BRYAN SCHWAB HARRY DODSON - ACCELERATED TESTING - A Practitioner's Guide To Accelerated and Reliability testing.-SAE SOC OF AUTOMOTIVE ENG (2021) PDF

Accelerated Testing:
A Practitioner’s Guide
to Accelerated and
Reliability Testing
Second Edition
Bryan L. Dodson | Harry Schwab

to Accelerated and
Reliability Testing
Second Edition
to Accelerated and
Reliability Testing
Second Edition
BRYAN L. DODSON AND HARRY SCHWAB
Warrendale, Pennsylvania, USA

400 Commonwealth Drive
Warrendale, PA 15096-0001 USA
E-mail: CustomerService@sae.org
Phone: 877-606-7323 (inside USA and
Canada)
724-776-4970 (outside USA)
FAX: 724-776-0790
Copyright © 2021 SAE International. All rights reserved. Chief Growth Officer
Frank Menchaca
No part of this publication may be reproduced, stored in a retrieval
system, or transmitted, in any form or by any means, electronic, mechan- Publisher
ical, photocopying, recording, or otherwise, without the prior written Sherry Dickinson Nigam
permission of SAE International. For permission and licensing requests,
contact SAE Permissions, 400 Commonwealth Drive, Warrendale, PA Development Editor
15096-0001 USA; e-mail: copyright@sae.org; phone: 724-772-4028. Publishers Solutions, LCC
Albany, NY
Library of Congress Catalog Number 2021942286
http://dx.doi.org/10.4271/9781468603514 Director of Content
Management
Information contained in this work has been obtained by SAE Kelli Zilko
International from sources believed to be reliable. However, neither SAE
International nor its authors guarantee the accuracy or completeness Production and
of any information published herein and neither SAE International nor Manufacturing Associate
its authors shall be responsible for any errors, omissions, or damages Erin Mendicino
arising out of use of this information. This work is published with the
understanding that SAE International and its authors are supplying
information but are not attempting to render engineering or other
professional services. If such services are required, the assistance of an
appropriate professional should be sought.
ISBN-Print 978-1-4686-0350-7
ISBN-PDF 978-1-4686-0351-4
ISBN-ePub 978-1-4686-0352-1
To purchase bulk quantities, please contact: SAE Customer Service
E-mail: CustomerService@sae.org
Phone: 877-606-7323 (inside USA and Canada)
724-776-4970 (outside USA)
Fax: 724-776-0790
Visit the SAE International Bookstore at books.sae.org
For access to free resources to accompany this book, please visit

sae.org/r518-templates
Contents
Preface xi
Acknowledgments xiii
CHAPTER 1
Introduction 1
Purpose of Accelerated Testing 1
Design Life 2
Statistical Sample Size Determination 5
Tolerances in Accelerated Testing 5
Financial Considerations 8
Response from Finance 11
Response from Engineering 12
Response from Manufaturing 12
Summary 16
CHAPTER 2
Probability Fundamentals 17
Sampling 17
Probability Density Function 19
Cumulative Distribution Function 24
Reliability Function 25
Hazard Function 26
Expectation 27
Summary 29
©2021 SAE International v

vi Contents
CHAPTER 3
Distributions 31
Continuous Modeling Distributions 31
Weibull Distribution 32
Normal Distribution 37
Lognormal Distribution 43
Exponential Distribution 47
Discrete Modeling Distributions 51
Poisson Distribution 51
Binomial Distribution 53
Hypergeometric Distribution 54
Geometric Distribution 57
Identifying the Correct Discrete Distribution 58
Sampling Distributions 59
Chi-Square Distribution 59
t-Distribution 60
F-Distribution 62
Bayes’ Theorem 64
Summary 67
CHAPTER 4
Parameter Estimation 69
Maximum Likelihood Estimation 70
Probability Plotting 70
Hazard Plotting 70
Hazard Plotting 74
Normal Distribution 79
Hazard Plotting 83
Lognormal Distribution 87
Contents vii
Hazard Plotting 90
Nonparametric Confidence Intervals 96
Summary 96
CHAPTER 5
Accelerated Test Plans 97

Mean Time to Fail 97
Test Plan Problems 98
Zero-Failure Testing 103
Bogey Testing 103
Bayesian Testing 103
Effect of the Shape Parameter 103
Estimating the Shape Parameter 106
Determining Test Parameters 106
Guidelines for Zero-Failure Test Planning 108
Sequential Testing 111
Pass-Fail Testing 111
Randomization of Load Cycles 122
Reliability Growth
124
Reliability Growth Process 125
Reliability Growth Models 126
Duane Model 126
AMSAA Model 128
Summary 130
CHAPTER 6
Accelerated Testing Models 131

Linear Acceleration 131
Field Correlation 132
Arrhenius Model 138
viii Contents
Eyring Model 145

Voltage Models 155
Mechanical Crack Growth 156
Degradation Testing 157
Qualitative Tests 162
Step-Stress Testing 162
Elephant Tests 163
HALT and HASS 164
Summary 165
CHAPTER 7
Environmental Stress Screening 167

Stress Screening Theory 167
Product Reliability Equation 168
What Is ESS? 170
Evolution of ESS 170
Misconceptions about ESS 172
Types of Environmental Stress 173
Temperature Cycling 173
Random Vibration 176
High-Temperature Burn-In 177
Electrical Stress 177
Thermal Shock 177
Sine Vibration, Fixed Frequency 178
Low-Temperature Screen 178
Sine Vibration, Swept Frequency 178
Combined Environment 178
Advantages of Temperature Cycling 178
Levels of Screen Complexity 181
Failure Analysis 182
Case Histories 182
Analogic Corporation 182
Bendix Corporation 183
Hewlett-Packard (HP) 183
Implementing an ESS Program 183
Equipment 186
Burn-In Optimization 187
Summary 189
Contents ix
CHAPTER 8
Test Equipment Methods and Applications 191

Simulation versus Stimulation 193
Simulation 193
Application 193
Typical Test Types 193
Defining the Real World 193
Stimulation 195
Application 195
Typical Test Types 196
Types of Stresses 196
Temperature 196
Steady State 196
Temperature Cycling 197
Thermal Shock 197
Vibration 198
Analysis Tools 198
Mechanical and Pneumatic 199
Loose Cargo Testing 199
Cyclic Testing 199
Electrodynamic 200
Hydraulic 200
Mechanical Shock 201
Electrical Stress 202
Combined Environments 202
Other Types of Stress 202
Summary 203
Appendix A: Statistical Tables 205
Appendix B: Government Documents 217
Appendix C: Glossary 219
Appendix D: List of Acronyms 225
Bibliography 227
About the Authors 229
Index 231
Preface
T
his book presents accelerated testing from a practical viewpoint. The material
is presented with the practitioner in mind; thus, proofs and derivations have
been omitted, and numerous examples have been included. In addition, most
examples have been worked in Microsoft Excel and are included on the accompa-
nying website. For those desiring proofs and derivations, references are provided.
Our goal is that practicing engineers will be able to apply the methods presented
after studying this text.
Practitioners will find this text valuable as a comprehensive reference book, but
this book is also ideal for use in college courses. In particular, it is recommended that
this text be used for one-semester college courses. Students should have a familiarity
with basic probability and statistics before attempting this material.
The text consists of eight chapters. Chapter 1 provides an introduction and
overview of the limitations of accelerated testing. Chapters 2, 3, and 4 describe the
fundamentals of statistical distributions, the most commonly used distributions in
accelerated testing, and parameter estimation methods. Chapter 5 describes test plans
for accelerated testing, including reliability growth. Chapter 6 explains models for
accelerated aging, along with qualitative methods of accelerated testing. Chapter 7
explains environmental stress screening (ESS), and Chapter 8 presents the equipment
and methods used in accelerated testing.
The website (sae.org/r518-templates) contains the following content:
•• Examples—The examples presented in the text are worked in Microsoft Excel
templates. These templates will be useful when applying the material to real-
world problems.
•• Statistical Tables—The statistical tables included in the appendices of books
are holdovers to times when computers were not available. These tables give
solutions to closed integrals of functions that could not be solved implicitly
and required numerical methods to solve. These functions are now included
in electronic spreadsheets. When the text references a value available in
Appendix A, use the Microsoft Excel templates available on the website.
•• Burn-In Optimization—This is a Microsoft Excel template for determining
the optimum burn-in duration based on the cost of burn-in time, burn-in
failures, and field failures.
©2021 SAE International xi

xii Preface
•• Random Number Generator—This Microsoft Excel template generates

random numbers that can be used for simulations. There are random number
generators for the Weibull, normal, lognormal, and exponential distributions.
•• Government Documents—This page contains many documents in PDF
format. There are numerous military standards and handbooks related to
reliability and accelerated testing.
•• AMSAA Reliability Growth Handbook—This page provides the AMSAA
Reliability Growth Handbook in Microsoft Word format.
Acknowledgments
W
e are grateful to many individuals for helping prepare this book. Most
notable are the reviewers. Several reviewers were anonymous, but
you know who you are, and we thank you for your comments. Special
thanks go to Lois Dodson and Matthew Dodson for creating the accompanying
website for the first edition. We greatly appreciate Thermotron for providing the ESS
material in Chapter 8 and the Quality Council of Indiana for allowing us to use some
previously published material in Chapters 2 and 3.
©2021 SAE International xiii

1
Introduction
T
he scientific theory of accelerated testing is highly developed, but the application
of this theory has proven difficult, especially in the mobility industries. The
required design life for many components exceeds 10 years, and the application
environment is harsh and highly variable. Vehicles must operate reliably in arctic
conditions and desert conditions. Driver profiles range from the 16-year-old male
to the 90-year-old female. An airliner may fly long-haul ocean routes for 20 years,
while an identical model may fly short-range routes that result in many more takeoffs
and landings over the life of the aircraft. Combining this variety into a realistic test
method that can be completed in a reasonable time frame with a reasonable budget
is difficult and requires compromises.
Purpose of Accelerated Testing

Ultimately, the only purpose of accelerated testing is cost reduction. The costs for
accelerated testing—components, monitoring equipment, labor, test equipment, and
so forth—must be recovered through lower warranty and the positive financial impact
of customer satisfaction.
Accelerated tests fall into two categories: (1) development tests and (2) quality
assurance tests. During the research, short inexpensive tests are needed to evaluate
and improve performance. The progress of a product in these development tests is
©2021 SAE International 1

2 Accelerated Testing
TABLE 1.1 Reliability demonstration sample sizes.
Reliability (%) Confidence (%) Sample size

99 95 299
99 90 229
99 50 69
95 95 59
95 90 45
95 80 31
90 90 22
90 80 16
often monitored statistically with a reliability growth program. Some quality assur-
ance tests are as follows:
•• Design verification
•• Production validation
•• Periodic requalification
Quality assurance tests are often tied to statistical sampling plans with require-
ments such as a demonstrated reliability of at least 95% at 10 years in service with a
confidence level of 90%. Statistically, a 95% reliability with 90% confidence can
be demonstrated by testing 45 units to the equivalent of 10 years in service. Table 1.1
gives the required sample sizes for some common reliability requirements.
Before proceeding with a test of 299, 45, or even 16 samples, the purpose of the
test should be investigated. What does it cost to test 299 units? The following costs
should be considered:
•• Prototype costs
•• Instrumentation costs (results monitoring)
•• Setup costs
•• Labor costs
•• Laboratory costs (many tests take two or more months to complete)
Using the sample sizes shown in Table 1.1 allows no failures. If a failure occurs,
do timing and budget constraints allow changes and a repeat of the test? What are
the implications of bringing a product to market if that product did not demonstrate
the required reliability with an appropriate level of confidence?
Design Life
Determining the design life to be simulated with an accelerated test can be difficult.
Many automobile manufacturers specify a design life of 10 years for brake systems
(excluding pads), but how do 10 years in service translate to an accelerated test?
Accelerated Testing 3
According to a study of brake system usage for minivans in 1990, the following state-
ments are true:
•• The median number of brakes applied was 322,000 for 10 years in service.
•• Five percent of the vehicles had more than 592,000 brakes applied in 10 years
of service.
•• One percent of the vehicles had more than 709,000 brakes applied in 10 years
of service.
•• The force of the brakes applied ranged from 0.1 to 1.0 g-force, with the shape
of the distribution shown in Figure 1.1.
How many times should the brake system be cycled in the accelerated test repre-
senting 10 years? Engineers design for the most stressful conditions; therefore, does
this mean that the number of cycles is determined by the most stressful driver?
User profiles are often defined by percentile. The 95th percentile point is the point
with 5% of the users having a more stressful profile. One percent of the users have a
driving profile that is more stressful than the 99th percentile driver. Table 1.2 gives
the number of brake applications as a function of the user percentile; these data also
are shown in Figure 1.2.
As shown in Figure 1.2, the number of brake applications increases dramatically
as the percent of the population covered nears 100%. This is typical of many other
applications, such as door slams, ignition cycles, and trunk release cycles. Increasing
the percent of the population covered from 75% to 99.9% requires an approximate
doubling of the number of cycles in the accelerated test. Not only does this increase
FIGURE 1.1 Distribution of brake force application.

