NOTES ON

RELIABILITY
THEORY

BY TANUJIT CHAKRABORTY, RESEARCH SCHOLAR, ISI KOLKATA

Content Page No

Introduction to Reliability 1
Hazard Function 8
Bath Tub Curve 15
Constant Failure Rate Model 22
Time-Dependent Failure Models 24
System Reliability 40
Fault Tree Analysis 55
Life Tests 59
FMEA 71
Sequential Life Testing 73
Reliability, Availability, Serviceability 79
Warranty Analysis 86
Stress Strength Models 89
Accelerated Life Testing 97
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
Scanned by CamScanner
 89

Stress Strength Models

INTRODUCTION

The purpose of studying Stress-Strength Models is to determine the probability that a

component, a subsystem, or a system fails when the stress, in general, exceeds the strength.

In order to compute the reliability we have to know the nature of stress (S) and strength (T)
random variables. Our focus of this session is to show how to compute reliability of a component
when the density functions for the stress and the strength are known.

Stress-strength analysis is highly useful in mechanical component design.

Our objective is to find the reliability of a component when the density functions for the stress
and strength random variables are known.

We can find expression for reliability when stress & strength following different distributions
such as Normal, Exponential, Lognormal, Gamma, Weibull distribution.

Examples of stress-related failures include the following:

1. Misalignment of a journal bearing, lack of lubricants, or incorrect lubricants generate an

internal load (mechanical or thermal stress) that causes the bearing to fail.

2. The voltage applied to transistor gate is too high, causing a high temperature that melts
the transistor’s semiconductor material.

3. Cavitation causes pump failure, which in turn causes a violent vibration that ultimately
breaks the rotor.

4. Lack of heat removal from a feed pump in a power plant results in overheating of the
pump seals, causing the seals to break.

General Expression for Reliability

Let the density function for the stress (S) be denoted by , and that for strength (T) by .
Then by definition,

 The probability that the strength T is greater than certain stress is given by
 90

 Now the reliability of the component is the probability that the strength T is greater than
the stress S for all possible values of the stress S and is given by
∫ *∫ +
 Reliability can also be computed on the basis that the stress remains less than the strength
and the probability of the stress being less than is given by
∫
 Hence the reliability of the component for all the possible values of the strength T is
∫ *∫ +
 Some other expressions of unreliability. Let unreliability be denoted by ̅ . ̅
.

Substituting for R from equation (2) we have

̅ ∫ *∫ +

∫ [ ]

∫ (4)

Alternatively using equation (3) we have

̅ ∫ [
 Define . Y is called the interference random variable. We can define reliability
as

Assuming T and S are independent random variables and greater than equal to zero the
density of y is given by

∫
{

Hence the probability of failure is given by

̅ ∫ ∫ ∫
 91

and the reliability by

∫ ∫ ∫

Reliability Computation for Normally Distributed Strength and Stress

The probability density function for a normally distributed stress S is given by

* ( ) +
√
The probability density function for a normally distributed stress S is given by
[ ]
√

where

Let us define Y=T-S. It is well known that the random variable Y is normally distributed with a
mean of

and a standard deviation of

√
The reliability R can now be expressed in terms of Y as

∫ [ ( ) ]
√
If we let then

When , the lower limit of is given by

and when , the upper limit of . Therefore,

⁄
∫
√
√

Clearly the random variable is the standard normal variable. Clearly reliability
can be found by merely referring to the normal tables.
 92

Equation (12) can be rewritten as

( )
√

Example 1
An automotive component has been designed to withstand certain stresses It is known from
the past experience that, because of variation in loading, the stress on the component is
normally distributed with a mean of 30,000 kPa and a standard deviation of 3000 kPa. The
strength of the component is also random because of variations in the material
characteristics and the dimensional tolerances. It has been found that the strength is
normally distributed with a mean of 40,000 kPa and a standard deviation of 4000 kpa.
Determine the reliability of the component.

Sol: We are given that

Then the lower limit of the integral for R is given by

√
and hence from the normal tables R=0.977

Example 2
The stress developed in an engine component is known to be normally distributed with a
mean of 350.00 Mpa and a standard deviation of 40.00 Mpa. The material strength
distribution, based on the expected temperature range and various other factors, is known
to be normal with a mean of 820.00 Mpa and a standard deviation of 80.00 Mpa.

