Introduction to reliability

Reliability has gained increasing importance in the last few years in manufacturing organisations, the
government and civilian communities. With recent concern about government spending, agencies
are trying to purchase systems with higher reliability and lower maintenance costs. As consumers,
we are mainly concerned with buying products that last longer and are cheaper to maintain, i.e.,
have higher reliability. The reasons for wanting high product or component or system reliability are
obvious:
Higher customer satisfaction
Increased sales
Improved safety
Decreased warranty costs
Decreased maintenance costs, etc.
This document covers

The definition and measurement of reliability and failure rates

Modelling reliability with the exponential distribution
Mean time to and before failure
The addition and multiplication rules of probability
System reliability and fault tree analysis

Obviously, we only cover a few of the basics here. For more detail see Besterfield (2013), OConnor
(2002), or many other books and websites. There is also a presentation by Professor Ashraf Labib
covering asset management, maintenance management and various other practical aspects of the
process of managing reliability at http://woodm.myweb.port.ac.uk/q/AshrafHotTopic2012.pdf.
It is important to work through the exercises, and check your answers against the notes on the
answers at the end of the document.

What is reliability
The reliability of a product (or system) can be defined as the probability that a product will perform a
required function under specified conditions for a certain period of time. If we have a large number
of items that we can test over time, then the Reliability of the items at time t is given by

R (t ) =

number of survivors at time t

number of items put on test at time t = 0

At time t = 0, the number of survivors is equal to number of items put on test. Therefore, the
reliability at t = 0 is
R(0) = 1 = 100%

Introduction to reliability (Portsmouth Business School, April 2012)

After this, the reliability, R(t), will decline as some components fail (to perform in a satisfactory
manner).

The failure rate

The failure rate (usually represented by the Greek letter ) is a very useful quantity. This is defined as
the probability of a component failing in one (small) unit of time.
Let NF = number of failures in a small time interval, say, t.
NS = number of survivors at time t.
The failure rate can then be calculated by the equation:

(t ) =

NF
N S * t

For example, if there are 200 surviving components after 400 seconds, and 8 components fail over
the next 10 seconds, the failure rate after 400 seconds is given by
(400) = 8 / (200 x 10) = 0.004 = 0.4%
This simply means that 0.4% of the surviving components fail in each second.
(You may wonder why the above equation defines (400) and not (410). The reason is that t is a
small time interval, so it is reasonable to assume that the failure rate will not change appreciably
during the interval. We then define the failure rate using the beginning of the interval for
convenience. In the extreme we can make t infinitesimally smallwhich is the basis of the
differential calculus.)
We can define failure rates for individual components, and also for complex products like cars or
washing machines. In the latter case we need to be clear about what is meant by a failure for a car,
for example, these can range from complete breakdown to relatively minor problems. And we
should also remember that failures in complex products are generally repairable, whereas this may
not be true for individual components.

The Concept of the Bath-tub Curve

The so-called bath-tub curve represents the pattern of failure for many products especially
complex products such as cars and washing machines. The vertical axis in the figure is the failure rate
at each point in time. Higher values here indicate higher probabilities of failure.
The bath-tub curve is divided into three regions: infant mortality, useful life and wear-out.

Introduction to reliability (Portsmouth Business School, April 2012)

Bath tub curve

Infant Mortality: This stage is also called early failure or debugging stage. The failure rate is high but
decreases gradually with time. During this period, failures occur because engineering did not test
products or systems or devices sufficiently, or manufacturing made some defective products.
Therefore the failure rate at the beginning of infant mortality stage is high and then it decreases with
time after early failures are removed by burn-in or other stress screening methods. Some of the
typical early failures are:

poor welds
poor connections
contamination on surface in materials
incorrect positioning of parts, etc.

Useful life: This is the middle stage of the bath-tub curve. This stage is characterised by a constant
failure rate. This period is usually given the most consideration during design stage and is the most
significant period for reliability prediction and evaluation activities. Product or component reliability
with a constant failure rate can be predicted by the exponential distribution (which we come to
later).
Wear-out stage: This is the final stage where the failure rate increases as the products begin to wear
out because of age or lack of maintenance. When the failure rate becomes high, repair, replacement
of parts etc., should be done.

Exercises
1

Would you expect the bath tub curve to apply to a car? What about a human being?

One thousand transistors are placed on life test, and the number of failures in each time

Introduction to reliability (Portsmouth Business School, April 2012)

interval are recorded. Find the reliability and the failure rate at 0, 100, 200, etc hours. (You
may find it helpful to set this up on a spreadsheet.)
Time interval
0-100
100-200
200-300
300-400
400-500
500-600
600-700
700-800
800-900
900-1000

Number of failures
160
86
78
70
64
58
52
43
42
36

Draw a graph to show the change in the failure rate as the transistors get older.
Do you think this component shows the bath tub pattern of failure?
Draw a graph to show how the reliability changes over time.

