bathtub Curve (see
below). This function can also be modeled using the Exponential distribution.
Bathtub-shaped hazard function
If the hazard function is bathtub-shaped, the risk of failure is high at the start, decreases rapidly, levels off and
remains fairly constant, and then increases rapidly at the end of the products life. Televisions and handheld
calculators are two products that commonly exhibit a bathtub-shaped hazard function. Another example is a
microprocessor, which may fail soon after being put into a computer system.

Figure 1: Bathtub shaped hazard function

The Weibull distribution can be used to model every phase of the bathtub-shaped hazard function. In the early
failure period, the risk of failure is initially high, then decreases rapidly and levels off. Typically, the decreasing
risk occurs over a period of several weeks to a few months. This early infant mortality phase can be modeled
using a Weibull shape parameter, , between 0 and 1.
Products that are initially prone to manufacturing, design, or component defects often have a high risk of failure
at the onset. Defects of this type include poor soldering of parts, faulty mechanical attachments, weak wire
bonding, and poor die attachment. To reduce the risk of their infant mortality, these products often require
initial stress testing such as burn-in (to stress devices under constant operating conditions); power cycling (to
stress devices under the surges of turn-on and turn-off); and temperature cycling (to mechanically and
electrically stress devices over temperature extremes).
After the early high-risk period is passed, the product is then likely to fail only from normal wear-out, near the
end of its expected life, at which time the risk of failure sharply increases. This last stage of product life can be
modeled using a Weibull shape parameter, , greater than 3.
Rayleigh distribution
When the Weibull distribution has a shape parameter of 2, it is known as the Rayleigh distribution. This
distribution is often used to describe measurement data in the field of communications engineering, such as
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 4

Technical Support Document

Distribution Models for Reliability Data

measurements for input return loss, modulation side-band injection, carrier suppression, and RF fading. This
distribution is also commonly used in the life testing of electrovacuum devices.
Weakest link model
The Weibull distribution can also model a life distribution with many identical and independent processes
leading to failure, in which the first to reach a critical stage determines the failure time. Extreme value theory
serves as the basis for this weakest link model, where many flaws compete to be the eventual site of failure.
Because the Weibull distribution can be theoretically derived from the Smallest Extreme Value Distribution, it
can also provide an effective model for weakest-link applications such as capacitor, ball bearing, relay and
material strength failures. However, if the variable of interest can take negative values, the smallest extreme
value distribution is preferable, because the Weibull distribution can only model positive values due to its lower
bound of 0.
In short, the Weibull distribution is extremely versatile and can be used to model a wide range of reliability
data. Use this distribution to answer questions such as:

What percent of items are expected to fail during the burn-in period? E.g. What percent of fuses are
expected to fail during the 8-hr burn-in period?

How many warranty claims can be expected during the useful life phase? E.g. How many warranty
claims do you expect to see over the 50,000 mile useful life of this tire?

When is rapid wear-out expected to occur? E.g. When should maintenance be regularly scheduled to
prevent engines from entering their wear-out phase?

The Weibull distribution may not work as effectively for product failures that are caused by chemical reaction
or a degradation process like corrosion, which can occur with semiconductor failures. Typically, these types of
situations are usually modeled using the lognormal distribution.
Example 1: Capacitors
Capacitors were tested at high stress to obtain failure data (in hours). The lifetime data were modeled by a
Weibull distribution.
Capacitor Failure Times
6

Frequency

5
4
3
2
1
0

400

800

1200
1600
Hours to Failure

2000

2400

2800

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 5

Technical Support Document

Distribution Models for Reliability Data

Example 2: Tensile strength
Engineers tested and recorded tensile strength from a sample of alloys. Distribution of the data is skewed to left
and modeled with a 2-parameter Weibull.
Histogram of Strength (ksi)
25

Frequency

20

15

10

80

82

84
Strength (ksi)

86

88

The Exponential Distribution

The exponential distribution is a simple distribution with only one parameter and is commonly used to model
reliability data. This distribution is actually a special case of the Weibull distribution (see example of Weibull
with = 1).
An important property of the exponential distribution is that it is memoryless. The memoryless property
indicates that the remaining life of a component is independent of its current age. For example, a system that
experiences wear and tear and therefore becomes more likely to fail later in its life is not memoryless.
Therefore, this distribution should be used when the failure rate is constant over the entire life of the product.
The number of failures per unit in time is typically expressed in failures per unit of time, such as percent of
failure per thousand-hour units.
Probability Distribution Function
Skewed right