TABLE 1.2 Percentiles for minivan brake applications.
Percentile Number of brake applications

50th 321,891
60th 361,586
70th 405,155
75th 429,673
80th 457,241
85th 489,671
90th 530,829
95th 592,344
97.5th 646,007
99th 708,571
99.9th 838,987
FIGURE 1.2 Percentiles for minivan brake applications.
the cost and duration of the test, but the cost of the component increases because the
number of cycles in the test is part of the design requirement.
The percent of the population covered is a compromise among the development
cost, development time, component cost, and field performance of the component.
For safety-critical items, the user percentile may exceed 100% to allow a safety margin.
For other items, such as glove box latches, the user percentile may be as low as 80%.
In reality, there is no 95th percentile user. There is a 95th percentile user for the
number of cycles, a 95th percentile user for temperature, a 95th percentile user for
the number of salt exposure, a 95th percentile user for vibration, and so forth.
However, determining the 95th percentile user for the combination of conditions
is unrealistic.
The worst-case user profile may not be at the high end for the number of cycles
of operation. Consider a parking brake. The worst-case scenario may be a brake that
is used for the first time after the vehicle is 10 years old. This type of user profile must
be incorporated into a test separate from a test utilizing the 95th percentile of parking
brake applications.
Accelerating a test by eliminating the time between cycles can introduce unre-
alistic conditions. Consider a durability test for an automobile door. The door is
opened and closed 38,000 times in 12 hours. Opening and closing the door this quickly
does not allow the door hinges or latches to cool, nor does it give any contaminants
that may be introduced in the hinges time to form corrosion. Consider an automobile
engine: the 95th percentile user profile for engine-on time is approximately 7,000
hours. Does running the engine for 7,000 consecutive hours approximate 7,000 hours
of operation over 10 years? Consider an automobile starter: the 95th percentile user
profile for the number of engine starts is approximately 4,000. Starting the engine
4,000 times as quickly as possible does not stress the starter as much as the actual
usage conditions because the engine would be warm for nearly every engine start.
To more adequately represent the true usage conditions, the engine would need to
be cooled for some of the starts.
Statistical Sample Size Determination

The sample sizes given in Table 1.1 are based on statistical sampling. Statistical confi-
dence assumes a random sample representative of the population. Obtaining a random
sample representative of the population requires all sources of variation to be present,
such as the following:
•• Variation from multiple production operators
•• Variation from multiple lots of raw materials
•• Variation from tool wear
•• Variation from machine maintenance
•• Variation from seasonal climatic changes
•• Variation from supplier changes
It may be possible to obtain a random sample representative of the population
for periodic requalifications, but it is nearly impossible for new product development.
Thu, designing tests to demonstrate reliability with statistical confidence is not always
possible. The best alternative is to test with worst-case tolerances.
Tolerances in Accelerated Testing

Determining the worst-case combination of tolerances can be difficult. Consider the
simple system shown in Figure 1.3. Component A is inserted into Component B and
rotates during operation. The worst-case tolerance is either Component A at the
FIGURE 1.3 Simple tolerancing example.
maximum diameter and the inner diameter of Component B at a minimum or

Component A at a minimum diameter and the inner diameter of Component B at
a maximum.
But even with this simple system, other tolerances must be accounted for, such
as the following:
•• Surface finish (for both components)
•• Volume of lubricant
•• Viscosity of lubricant
•• Roundness (for both components)
•• Hardness (for both components)
The number of tolerance combinations can become unmanageable. Table 1.3
shows the number of possible tolerance combinations as a function of the number of
dimensions. With 10 characteristics to consider for worst-case tolerancing in this
TABLE 1.3 Number of tolerance combinations.
Number of characteristics Number of tolerance combinations

2 4
3 8
4 16
5 32
10 1,024
20 1,048,576
50 1.126 (1,015)
100 1.268 (1,030)
simple two-component system, there are more than 1,000 combinations of tolerances
to consider. Determining which of these 1,000 combinations is the worst case is
often difficult.
Confounding the problem is the fact that the worst-case tolerance combination
for a specific environmental condition may be the best-case tolerance combination
for another environmental condition. Manufacturing capabilities also complicate
testing at worst-case tolerance combinations. It is often impossible or prohibitively
expensive to produce parts at the desired tolerance level. In these cases, a compromise
is made by using a dimension as close as possible to the desired value.
Ideally, if all characteristics are within tolerance, the system would work perfectly
and survive for the designed life. And if one or more characteristics are out of toler-
ance, the system would fail. Reality demonstrates that a component with a characteristic
slightly out of tolerance is nearly identical to a component with the same characteristic
slightly within tolerance. Tolerances are not always scientifically determined because
time and budget do not always allow for enough research. There is a strong correlation
between the defect rate in the manufacturing facility and field reliability. A portion of
the reduction in defect rate has been due to a reduction in manufacturing variability.
As manufacturing variability is reduced, characteristics are grouped closer to the target.
Consider a motor with its long-term durability dependent on the precision fit of
three components in a housing. The three components are stacked in the housing;
historically, the tolerance stackup has caused durability problems, and the maximum
stackup of the three components has been specified at 110. To meet this requirement,
an engineer created the specifications shown in Table 1.4.
If the three components are manufactured to the target, the total stackup is 60.
However, there is always variance in processes, so the engineer specifies a maximum
allowable size. If the manufacturing capability (Cpk) for each of the components is
0.5 (a defect rate of 67,000 parts per million), the process will produce the results
shown in Figure 1.4 for the stackup of the system.
By increasing the manufacturing capability for each of the components to 0.83
(a defect rate of 6,200 parts per million), the process will produce the results shown
in Figure 1.5 for the stackup of the system.
The motor housing has a perfect fit with the three components if the stackup is
60. Any deviation from 60 will reduce the life of the motor. As long as the total stackup
is less than 110, the motor will have an acceptable life; however, motors with a stackup
closer to 60 will last longer. It is easy to see that the reduced variance in manufacturing
will increase the life of the motors.
Manufacturing capability cannot be overlooked by reliability engineers. First-
time capability verification, statistical process control (SPC), and control plans are
essential to providing highly reliable products. Without capable manufacturing, all
previous reliability efforts will provide little or no benefit.
TABLE 1.4 Motor component tolerances.
Component A B C Total
Target size 30 20 10 60
Maximum size 50 30 15 95
FIGURE 1.4 Tolerance stackup of 3 components with Cpk of 0.50.
FIGURE 1.5 Tolerance stackup of 3 components with Cpk of 0.83.
Financial Considerations
Priorities for a reliability program are determined the same way as most other
programs. The number one priority is an emergency. If there is a hole in one of the
water pipes in your home, you will find a way to fix it, regardless of your
financial situation.
The next level of priority is an obvious need that can be delayed with a risk.
Consider again the leaking water pipe. If the water pipe is patched with duct tape and
epoxy, and while fixing the pipe it is determined that all the water pipes in the home
are in danger of bursting, then obviously there is a need to replace all the water pipes
in the house. However, this is an expensive task and can be delayed. There is no
immediate crisis, but by delaying the repair, there is a risk of an expensive accident.
If a water pipe bursts, thousands of dollars of damage will result. This risk is tolerated
because the immediate expense of correcting the problem is perceived to be greater
than the cost of the water pipes bursting weighted by the probability of the water
pipes bursting.
The most dangerous priority is one that is not known. Consider a home that is
being consumed by termites without the owner’s knowledge. Nothing is done to
correct the problem because the owner is ignorant of the problem. For reliability
programs, the largest expenses are often overlooked.
A typical reliability program contains the following eight elements:
1. Understand your customer’s requirements.
a. Environment
i. Vibration (g-force root mean square [GRMS] and frequency profile)
ii. Humidity
iii. Temperature (absolutes and cycling rates)
iv. Electrical stimulus (electrostatic discharge [ESD])
v. Contamination (salt, brake fluid, transmission fluid, spilled
milkshakes or sodas)
b. Duty cycle
i. Number of usage cycles
ii. Required life in months or years
c. Load
i. Pounds of force
ii. Pressure
iii. Voltage
iv. Current
d. Reliability goals
2. Feedback from similar components.
a. Failure rate analysis and corrective action system (FRACAS)—Parts from
test failures, internal production failures, external production failures,
and field returns must be analyzed and cataloged.
b. J.D. Power and Associates.
c. Warranty (return rates, feedback from customers and technicians).
d. Development tests, design verification tests, and production
validation tests.
3. Begin the failure modes and effects analysis (FMEA) process. The FMEA will
be updated during the entire process.
4. Intelligent design.
a. Use design guides—All lessons from previous incidents must be captured
in design guides. This includes all information from the FRACAS.
b. Parameter design—Choose the design variable levels to minimize the
effect of uncontrollable variables.
c. Tolerance design—Scientifically determine the correct
drawing specifications.
d. Hold periodic design reviews.
e. Design with information from developmental activities.
i. Sneak circuit analysis
ii. Highly accelerated life testing (HALT)
iii. Step-stress tests to failure
iv. Worst-case tolerance analysis
5. Concept validation (includes parts supplied by outside sources).
a. Early in the development phase, have short, inexpensive tests to provide
approximate results. The purpose of these tests is to provide
engineering feedback.
b. Every concept must pass an independent (i.e., not conducted by
engineering) verification test. The concept should include design limits.
For example, the component has been validated to operate up to 85°C,
withstand brake fluid, and 4.8 GRMS random vibrations in a frequency
range from 0 to 800 Hz.
c. A standard list of validated subsystems and components should
be maintained. This includes parts supplied from outside sources.
6. Design validation (includes parts supplied by outside sources).
a. Early in the development phase, have short, inexpensive tests to provide
approximate results. The purpose of these tests is to provide
engineering feedback.
b. Every design must pass an independent (not conducted by engineering)
verification test. Be careful not to burden the company with timing and
cost issues when specifying the test. Build on the results of the concept
verification and any other implementations of the concept.
c. System simulation.
7. Manufacturing.
a. Parts from production-intent tooling must pass the design validation test.
b. The production process is qualified by producing a specified number of
parts at production rates, and obtaining a minimum Cpk of 1.67 for every
drawing specification.
c. Ensure compliance with an SPC program.
i. All drawing specifications will be monitored with an electronic
SPC system.
ii. For the first week of production, the sampling rate is 100%.
iii. If a Cpk of 1.67 is achieved for the first week, the sampling rate may
be reduced.
iv. Each drawing specification must have a control plan that details the
critical processes affecting the drawing specification. Each of these
processes also must be monitored with an SPC.
v. For each measurement system:
1. Provide a control plan to ensure stability and bias control.
2. The measurement error should be an acceptable percentage of the
tolerance. What is acceptable is a function of the process Cpk.
d. Qualify the storage, transportation, and installation systems.
i. Can parts be contaminated during storage?
ii. Can parts degrade during storage?
1. Rubber ages.
2. Gas charge levels decrease.
iii. Are the temperature and vibration profiles during transportation
significantly different from those of the vehicle specification?
iv. Is the part protected from corrosion caused by salt in the air during
transportation on an ocean liner?
v. Requalify if the transportation system is changed.
8. Change control—Any changes in engineering, production, or the supply base
must be qualified.
The following is a consolidated set of responses based on actual experiences.
Response from Finance

After analyzing the reliability program, there appear to be many opportunities
for savings.
1. The current understanding of the customer’s environment is adequate. There
is no need to collect additional customer usage or environmental data.
2. How many prototypes will be used during the parameter design and
tolerance design processes? What is the cost of a prototype?
3. What is the cost of testing?
a. HALT
b. Design verification
c. Production validation
d. Step-stress testing
4. The manufacturing qualification plan appears to be excessive.
We recommend reducing the number of samples required for initial
validation and reducing the number of characteristics that require
control plans.
Response from Engineering

Although we agree that the approach of Reliability is correct, we are supporting
Finance. The program is behind schedule, and reducing the reliability effort will
improve program timing. In addition, with recent cutbacks (or the recent increase
in business), Engineering lacks the resources to complete the entire reliability program.
Response from Manufaturing

The program proposed by Reliability will ensure the initial quality, but it is too costly.
With recent production increases (or other convenient excuses), Manufacturing cannot
meet the initial production schedule with the restrictions proposed by Reliability. The
reliability program would also require additional personnel to maintain the SPC program.
These responses result in an organization with a strong incentive to gamble that
there will be no consequences for abandoning a thorough reliability program in favor
of a program that is less expensive. This behavior suboptimizes the finances of the
company by assuming any potential failure costs are near zero. To be effective, a good
reliability program must include a financial assessment of the risks involved if reli-
ability activities are not completely executed.
The urge to reduce the investment in the reliability program can be combatted
by visualizing the failure costs that the reliability program is designed to prevent. In
addition to warranty costs, other failure costs are as follows:
•• Customer returns
•• Customer stop shipments
•• Retrofits
•• Recalls
These costs often are ignored because they are not quantified. An effective method
for quantifying these costs is to record a score for each element in the reliability
program and compare this to the field performance of the product. This can be done
by auditors using a grade scale of A through F for each element of the program. The
grades for each element can be combined into a grade point average (GPA) for the
program using 4 points for an A, 3 points for a B, and so forth.
Table 1.5 gives an example of how a reliability program may be scored, and
Table 1.6 shows how the field performance is recorded. When quantifying these costs,
TABLE 1.5 Example reliability program scores.
Reliability program item Score (GPA)

Understanding of customer requirements B-3
FMEA A-4
FRACAS C-2
Verification C-2
Validation D-1
Manufacturing B-3
Overall program score 2.33
TABLE 1.6 Example field reliability performance.
Reliability performance item Cost