Measuring reliability
To see the level and pattern of the reliability of a product or component in practice, it is necessary to
make some measurements. The simplest way to do this is to test a large number of products or
components until they fail, and then analyse the resulting data. Exercise 2 above shows how this
works. This enables us to estimate the failure rate and reliability after different lengths of time - and
decide, empirically, if the bath tub curve applies, or if the failure rate shows some other pattern.
There are a number of obvious difficulties which may arise. If the useful life is large it may be not be
practical to wait until products or components fail. It may be too expensive to test large samples, so
small samples may have to suffice. And for some products (e.g. space capsules) it may be difficult to
simulate operating conditions at all closely. There are a number of approaches to these difficulties most of these are beyond the scope of this unit, but they are discussed in the reading suggested for
this session. (An exception is the prediction of the reliability of a product from information about the
reliability of its components - this avoids the necessity to test the whole product - which is discussed
in a later section of the notes for this session).
It is also possible that the failure rate will depend on environmental conditions in a predictable way.
For example, one of the key factors affecting the reliability of electronic components and systems is
temperature basically the higher the temperature of the device the higher the failure rate. Most
computer equipment therefore has some form of cooling, ranging from a simple fan to forced chilled
air cooling.

Reliability Distributions
There are many statistical distributions used for reliability analysisfor example, the exponential
Introduction to reliability (Portsmouth Business School, April 2012)

distribution, the Weibull distribution, the normal distribution, the lognormal distribution, and the
gamma distribution. Here we look at the exponential distribution only, as this is the simplest and the
most widely applicable.

Reliability Prediction Using the Exponential Distribution

The exponential distribution applies when the failure rate is constant - the graph is a straight
horizontal line, instead of a bath tub. (It can be used to analyse the middle phase of a bath tub e.g. the period from 100 to 1000 hours in Exercise 2 above.) It is one of the most commonly used
distributions in reliability, and is used to predict the probability of survival to a particular time. If is
the failure rate and t is the time, then the reliability, R, can be determined by the following equation:
R(t) = e - t
There is a brief note on the mathematical background to this equation in the Appendix.
To see that it gives sensible results, imagine that there are initially 1000 components and that is
10% (0.1) per hour. After one hour about 10% of the original 1000 components will have failed leaving about 900 survivors. After two hours, about 10% of the 900 survivors will have failed leaving
about 810. Similarly there will be about 729 survivors after the third hour, which means that the
reliability after 3 hours is 0.729.
Using the exponential distribution the reliability after 3 hours, with =0.1, is given by
R(t) = e -3 = e -0.3 = 0.741
(You can work this out using a calculator or a spreadsheetsee the mathematical appendix for more
details.)
This is close to the earlier answer as we should expect. The reason it is not identical is that the
method of subtracting 10% every hour to obtain 900, 810, etc ignores the fact that the number of
survivors is changing all the time, not just every hour. The exponential formula uses calculus to take
this into account.
If the failure rate is small in relation to the time involved a much simpler method will give reasonable
results. Lets suppose that =0.01 (1%) in the example above. The formula now gives the reliability
after three hours as
R(t) = e -3 = e -0.03 = 0.9704
A simpler way of working this out would be just to say that if the failure rate is 0.01 per hour the
total proportion of failures in 3 hours will be 0.03 (3%) so the reliability after three hours is simply
R(t) = 1 0.03 = 0.97 (or 100% 3% = 97%)

Introduction to reliability (Portsmouth Business School, April 2012)

This is not exactly right because in each hour the expected number of failures will decline as the
surviving pool of working components gets smaller. But when the failure are is 1% and we are
interested in what happens after three hours, the error is negligible. On the other hand, if we want
to know what happens after 300 hours the simple method gives a silly answer (the reliability will be
negative!) and we need to use the exponential method. You should be able to check this answer
with your calculator or a computer (the answer should be 0.0498 or about 5%).

Mean time to Failure (MTTF) and Mean time between Failures (MTBF)
MTTF applies to non-repairable items or devices and is defined as "the average time an item may be
expected to function before failure". This can be estimated from a suitable sample of items which
have been tested to the point of failure: the MTTF is simply the average of all the times to failure.
For example, if four items have lasted 3,000 hours, 4000, hours, 4000 hours and 5,000 hours, the
MTTF is 16,000/4 or 4,000 hours.
The MTBF applies to repairable items. The definition of this refers to between failures for obvious
reasons. It should be obvious that
MTBF = Total device hours / number of failures
For example, consider an item which has failed, say, 4 times over a period of 16,000 hours. Then
MTBF is 16,000/4 = 4,000 hours. (This is, of course, just the same method as for MTTF.)
For the particular case of an exponential distribution,
= 1/MTBF (or 1/MTTF)
where is the failure rate.
For example, the item above fails, on average, once every 4000 hours, so the probability of failure
for each hour is obviously 1/4000. This depends on the failure rate being constant - which is the
condition for the exponential distribution.
This equation can also be written the other way round:
MTBF (or MTTF) = 1/
For example, if the failure rate is 0.00025, then
MTBF (or MTTF) = 1/0.00025 = 4,000 hours.