Sol: Conventional factor of safety, defined as the ratio of mean strength to mean stress, is given
by

To compute the reliability of the component we use the coupling equation:

√ √
Hence the reliability of the component is 0.9999999.

Now, suppose that poor heat treatment and larger variations in the environmental temperatures
cause the standard deviation for the strength of the component to increase to 150.00 Mpa. In that
case the factor of safety as defined before remains unchanged, but the reliability is altered. Using
the coupling equation,

√ √
 93

and the reliability of the component is found to be 0.99877. Thus we witness a downgrading of
reliability resulting from an increased variability in the strength of the component.

Example 3
A new component is to be designed. A stress analysis revealed that the component is
subjected to a tensile stress. But there are variations in the load and the tensile stress is
found to be normally distributed with a mean of 35,000 psi and a standard deviation of
4000 psi. The manufacturing operations create a residual compressive stress that is
normally distributed with a mean of 10,000 psi and a standard deviation of 1500 psi. A
strength analysis of the component showed that the mean value of the significant strength is
50,000 psi. The variations introduced by various by various strength factors are not clear at
the present time. The engineer wants to know the maximum value of the standard
deviation for the strength that will ensure that the component reliability does not drop
below 0.999.

Sol: We are given that

where is the tensile stress and is the residual compressive stress.

The mean effective stress ̅ is obtained by
̅ ̅ ̅
and its standard deviation by
√
√

From the normal tables, we find the value of z associated with a reliability of 0.999 to be .
Substituting in the coupling equation yields

√
Solving for we get
 94

Reliability Computation for Log Normally Distributed Strength and Stress

The standard form of a log normal density function is

[ ]
√

Where Y is the random variable. The parameters and are the mean and the standard
deviation, respectively, of the variable , which is normally distributed. First we develop those
relationships for the log normal distribution that are needed later in the analysis.

Let . Then ( ) From equation (14) we have

[ ]
√

and hence,

and

Now considering the exponent of in the expression

∫ { ( )( ) }
√
We have
( )

[ ]
Therefore
{ }
( )∫ * +
√
( )
To compute the variance of Y we observe that
∫ [ ]
√

Considering the exponent of in the expression for , we have

 95

[ ]

[ ]

which, when substituted back and simplified as before yields,

[ ]

Hence by the definition of variance we may write

[ ] , * +-
[ ][ ]

We now observe that

[ ]

which, after rearranging, leads to

[ ]
[ ]
 96

Example:
The strength T and stress S are log normally distributed for a component with the
following parameters:
E(T) = 100,000 kPa Standard Deviation of T = 10,000 kPa
E(S) = 60,000 kPa Standard Deviation of T = 20,000 kPa
Compute the Reliability of the component.
 97

Accelerated Life Testing

INTRODUCTION
The development of new products in a short time has motivated the development of new
methods such as robust design, just−in−time manufacturing, and design for manufacturing and
assembly.
More importantly, both producers and customers expect that the product will perform the
intended functions for extended periods of time. Hence, extended warranties and similar
assurances of product reliability have become standard features of the product.
These requirements have increased the need for providing more accurate estimates of
reliability by performing testing of materials, components, and systems at different stages of
product development. Testing under normal operating conditions requires a very long time
possibly years and the use of an extensive number of units under test, so it is usually costly and
impractical to perform reliability testing under normal conditions.
This has led to the development of accelerated life testing (ALT), where units are subjected
to a more severe environment (increased or decreased stress levels) than the normal operating
environment so that failures can be induced in a short period of test time.
Information obtained under accelerated conditions is then used in conjunction with a
reliability prediction (inference) model to relate life to stress and to estimate the characteristics
of life distributions at design conditions (normal operating conditions).
Conducting an accelerated life test requires careful allocation of test units to different stress
levels so that accurate estimation of reliability at normal conditions can be obtained using
relatively small units and short test durations.

Design of Accelerated Life Testing Plans

A detailed test plan is usually designed before conducting an accelerated life test. The plan
requires determination of the type of stress, methods of applying stress, stress levels, the
number of units to be tested at each stress level, and an applicable accelerated life testing
model that relates the failure times at accelerated conditions to those at normal conditions.