Customer returns $8,245
Customer stop shipments $0
Retrofits $761,291
Recalls $0
Overall program unreliability cost $769,536
FIGURE 1.6 Reliability program execution score versus failure costs.
be sure to include all associated labor and travel costs. For example, a customer may
charge $500 for returning a single part; however, the associated paperwork, travel,
and investigation could easily be several thousand dollars.
Figure 1.6 is a scatter chart of the results of several programs. The slope of the
trend line quantifies the loss when the reliability program is not fully executed. For
this example, moving the GPA of the overall reliability program by one point is
expected to result in a savings of $755,000 in failure costs. These savings can be used
to financially justify the investment in the reliability program.
These failure costs are similar to the cost of water pipes bursting in your house.
You know of the risk, and you decide to act on the risk or tolerate the risk, based on
the finances of the situation.
Another method to focus the management’s attention on reliability is by
presenting the effects of the data shown in Table 1.6 on corporate profits. The data in
Figure 1.7 are examples of the effects of a poor reliability program. Money was saved
years earlier by gambling with a substandard reliability program, but as shown in
Figure 1.7, the short-term gain was not a good long-term investment.
FIGURE 1.7 Effect of poor reliability on company profits.
Similar to termites damaging a home without the owner’s knowledge, hidden

reliability costs are causing poor decisions to be made and are damaging profits. The
losses caused by these hidden costs can be orders of magnitude greater than warranty
costs. To illustrate this concept, consider the automotive industry.
For the model year 1998, the average vehicle manufactured by General Motors,
Ford, or Chrysler (the “Big Three”) required $462* in repairs. These automakers sell
approximately 13 million vehicles in North America annually, resulting in a total
warranty bill of $6 billion. That may sound like a lot of money, but it is by far the
smallest piece of the cost of poor reliability. Table 1.7 illustrates the retail value for
several 1998 model year vehicles with sticker prices all within a $500 range.
For lease vehicles, the manufacturer absorbs the $5,715 difference in resale value
between Vehicle B and Vehicle H. For non-lease vehicles, the owner of Vehicle B
absorbs the cost. But this does not mean the manufacturer is not impacted. The
reduced retail value is reflected in the ability of the manufacturer to price new vehicles.
The manufacturer of Vehicle H can charge more for new vehicles because they depre-
ciate more slowly. Considering that sales for many of these midsize sedans topped
200,000 units, the $5,715 difference in resale value is worth more than $1 billion
annually. Figure 1.8 shows the correlation between the reliability of a vehicle and its
resale value. Using the slope of the regression line shown in Figure 1.8, a single point
in Consumer Reports’ reliability rating is worth $51.58.
When faced with pressures to compromise on your reliability program, consider
the system cost. For example, Design A uses a material that is a nickel less expensive
than Design B, and the component is expected to be installed on 300,000 vehicles.
* Day, Joseph C., address to the Economic Club of Detroit, December 11, 2000.
TABLE 1.7 Vehicle resale value.
Consumer reports reliability

Vehicle (1998 model year) Retail value as of July 2001 rating*
A $8,430 −45
B $9,500 20
C $9,725 18
D $11,150 25
E $11,150 30
F $13,315 −5
G $14,365 55
H $15,215 50
* The Consumer Reports’ scale is from −80 to 80, with −80 being the worst and 80 being the
best.
FIGURE 1.8 Automobile resale value versus reliability rating.
If Design A is chosen, the savings is $15,000 in material cost. This is a good decision
only if the expected impact on the warranty, resale value, and market share is less
than $15,000. Before using Design A, verify the reliability of the design by doing
the following:
•• Compare the designs using a durability test.
•• Ensure your durability test will incorporate any potential failure modes
associated with material changes.
•• Verify key performance issues with computer modeling.
•• Check for any potential system failures that could be caused by a material
change.
•• Explore the impact on manufacturing. What is the impact of material changes,

processing changes, and so forth? Are there issues such as dunnage issues,
environmental issues, and safety hazards?
Summary
The theoretical science of accelerated testing is exact, but implementing an accelerated
testing plan requires several compromises. Ultimately, the science involved is inexact
and serves only as a guideline for engineering judgment. Reliability demonstration
based on statistically determined sample sizes is often invalid because the samples
could not be a random representation of production parts. Testing to worst-case
tolerance limits is difficult because of the number of combinations and the difficulty
of producing parts at the desired tolerance level.
Automobile manufacturers often specify a design life of 10 years for automobiles.
Many types of aircraft have a design life of 25 years, and some B-52 bombers have
been in service for more than 40 years. Accelerating a test for an automobile compo-
nent by a factor of 10 would yield a test with one-year duration. This obviously is
unacceptable. Obtaining a test with a duration of one month requires an acceleration
factor of 120. A test lasting 24 hours would have an acceleration factor of 3,653.
Is this type of acceleration possible? Although this is a slightly controversial
subject, many experts suggest it is impossible to accelerate a test by more than a
factor of 10 without losing some correlation to real-world conditions. This is another
of the ambiguities faced when accelerating testing conditions. A test is required, and
the test is useless if it cannot be completed in a reasonable time frame. However, the
greater the acceleration the less realistic the test.
Accelerated testing is a balance between science and judgment. Do not let the
science cause bad decisions to be made. For example, if demonstrating 95% reliability
at 150,000 miles in service calls for testing 45 units for 2,000 hours without failure,
do not be concerned if only 35 units can be tested. The sample size of 45 assumes
random sampling from a population representative of production. Violating this
assumption is more important than testing with a reduced sample size. In this situ-
ation, try to ensure that the test is representative of real-world conditions and secure
35 samples with key characteristics set at worst-case tolerance levels.
2
Probability Fundamentals
S
tatistical methods are based on random sampling. Four functions are used to
characterize the behavior of random variables:
1. Probability density function

2. Cumulative distribution function
3. Reliability function
4. Hazard function
If any one of these four functions is known, the others can be derived. This
chapter describes these functions in detail.
Sampling
Statistical methods are used to describe populations by using samples. If the popula-
tion of interest is small enough, statistical methods are not required; every member
of the population can be measured. If every member of the population has been
measured, a confidence interval for the mean is not necessary because the mean is
known with certainty (ignoring measurement error).
For statistical methods to be valid, all samples must be chosen randomly. Suppose
the time to fail for 60-W light bulbs is required to be greater than 100 hours, and the
customer verifies this on a monthly basis by testing 20 light bulbs. If the first 20 light
bulbs produced each month are placed on the test, any inferences about the reliability
of the light bulbs produced during the entire month would be invalid because the

TABLE 2.1 Time to fail for light bulbs.
23 63 90
39 72 96
41 79
58 83
10 units survive for 96 hours without failing.
20 light bulbs tested do not represent the entire month of production. For a sample
to be random, every member of the population must have an equal chance of being
selected for the test.
Now suppose that when 20 light bulbs, randomly selected from an entire month’s
production, are placed on a test stand, the test is ended after 10 of the light bulbs fail.
Table 2.1 gives an example of what the test data might look like.
What is the average time to fail for the light bulbs? It obviously is not the
average time to fail of the 10 failed light bulbs, because if testing were continued
until all 20 light bulbs failed, 10 data points would be added to the data set that
are all greater than any of the 10 previous data points. When testing is ended before
all items fail, the randomness of the sample has been destroyed. Because only 10
of the light bulbs failed, these 10 items are not a random sample that is representa-
tive of the population. The initial sample of 20 items is a random sample represen-
tative of the population. By ending the testing after 10 items have failed, the
randomness has been destroyed by systematically selecting the 10 items with the
smallest time to fail.
This type of situation is called censoring. Statistical inferences can be made using
censored data, but special techniques are required. The situation described above is
the right censoring; the time to fail for a portion of the data is not known, but it is
known that the time to fail is greater than a given value. Right-censored data may
be either time censored or failure censored. If testing is ended after a predetermined
amount of time, it is time censored. If testing is ended after a predetermined number
of failures, the data are failure censored. The data in Table 2.1 are failure censored.
Time censoring is also known as Type I censoring, and failure censoring is also known
as Type II censoring.
The opposite of right censoring is left censoring. For a portion of the data, the
absolute value is not known, but it is known that the absolute value is less than a given
value. An example of this is an instrument used to measure the concentration of
chemicals in a solution. If the instrument cannot detect the chemical below certain
concentrations, it does not mean there is no chemical present but that the level of the
chemical is below the detectable level of the instrument.
Another type of right censoring is multiple censoring. Multiple censoring occurs
when items are removed from testing at more than one point in time. Field data are
often multiple censored. Table 2.2 gives an example of multiple censoring. A “+” next
to a value indicates the item was removed from testing at that time without failing.
There are two types of data: (1) continuous and (2) discrete. Continuous variables
are unlimited in their degree of precision. For example, the length of a rod may be 5,
5.01, or 5.001 in.
TABLE 2.2 Multiple-censored data.
112 172 220

145+ 183 225+
151 184+ 225+
160 191 225+
166+ 199 225+
It is impossible to state that a rod is exactly 5 in. long—only that the length of
the rod falls within a specific interval. Discrete variables are limited to specific values.
For example, if a die is rolled, the result is either 1, 2, 3, 4, 5, or 6. There is no possibility
of obtaining any value other than these six values.
Probability Density Function

The probability density function, f(x), describes the behavior of a random variable.
Typically, the probability density function is viewed as the “shape” of the distribution.
Consider the histogram of the length of fish as shown in Figure 2.1.
FIGURE 2.1 Example histogram.

FIGURE 2.2 Example histogram with an overlaid model.
A histogram is an approximation of the shape of the distribution. The histo-

gram shown in Figure 2.1 appears symmetrical. Figure 2.2 shows this histogram
with a smooth curve overlying the data. The smooth curve is the statistical model
that describes the population—in this case, the normal distribution. When using
statistics, the smooth curve represents the population, and the differences between
the sample data represented by the histogram and the population data represented
by the smooth curve are assumed to be due to sampling error. In reality, the differ-
ences could also be caused by the lack of randomness in the sample or an
incorrect model.
The probability density function is similar to the overlaid model shown in
Figure 2.2 . The area below the probability density function to the left of a given value
x is equal to the probability of the random variable represented on the x-axis being
less than the given value x. Because the probability density function represents the
entire sample space, the area under the probability density function must equal 1.
Negative probabilities are impossible; therefore, the probability density function, f(x),
must be positive for all values of x. Stating these two requirements mathematically,
∞
∫ f ( x ) = 1 (2.1)
−∞
and f(x) ≥ 0 for continuous distributions. For discrete distributions,

n
∑ f ( x ) = 1 (2.2)
for all values of n, and f(x) ≥ 0.
The area below the curve in Figure 2.2 is greater than 1; thus, this curve is not a
valid probability density function. The density function representing the data in
Figure 2.2 is shown in Figure 2.3.
Figure 2.4 demonstrates how the probability density function is used to compute
probabilities. The area of the shaded region represents the probability of a single fish, drawn
randomly from the population, having a length less than 185. This probability is 15.9%.
Figure 2.5 demonstrates the probability of the length of one randomly selected
fish having a length greater than 220.
The area of the shaded region in Figure 2.6 demonstrates the probability of the
length of one randomly selected fish having a length greater than 205 and less than
215. Note that as the width of the interval decreases, the area, and thus the probability
of the length falling in the interval, decreases. This also implies that the probability
of the length of one randomly selected fish having a length exactly equal to a specific
value is 0. This is because the area of a line is 0.
FIGURE 2.3 Probability density function for the length of fish.

FIGURE 2.4 The probability of the length being less than 185.
FIGURE 2.5 The probability of the length being greater than 220.
FIGURE 2.6 The probability of the length being between 205 and 215.
Example 2.1: A probability density function is defined as
a
f ( x )= , 1 ≤ x ≤ 10
x
For f(x) to be a valid probability density function, what is the value of a?
Solution: To be a valid probability density function, all values of f(x) must be positive,
and the area beneath f(x) must equal 1. The first condition is met by restricting a
and x to positive numbers. To meet the second condition, the integral of f(x) from
1 to 10 must equal 1.
10
a
∫x =1
1
10
a ln x 1 = 1
a ln ( 10 ) − a ln( 1) =
1
1
a=
ln( 10 )
Cumulative Distribution Function

The cumulative distribution function, F(x), denotes the area beneath the probability
density function to the left of x, as shown in Figure 2.7.
The area of the shaded region of the probability density function in Figure 2.7 is
0.2525. This is the corresponding value of the cumulative distribution function at
x = 190. Mathematically, the cumulative distribution function is equal to the integral
of the probability density function to the left of x.
x
F (x) = ∫ f ( τ )d τ (2.3)
−∞
Example 2.2: A random variable has the probability density function f(x) = 0.125x,
where x is valid from 0 to 4. The probability of x being less than or equal to 2 is
2 2
0.125x 2 2
F ( 2)
= ∫ 0.125
= x dx = 0.0625 =x2 0.25
0 2 0
0
FIGURE 2.7 Cumulative distribution function.

Example 2.3: The time to fail for a transistor has the following probability density
function. What is the probability of failure before t = 200?
f (t ) = 0.01e −0.01t
Solution: The probability of failure before t = 200 is
200
P (t < 200 ) =
∫ 0.01e dt
−0.01t
0
0
= −e −0.01t |
2

( )
=−e −2 − −e 0 =1 − e −2 =0.865
Reliability Function
The reliability function is the complement of the cumulative distribution function.
In modeling the time to fail, the cumulative distribution function represents the
probability of failure, and the reliability function represents the probability of survival.
Thus, the cumulative distribution function increases from 0 to 1 as the value of x
increases, and the reliability function decreases from 1 to 0 as the value of x increases.
This is shown in Figure 2.8.
FIGURE 2.8 The reliability function.