Exercises
3

An industrial machine compresses natural gas into an interstate gas pipeline. The
compressor is on line 24 hours a day. (If the machine is down, a gas field has to be shut
down until the natural gas can be compressed, so down time is very expensive.) The vendor

Introduction to reliability (Portsmouth Business School, April 2012)

knows that the compressor has a constant failure rate of 0.000001 failures/hr. What is the
operational reliability after 2500 hours of continuous service?
4

What is the highest failure rate for a product if it is to have a reliability (or probability of
survival) of 98 percent at 5000 hours? Assume that the time to failure follows an
exponential distribution.

Suppose that a component we wish to model has a constant failure rate with a mean time
between failures of 25 hours? Find:(a)

The reliability function.

(b)

The reliability of the item at 30 hours.

The equipment in a packaging plant has a MTBF of 1000 hours. What is the probability that
the equipment will operate for a period of 500 hours without failure?

TALCO manufactures microwave ovens. In order to develop warranty guidelines, TALCO

randomly tested 10 microwave ovens continuously to failure. The failure information of the
10 ovens is shown in the table.
Microwave

Hours

2300

2150

2800

1890

2790

1890

2450

2630

2100

10

2120

What is the mean time to failure of the microwave ovens?

Does the evidence suggest that the reliability of the ovens follows the exponential distribution (with
a constant failure rate)?

The addition and multiplication rules of probability

The next aspects of reliability theory we will consider depend on some probability theorythe
addition and multiplication rules. I will explain these in terms of dice and cards.

Introduction to reliability (Portsmouth Business School, April 2012)

Suppose that you throw a single dice. The probability of getting a 6 is

P(6) = 1/6
And the probability of getting a 5 is also
P(5) = 1/6
Equally obvious is that the probability getting a 5 or a 6 is
P(5 or 6) = 2/6 = 1/3
And that
P(5 or 6) = P(5) + P(6)
Similarly if you choose a single card from a pack of (52) cards, the probability of getting an ace or a
picture card (jack, queen or king) is
P(ace or picture) = 4/52 + 12/52 = 16/52
because 4 of the 52 cards are aces, and another 12 are picture cards.
These examples suggest that if we want the probability of something or something else happening,
we can add the probabilities.
But what about the probability of an ace or a red card (hearts or diamonds). Can we say that
P(ace or red) = P(ace) + P(red) = 4/52 + 26/52 = 30/52 ??
This is obviously wrong because two of the aces are also red, so we are in effect double counting
these aces if we add the probabilities. Before adding the probabilities you need to check that the
two events cannot both occuri.e. they do not overlap or are mutually exclusive (i.e. each excludes
the other).
The complete addition rule is
P(A or B) = P(A) + P(B) if A and B are mutually exclusive (i.e. they dont overlap).
It can easily be extended to three or more events:
P(A or B or C or ) = P(A) + P(B) + P(C) + if A, B, C are mutually exclusive.
To explain the multiplication rule we need to do more than one thing, so lets throw the dice and
draw a card from the pack. How can we work out the probability of getting a 6 and a spade?
This is a little more complicated than the addition rule. It helps to imagine doing the experiment lots
of timessay 1000 times.
Introduction to reliability (Portsmouth Business School, April 2012)

Obviously you will get a 6 on about 1/6 of these thousand timesi.e. about 167 times. Now think
about how many times you will get a spade as well. This will happen on about of these 167 times,
or about 42 times out of the 1000 times we imagined doing the experiment. So the probability is
about 42/1000.
Now do the same thought experiment, but working in terms of the probabilities this time. We will
get a 6 on about one sixth of the times, and a spade as well on about one quarter of this one sixth.
One quarter of one sixth means the same as one quarter times one sixth, so we simply multiply the
probabilities:
P(6 and spade) = P(6)*P(spade) = 1/6 * 1/4 = 1/24
which is, of course, about 42/1000 as before.
Just like the addition rule there is an important assumption here. We assumed that the two events
are statistically independent. This means that the two probabilities are independent of each other:
knowing whether one has happened is of no help in assessing the probability of the other
happening. In the example above, knowing that weve got a 6 is of no relevance to what will happen
with the cardsso these events are statistically independent.
But in some situations you need to be very careful about this assumption. Suppose you know that in
a particular place the probability of rain falling on a given day is 1/3. Can you say that
P(rain today and rain tomorrow) = 1/3 * 1/3 = 1/9 ??
In practice this is likely to be wrong because the two events are not likely to be independent. In
England certainly, if it rains today the probability of rain tomorrow is likely to be higher than if it did
not rain today because the weather tends to set in to wet or dry spells. This means that the second
probability should probably be rather more than 1/3, so the result is likely to be substantially more
than 1/9.
The multiplication rule, then, also has an important condition:
P(A and B) = P(A)*P(B) if A and B are statistically independent
And, just as before, it can be extended:
P(A and B and C and ) = P(A)*P(B)*P(C)* if A, B, C, are statistically independent