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 6

Technical Support Document

Distribution Models for Reliability Data

Hazard function

Risk of failure is constant

Applications of the Exponential Distribution

The exponential distribution provides a good model for a product or item that is just as likely to fail at any time,
regardless of whether it is brand new, a year old, or several years old. In other words, the item should not age or
wear out over its expected application.
For this reason, the exponential distribution is often used to model electronic components that typically do not
wear out until long after the expected life of the product in which they are installed. Examples include
components of high-quality integrated circuits, such as diodes, transistors, resistors, and capacitors.
The exponential distribution is also considered an excellent model for the long, flat(relatively constant) period
of low failure risk that characterizes the middle portion of the Bathtub Curve. This phase corresponds with the
useful life of the product and is known as the intrinsic failure portion of the curve.
However, the exponential distribution should not be used to model mechanical or electric components that are
expected to show fatigue, corrosion, or wear before the expected life of the product is complete, such as ball
bearings, or certain laser or filament devices.
Example: Transistors
An electronic component is known to have a constant failure rate over the expected life of a product. Engineers
record the time to failure of the component under normal operating conditions.

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 7

Technical Support Document

Distribution Models for Reliability Data

Component failure
7
6

Frequency

5
4
3
2
1
0

160

320
Hours

480

640

Example 2: Filaments
A light bulb company makes an incandescent filament that is not expected to wear out during an extended
period of normal use. They want to guarantee it for 10 years of operation. Engineers stress the bulbs to simulate
long-term use and record the months until failure for each bulb.
Bulb Failure
50

Percent

40

30

20

10

200

400
600
Month to Failure

800

1000

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 8

Technical Support Document

Distribution Models for Reliability Data

The Lognormal Distribution
The lognormal distribution is a flexible distribution that is closely related to the normal distribution. This
distribution can be particularly useful for modeling data that are roughly symmetric or skewed to the right.
Probability Density Function
Skewed right

Hazard Function
Increases to a maximum then decreases

Like the Weibull distribution, the lognormal distribution can take on markedly different appearances depending
on its shape parameter.
Lognormal distributions

Density

Scale
0.2
0.5
1
5

In fact, the lognormal model and the Weibull model may sometimes fit a given set of life test data equally well.
However, there is one important difference to consider. When using these distributions to extrapolate beyond
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 9

Technical Support Document

Distribution Models for Reliability Data

the range of sample data, the lognormal will predict lower average failure rates at earlier times than the Weibull
distribution.
Applications of the Lognormal Distribution
The lognormal distribution has been called the most commonly used life distribution model for many hightechnology applications. The distribution is based on the multiplicative growth model, which means that at any
instant of time, the process undergoes a random increase of degradation that is proportional to its present state.
The multiplicative effect of all these random independent growths accumulates to trigger failure. Therefore, the
distribution is often used to model parts or components that fail primarily due to stress or fatigue, including:

Failure due to chemical reactions or degradation, such as corrosion, migration, or diffusion, which is
common with semiconductor failure

Time to fracture in metals subject to the growth of fatigue cracks

Electronic components that exhibit decreased risk of failure after a certain time

However, if components are not expected to fail until well after the technological life of the product in which
they are installed is complete (that is, the failure rate a component is constant over its expected lifetime), an
exponential distribution may be more appropriate.
Example 1: Electronic Components
Engineers record the time to failure of an electronic component under normal operating conditions. The
component shows a decreased risk of failure over time.
Component life
12
10

Frequency

8
6
4
2
0

30

45

60

75
90
Time to failure

105

120

135

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 10

Technical Support Document

Distribution Models for Reliability Data

Example 2: Diesel Generator Fans
Time until failure was tracked over the life of diesel generator fans.
Histogram of Hours
16
14

Frequency

12
10
8
6
4
2
0

2000

4000

6000
Hours

8000

10000

12000

Smallest Extreme Value Distribution

The smallest extreme value distribution is a limiting distribution for the minimum of a very large collection of
random observations from the same arbitrary distribution. Extreme value theory is a useful model in situations
where many identical and independent processes can lead to failure and the first one to fail determines the
failure time. In this scenario, sometimes referred to as the worst or weakest link, the distribution is typically
skewed to the left.
Probability Distribution Function
Smallest Extreme Value Distribution
Scale
5
6
7

Density

The relationship between the extreme value distributions and the Weibull distribution is similar to that between
the normal and lognormal distributions. Specifically, the log (base e) of a variable following a Weibull
distribution has a smallest extreme value distribution.
Despite this equivalence, the distributions are not strictly interchangeable in their applications. The National
Institute of Standards and Technology recommends trying the extreme value distribution in any modeling
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 11

Technical Support Document

Distribution Models for Reliability Data

application for which the variable of interest is the minimum of many random factors, all of which can take
positive or negative values.
The hazard function of the smallest extreme value distribution shows a risk of failure that is exponentially
increasing.