As seen from Figure 2.8, for the probability that the time to fail is greater than
190, the reliability is 0.7475. For the probability that the time to fail is less than 190,
the cumulative distribution function is
1 − 0.7475 = 0.2525
Mathematically, the reliability function is the integral of the probability density
function from x to infinity.
∞
R ( x ) = ∫ f ( τ )d τ (2.4)
x
Hazard Function
The hazard function is a measure of the tendency to fail; the greater the value of the
hazard function the greater the probability of impending failure. Technically, the
hazard function is the probability of failure in the very small time interval, x0 to x0 + δ,
given a survival until x0. The hazard function is also known as the instantaneous
failure rate. Mathematically, the hazard function is defined as
f (x)
h( x ) = (2.5)
R(x)
Using Equation (2.5) and Equations (2.6) and (2.7), if the hazard function, the
reliability function, or the probability density function is known, the remaining two
functions can be derived.
x
∫
− h ( τ )dτ
R(x) = e −∞
(2.6)
∫
− h( τ )dτ
f ( x ) = h ( x )e −∞
(2.7)
Example 2.4: Given the hazard function, h(x) = 2x, derive the reliability function
and the probability density function.
Solution: The reliability function is
x
− ∫ 2xdx
R (x ) = e −∞
R ( x ) = e −x
2
The probability density function is
f ( x ) h=
( x )R ( x ) 2xe −x
2
=
Expectation
Several terms are used to describe distributions. The most common terms are the
mean, variance, skewness, and kurtosis. These descriptors are derived from moment-
generating functions. Readers with an engineering background may recall that the
center of gravity of a shape is
cog =
∫ xf ( x )dx (2.8)
−∞
∞
∫ f ( x )dx
−∞
The mean, or average, of a distribution is its center of gravity. In Equation (2.8),

the denominator is equal to the area below f(x), which is equal to 1 by definition for
valid probability distributions. The numerator in Equation (2.8) is the first moment-
generating function about the origin. Thus the mean of a distribution can be deter-
mined from the expression
∞
E ( x ) = ∫ xf ( x )dx (2.9)
−∞
The second moment-generating function about the origin is
∞
E ( x 2 ) = ∫ x 2 f ( x )dx (2.10)
−∞
The variance of a distribution is equal to the second moment-generating function

about the mean, which is
∞
σ2 = ∫ x 2 f ( x )dx − µ2 (2.11)
−∞
The variance is a measure of the dispersion in a distribution. In Figure 2.9, the

variance of Distribution A is greater than the variance of Distribution B.
The skewness of a distribution is equal to the third moment-generating function
about the mean. If the skewness is positive, the distribution is right skewed. If the
skewness is negative, the distribution is left skewed. Right and left skewness are
demonstrated in Figure 2.10.
Kurtosis is the fourth moment-generating function about the mean and is a
measure of the peakedness of the distribution.
FIGURE 2.9 Distribution variance.
FIGURE 2.10 Right and left skewness.

Summary
A statistical distribution can be described by the following functions:
•• Probability density function, f(x)
•• Cumulative distribution function, F(x)
•• Reliability function, R(x)
•• Hazard function, h(x)
When one of these functions is defined, the others can be derived. These func-
tions, with characteristics concerning central tendency, spread, and symmetry,
provide a description of the population being modeled.
3
Distributions
S
tatistical distributions fall into two categories: (1) modeling distributions and
(2) sampling distributions. Modeling distributions are used to describe data sets
and are divided into two classes: (1) continuous distributions and (2) discrete
distributions. Sampling distributions are used to construct confidence intervals and
to test hypotheses.
Continuous Modeling Distributions

The four most common reliability modeling distributions are
1. Weibull distribution
2. Normal distribution
3. Lognormal distribution
4. Exponential distribution
Each of these distributions is described in detail in the following sections.

Weibull Distribution
The Weibull distribution is one of the most commonly used distributions in reliability
to model time to fail, time to repair, and material strength. The Weibull probability
density function is
β
(β −1)  x − δ  
β x −δ  −  
f (x ) =   e  θ  
, x ≥ δ (3.1)
θ θ 
where
β is the shape parameter
θ is the scale parameter
δ is the location parameter
The shape parameter gives the Weibull distribution its flexibility. By changing the
value of the shape parameter, the Weibull distribution can model a variety of data. If
β = 1, the Weibull distribution is identical to the exponential distribution. If β = 2, the
Weibull distribution is identical to the Rayleigh distribution. If β is between 3 and 4, the
Weibull distribution approximates the normal distribution. The Weibull distribution
approximates the lognormal distribution for several values of β. For most populations,
more than 50 samples are required to differentiate between the Weibull and lognormal
distributions. Figure 3.1 shows a sample of the flexibility of the Weibull distribution.
FIGURE 3.1 Weibull probability density functions with θ = 100 and δ = 0.

The scale parameter determines the range of the distribution. The scale parameter
is also known as the characteristic life if the location parameter is equal to 0. If δ ≠ 0,
the characteristic life is θ + δ ; 63.2% of all values fall below the characteristic life,
regardless of the value of the shape parameter. Figure 3.2 shows the effect of the scale
parameter of the probability density function.
The location parameter is used to define a failure-free zone. The probability of
failure when x < δ is 0. When δ > 0, there is a period when no failures can occur.
When δ < 0, failures have occurred before time equals 0. At first, this seems ridiculous;
however, a negative location parameter is caused by shipping failed units, failures
during transportation, and shelf-life failures. Generally, the location parameter is
assumed to be zero. Figure 3.3 shows the effect of the location parameter.
The Weibull hazard function is determined by the value of the shape parameter,
(β −1)
β x −δ 
h( x ) =   (3.2)
θ θ 
When β < 1, the hazard function is decreasing; this is known as the infant
mortality period. When β = 1, the failure rate is constant. When β > 1, the failure rate
is increasing; this is known as the wear-out period. Figure 3.4 shows the Weibull
hazard function.
When β < 1, the hazard function decreases more steeply as β decreases. When
β > 1, the rate of increase for the hazard function increases as β increases. As seen
from Figure 3.4, when β = 2, the hazard function increases linearly.
FIGURE 3.2 Effect of the Weibull scale parameter.

FIGURE 3.3 Effect of the Weibull location parameter.
FIGURE 3.4 The Weibull hazard function.

The Weibull reliability and cumulative distribution functions are respectively

  t −δ  β 
− −  
R (t ) = e
  θ  
(3.3)
and
  t −δ  β 
− −  
F (t ) = 1 − e
  θ  
(3.4)
Example 3.1: The time to fail for a flexible membrane follows the Weibull distribu-
tion with β = 2 and θ = 300 months. What is the reliability at 200 months? After
how many months is 90% reliability achieved?
Solution: After 200 months, the reliability of the flexible membrane is
  200 2 
 −  
( 200) e= 0.6412
R=
  300  
By manipulating the expression for reliability, 90% reliability is achieved after
1
t =θ ( − ln R ) β 
 

t =300 − ln ( 0.9 ) =97.38months
The mean and variance of the Weibull distribution are computed using the
gamma distribution, which is given in Table A.1 of Appendix A.1 The mean of the
Weibull distribution is
 1
µ = θΓ  1 +  (3.5)
 β
The mean of the Weibull distribution is equal to the characteristic life if the shape
parameter is equal to 1. Figure 3.5 shows the mean as a function of the shape parameter.
The variance of the Weibull distribution is
  2  1 
σ2 = θ2 Γ  1 +  − Γ2  1 +   (3.6)
  β  β 
The variance of the Weibull distribution decreases as the value of the shape
parameter increases, as shown in Figure 3.6.
The reliability measure of many components is the mean time to fail. Consumer
electronics companies often advertise the mean time to fail of the products on
television. The mean time to fail is a deceptive measure because the variance of the
1
The gamma function is available in Microsoft® Excel. The function is = GAMMA(x).
FIGURE 3.5 The Weibull mean.
FIGURE 3.6 The Weibull variance.

time-to-fail distribution is not considered. To achieve the same reliability with a larger
variance requires a larger mean.
Consider two components, A and B. Component A has a mean time to fail of 4,645
hours, and Component B has a mean time to fail of 300 hours. If both components sell
for the same price, which component should be used to maximize reliability at
100 hours?
This question cannot be answered without knowing more information about the
distribution of the time to fail. Component A has a mean of 4,645 hours and a Weibull
time-to-fail distribution with a shape parameter of 0.8. Using the mean and the shape
parameter, the scale parameter of Component A can be computed to be 4,100 hours.
The reliability at 100 hours is
0.8
 100 
− 
R (100 ) = e  4 ,100 
= 0.95
Component B has a mean of 300 hours and a Weibull time-to-fail distribution
with a shape parameter of 3. Using the mean and the shape parameter, the scale
parameter of Component B can be computed to be 336 hours. The reliability at 100
hours is
3
 100 
− 
R (100 ) = e  336 
= 0.974
Although the mean of Component A is more than 10 times as large as the mean
of Component B, the reliability of Component B is greater than the reliability of
Component A at 100 hours. Continuing with this example, if the reliability at 1,000
hours is to be maximized, Component A has a reliability of 0.723, and Component
B has a reliability of approximately 0.
Normal Distribution
Whenever several random variables are added together, the resulting sum tends to
normal, regardless of the distribution of the variables being added. Mathematically, if
y = x1 + x2 + x3 + + xn (3.7)
then the distribution of y becomes normal as n increases. If the random variables
being summed are independent, the mean and variance of y are respectively
µ y = µ x 1 + µ x 2 + µ x 3 + + µ xn (3.8)
and
σ2y = σ2x 1 + σ2x 2 + σ2x 3 + + σ2xn (3.9)
When several random variables are averaged, the resulting average tends to normal,
regardless of the distribution of the variables being averaged. Mathematically, if
x1 + x2 + x3 + + xn
y= (3.10)
n
then the distribution of y becomes normal as n increases. If the random variables

being averaged have the same mean and variance, then the mean of y is equal to the
mean of the individual variables being averaged, and the variance of y is
σ2
σ2y = (3.11)
n
where σ2 is the variance of the individual variables being averaged.
The tendency of sums and averages to become normally distributed as the number
of variables being summed or averaged becomes large is known as the central limit
theorem or the theory of large numbers. For distributions with little skewness,
summing or averaging as few as three or four variables will result in a normal distri-
bution. For highly skewed distributions, more than thirty variables may have to be
summed or averaged to obtain a normal distribution.
The normal probability density function is
2
1  x −µ 
1 −  
f (x ) = e 2 σ 
, − ∞ < x < ∞ (3.12)
σ 2Π
where
μ is the mean
σ is the standard deviation
The normal probability density function is not skewed and is shown in Figure 3.7.
The density function shown in Figure 3.7 is the standard normal probability
density function. The standard normal probability density function has a mean of 0
FIGURE 3.7 The standard normal probability density function.

and a standard deviation of 1. The normal probability density function cannot be

integrated implicitly. Because of this, historically a transformation to the standard
normal distribution is made, and the normal cumulative distribution function or
reliability function is read from a table. Table A.2 in Appendix A gives the standard
normal reliability function. If x is a normal random variable, it can be transformed
to standard normal using the expression
x −µ
z= (3.13)
σ
Example 3.2: The tensile strength of a metal extrusion is normally distributed with
a mean of 300 and a standard deviation of 5. What percentage of extrusions has
a strength greater than 310? What percentage of extrusions has a strength less
than 295? What percentage of extrusions has a strength between 295 and 310?
Solution: The shaded area in the top graph in Figure 3.8 represents the probability
of an extrusion being greater than 310. The shaded area in the bottom graph in
Figure 3.8 represents the area under the standard normal distribution to the right
of z = 2, which is the same as the probability of an extrusion being greater than 310.
Transforming to standard normal,
310 − 300
=z = 2
5
To determine the area under the standard normal probability density function to
the right of z = 2, look up z = 2 in Table A.1 of Appendix A, which is 0.0228. The
percentage of extrusions with strength greater than 310 is 0.0228.
FIGURE 3.8 Probability of strength being greater than 310.

FIGURE 3.9 Probability of strength being less than 295.
The shaded area in the top graph in Figure 3.9 represents the probability of an
extrusion being less than 295. The shaded area in the bottom graph in Figure 3.9
represents the area under the standard normal distribution to the left of z = −1, which
is the same as the probability of an extrusion being less than 295. Transforming to
standard normal,
295 − 300
z= = −1
5
From Table A.1 of Appendix A, the area to the left of z = −1 is equal to 0.1587.
The probability of the strength being between 295 and 310 is
1 − 0.0228 − 0.1587 = 0.8185

The transformation to standard normal is no longer required. Electronic spread-
sheets have functions to evaluate the normal probability density function. The area
to the left of x under the normal probability density function with a mean of μ and
a standard deviation of σ is found using the Microsoft Excel expression
= NORM.DIST ( x ,µ,σ,1)
The inverse of this function is also available in Microsoft Excel. Given a normal
distribution with a mean of μ and a standard deviation of σ, the value of y having an
area of x under the probability density function to the left of y is
= NORM.INV ( x ,0,1)
Example 3.3: Repeat Example 3.2 using Microsoft Excel.

Solution: The percentage of extrusions having a strength less than 310 is found
using the expression
= NORM.DIST ( 310,300,5,1)
The percentage of extrusions having a strength greater than 310 is found using
the expression
= 1 − NORM.DIST ( 310,300,5,1)
The percentage of extrusions having a strength less than 295 is found using the
expression
= NORM.DIST ( 295,300,5,1)
The percentage of extrusions having a strength between 295 and 310 is found
using the expression

= NORM.DIST ( 310,300,5,1) − NORM.DIST ( 295,300,5,1)
Example 3.4: A type of battery is produced with an average voltage of 60 with

a standard deviation of 4 V. If nine batteries are selected at random, what is the
probability that the total voltage of the nine batteries is greater than 530? What
is the probability that the average voltage of the nine batteries is less than 62?
Solution: The expected total of the voltage of nine batteries is 540. The expected
standard deviation of the voltage of the total of nine batteries is
=σ
(9=
)( 4 ) 12
9
Transforming to standard normal,
530 − 540
z= = −0.833
12
From Table A.1 of Appendix A, the area to the left of z = −0.833 is 0.2024. The area
to the right of z = −0.833 is 1 − 0.2024 = 0.7976. This value can also be computed
using the Microsoft Excel function
= 1 − NORM.DIST ( 530,540,12,1)
The probability density function of the voltage of the individual batteries and
of the average of nine batteries are shown in Figure 3.10. The distribution of the
averages has less variance because the standard deviation of the averages is equal to
the standard deviation of the individuals divided by the square root of the sample size.
The probability of the average voltage of nine batteries being less than 62 is equal
to the probability of a standard normal variable being less than z = 1.5. This is shown
in the expression
62 − 60
Z= = 1.5
4/ 9
FIGURE 3.10 Distribution of average voltage.
From Table A.1 of Appendix A, the area under the standard normal curve to the
left of z = 1.5 is 0.9332, which is the probability of the average voltage of nine batteries
being less than 62. This value can also be computed using the Microsoft Excel function
= NORM.DIST ( 62, 60,1.333,1)
Example 3.5: A type of battery is produced with an average voltage of 12 with a

standard deviation of 1.4 V. Ninety percent of the batteries have a voltage greater
than what value?
Solution: From Table A.1 in Appendix A, 90% of the area under the standard normal
probability density function is to the right of z = 1.282. Transforming from standard
normal to a normal distribution with a mean of 12 and a standard deviation of 1.4
gives
X − 12
1.282 =
1.4
X = 10.21
The voltage of 90% of the batteries is greater than 10.21. The value can be found
using the Microsoft Excel function
= NORM.INV ( 0.1,12,1.4 )
FIGURE 3.11 Normal hazard function.
The normal hazard function is
 x −µ 
φ 
h( x ) =  σ 
  x − µ   (3.14)
σ 1 − Φ  
  σ 
where
ϕ(x) is the standard normal probability density function
Φ(x) is the standard normal cumulative distribution function
The hazard function for the normal distribution is monotonically increasing, as

shown in Figure 3.11.
Lognormal Distribution
If a data set is known to follow a lognormal distribution, transforming the data by
taking a logarithm yields a data set that is normally distributed, as shown in Table 3.1.
The most common transformation is made by taking the natural logarithm, but
any base logarithm, such as base 10 or base 2, also yields a normal distribution. The
remaining discussion will use the natural logarithm denoted as “ln.”
TABLE 3.1 Transformation of lognormal data.