Analyzing and improving the reliability of Systems

It is often useful to be able estimate the reliability of a whole system from the reliability of the
individual components. (The system in question might be a satellite to be sent into space, or a
computer, or a process for responding to accidents.) This enables designers to predict reliability
levels without having to build and test the whole system. It also enables them to see weak spots in
Introduction to reliability (Portsmouth Business School, April 2012)

the system, and to experiment with ways of improving these - perhaps by installing backups for
critical components.
There are two established ways of analyzing the reliability of a system. The first uses reliability block
diagrams to model the reliability of individual components and to predict the reliability of the whole
system. The second uses fault tree analysis to model the probabilities of failure of individual
components and so predict the probability of the whole system failing. Since the reliability of the
system is 100% minus the probability of failure (as a percentage), the two approaches are equivalent
so there is no point in doing both. Here we will look at fault tree analysis. (Reliability block diagrams
do the same thing in a slightly different way but come to the same conclusions.)

Fault Tree Analysis

This is a commonly used technique in industry to evaluate reliability of a system. It was developed in
the early 1960s to evaluate reliability of the Minuteman Launch Control System. Since then it has
gained favour especially when analysing complex systems. Fault tree analysis begins by identifying
the top event, known as the undesirable event of the system. The undesirable event of the system is
caused by events generated and connected by logic gates such as AND, OR, etc (as you will see
below). The following basic steps are involved in performing fault tree analysis:
i

Establishing system definition.

ii

Constructing the fault tree.

iii

Evaluating the fault tree qualitatively.

iv

Collecting basic data such as components' failure rates, repair rates, and failure occurrence
probability.

Evaluating fault tree quantitatively.

vi

Recommending corrective measures.

This method is frequently used as a qualitative evaluation method in order to assist the designer,
planner or operator in deciding how a system may fail and what remedies may be used to overcome
the causes of failure. The method can also be used for quantitative evaluation, in which case the
causes of system failure are gradually broken down into an increasing number of hierarchical levels
until a level is reached at which reliability data is sufficient or precise enough for a quantitative
assessment to be made. The appropriate data is then inserted into the tree at this hierarchical level
and combined together using the logic of the tree to give the reliability assessment of the complete
system being studied.
In order to illustrate the application of this method, consider the electric power requirements of the
system in the following example. In this example, the failure event being considered is loss of the
electric power. In practice the electric power requirements are both a.c. power, to supply energy for
Introduction to reliability (Portsmouth Business School, April 2012)

10

prime movers, and d.c. power, to operate relays and contractors, both of which are required to
ensure the successful operation of the electric power. Consequently the event 'loss of electric
power' can be divided into two sub-events 'loss of a.c. power' and 'loss of d.c. power'. This is shown
in the following figure with the events being joined by an OR gate as failure of either, or both, causes
the system to fail.
A very simple fault tree

If this subdivision is insufficient, sub-events can be divided further. The event 'loss of a.c. power' may
be caused by 'loss of offsite power' (the grid supply) and by 'loss of onsite power' (standby
generators or similar devices). These are joined by an AND gate as they both have to fail for the a.c.
power to fail. This process can be continued downwards to any required level of subdivision. After
developing a fault tree, it is necessary to evaluate the probability of occurrence of the upper event
by combining component probabilities using basic rules of probability and the logic defined in the
fault tree.
In the present example, suppose the probabilities of the events at the bottom of the tree are:
Prob (Loss of offsite power)

= 0.067

Prob (Loss of onsite power)

= 0.075

Prob (Loss of dc power)

= 0.005

These probabilities for the faults at the bottom of the tree can now be combined using the addition
and multiplication rules of probability. For the AND gate at the bottom we multiply the probabilities
to work out
Prob (Loss of a.c. power)

= Prob(Loss of offsite power) x Prob (Loss of onsite power)

Introduction to reliability (Portsmouth Business School, April 2012)

11

= 0.067 x 0.075 = 0.005025.