Stress loadings
Stress in ALT can be applied in various ways. Typical loadings include constant, cyclic, step,
progressive, random stress loading, and combinations of such loadings.
Typical accelerated testing plans allocate equal units to the test stresses. However, units
tested at stress levels close to the design or operating conditions may not experience enough
failures that can be effectively used in the acceleration models.
Therefore, it is preferred to allocate more test units to the low stress conditions than to the
high stress conditions so as to obtain an equal expected number of failures at both conditions.

Types of Stress
The type of applied stress depends on the intended operating conditions of the product
and the potential cause of failure. We classify the types of the stresses as follows:
1. Mechanical stresses: Fatigue stress is the most commonly used accelerated test for
mechanical components. When the components are subject to elevated temperature, then
 98

creep testing (which combines both temperature and load) should be applied. Shock and
vibration testing is suitable for components or products subject to such conditions as in the
case of bearings, shock absorbers, tires and circuit boards in airplanes and automobiles.
2. Electrical stresses: These include power cycling, electric field, current density, and
electromigration. Electric field is one of the most common electrical stresses, as it induces
failures in relatively short times; its effect is also significantly higher than other types of
stress.
3. Environmental stresses: Temperature and thermal cycling are commonly used for
most products. Of course, it is important to use appropriate stress levels that do not induce
different failure mechanisms than those under normal conditions. Humidity is as critical as
temperature, but its application usually requires a very long time before its effect is noticed.
Other environmental stresses include ultraviolet light, sulfur dioxide, salt and fine particles,
and alpha and gamma rays.

Accelerated Life Testing Models

Elsayed classified the inference procedures (or models) that relate life under stress
conditions to life under normal or operating conditions into three types:
• statistical−based models

• physics−statistics−based models

• physics−experimental−based models

The underlying assumption in relating the failure data, when using any of the models, is that
the components/products operating under normal conditions experience the same failure
mechanism as those occurring at the accelerated conditions.
The statistics−based models are further classified as parametric models and non−parametric
models. We are here going to discuss about Parametric-Statistics-based Models.

Statistics−based models are generally used when the exact relationship between the
applied stresses and the failure time of the component or product is difficult to determine
based on physics or chemistry principles. In this case, components are tested at different stress
levels and the failure times are then used to determine the most appropriate failure time
distribution and its parameters.
The most commonly used failure time distributions are the exponential, Weibull, normal,
lognormal, gamma, and the extreme value distributions. The failure times follow the same
general distributions for all different stress levels, including the normal operating conditions.
 99

Parametric Statistics−based Models

As stated above, statistics−based models are generally used when the exact relationship
between the applied stresses (temperature, humidity, voltage, etc.) and the failure time of the
component (or product) is difficult to determine based on physics or chemistry principles. In
this case, components are tested at different accelerated stress levels . The failure
times at each stress level are then used to determine the most appropriate failure time
probability distribution, along with its parameters. Under the parametric statistics−based model
assumptions, the failure times at different stress levels are linearly related to each other.
Moreover, the failure time distribution at stress level is expected to be the same at
different stress levels as well as under the normal operating conditions. In other words,
the shape parameters of the distributions are the same for all stress levels (including normal
conditions) but the scale parameters may be different.
 Failure times

Where is the failure time under operating conditions, is the failure time under
operating conditions, is the failure time under stress conditions, and is the acceleration
factor (the ratio between product life under normal conditions and life under accelerated
conditions);
 Cumulative distribution functions (CDFs)

( )
 Probability density functions

( ) ( )
 Failure rates

( ) ( )
The most widely used parametric models are the exponential and Weibull models.
Therefore, we derive the above equations for both models and demonstrate their use.