Lognormal data Normal data
12 ln(12)
16 ln(16)
28 ln(28)
48 ln(48)
87 ln(87)
143 ln(143)
When random variables are summed, as the sample size increases, the distribu-
tion of the sum becomes a normal distribution, regardless of the distribution of the
individuals. Because lognormal random variables are transformed to normal random
variables by taking the logarithm, when random variables are multiplied, as the
sample size increases, the distribution of the product becomes a lognormal distribu-
tion, regardless of the distribution of the individuals. This is because the logarithm
of the product of several variables is equal to the sum of the logarithms of the indi-
viduals. This is shown as
y = x1 x2 x3
ln y = ln x1 + ln x2 + ln x3
The lognormal probability density function is
2
1  ln x − µ 
1 −  
f (x ) = e 2 σ 
, x > 0 (3.15)
x σ 2Π
where
μ is the location parameter or log mean
σ is the scale parameter or log standard deviation
The location parameter is the mean of the data set after transformation by taking
the logarithm, and the scale parameter is the standard deviation of the data set
after transformation.
The lognormal distribution takes on several shapes, depending on the value of
the shape parameter. The lognormal distribution is skewed right, and the skewness
increases as the value of σ increases, as shown in Figure 3.12.
The lognormal cumulative distribution and reliability functions are respectively
 ln x − µ 
F (x ) = Φ 
 σ 
 ln x − µ 
R (x ) = 1− Φ 
 σ 
where Ф(x) is the standard normal cumulative distribution function.
FIGURE 3.12 Lognormal probability density function.
Example 3.6: The following data are the time to fail for four light bulbs and are
known to have a lognormal distribution. What is the reliability at 100 hours?
115 hours 155 hours

183 hours 217 hours
Solution: Table 3.2 shows this data and the transformation to normal.
The parameters of the lognormal distribution are found by computing the mean
and the standard deviation of the transformed data in the second column of Table 3.2.
TABLE 3.2 Transformation to normal.

Time to fail ln(time to fail)
115 4.7449
155 5.0434
183 5.2095
217 5.3799
The mean of the transformed data is
4.7449 + 5.0434 + 5.2095 + 5.3799

µ= = 5.0944
4
The sum of the second column of Table 3.2 is 20.3777. If each value in the second
column of Table 3.2 is squared, the sum of the squared values is 104.0326. Using these
values, the sample standard deviation of the values in the second column of
Table 3.2 is
(∑ x )
2
n ∑ i =1xi2 −
n n
i =1 4 (104.0326 ) − 20.37772
s= = = 0.2708
n ( n − 1) 4 ( 4 − 1)
An easier method of computing the standard deviation is to use the Microsoft

Excel function
= STDEV.S ( )
The reliability at 100 hours is
 ln (100 ) − 5.0944 
R (100 ) = 1 − Φ   = 1 − Φ ( −1.807 )
 0.2708 
From Table A.1 of Appendix A, the standard normal cumulative distribution

function at z = −1.807 is 0.0354. Thus the reliability at 100 hours is
R (100 ) = 1 − 0.0354 = 0.9646
The reliability at 100 hours can also be computed using the Microsoft
Excel function
= 1 − LOGNORM.DIST (100,5.0944, 0.2708,1)
The location parameter, or log mean, is often mistaken for the mean of the
lognormal distribution. The mean of the lognormal distribution can be computed
from its parameters
 σ2 
 µ + 
2 
mean = e 

The variance of the lognormal distribution is
variance = e ( ( ) ) (e
2 µ + σ2 σ2
)
−1
The lognormal hazard function has a unique behavior; it increases initially, then
decreases, and eventually approaches zero. This means that items with a lognormal
distribution have a higher chance of failing as they age for some period of time, but
after survival to a specific age, the probability of failure decreases as time increases.
Figure 3.13 shows the lognormal hazard function.
In Figure 3.13, when σ = 2, the hazard function increases so quickly that it cannot
be seen on the graph.
FIGURE 3.13 The lognormal hazard function.
Exponential Distribution
The exponential distribution is used to model items with a constant failure rate,
usually electronics. The exponential distribution is closely related to the Poisson
distribution. If a random variable x is exponentially distributed, then the reciprocal
1
of x, y = , follows a Poisson distribution. Likewise, if x is Poisson distributed, then
1 x
y = is exponentially distributed. Because of this behavior, the exponential distribu-
x
tion is usually used to model the mean time between occurrences, such as arrivals
or failures, and the Poisson distribution is used to model occurrences per interval,
such as arrivals, failures, or defects.
The exponential probability density function is
f ( x ) = λe − λ x , x > 0 (3.16)
where λ is the failure rate.

The exponential probability density function (Figure 3.14) is also written as
1 −x
f ( x ) = e θ , x > 0 (3.17)
θ
1
From these equations, it can be seen that λ = . The variance of the exponential
distribution is equal to the mean squared, θ
1
σ 2 = θ2 = (3.18)
λ2
FIGURE 3.14 Exponential probability density function.
The exponential reliability function is

x
−
R( x ) = e θ
= e − λx , x > 0 (3.19)
The exponential hazard function is
1
h(x ) = = λ (3.20)
θ
Figure 3.15 shows the exponential hazard function.
Example 3.7: A resistor has a constant failure rate of 0.04 per hour. What is the
reliability of the resistor at 100 hours? If 100 resistors are tested, how many would
be expected to be in a failed state after 25 hours?
Solution: The reliability at 100 hours is
R ( 100 ) e=( ) −0.04 100

= 0.0183
The probability of failing before 25 hours is given by the cumulative distribution

function, which is equal to 1 minus the reliability function,
F ( 25 ) =
1 − R ( 25 ) =
0.04( 25 )
1−e = 0.632
1 − 0.368 =
The expected number of resistors in a failed state is 100(0.632) = 63.2.

FIGURE 3.15 Exponential hazard function.
The exponential distribution is characterized by its hazard function, which is

constant. Because of this, the exponential distribution exhibits a lack of memory, that
is, the probability of survival for a time interval, given survival to the beginning of
the interval, is dependent only on the length of the interval and not on the time of
the start of the interval. For example, consider an item that has a mean time to fail
of 150 hours that is exponentially distributed. The probability of surviving through
the interval 0 to 20 hours is
20
−
R ( 20 ) = e 150
= 0.8751
The probability of surviving the interval 100 hours to 120 hours is equal to
120
−
R (120 ) e 150
0.4493
R (120, given survival to t = 100 ) = = = = 0.8751
R (100 ) −
100
0.5134
e 150
The mean of the exponential distribution is estimated using the expression
T
θ= (3.21)
r
where T represents the total test time for all items, both failed and unfailed.
A (1 − α)% confidence limit for θ is
2T 2T
≤θ≤ 2 (3.22)
X α2 , 2r + 2 X 1− α , 2 r
Equation 3.22 assumes time censoring. If the data are failure censored, the critical
value of the chi-square statistic has 2r degrees of freedom instead of 2r + 2,
2T 2T
2
≤θ≤ 2 (3.23)
X a ,2 r X 1− a , 2 r
Confidence limits for reliability can be found by substituting the lower and upper
limits for θ into the exponential reliability function.
Example 3.8: Five items were tested with failures after 43 hours, 57 hours, and 80
hours. The remaining two items were removed from testing after 80 hours without
failing. Find the 90% confidence interval for reliability at 130 hours.
Solution: Because the items that did not fail were removed from testing at the
same time as the last failure, this is failure censoring. To determine the confidence
limits for reliability, the confidence limits for the mean must be determined first. The
critical value of the chi-square distribution with α = 0.05 and 6 degrees of freedom
is 2.592. This is found from Appendix A or from the Microsoft Excel expression
= CHISQ.INV ( 0.05,6 )
The critical value of the chi-square distribution with α = 0.95 and 6 degrees of
freedom is 1.635. This is found from Appendix A or from the Microsoft Excel
expression
= CHISQ.INV ( 0.95,6 )
The 90% confidence interval for the mean is
2 ( 43 + 57 + 80 + 80 + 80 ) 2 ( 43 + 57 + 80 + 80 + 80 )
≤θ≤
12.592 1.635
54.0 ≤ θ ≤ 415.9
The 90% confidence interval for reliability is

130 130
− −
e 54.0
≤e 415.9

0.090 ≤ 0.732
Example 3.9: Five items were tested with failures after 43 hours, 57 hours, and
80 hours. The remaining two items survived for 100 hours without failing. Find the
90% confidence interval for reliability at 130 hours.
Solution: Because the items that did not fail were not removed from testing at
the same time as the last failure, this is time censoring. The critical value of the
chi-square distribution for the lower limit has 2r + 2 degrees of freedom, and the
critical value for the upper limit has 2r degrees of freedom. The critical value of
the chi-square distribution with α = 0.05 and 8 degrees of freedom is 15.507. The
critical value of the chi-square distribution with α = 0.95 and 6 degrees of freedom
is 1.635. The 90% confidence interval for the mean is
2 ( 43 + 57 + 80 + 100 + 100 ) 2 ( 43 + 57 + 80 + 100 + 100 )

≤θ≤
15.507 1.635
49.01 ≤ θ ≤ 464.8
The 90% confidence interval for reliability is

130 130
− −
e 49.01
≤e 464.8

0.0705 ≤ 0.756
Discrete Modeling Distributions

The Poisson, binomial, hypergeometric, and geometric distributions are used to model
discrete data. Time to fail is continuous data, but some situations call for discrete
data, such as tests that are either pass or fail.
Poisson Distribution
The Poisson distribution is used to model rates, such as rabbits per acre, defects per
unit, or arrivals per hour. The Poisson distribution is closely related to the exponential
1
distribution. If x is a Poisson-distributed random variable, then is an exponential
x
1
random variable. If x is an exponential random variable, then is a Poisson random
x
variable. For a random variable to be Poisson distributed, the probability of an occur-
rence in an interval must be proportional to the length of the interval, and the number
of occurrences per interval must be independent.
The Poisson probability density function is
e −µ µ x
p ( x ,µ ) = (3.24)
x!
The term p(x, μ) represents the probability of exactly x occurrences in an interval

having an average of μ occurrences. The mean and variance of the Poisson distribu-
tion are both equal to μ. The Poisson cumulative distribution function is simply the
sum of the Poisson probability density function from 0 to x,
x
e − µ µi
P ( x ,µ ) = ∑ (3.25)
i =0 i!
The cumulative Poisson distribution has been computed in Appendix A to elimi-

nate the need for tedious calculations, but most people prefer to use the Microsoft
Excel formula
= POISSON.DIST ( x ,µ,1)
Example 3.10: A complex software system averages 7 errors per 5,000 lines of
code. What is the probability of exactly 2 errors in 5,000 lines of randomly selected
lines of code?
Solution: The probability of exactly 2 errors in 5,000 lines of randomly selected
lines of code is
e −7 72
p (=
2,7 ) = 0.022
2!
code. What is the probability of exactly 3 errors in 15,000 lines of randomly selected
lines of code?
The average number of errors in 15,000 lines of code is
 7 
=µ =  ( 15,000 ) 21
 5,000 
Solution: The probability of exactly 3 errors in 15,000 lines of randomly selected

lines of code is
e −21 213
( 3,21)
p= = 0.00000117
3!
code. What is the probability of fewer than 3 errors in 2,500 lines of randomly
selected lines of code? What is the probability of more than 2 errors in 2,500 lines
of randomly selected lines of code?
Solution: The average number of errors in 2,500 lines of code is μ = 3. The prob-
ability of fewer than 3 defects is equal to the probability of exactly 0 defects plus
the probability of exactly 1 defect plus the probability of exactly 2 defects. Entering
the cumulative Poisson table in Appendix A with r = 2 and μ = 3 gives the probability
of 2 or fewer defects, which is 0.4232. This value can also be computed manually.
The same solution is found using the Microsoft Excel formula
= POISSON.DIST ( 2,3,1)
The “1” at the end of this formula gives the cumulative Poisson.
The probability of more than 2 errors is equal to the probability of exactly 3 plus
the probability of exactly 4 plus the probability of exactly 5, and so forth. A simpler
approach is to consider that the probability of more than 2 errors is equal to 1 minus
the probability of 2 or fewer errors. Thus, the probability of more than 2 errors is
0.5768
1 − 0.4232 =
Binomial Distribution
The binomial distribution is used to model situations having only two possible
outcomes, usually labeled as success or failure. For a random variable to follow a
binomial distribution, the number of trials must be fixed, and the probability of
success must be equal for all trials. The binomial probability density function is
n
p ( x ,n,p ) =   p x (1 − p ) (3.26)
n−x

x
where p(x, n, p) is the probability of exactly x successes in n trials with a probability
of success equal to p on each trial. Note that
n n!
 = (3.27)
 x  x !( n − x )!
This notation is referred to as “n choose x” and is equal to the number of combi-
nations of size x made from n possibilities. This function is found on most calculators.
The binomial cumulative distribution function is
x
n
P ( x ,n ,p ) = ∑   p i (1 − p ) (3.28)
n −i

i =0  i 
where P(x, n, p) is the probability of exactly x or fewer successes in n trials with a

probability of success equal to p on each trial.
The mean and variance of the binomial distribution are respectively
µ = np (3.29)
σ 2 = np (1 − p ) (3.30)
Example 3.13: The probability of a salesman making a successful sales call is 0.2,
and 8 sales calls are made in a day. What is the probability of making exactly two
successful sales calls in a day? What is the probability of making more than two
successful sales calls in a day?
Solution: The probability of exactly 2 successes in 8 trials is
8
p ( 2,8,0.2=
)   0.22 ( 1 − 0.2) = 0.2936
8−2
2
 