For the OR gate we add the probabilities to get the probability of the top event:
Prob (Loss of electric power) = Prob (Loss of a.c. power) + Prob (Loss of d..c power)
= 0.005025 + 0.005 = 0.010025.
This means that the probability of the top, undesirable, event Loss of electric power is about 1%.
The calculation make two assumptions.
First, to multiply the two probabilities at the bottom of the tree we must assume they are
statistically independent. This is reasonable if the onsite and offsite systems are separate so that the
failure of one is independent of the other. Notice that the combined probability (Loss of a.c. power)
is much less than the probability of either of the two probabilities in the AND gate. The principle
here is that of reducing the risk of a fault by using a backup system: this will not work, of course, if
the two systems are dependent so that if one fails there is an increased chance that other will fail
too.
The second assumption is the assumption when we add the probabilities for the OR gate that the
events are mutually exclusive (i.e. they dont overlap). With small probabilities, like we have here,
this is reasonable because the probability of more than one fault occurring is small.
The case study and exercises below contain further examples of fault trees.

Case study: healthy laptop

This case study explores some of the concepts of reliability. It concerns a private provider of health
care that runs a healthcare system with 8 acute care hospitals, 100 clinics, and home healthcare.
The home healthcare (HHC) division provides nurses (i.e. district nurses) to home care patients. It
employs 125 nurses and each one is provided with a laptop that allows them to maintain patient
records, point of care documentation, case loads, and payroll entry. The nurses maintain a remote
connection with the central server to access new cases and upload any historical patient data gained
throughout the day. The availability of the system, quality of the data and timeliness of transcribing
information are critical to the nurses.
The nurses have been complaining that they are having a lot of problems with the laptops and the
connection to the network and server. Reported laptop performance is very poor and 60% of the
nurses have complained of problems of one sort or another. All the laptops are from the same
manufacturer and have a similar configuration. Whenever a failure occurs the nurse has to return
the laptop to the office, obtain a loan machine and then return that to the office when their own
Introduction to reliability (Portsmouth Business School, April 2012)

12

laptop had been repaired. The company knows how many machines have been returned to the
manufacturer for repair but did not itself keep any records of failure symptoms and causes that is
done by the vendors field service group who made the repairs.
Discussions between HHCs IT Director and the manufacturers Field Support Manager are getting
progressively more heated with HHC threatening to switch suppliers completely with serious
consequences for the manufacturer who actually supplied all HHCs IT hardware. Losing the laptop
contract is likely could result in the entire customer being lost.
In order to try and bring the relationship back onto a more business like footing HHC decided to
bring in an independent reliability specialist to arbitrate.
The manufacturer did not, and would not, publish reliability data for its products, saying that
We do not quote MTBF (Mean Time Before Failure) numbers for our products or believe this
type of information should be used as a meaningful description of the quality of our systems or
component parts within those systems. MTBF is an older industry term that today has very little
value and is mostly misinterpreted and misunderstood. MTBF is the point at which 63.2% of the
population (all components of the given type built by a manufacturer), will fail. So for example,
disk drives that claim 1,000,000 hour MTBF with 720 power on hours per month would take 115
years to reach the 63.2% mark. It does NOT mean that no failure should occur for 115 years. This
is a total population statement from the manufacturers point of view not from the customer
point of view. As a Manufacturer, we know that some of that product will fail but we have no
idea which ones, and we do not know the distribution of good or bad product shipped to any
that may fail earlier than expected and therefore may achieve better or worse results and still
meet the MTBF criteria.
The manufacturers were willing to show the defect data for HHCs laptops, but once again would not
compare those results with a control sample of the same number of the same product. This laptop
failure data (a Pareto diagram) was produced at the next meeting, as in the diagram below.

Introduction to reliability (Portsmouth Business School, April 2012)

13

Laptop Failure causes

25
20
15
10
5
0

The average age of the machines was three years. This prompted some discussion that they were
therefore fully depreciated and should be written off and replacements bought. The consultant
thought that this was an accountants view of the situation and was not a view that could be
supported from a quality and reliability perspective. At three years the machines should be at the
most reliable part of the reliability (bathtub) curve well past early life failure and not yet
approaching the wear-out part of the curve when failures start to increase. The high number of
problems due to damaged covers, hard drives and CD failures made him suspect that part of the
problem lay in the environment in which the laptops were used and the treatment that they
received. If this were the case then replacing them with new machines would not solve the problem
as the new ones would ultimately suffer the same fate.
The diagram below shows a fault tree for one of these problem categories hard drive failures. (In
order to make it legible it has not been fully developedseek and logic failures have not been
explored.)

Introduction to reliability (Portsmouth Business School, April 2012)

14

Exercises
8

In relation to the Healthy Laptop case study:

(a)

Comment on the manufacturers refusal to publish MTBF data is their position

reasonable?
(b)

Where does the 63.2% come from?

(c)

How can the fault tree be used to analyse and help prevent failures? Can it show

how misuse can contribute to the failure of the component under review?