Acceleration Model for the Exponential Model

This is the case where the time to failure under stress conditions is exponentially distributed
with a constant failure rate . The CDF at stress s is

And the CDF under normal conditions is

( )

The failure rates are related as

 100

Table 22.1 Failure times of the capacitors in hours

Temperature Temperature Temperature
145 240 305
75 179 116
359 407 189
701 466 300
722 571 305
738 755 314
1015 768 403
1388 1006 433
2285 1094 440
3157 1104 468
3547 1493 609
3986 1494 634
4077 2877 640
5447 3001 644
5735 3160 699
5869 3283 781
6242 4654 813
7804 5259 860
8031 5925 1009
8292 6229 1176
8506 6462 1184
8584 6629 1245
11512 6855 2071
12370 6983 2189
16062 7387 2288
17790 7564 2637
19767 7783 2841
20145 10067 2910
21971 11846 2954
30438 13285 3111
42004 28762 4617
Mean 9287 Mean 5244 Mean 1295

Table 22.2 Temperatures and the 50th percentiles

Temperature ( ) 145 240 305

50th percentile 6437 3635 898
 101

Example1. In recent years, silicon carbide (SiC) is used as an optional material for
semiconductor devices, especially for those devices operating under high temperatures and
high electric fields conditions. An extensive accelerated life experiment is conducted by
subjecting 6H−SiC metal−oxide−silicon (MOS) capacitors to temperatures of 145, 240, and
305 . The failure times are recorder in Table 22.1. Determine the mean time to failure
(MTTF) of the capacitors at 25 and plot the reliability function.

Solution: The data for every temperature are fitted using an exponential distribution and
the means are shown in Table 22.1. In order to estimate the acceleration factor we chose some
percentile of the failed population, which can be done non−parametrically using the rank
distribution or a parametric model. In the example, the exponential distribution is used and the
time at which 50 of the population fails is

The 50th percentiles are given in Table 22.2.

We use the Arrhenius model to estimate the acceleration factor

Where t is the time at which a specified portion of the population fails, k and c are
constants and T is the absolute temperature (measured in degrees Kelvin). Therefore

Using the values in table 22.2 and least−squares regression we obtain

. Therefore, the estimated 50th percentile at 25 is

The acceleration factor at 25 is

And the failure rate under normal operating conditions is

failures/h, the mean time to failure is 30419h and the plot of the reliability function is
shown in Figure 22.3.
 102

Figure 22.3. Reliability function for the capacitors

Acceleration Model for the Weibull Model

Again, we consider the true linear acceleration case. Therefore, the relationships between the
failure time distributions at the accelerated and normal conditions can be derived using
Equations 22.2−22.4. Thus
( )

where is the shape parameter of the Weibull distribution under stress conditions and
is the scale parameter under stress conditions. The CDF under normal operating conditions is
[ ]
( )

[ ]

The underlying failure time distributions under both the accelerated stress and operating
conditions have the same shape parameters, i.e., . If the shape
parameters at different stress levels are significantly different, then either the assumption of
true linear acceleration is invalid or the Weibull distribution is inappropriate to use for analysis
of such data.

Let . Then the probability density function under normal operating

conditions is
 103

[ ]
( )

The MTTF under normal operating conditions is

( )
The failure rate under normal operating conditions is
( )

Table 22.3. Time (hours) to detect leak

100 psi 120 psi 140 psi
1557 1378 215
4331 2055 426
5725 2092 431
5759 2127 435
6207 2656 451
6529 2801 451
6767 3362 496
6930 3377 528
7146 3393 565
7277 3433 613
7346 3477 651
7668 3947 670
7826 4101 708
7885 4333 710
8095 4545 743
8468 4932 836
8871 5030 865
9652 5264 894
9989 5355 927
10471 5570 959
11458 5760 966
11728 5829 1067
12102 5968 1124
12256 6200 1139
12512 6783 1158
13429 6952 1198
13536 7329 1293
14160 7343 1376
14997 8440 1385
17606 9183 1780
 104

Table 22.4 Percentiles at different pressures

Pressure (psi) 100 120 140

50th percentile 9050 4681 821

Example2. A manufacturer of Bourdon tubes (used as a part of pressure sensors in

avionics) wishes to determine its MTTF. The manufacturer defines the failure as a leak in the
tube. The tubes are manufactured from 18 Ni (250) maraging steel and operate with dry
99.9 nitrogen or hydraulic fluid as the internal working agent. Tubes fail as a result of
hydrogen embrittlement arising from the pitting corrosion attack. Because of the criticality of
these tubes, the manufacturer decides to conduct ALT by subjecting them to different levels
of pressures and determining the time for a leak to occur. The units are continuously
examined using an ultrasound method for detecting leaks, indicating failure of the tube. Units
are subjected to three stress levels of gas pressures and the times for tubes to show leak are
recorded in Table 22.3.
Determine the mean lives and plot the reliability function for design pressures of 80 and
90 psi.