This solution is found using the Microsoft Excel function
= BINOM.DIST ( 2,8,0.2,0 )
The probability of more than two successes is equal to one minus the probability
of two or fewer successes. The binomial probability density function can be used
to compute the probability of exactly 0 successes and the probability of exactly 1
success, but it is easier to use the Microsoft Excel function
= BINOM.DIST ( 2,8,0.2,1)
The “1” at the end of this function returns the cumulative binomial distribution.
The probability of 2 or fewer successes is 0.7969. The probability of more than 2
successes is
1 − 0.7969 =0.2031
Before electronic spreadsheets were common, the Poisson distribution was used to
approximate the binomial distribution because Poisson tables were more accommo-
dating than binomial tables. This approximation is now useless; why approximate a
value when you can get an exact answer? The requirement for a valid approximation
is that p must be small and n must be large. The approximation is done by using np
as the mean of the Poisson distribution.
Hypergeometric Distribution
The hypergeometric distribution is similar to the binomial distribution. Both are
used to model the number of successes, given the following:
•• A fixed number of trials
•• Two possible outcomes on each trial
The difference is that the binomial distribution requires the probability of success
to be the same for all trials, whereas the hypergeometric distribution does not. Consider
drawing from a deck of cards. If five cards are drawn, the probability of getting exactly
two hearts can be computed using the binomial distribution if after each draw the card
is replaced in the deck and the cards are shuffled. By replacing the cards and shuffling,
13
the probability of getting a heart on each of the 5 draws remains fixed at . If each
52
card is not replaced after it is drawn, the probability of getting a heart on the first draw
13
is , but the probability of getting a heart on the second draw is dependent on the
52
outcome of the first draw. If the first draw resulted in a heart, the probability of getting
a heart on the second draw is 12/51. If the first draw did not result in a heart, the prob-
ability of getting a heart on the second draw is 13/51. The hypergeometric distribution
is used to model this situation. This is also why the hypergeometric distribution is
referred to as the distribution that models sampling without replacement.
The hypergeometric probability density function is
 m  N − m 
  
x n−x 
p ( x ,N ,n,m ) =    (3.31)
N 
 
n
where p(x, N, n, m) is the probability of exactly x successes in a sample of n drawn

from a population of N containing m successes. The hypergeometric cumulative
distribution function is
 m  N − m 
  
x
i n −i 
P ( x ,N ,n ,m ) = ∑   
i =0 N  (3.32)
 
n
The mean and the variance of the hypergeometric distribution are respectively
nm
µ= (3.33)
N
 nm   m   N − n 
σ2 =  1 −    (3.34)
 N  N  N − 1 
Example 3.14: Fifty items are submitted for acceptance. If it is known that there
are four defective items in the lot, what is the probability of finding exactly one
defective item in a sample of five? What is the probability of finding fewer than two
defective items in a sample of five?
Solution: For this example,
x = 1
N = 50
n = 5
m = 4
The probability of finding exactly 1 defective item in a sample of 5 is
 4  50 − 4 
  
p ( 1,50,5,4 ) =
=
1  5 − 1  (=
4 )( 163,185 )
0.30808
 50  2,118,760
 
 5
This answer can be found using the Microsoft Excel function
= HYPGEOM.DIST ( 1,5,4,50 )
The probability of finding fewer than 2 defective items in a sample of 5 is equal to the
probability of exactly 0 plus the probability of exactly 1. The probability of exactly 0 is
 4  50 − 4 
  
p ( 0,50,5,4 ) =
=
0  5 − 0  ( 1)(=
1,370,754 )
0.64696
 50  2,118,760
 
 5
The probability of finding fewer than 2 defective items in a sample of 5 is
0.95504
0.30808 + 0.64696 =
This can also be found using the cumulative form of the hypergeometric distribu-
tion in Excel.
The binomial distribution can be used to approximate the hypergeometric distri-

bution when the population is large with respect to the sample size. When N is much
larger than n, the change in the probability of success on a single trial is too small to
significantly affect the results of the calculations. Again, it is silly to use approxima-
tions when exact solutions can be found with electronic spreadsheets. These approxi-
mations were useful for the engineer toiling with a slide rule but are of little use now.
Geometric Distribution
The geometric distribution is similar to the binomial distribution, in that the prob-
ability of occurrence is constant from trial to trial and the trials are independent. The
binomial distribution models situations where the number of trials is fixed, and the
random variable is the number of successes. The geometric distribution requires
exactly 1 success, and the random variable is the number of trials required to obtain
the first success. The geometric distribution is a special case of the negative binomial
distribution. The negative binomial distribution models the number of trials required
to obtain m successes, and m is not required to be equal to 1.
The geometric probability density function is
( x −1)
p ( x ,p ) = p (1 − p ) (3.35)
where p(x, p) is the probability that the first success occurs on the xth trial given a
probability of success on a single trial of p. The probability that more than n trials
are required to obtain the first success is
p ( x > n ) = (1 − p ) (3.36)
n

The mean and variance of the geometric distribution are, respectively,
1
µ = (3.37)
p
1− p
σ2 = (3.38)
p2
Example 3.15: The probability of an enemy aircraft penetrating friendly airspace

is 0.01. What is the probability that the first penetration of friendly airspace is
accomplished by the 80th aircraft to attempt the penetration of friendly airspace?
Solution: The probability that the first success occurs on the 80th trial with a prob-
ability of success of 0.01 on each trial is
( 80−1)
p ( 80,0.01) =
0.01( 1 − 0.01) 0.0452
=
Example 3.16: The probability of an enemy aircraft penetrating friendly airspace

is 0.01. What is the probability that it will take more than 80 attempts to penetrate
friendly airspace?
Solution: The probability that it will take more than 80 attempts to penetrate
friendly airspace with a probability of success of 0.01 on each trial is
p ( x > 80 ) =−
( 1 0.01) =
80
0.4475
Identifying the Correct Discrete Distribution

It is often difficult to determine the correct discrete distribution when faced with a
statistical problem. To make this task easier, ask the following questions.
1. Is a rate being modeled, such as defects per car or rabbits per acre, and is
there no upper bound on the number of possible occurrences? (For example,
is it conceivable that there could be 10,000 defects on a car? There is no upper
limit imposed by a specific number of trials. If there are 5 trials, the upper
limit on the number of successes is 5.) If the answer is yes, the Poisson
distribution is probably the appropriate distribution. If the answer is no, go
to question 2.
2. Is there a fixed number of trials? If yes, go to question 3. If no, is there a fixed
number of successes, with the number of trials being the random variable? If
the answer is yes, then if the number of successes is fixed at 1, use the
geometric distribution; if the number of successes is fixed at a value greater
than 1, use the negative binomial distribution.
3. Is the probability of success the same on all trials? If yes, use the binomial
distribution; if no, use the hypergeometric distribution.
The flowchart shown in Figure 3.16 summarizes these questions.
FIGURE 3.16 Choosing a discrete distribution.

Sampling Distributions
The chi-square, t-, and F-distributions are formed from combinations of random
variables. Because of this, they generally are not used to model physical phenomena,
such as time to fail, but are used to make decisions and to construct
confidence intervals.
Chi-Square Distribution
The chi-square distribution is formed by summing the square of standard normal
random variables. For example, if z is a standard normal random variable, then
y = z12 + z 22 + z 32 + + z n2 (3.39)
is a chi-square random variable with n degrees of freedom.

A chi-square random variable is also created by summing two or more chi-square
random variables. A distribution having this property is regenerative. The chi-square
distribution is a special case of the gamma distribution with a failure rate of 2, and
degrees of freedom equal to 2 divided by the number of degrees of freedom for the
corresponding chi-square distribution.
The chi-square probability density function is
v  x
 −1  −
x  2 e 2
f (x ) = v , x > 0 (3.40)
v 
2 Γ 
2
2
where ν is the degrees of freedom, and Г(x) is the gamma function. Figure 3.17 shows
the chi-square probability density function.
The critical value of the chi-square distribution is given in Appendix A, but it is
easier to use the Microsoft Excel function
= CHISQ.INV ( probability,degrees − freedom )
Example 3.17: A chi-square random variable has 7 degrees of freedom. What is the
critical value if 5% of the area under the chi-square probability density is desired
in the right tail?
Solution: When hypothesis testing, this is commonly referred to as the critical value
with 5% significance or α = 0.05. From Appendix A, this value is 14.067. This value
is also found using the Microsoft Excel function
= CHISQ.INV ( 0.05,7 )
FIGURE 3.17 Chi-square probability density function.
t-Distribution
The t-distribution is formed by combining a standard normal random variable and
a chi-square random variable. If z is a standard normal random variable, and X 2 is a
chi-square random variable with ν degrees of freedom, then a random variable with
a t-distribution is
z
t=
X 2 (3.41)

v
The t-distribution is equivalent to the F-distribution with 1 and ν degrees of
freedom. The t-distribution is commonly used for hypothesis testing and constructing
confidence intervals for means. It is used in place of the normal distribution when
the standard deviation is unknown. The t-distribution compensates for the error in
the estimated standard deviation. If the sample size is large, n > 100, the error in the
estimated standard deviation is small, and the t-distribution is approximately normal.
The t-probability density function is
 ( ν + 1) 
Γ  −
 ν +1 
 2   x 2   2 
f (x ) = 1 +  , − ∞ < x (3.42)
ν ν 
Γ   Πν 
2
FIGURE 3.18 The t-probability density function.
where ν is the degrees of freedom. Figure 3.18 shows the t-probability density function
The mean and variance of the t-distribution are respectively
µ = 0 (3.43)
v
σ2 = , v ≥ 3 (3.44)
v −2
From a random sample of n items, the probability that
x −µ
t=
s (3.45)
n
falls between any two specified values is equal to the area under the t-probability
density function between the corresponding values on the x-axis with n − 1 degrees
of freedom.
Example 3.18: The burst strength of 15 randomly selected seals is given below.
What is the probability that the average burst strength of the population is greater
than 500?
480 489 491 508 501

500 486 499 479 496
499 504 501 496 498
Solution: The mean of these 15 data points is 495.13. The sample standard devia-
tion of these 15 data points is 8.467. The probability that the population mean is
greater than 500 is equal to the area under the t-probability density function, with
14 degrees of freedom, to the left of
495.13 − 500
t= = −2.227
8.467
15
From Appendix A, the area under the t-probability density function, with 14 degrees
of freedom, to the left of −2.227, is 0.0215. This value must be interpolated using
Appendix A but can be computed directly in Microsoft Excel using the function
= T.DIST ( −2.227,14,1)
Simply stated, making an inference from the sample of 15 data points, there is a
2.15% that the true population mean is greater than 500.
F-Distribution
If X is a chi-square random variable with ν1 degrees of freedom, Y is a chi-square
random variable with ν2 degrees of freedom, and X and Y are independent, then
X
ν1
F = (3.46)
Y
ν2
is F-distributed with ν1 and ν2 degrees of freedom.

The F-distribution is used extensively to test for equality of variances from two
normal populations. If U and V are the variances of independent random samples of
size n and m taken from normally distributed populations with variances of w and z, then
F = w (3.47)
V
z
FIGURE 3.19 The F-probability density function.
is a random variable with an F-distribution with ν1 = n − 1 and ν2 = m − 1. The

F-probability density function is
 ν1
 
 Γ  ν1 + ν 2   ν1  2   ν1  
  2   ν 2    −1 
x 2  
f (x ) =   ( ν1 + ν2 )
 , x > 0 (3.48)
 Γ  ν1  Γ  ν 2   
      
 1 + ν x  2

 2  2  
1
 
   ν2  
Figure 3.19 shows the F-probability density function.