Two football teams are due to play each other three times in the next few weeks. The data

from their matches in the past suggests that the probability of Team A winning is 50%, and the
probability of Team B winning is 30%
a) What is the probability of a draw in the first match?
b) Assuming that the results of the three matches are statistically independent what is the
probability of Team B winning all three matches?
c) Assuming that the results of the three matches are statistically independent what is the
probability of Team B winning at least one of the matches?
d) Do you think the assumption of statistical independence is justified?

Introduction to reliability (Portsmouth Business School, April 2012)

15

10

During World War II, the failure rate for bomber aircraft flying over Europe was believed to
be a 4% chance of non-return from each mission, however experienced the pilot was.
Calculate the probability that a crew member will survive 25 missions.

11

The diagram below is a cross-section diagram of a booster rocket outer shell at a joint
between two stages. The system will fail only if both o-rings fail.

Fuel
Atmosphere

O-rings
Vent

(a)

Suppose the reliability of each o-ring is 0.95 during the most critical phase of flight.
What is the system reliability? (You will need to work out the probability of failure
for each O-ring.)

(b)
Do you think the two reliabilities are likely to be statistically independent? Would this
make any difference to your answer for the system reliability?
12

A certain electronic component has an exponential failure time with a mean of 50 hours.
(a)

What is the rate of this component?

(b)

What is the reliability of this component at 100 hours?

(c)

13

What is the minimum number of these components that should be placed in parallel
if we desire a reliability of 0.90 at 100 hours? (The idea of placing extra components in
parallel is to provide a backup if the first component fails.)

A risk assessment of a proposed new railway included the following results:

Passenger train derailment would occur if
derailment occurred due to over-speeding (4.3 x 10-11) or
derailment occurred due to rail faults (6.4 x10-9) or
derailment occurred due to rolling stock faults (3.9 x10-9) or
derailment occurred due to running into obstructions (7.1 x 10-9)

The figures in brackets represent the estimated probability of occurrence of each of these per train
Introduction to reliability (Portsmouth Business School, April 2012)

16

kilometre. For example the first probability means that the probability for the event in question is
4.3 x 10-11 for each train for each kilometre. Clearly the probability for six trains, each travelling 100
km, would be 600 times as great.
The probabilities were estimated by breaking the events down into more detail and using historical
data. For example, derailment occurred due to rolling stock faults was broken down into seven
events which would lead to rolling stock faults - these included wheel failure and suspension system
failure. Furthermore, suspension system failure would occur if
the suspension system deflated (1.5 x 10-6), and
the emergency springs failed (1.0 x 10-4)
The assigned task is:
(a)

Draw a fault tree from the above information.

(b)

Estimate the probability (per train kilometre) of suspension system failure and explain the
assumptions on which your estimate is based.

(c)

Estimate the probability (per train kilometre) of passenger train derailment (the top event of
the fault tree) and explain the assumptions on which your answer is based.

(d)

Assuming that the railway is 100 km long and there are 20 trains each way every day,
estimate the mean time between passenger train derailments. Do you think this represents
a satisfactory level of risk?

(e)

What do you think of the accuracy and usefulness of this type of analysis?

(The events and probabilities are taken from a presentation by C. Leighton & C. Dennis at a
conference on Risk Analysis and Assessment, Edinburgh, 1994.)

Notes on answers to exercises

1

The bath tub pattern would apply to a human being. Failure rates, in the sense of death, are
higher in early infancy and old age. And I think it might apply to a car. Old cars are probably
more liable to failure than new ones, and there is also a possibility that when a car is new
failure rates may be higher as faults left over from the manufacturing process are found and
remedied. On the other hand, if the manufacturer has checked carefully for the early faults
the infant mortality stage may not occur.

My answers are below.

Time Survivors Failures in
next 100 h
0
1000
160
100 840
86
200 754
78
300 676
70
400 606
64

Reliability Failure rate

(per hr)
100.00% 0.16%
84.0%
0.10%
75.4%
0.10%
67.6%
0.10%
60.6%
0.11%

Introduction to reliability (Portsmouth Business School, April 2012)

17

500
600
700
800
900
1000

542
484
432
389
347
311

58
52
43
42
36

54.2%
48.4%
43.2%
38.9%
34.7%
31.1%

0.11%
0.11%
0.10%
0.11%
0.10%

Failure rate (per hr)

0.20%
0.15%
0.10%
0.05%
0.00%
0

200

400

600

800

1000

Time

Note that the failure rate is per hour. (This is why the figures are so small.) A failure rate of
0.10% means that the probability of a failure in a given hour is 0.10%.
Note also that I am using data from the whole interval from 0-100 hours to calculate the
failure rate at 0 hours. This is an approximation but it is all we can do. In the table I have
assumed that the first failure rate (0.16%) applies to 0 hours but I might equally well have
said that it applies to 50 hours.
This result shows that the failure rate is initially 0.16% per hour, and then drops to around
0.10% to 0.11%. There is an infant mortality phase, a useful life phase with a fairly constant
failure rate, but no wear out phase. However, there are still 311 components left after the
1000 hour test - the wear out phase may become apparent if the test were to be prolonged.
Obviously, the best way to show this pattern would be to draw a graph of failure rate against
time.