This hazard function suggests that the smallest extreme value distribution is suitable for modeling the life of a
product that experiences very rapid wear-out after a certain age. This includes the final stage of the Bathtub
Curve, known as the wear-out period.
Applications of the Smallest Extreme Value Distribution
The smallest extreme value distribution is often appropriate for product failures related to load and strength.
The extreme value distribution is used to model minimum values. When using this distribution, you are
typically not concerned with the distribution of variables that describes most of the population but only with the
extreme values that can lead to failure. In other words, you are investigating imperfections in certain materials
that can cause nonuniform stress under a load. The strength of the material is therefore related to the effect of
the imperfection causing the greatest reduction in strength (the weakest link).
One common application is capacitor dielectric breakdown, where many flaws compete to be the eventual site
of failure. Another example is semiconductor wire bonds, which typically do not fracture or overheat under
normal operating conditions, unless they are subject to extreme electrical load or extremely low bond strength.
Similarly, coolant tubes have a minimum thickness to provide sufficient heat transfer to coolant liquid. But a
failure occurs if the hot combustion gases burn pinholes through any point on the tubes.
Use the smallest extreme value distribution to answer questions such as:
Which material can withstand the largest load?
How many items are expected to break during the warranty period?
What is the minimum force needed to break a pouch when multiple strength tests are performed on different
section of each part?
Which brand of cable is better able to withstand a load of 1,000 pounds?
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 12

Technical Support Document

Distribution Models for Reliability Data

Example 1: Wire Strength
Wire samples of equal length are tested for breaking strength.
Histogram of Strength
4

Frequency

48

56

64
Strength

72

80

Example 2: Alloy Specimens

Engineers subject an alloy specimen to a total of 300,000 cycles and measure the number of cycles until failure.
14
12

Frequency

10
8
6
4
2
0

120

160
200
240
Number of Cycles Until Failure

280

The Normal Distribution

This well-known distribution, which is symmetric and bell-shaped, is widely used in statistics. Industrial
applications often generate data that are normally distributed.
Historically, the normal distribution has not been as commonly used to model reliability data as other
distributions, partially because its left tail extends to negative infinity, which could result in erroneous modeling
of negative times to failure. Most reliability data are modeled using distributions for positive random variables,
such as exponential, Weibull, gamma, and lognormal. Fewer applications use the normal distribution as a model
Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 13

Technical Support Document

Distribution Models for Reliability Data

for product life. However, if the mean of the data is greater than 0 and its variation relatively low, the normal
distribution can be useful for modeling certain types of life data.
Probability Distribution Function

Symmetric and bell-shaped

Hazard Function

The risk of failure is strictly increasing

Note that the normal distribution closely approximates the Weibull distribution when 3 < < 4.
Applications of the Normal Distribution
The normal distribution can sometimes be used to model the life of consumable items, in which the risk of
failure is always increasing. Electric filament devices, such as incandescent light bulbs and toaster heating
elements, have been given as examples of items whose failure data may follow a normal distribution. The
strength of a wire bond in integrated circuits is another example.

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 14

Technical Support Document

Distribution Models for Reliability Data

Example 1: Beverage Shelf Life
To evaluate the shelf life of beverages, analysts record the number of days before a bottled beverage discolors.
Histogram of Shelf Life
9
8
7

Frequency

6
5
4
3
2
1
0

80

100

120
Shelf Life

140

160

Example 2: Toaster Reliability

Engineers perform life tests on a toaster with a new component design.
Histogram of Hours
20

Frequency

15

10

150

300

450
600
Hours

750

900

1050

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 15

Technical Support Document

Distribution Models for Reliability Data

References
National Institute of Standard and Technology Engineering Statistics Handbook. Chapter 8: Assessing Product
Reliability. Available online at: http://itl.nist.gov/div898/handbook/toolaids/pff/8-apr.pdf
Patrick O Connor. Practical Reliability Engineering. 3rd edition. New York: John Wiley and Sons, 1991.
Russell J Hoppenstein. Statistical reliability analysis on Rayleigh probability distributions. RF design tutorial.
October 2000. Available at:
http://s3.amazonaws.com/zanran_storage/www.frbb.utn.edu.ar/ContentPages/42682870.pdf
Lawrence M. Leemis. Reliability: Probabilistic Models and Statistical Methods. Upper Saddle River, NJ:
Prentice Hall, 1995.
David K Lloyd and Myron Lipow. Reliability: Management, Methods, and Mathematics. Englewood Cliffs, NJ:
Prentice-Hall, 1962.
William Q. Meeker and Luis A. Escobar. Statistical Methods for Reliability Data. New York: John Wiley and
Sons, 1998.
Paul A. Tobias and David C. Trinidade. Applied Reliability. 2nd edition. New York: Chapman and Hall/CRC,
1995.
Raytheon Reliability Analysis Laboratory. Finding Defects Using Burn-in Analysis. Joe Dzekevich. Available
at: https://www.reliabilityanalysislab.com/0208_BurnIn.asp

Knowledgebase ID 2716: http://www.minitab.com/support/answers/answer.aspx?id=2716

Page 16