The F-cumulative distribution function is given in Appendix A. Only the upper
tail is given, but the other tail can be computed using the expression
1
Fα ,v1 ,v2 = (3.49)
F1−α ,v2 ,v1
It is easier to find critical values of the F-distribution using the Microsoft

Excel function
= FINV (α ,v1 ,v 2 )
Example 3.19: Find F0.05 with ν1 = 9 and ν2 = 10, and find F0.95 with ν1 = 10 and ν2 = 9.
Solution: From Appendix A, F0.05 with ν1 = 9 and ν2 = 10 is 3.02. F0.95 with ν1 = 10 and
ν2 = 9 is equal to the inverse of F0.05 with ν1 = 9 and ν2 = 10. Thus
1 1
=
F0.95,10,9 = = 0.331
F0.05,9,10 3.02
The following Microsoft Excel functions provide the solutions respectively
= F.INV.RT ( 0.05,9,10 )
= FI.NV.RT ( 0.95,10,9 )
Bayes’ Theorem
Previous discussions of probability distributions in this chapter assumed no prior
knowledge of events. The event being modeled, such as time to fail, was not influenced
by the knowledge of some other event. Bayes’ theorem is based on a prior distribution
and a posterior distribution.
Consider the life of a human being. If a person is selected at random, what is the
probability that this person will live to be 80 years old? Now if it is known that this
randomly selected person has heart disease, the probability that this person survives
to be 80 years old changes. With no knowledge of the person, estimates of life expec-
tancy are based on the prior distribution. After the influence of the knowledge that
the person has heart disease is considered, estimates are made using the
posterior distribution.
For example, an item has a normally distributed time-to-fail distribution with
a mean of 800 hours and a standard deviation of 200. What is the probability of
an item surviving for 900 hours? Given an item has survived for 850 hours, what
is the probability of surviving for 900 hours? The first question is answered using
the prior distribution, and the second question is answered using the posterior
distribution. Figure 3.20 shows the prior and posterior probability density functions.
Figure 3.21 shows the prior and posterior reliability functions.
Given a prior probability density function of f(x) and survival until x0, the poste-
rior probability density function is
fprior ( x )
f posterior ( x ) = (3.50)
R ( x0 )
For the previous example, the probability of survival for 850 hours is equal to the
area under the standard normal probability density function to the right of z = 0.25,
x − µ 850 − 800
z= = = 0.25
σ 200
FIGURE 3.20 Prior and posterior probability density functions.
FIGURE 3.21 Prior and posterior reliability functions.

From Appendix A, the area under the standard normal probability density
function to the right of z = 0.25 is 0.4013.2 Thus, given the survival for 850 hours, the
posterior probability density function for the previous example is
φ (x)
f posterior ( x ) =

0.4013
where ϕ(x) is the standard normal probability density function.
Given a prior reliability function of f(x) and survival until x0, the posterior reli-
ability function is
Rprior ( x )
Rposterior ( x ) = (3.51)
R ( x0 )
Stated another way, if an item has survived until x0, the probability of survival
until x1 is
R ( x1 )
R ( x 1 \x 0 ) = (3.52)
R (x0 )
Continuing with the example, the probability of survival for 850 hours is 0.4013.
The probability of survival for 900 hours given survival for 850 hours is equal to the
reliability at 850 hours divided by the reliability at 900 hours. The probability of
survival for 900 hours is equal to the area under the standard normal probability
density function to the right of z = 0.5,
x − µ 900 − 800
z= = = 0.5
σ 200
From Appendix A, the area under the standard normal probability density
function to the right of z = 0.5 is 0.3085. Thus, given the survival for 850 hours, the
probability of survival for 900 hours is
R ( 900 ) 0.3085
R ( 900\800 ) = = = 0.76
R ( 850 ) 0.4013
Example 3.20: A resistor has an exponential time-to-fail distribution with a failure

rate of 0.02 hours. A resistor has survived for 100 hours. What is the probability of
survival for 200 hours?
Solution: The probability of survival for 200 hours, given the survival for 100 hours, is
R ( 200 ) e −(0.02)( 200) 0.0183156
R ( 200=
\ 100 ) = = = 0.1353
R ( 100 ) e −(0.02)( 100) 0.135335
Note that the probability of survival in the interval from 0 to 100 hours is equal to
the probability of survival in the interval from 100 hours to 200 hours, given the
survival for 100 hours. This is known as the “lack of memory” property, which is
unique to the exponential and Poisson distributions.
2
This value can be found using the Microsoft Excel function = −NORM.DIST (850,800,200,1).
A more general form of Bayes’ theorem is
P ( Ai ) P ( B \ Ai )
P ( Ai \B ) = (3.53)
∑ i =1P ( Ai ) P ( B \ Ai )
n
Example 3.21: Three jars each contain three coins. Jar 1 contains three quarters; Jar
2 contains one quarter, one nickel, and one penny; and Jar 3 contains three pennies.
If a coin is drawn from a randomly selected jar, what is the probability the coin was
drawn from Jar 3 if the selected coin was a penny?
Solution: The probability of selecting any of the three jars with no knowledge of
1
the coin drawn is . Given the coin was selected from Jar 1, the probability of the
3
coin being a penny is 0; there are no pennies in Jar 1. Given the coin was selected
1
from Jar 2, the probability of the coin being a penny is ; there are three coins in
3
Jar 2, and 1 of the 3 coins is a penny. Given the coin was selected from Jar 3, the
probability of the coin being a penny is 1; all three coins in Jar 3 are pennies. The
probability that the coin was drawn from Jar 3, given the coin was a penny, is
1
 3  ( 1) 3
P ( Jar
= 3\apenny was drawn) =  
1  1  1   1  4
 3  ( 0 ) +  3  3  +  3  ( 1)
      
Summary
Distributions can be used to model (Weibull, exponential, normal, lognormal,
Poisson, binomial, hypergeometric, geometric) or for statistical inference (normal, t,
chi-square, F).
It is important to verify that the correct distribution is being used. Improper
distribution choice will probably result in errors. The errors may be small or large,
depending on the situation. Understand the statistical distributions, and verify the
distribution fit and any assumptions before proceeding.
4
Parameter Estimation
M
any methods are available for parameter estimation. In reliability engineering,
the most popular methods are the following:
•• Maximum likelihood estimation

•• Probability plotting
•• Hazard plotting
It is desirable for a parameter estimator to have the following properties:
1. Lack of Bias—If the expected value of the estimator is equal to the true value
of the parameter, it is said to be unbiased.
2. Minimum Variance—The smaller the variance of the estimate the smaller
the sample size required to obtain the level of accuracy desired, and the more
efficient the estimator. The most efficient estimator is the estimator with
minimum variance.
3. Consistency—As the sample size is increased, the value of the estimated
parameter becomes closer to the true value of the parameter.
4. Sufficiency—The estimator uses all information available in the data set.
The method used to estimate parameters depends on the type of data or testing
involved and the distribution of interest. In addition to these parameter estimation
methods, this chapter describes censored data and presents methods of parameter
estimation for the exponential, normal, lognormal, and Weibull distributions.
Methods are presented for complete and censored data.
Maximum Likelihood Estimation

Maximum likelihood is the most widely used method of generating estimators. It is
based on the principle of determining the parameter(s) value(s) that maximize(s) the
probability of obtaining the sample data.
The likelihood function for a given distribution is a representation of the prob-
ability of obtaining the sample data. Let x1, x2,…, xn be independent random variables
from the probability density function f(x, θ), where θ is the single distribution param-
eter. Then
L ( x 1 ,x 2 ,¼,x n ;q ) = f ( x 1 ,q ) f ( x 2 ,q )¼f ( x n ,q ) (4.1)
is the joint distribution of the random variables, or the likelihood function. The
maximum likelihood estimate, q̂, maximizes the likelihood function. This estimate
is asymptotically normal. Often, the natural logarithm of the likelihood function is
maximized to simplify computations.
The variances of the estimates can be found by inverting the matrix of the negative
of the second partial derivatives of the likelihood function, also known as the local
information matrix. These estimates are asymptotically normal, and the variances
obtained from the local information matrix are used to calculate the confidence intervals.
Probability Plotting
Probability plotting is a graphical method of parameter estimation. For the assumed
distribution, the cumulative distribution function is transformed to a linear expres-
sion, usually by a logarithmic transformation, and plotted. If the plotted points form
a straight line, the assumed distribution is acceptable, and the slope and the intercept
of the plot provide the information needed to estimate the parameters of the distribu-
tion of interest. The median rank is usually used to estimate the cumulative distribu-
tion function, although there are several alternatives such as the mean rank and the
Kaplan-Meier product limit estimator.
In manually constructing a probability plot, a distribution-specific probability
paper is required. By using a probability paper, the failure times and cumulative
distribution function estimates can be plotted directly. With the power of personal
computers and electronic spreadsheets, a specialized graph paper is no longer needed
because the necessary transformations can be made quickly and easily.
Hazard Plotting
Hazard plotting is a graphical method of parameter estimation. The cumulative
hazard function is transformed to a linear expression, usually by a logarithmic trans-
formation, and plotted. The slope and the intercept of the plot provide the information
needed to estimate the parameters of the distribution of interest.
In manually constructing a hazard plot, a distribution-specific hazard paper is

required. By using a hazard paper, the failure times and cumulative hazard function
estimates can be plotted directly. With the power of personal computers and electronic
spreadsheets, a specialized graph paper is no longer needed because the necessary
transformations can be made quickly and easily.
The simplest method of parameter estimation for the exponential distribution is the
method of maximum likelihood. Maximum likelihood provides an unbiased estimate
but no indication of goodness of fit. Graphical methods, although more involved,
provide a visual goodness-of-fit test. Often, graphical methods will be used in conjunc-
tion with maximum likelihood estimation.

The exponential probability density function is
1 -x
f (x) = e q , x ³ 0 (4.2)
q
The maximum likelihood estimation for the parameter θ is
å
n
x
q= i =1 i
(4.3)
r
where
xi is the ith data point (this may be a failure or a censoring point)
n is the total number of data points (both censored and uncensored)
r is the number of failures
This estimate is unbiased and is the minimum variance estimator.
Example 4.1: The cycles to fail for seven springs are

30,183 14,871 35,031 76,321
43,891 31,650 12,310
Assuming an exponential time-to-fail distribution, estimate the mean time to fail
and the mean failure rate.
Solution: The mean time to fail is
ˆq = 30,183 + 14,871 + 35,031 + 76,321 + 43,891 + 31,650 + 12,310 = 244,257 = 34,893.9 cycles
7 7
The mean failure rate is the inverse of the mean time to fail,
1
ˆ=
l = 0.0000287 failures per cycle
34,893.9
Example 4.2: Assume the data in Example 4.1 represent the cycles to fail for seven
springs, but an additional 10 springs were tested for 80,000 cycles without failure.
Estimate the mean time to fail and the mean failure rate.
Solution: The mean time to fail is
244,257 + 10 ( 80,000 )
ˆ=
q = 149,179.6 cycles
7
The mean failure rate is
1
l= = 0.0000067 failures per cycle
149,179.6
For a time truncated test, a confidence interval for θ is
2å i = 1 x i 2å i = 1 x i
n n
£q £ (4.4)
x(2a /2,2 r +2) x(12 -a /2,2 r )
Note that the X 2 degrees of freedom differ for the upper and lower limits.
Example 4.3: Fifteen items were tested for 1,000 hours. Failures occurred at times
of 120 hours, 190 hours, 560 hours, and 812 hours. Construct a 90% confidence
interval for the mean time to fail and the failure rate.
Solution: This is a time truncated test. The mean life estimate is
120 + 190 + 560 + 812 + 11( 1,000 ) 12,682

ˆ=
q = = 3,170.5
4 4
For a 90% confidence interval,
a = 0.1
X (20.05,10) = 18.307
X (20.95,8) = 2.733
The 90% confidence interval for θ is
2 ( 12,682 ) 2 ( 12,682 )
£q£
18.307 2.733
1,385.5 £ q £ 9,280.6
The confidence interval for the failure rate is the inverse of the confidence interval
for the mean time to fail,
1 1
£l£
9,280.6 1,385.5
0.0001077 £ l £ 0.0007217
For a failure truncated test and multiple-censored data, a confidence interval for θ is
2å i =1xi 2å i =1xi
n n
£q£ (4.5)
c2æ a ö
c2æ a ö
ç ,2r ÷ ç 1- ,2r ÷
è2 ø è 2 ø
Note that the X 2 degrees of freedom are the same for the upper and lower limits.
Example 4.4: Twelve items were tested with failures occurring at times of 43 hours,
67 hours, 92 hours, 94 hours, and 149 hours. At a time of 149 hours, the testing
was stopped for the remaining seven items. Construct a 95% confidence interval
for the mean time to fail.
Solution: This is a failure truncated test. The mean life estimate is
43 + 67 + 92 + 94 + 149 + 7 ( 149 ) 1,488

ˆ=
q = = 297.6
5 5
For a 95% confidence interval,

a = 0.05
X (20.025,10) = 20.483
X (20.975,10) = 20.483
The 95% confidence interval for θ is
2 ( 1,488 ) 2 ( 1,488 )
£q£
20.483 3.247
145.3 £ q £ 916.5
For failure-free testing, the one-sided lower confidence limit simplifies to

-nt
£ q (4.6)
ln a
where
t is the testing time
ln is the natural logarithm
α is the significance (α = 0.05 for a 95% limit)
Example 4.5: Twenty items are tested for 230 hours without failure. Determine a
90% lower confidence limit for θ.
Solution:
-20 ( 230 )
= 1,997.8hours
ln ( 0.1)
A confidence interval for reliability is

x x
- -
e qL
£ R (x ) £ e qU
(4.7)
where
θL is the lower confidence limit for the mean time to fail
θU is the upper confidence limit for the mean time to fail
A confidence interval for percentiles is
-qL ln ( 1 - P ) £ x £ -qU ln ( 1 - P ) (4.8)
where P is the probability of failure prior to time = x.
Example 4.6: Twenty items are tested for 230 hours without failure. Determine a
90% lower confidence limit for reliability at time = 1,000.
Solution: The lower 90% confidence limit for the mean is
-20 ( 230 )
= 1,997.8hours
ln ( 0.1)
The lower 90% confidence limit for reliability at time = 1,000 is
1,000
-
RL ( 1,000 ) = e 1,997.8
= 0.606
Hazard Plotting
The exponential cumulative hazard function is
x
H ( x ) = (4.9)
q
If a data set is exponentially distributed, a plot of the exponential cumulative