Introduction to reliability (Portsmouth Business School, April 2012)

18

Reliability
120.00%
100.00%
80.00%
60.00%
40.00%
20.00%
0.00%
0

200

400

600

800

1000

1200

Time

This graph is very similar to the exponential distribution except that it is a bit steeper for the
first 100 hours because of the higher failure rate.
3

The compressor has a constant failure rate and therefore the reliability follows the
exponential distribution:
R = e - t
Failure rate = 0.000001 f/hr., operational time t = 2500 hours
Reliability = e-(0.000001 * 2500) = 0.9975 or 99.75%

The reliability of the product is given to be 0.98. The reliability of an exponential distribution
is given by:
R = e t, so
0.98 = e(- * 5000)
Taking natural logarithms on both sides (see Appendix), we get,
-0.02020 = - * 5000
Therefore = 4.04 * 10-6 = 0.00000404 f/hr
(Alternatively, you could use a trial and error process with your calculator to find a negative
number which gives you 0.98 when you press the ex button. This number must be equal to
*5000, so you can divide the number by 5000 to find .)

(a)
Since the failure rate is constant, we will use the exponential distribution. Also, the
MTBF = 25 hours. We know, for an exponential distribution, MTBF = 1/.
Therefore = 1/25 = 0.04
The reliability function is given by: R(t) = e-t = e- (0.04 * t)

Introduction to reliability (Portsmouth Business School, April 2012)

19

(b)
6

The reliability of the item at 30 hours = e-0.04 * 30 = 0.3012

Assuming the exponential model, the failure rate is

1/MTBF = 0.001
So

R(500) = e-500*0.001 = e-0.5 = 0.61, so the probability of the equipment operating for 500
hours without failure is 61%.
7
The MTTF is 2312 hours (simply the average). The evidence suggests that the ovens do not
follow the exponential distribution, because the failure rate is obviously not constant: there are no
failures between in the first 1000 hours, 2 failures between 1000 and 2000 hours, and 8 between
2000 and 3000 hours. This gives an estimated failure rate of 0 for the first 1000 hours, 20%/1000 for
the next 1000 hours, and 100%/1000 for the next 1000 hours this certainly increases over time)
8
In these situations it is always very frustrating for customers and user groups to be told that
they will not be given Mean Time Between Failure (MTBF) and other failure rate data. From the
manufacturers perspective this is a sensible approach because there is always a danger that the
MTBF data will be misinterpreted by the customer. MTBF is a measure applied to the population as a
whole, not to a very small sample.
It is true that caution must exercised when applying averages to individual cases. However, provided
that the customer understands that the MTBF is an average figure, the MTBF would still be useful,
particularly for a customer such as HHC which buys large numbers of laptops. In this situation, a
significant departure from the average figures would indicate either a problem with the laptops
supplied, or that they are being used in unsuitable conditions.
The 63.2% comes from the exponential reliability distribution. If the MTBF is 1,000,000 hours, the
failure rate is 0.000001 per hour and the exponential distribution gives
R = e 0.000001 x 1000000 = e 1 = 0.368 = 36.8%
which means that the other 63.2% must have failed. The quotation from the manufacturer is
misleading in that it is not a definition of the MTBF, but rather a fact that follows from the
exponential distribution.
The fault tree is a good illustration of how these can help in the analysis and prevention of failures. It
shows how external factors such as mechanical shock and dirt can damage components. Other fault
trees should be constructed for the other sub-systems contained in the laptop schematic diagram
and then combined into a fault tree for the entire laptop system. The first stage of the analysis of the
fault tree is a qualitative evaluation of the subsystem to understand how failure occurs. The next
stage is to gather data on component failure rates and repair rates to allow a quantitative evaluation
where are the major sources of failure are, and the actions that can be taken to minimise or
eliminate them.
9a
The match must end in a win for Team A, or a win for Team B, or a draw. These three
probabilities are mutually exclusive so they must add up to 100%. This means the probability of a
Introduction to reliability (Portsmouth Business School, April 2012)

20

draw is 20% because 50%+30%+20%=100%.