Solution: We fit the failure times to Weibull distributions, which results in the following
parameters for pressure levels of 100, 120, and 140 psi.
For 100 psi:
For 120 psi:
For 140 psi:
Since , then the Weibull model is appropriate to describe the
relationship between failure times under accelerated conditions and normal operating
conditions. Moreover, we have a true linear acceleration. Following Example 1, we determine
the time at which 50 of the population fails as
[ ]
th
The 50 percentiles are shown in Table 22.4.
The relationship between the failure time t and the applied pressure P can be assumed to
be similar to the Arrhenius model;
Thus

Where k and c are constants. By making a logarithmic transformation, the above expression
can be written as

Using a linear regression model, we obtain . The estimated 50th

percentiles at 80 psi and 90 psi are 84361 h and 27332 h respectively. The corresponding
acceleration factors are 9.32 and 3.02. The failure rates under normal operating conditions are
( )
 105

Or

And

The reliability functions are shown in Figure 22.4.

The MTTFs for 80 and 90 psi are calculated as
( ) ( )

And

Figure 22.4. Reliability functions at 80 and 90 psi

The Arrhenius Model

Elevated temperature is the most commonly used environmental stress for accelerated life
testing of microelectronic devices. The effect of temperature on the device is generally
modeled using the Arrhenius reaction rate equation given by
( )

where r is the speed of reaction. A is an unknown non−thermal constant, is the

activation energy (i.e., energy that a molecule must have before it can take part in the
reaction), k is the Boltzmann constant (8.623 ), and T(K) is the temperature.
 106

Activation energy is a factor that determines the slope of the reaction rate curve with
temperature, i.e., it describes the acceleration effect that temperature has on the rate of a
reaction and is expressed in electron volts (eV). For most applications, is treated as a slope
of a curve rather than a specific energy level. A low value of indicates a small slope or a
reaction that has a small dependence on temperature. On the other hand, a large value of
indicates a high degree of temperature dependence.
Assuming that device life is proportional to the inverse reaction rate of the process, then
Equation 22.14 can be rewritten as
( )

The median lives of the units at normal operating temperature and accelerated
temperature are related by

Or
( )
The thermal acceleration factor is
( )

The calculation of the median life (or percentile of failed units) is dependent on the failure
time distribution. When the sample size is small it becomes difficult to obtain accurate results.
In this case, it is advisable to use different percentiles of failures and obtain a weighted average
of the median lives. One of the drawbacks of this model is the inability to obtain a reliability
function that relates the failure times under stress conditions to failure times under normal
condition. We can only obtain a point estimate of life. We now illustrate the use of the
Arrhenius model in predicting median life under normal operating conditions.

Table 22.5. Failure time data (hours) for oxide breakdown

Temperature Temperature
180 150
112 162
260 188
298 288
327 350
379 392
487 681
593 969
658 1303
 107

701 1527
720 2526
734 3074
736 3652
775 3723
915 3781
974 4182
1123 4450
1157 4831
1227 4907
1293 6321
1335 6368
1472 7489
1529 8312
1545 13778
2029 14020
1568 18640

Example 3. The gate oxide in MOS devices is often a source of device failure, especially for
high−density device arrays that require thin gate oxides. The reliability of MOS devices on
bulk silicon and the gate oxide integrity of these devices have been the subject of
investigation over the years. A producer of MOS devices conducts an accelerated test to
determine the expected life at 30 . Two samples of 25 devices each are subjected to stress
levels of 150 and 180 . The oxide breakdown is determined when the potential across the
oxide reaches a threshold value. The times of breakdown are recorded in Table 22.5. the
activation energy of the device is 0.42 eV. Obtain the reliability function of these devices.

Solution:

Figure 22.5. Weibull probability plot

 108

Figure 22.5 shows that it is appropriate to fit the failure data using Weibull distributions
with shape parameters approximately equal to unity. This means that they can be represented
by exponential distributions with means of 1037 and 4787 h for the respective temperatures of
180 and 150 respectively. Therefore, we determine the 50th percentiles for these
temperatures using Equation 22.8 as being 719 and 3318 respectively.
( )

Which results in