9b

P(B winning three matches)=P(B wins 1st match)*P(B wins 2nd match)*P(B wins 3rd)
=3025% * 30% * 30% = 27/1000 = 0.027 = 2.7%

so it is unlikely!
9c

This is a little more difficult. The easiest approach to at least one questions is to work out

the probability of the event never happening, and then subtract it from 1. In this case, the
probability of Team B failing to win each match is 70% (100%-30%), so the probability of failing to
win all three matches is
70% * 70% * 70% = 7/10 * 7/10 *7/10 = 343/1000 = 34.3%
In all other circumstances Team B will win at least one match so the probability is
P(B wins at least one match) = 100% - 34.3% = 65.7%
Which is fairly likely!
9d

This is a difficult question to which there is no easy answer. Perhaps if Team B wins the first

match they would become more confident and so more likely to win the next match? Or perhaps
Team A would get more determined.
10

The probability of surviving 10 missions is (1-0.04)25 = 0.9625 = 0.36 = 36%.

11

Assuming statistical independence we the probability of one failing is

0.05x0.05 = 0.0025

So the system reliability is 1-0.0025 = 0.9975 = 99.75% .

b
This answer assumes the two potential failures are statistically independent - which
may be unlikely in practice. A single factor - such as temperature - may be responsible for
failures in both. This means that, if we know one o-ring has failed, our estimate of the
probability of the other failing will be higher. In the extreme case, one o-ring fails whenever
the other does, and the reliability of the system is 0.95. In practice, the reliability of the whole
system is likely to be between 0.95 and 0.9975 (the answer based on the assumption of
independence).
11

1/50 = 2% per hour

R(100) = e-0.02x100 = 0.1353, which is not very good!

c The parallel system will only fail if all components fail. The probability of each failing is 1-0.1353
= 0.8647.
If there are n in parallel we need

Introduction to reliability (Portsmouth Business School, April 2012)

21

1 - 0.8647n = 0.9, or
0.8647n = 0.1
By trial and error (or see the Appendix), n = 16, so we need 16 components in parallel.
13 (a) The fault tree is on the next page.
(b)

Assuming the probabilities of the two events (suspension system deflated and emergency
springs failed) are independent, we can multiply the probabilities, so the estimate for the
probability of suspension system failure is 1.5 x 10-10. In practice, this assumption of
independence may not be fully realistic if some causes of the first event also make the second
more likely. If the events are not independent, the probabilities may be much higher.

(c)

To work out the probability of passenger train derailment we add the four probabilities to give
1.7 x 10-8. Note that the probability of the first event - over-speeding - is negligible compared
with the others.
This assumes that these are the only faults that will lead to derailment. Sabotage, for example,
does not seem to be included. (Strictly, adding probabilities presupposes the events are
mutually exclusive. It is possible that more than one event can occur. However, the
probability of this is so small - of the order of 10-18 - that it can reasonably be ignored.)

(d)

The probability of derailment per day is 100 x 20 x 2 x 1.7 x 10-8 = 6.8 x 10-5 This is, in effect,
a failure rate, so the mean time between derailments (failures) is 1/(6.8 x 10-5) or 1.5 x 104 or
15000 days or about 41 years. Derailments can be expected, on average, every 41 years.

(e)

The accuracy and usefulness of the model are obviously dependent on the data on which it is
built. In particular, it is obviously important that all input probabilities are accurate (e.g. of the
emergency springs failing), that assumptions about statistical independence are carefully
checked, and that all lists of events specifying the possible ways a fault can occur are as
complete as possible. This obviously requires systematic research.

Introduction to reliability (Portsmouth Business School, April 2012)

22

Derailment
OR

Over-speeding

Rail fault

Rolling stock
fault

Obstruction

OR

Wheel failure

Suspension
failure

AND

Suspension
deflated

Emergency springs
failed

References
Besterfield, D. H. (2013). Quality improvement. Boston: Pearson.
O'Connor, P. D. (2002). Practical reliability engineering, 4th edition. Chichester: Wiley.

Mathematical Appendix
Powers and roots
As you will doubtless know
pppp (four ps multiplied together) can be written as p4
You may also need to do this backwards. Suppose that
p4 = 0.7 (p4 is p^4 on a spreadsheet.)
Introduction to reliability (Portsmouth Business School, April 2012)

23

and you want to know p. p is the fourth root of 0.7 or

p = 0.71/4 = 0.70.25 = 0.915
To check, try 0.9154. This should come to 0.7 (approximately).
Powers are also defined for negative and any other fractional index. Experiment with your calculator
or spreadsheet.

Exponential functions and natural logarithms

ex is a function which arises from the mathematics of constant rates of growth. You should have a
button for it on your calculator, and on a spreadsheet the function will be EXP(X) or something
similar. Some examples (all rounded to two decimal places):
e1 = 2.72
e3 = 20.09
e-1.6 = 0.20
The inverse function to ex is the natural logarithm of x, loge(x) or ln(x). This is useful if you want to
know what x is if, for example,
ex = 0.5
You can find the answer by using natural logarithms:
x = loge(0.5) = -0.69315
This can be checked by working out e-0.69315. It should be very close to 0.5.

Introduction to reliability (Portsmouth Business School, April 2012)

24