1
hazard function yields a linear fit with a zero intercept and a slope of . To construct
q
a hazard plot, an estimate for H(x) is needed. The cumulative hazard function is
estimated by the cumulative of the inverse of the reverse ranks. For a data set of
n points ordered from smallest to largest, the first point has a rank of n, the second
n − 1, and so forth.
Example 4.7: Construct a hazard plot for the following failure data, given that an
additional 7 items were tested for 149 cycles without failure:
43,67,92,94,149
Solution: Table 4.1 shows the calculations for the values necessary for the hazard
plot. Figure 4.1 shows the plot. Note that censored data points are not plotted. The
slope of the best-fit straight line through the data with an intercept of 0 is 0.00333.
The estimated mean is
1
q= = 300.3
0.00333
TABLE 4.1 Tabulations for an exponential hazard plot.
Time to fail Reverse rank h(t) H(t)

43 12 0.0833 0.0833
67 11 0.0909 0.1742
92 10 0.1000 0.2742
94 9 0.1111 0.3854
149 8 0.1250 0.5104
149 c 7
149 c 6
149 c 5
149 c 4
149 c 3
149 c 2
149 c 1
c = Censored.
FIGURE 4.1 Hazard plot for exponential data.
The exponential cumulative distribution function is
x
-
F ( x ) = 1 - e q (4.10)
By manipulating this expression algebraically, it can be transformed to a

linear format,
1 x
ln = (4.11)
1- F (x) q
1 x
If a data set follows an exponential distribution, a plot of ln = versus
1- F (x) q
1
x will be linear with a 0 intercept and a slope of . Before a plot can be constructed,
q
an estimate for F(x) is needed. The cumulative distribution function, F(x), is usually
estimated from median rank, but other estimates such as the mean rank and the
Kaplan-Meier product limit estimator are also used. The median rank estimate for
F(x) is
O - 0.3
Fˆ ( x ) = i (4.12)
n + 0.4
where Oi is the modified order of failure of the ith data point.
A modified order of failure is needed only if censored data are involved; if not,
the original order of failure, i, is equivalent to the modified order of failure. The logic
for a modified order of failure is as follows. Consider three items: the first was tested
for three hours, and the test was stopped without failure; the second item was tested
and failed after four hours; and the third item was tested and failed after 4.5 hours.
For this data set, the failure order is unclear. The first item could have been either the
first failure, the second failure, or the third failure. Thus, it is not certain that the first
item to fail the second item is the first-order failure. The modified order of failure is
computed from the expression
Oi = Oi -1 + I i (4.13)
where Ii is the increment for the ith failure and is computed from the expression
( n + 1) - O p
Ii = (4.14)
1+ c
where
n is the total number of points in the data set (both censored and uncensored)
c is the number of points remaining in the data set (including the current point)
Op is the order of the previous failure
1
An alternative to plotting x versus ln on a conventional graph paper is
1- F (x)
to plot x versus F(x) on specialized probability paper. The advantage of probability
1
paper is that the values of ln do not have to be computed. With computers,
1- F (x)
this technique is obsolete.
Example 4.8: Construct a probability plot for the following failure data, given that
an additional 7 items were tested for 149 cycles without failure:
43, 67, 92, 94, 149
Solution: Table 4.2 contains the calculations necessary for plotting. Figure 4.2
shows the probability plot. The slope of the best-fit straight line through the origin
is 0.00304, which estimates the failure rate for the exponential distribution. The
mean of the distribution is
1
q= = 328.9
0.00304
TABLE 4.2 Tabulations for exponential probability plotting.
1 1
ln
Time to fail Oi Median rank, F(t) 1 - F ( t ) 1 - F (t )
43 1 0.0565 1.0598 0.0581
67 2 0.1371 1.1589 0.1475
92 3 0.2177 1.2784 0.2456
94 4 0.2984 1.4253 0.3544
149 5 0.3790 1.6104 0.4765
149 c
149 c
149 c
149 c
149 c
149 c
149 c
c = Censored.
FIGURE 4.2 Exponential probability plot.

Often, reliability confidence limits are added to probability plots. The upper and
lower confidence limits are approximated by 5% and 95% ranks. These ranks can
be taken from the expression
j
n - j +1
wa = (4.15)
j
F1- a ,2(n - j +1),2 j +
n - j +1
where
wα is the 100 (1 − α)% nonparametric confidence limit
j is the failure order
Fα, v1, v2 is the critical value from the F-distribution
When multiple-censored data are encountered, the modified failure orders will
not be integers, and the rank values will have to be interpolated. The rank values are
not plotted against the corresponding failure time. Any deviation of the failure time
from the best-fit straight line through the data is considered sampling error, and the
time against which the rank values are plotted is found by moving parallel to the
x-axis until the best-fit straight line is intersected. This plotting position is
æ 1 ö
xi = q ln çç ÷÷ (4.16)
è 1 - F ( xi ) ø
Normal Distribution
The normal probability density function is
1 é 1 æ x - m ö2 ù
f (x) = exp ê - ç ÷ ú , - ¥ < x < ¥ (4.17)
s 2p êë 2 è s ø úû
where
μ is the distribution mean
σ is the distribution standard deviation
If no censoring is involved, the distribution mean is estimated from the expression
å
n
x
i =1 i
m̂ = x = (4.18)
n
where n is the sample size.
If no censoring is involved, the distribution standard deviation is estimated from

the expression
(å x )
2
n å i =1xi2 -
n n
i =1 i
sˆ = (4.19)
n ( n - 1)
However, when censored data are involved, parameter estimation becomes

complicated. Three popular methods for parameter estimation for the normal distri-
bution when censored data are encountered are as follows:
1. Maximum likelihood estimation
2. Hazard plotting
3. Probability plotting
The following sections present each of these alternatives.

The maximum likelihood equations for the normal distribution are
¶L r é x - m k h ( x i ) ù
= ê +å ú = 0 (4.20)
¶m s ë s i =1 r û
¶L r é s + ( x - m ) z ( x i )h ( x i ) ù
2 2
k
= ê -1+ å ú = 0 (4.21)
¶s s êë s 2
i =1 r úû
where
k is the number of censored observations
x is the sample mean of the failures
s is the sample standard deviation for the failures
z(xi) is the standard normal deviate
xi - m
z ( xi ) =
s
h(xi) is the hazard function evaluated at the ith point
f ( z ( xi ) )
h ( xi ) =
s éë1 - F ( z ( xi ) ) ùû
where
ϕ(z(xi)) is the standard normal probability density function evaluated at the
ith point
Φ(z(xi)) is the standard normal cumulative distribution function evaluated at
the ith point
Note that if no censored data are involved, these expressions reduce to the sample
mean and the sample standard deviation.
Iterative techniques are necessary to these equations. A standard method based
on Taylor series expansions involves repeatedly estimating the parameters until a
desired level of accuracy is reached. Estimates of μ and σ are respectively given by
the expressions
mˆ i = mˆ i -1 + h (4.22)
sˆ i = sˆ i -1 + k (4.23)
where
h is a correction factor for the distribution mean
k is a correction factor for the distribution standard deviation
For each iteration, the correction factors are estimated from the expressions
¶2L ¶2L ¶L
h + k = (4.24)
¶m 2
¶m¶s ¶m
¶2L ¶ 2 L ¶L
h + k 2 = (4.25)
¶m¶s ¶s ¶s
where
¶2L r é k
Ai ù

¶m 2
=- 2
s ê å r ú (4.26)
1 +
ë i =1 û
¶2L r é 2 ( x - m ) k Bi ù
=- 2 ê + å ú (4.27)
¶m¶s s ë s i =1 r û
¶2L
é 2{
r ê3 s + ( x - m)
2
} k ù
Ci ú

¶s2
= -
s2 ê s2
- 1 + å
i =1 r
ú
(4.28)
êë úû
and
Ai = h ( xi ) éëh ( xi ) - z ( xi ) ùû (4.29)
Bi = h ( xi ) + z ( xi ) Ai (4.30)
Ci = z ( xi ) éëh ( xi ) + Bi ùû (4.31)
The estimated parameters are asymptotically normal. The variances of the esti-
mates can be found by inverting the local information matrix,
¶2L ¶2L
- -
¶m2 ¶m¶s
F=
¶2L ¶ 2 L (4.32)
- -
¶m¶s 2s 2
After inversion, the variances are
é var ( mˆ ) cov ( mˆ ,sˆ ) ù

F -1 = ê ú (4.33)
ëcov ( mˆ ,sˆ ) var ( sˆ ) û
Approximate (1 − α) 100% confidence intervals for the estimated parameters are
mˆ - K a var ( mˆ ) £ m £ mˆ + K a var ( mˆ ) (4.34)

2 2
æ K a var ( sˆ ) ö
sˆ ç ÷
£ s £ sˆ exp ç 2 ÷ (4.35)
æKa var ( sˆ ) ö è sˆ ø
ç ÷
exp ç 2 ÷
è sˆ ø
where K a is the inverse of the standard normal probability density function.

2
These confidence intervals are approximated but approach exactness as the
sample size increases. Confidence intervals for reliability can be found using
the expressions
æ var ( mˆ ) + zˆ 2 var ( sˆ ) + 2zˆ cov ( mˆ ,sˆ ) ö

var ( zˆ ) » çç ÷÷ (4.36)
è sˆ 2 ø
zˆ - K a var ( zˆ ) £ z £ zˆ + K a var ( zˆ ) (4.37)

2 2
æ ö æ ö
1 - F ç zˆ + K a var ( zˆ ) ÷ £ R ( x ) £ 1 - F ç zˆ - K a var ( zˆ ) ÷ (4.38)
è 2 ø è 2 ø
Confidence intervals for percentiles can be found using the expressions
var ( xˆ ) » var ( mˆ ) + zˆ 2 var ( sˆ ) + 2zˆ cov ( mˆ ,sˆ ) (4.39)
xˆ - K a var ( xˆ ) £ x £ xˆ + K a var ( xˆ ) (4.40)

2 2
Hazard Plotting
The normal cumulative hazard function is
é æ x - m öù
H ( x ) = - ln ê1 - F ç ÷ ú (4.41)
ë è s øû
where Ф(x) is the standard normal cumulative distribution function.

By rearranging Equation 4.41, the survival time can be represented as a function
of the cumulative hazard function,
x = m + sF -1 éë1 - e - H ( x ) ùû (4.42)
where Φ−1(x) is the inverse of the standard normal cumulative distribution function.
It can be seen that by plotting x versus Φ−1[1 − e−H(x)], the resulting y-intercept
equals μ and the resulting slope equals σ. The hazard function, h(x), is estimated from
the inverse of the reverse rank of the ordered failures, and H(x) is the cumulative of
the values of h(x). Censored data points are used to compute ranks but are not included
in hazard plots.
An alternative to plotting x versus Φ−1[1 − e−H(x)] on conventional graph paper is
to plot x versus H(x) on specialized hazard paper. The advantage of hazard paper is
that the values of Φ−1[1 − e−H(x)] do not have to be computed. Computers have made
this technique obsolete.
Time to fail
150 c 183 c 235
157 c 209 235
167 c 216 c 248 c
179 217 c 257
Example 4.9: Use hazard plotting to determine the parameters of the normal

distribution, given the following multiple censored data set. A “c” following an
entry indicates censoring.
Solution: Table 4.3 is constructed to obtain the necessary plotting data.
TABLE 4.3 Tabulations for a hazard plot for the normal distribution.
Time to fail Reverse rank h(t) H(t) 1 − e−H(t) Φ−1[1 −e−H(t)]

150 c 12
157 c 11
167 c 10
179 9 0.1111 0.1111 0.1052 −1.2527
183 c 8
209 7 0.1429 0.2540 0.2243 −0.7578
216 c 6
217 c 5
235 4 0.2500 0.5040 0.3959 −0.2640
235 3 0.3333 0.8373 0.5671 0.1691
248 c 2
257 1 1.0000 1.8373 0.8408 0.9976
c = Censored.
Now the five failure times can be plotted against Φ−1[1 − e−H(t)], as shown in
Figure 4.3. The best-fit straight line through the data points was found using linear
regression. The y-intercept of the best-fit straight line through the five points provides
an estimate of μ − 230.3 in this case, and the slope of the line provides an estimate
of σ − 32.9 in this case.
F -1 éë1 - e - H (t ) ùû
By rearranging the normal cumulative distribution function, a linear expression can
be obtained:
x = m + sF -1 éë F ( x ) ùû (4.43)
where
F(x) is the normal cumulative distribution function
Φ−1(x) is the inverse of the standard normal cumulative distribution function
It can be seen that by plotting x versus Φ−1[F(x)], the resulting y-intercept equals
μ, and the resulting slope equals σ. The cumulative distribution function, F(x), is usually
FIGURE 4.3 Hazard plot for the normal distribution.
estimated from the median rank, but other estimates such as the mean rank and the
Kaplan-Meier product limit estimator are also used. Median ranks are estimated using
the method detailed in the Exponential Probability Plotting section of this chapter.
An alternative to plotting x versus Φ−1[F(x)] on conventional graph paper is to
plot x versus F(x) on specialized probability paper. The advantage of probability paper
is that the values of Φ−1[F(x)] do not have to be computed. Computers have made this
technique obsolete.
Time to fail
150 c 183 c 235
157 c 209 235
167 c 216 c 248 c
179 217 c 257
Example 4.10: Use probability plotting to determine the parameters of the normal

distribution, given the following data set. A “c” following an entry indicates censoring.
Solution: Table 4.4 is constructed to obtain the data for plotting. These data are
plotted in Figure 4.4. The slope of this plot is 34.8, which is the estimated value of
σ. The y-intercept is 235.3, which is the estimated value of μ.
TABLE 4.4 Tabulations for a probability plot for normal data.
Median rank,
Time to fail Ii Oi F(x) Φ−1[F(x)]
150 c
157 c
167 c
179 1.3000 1.3000 0.0806 −1.4008
183 c
209 1.4625 2.7625 0.1986 −0.8467
216 c
217 c
235 2.0475 4.8100 0.3637 −0.3486
235 2.0475 6.8575 0.5288 0.0723
248 c
257 3.0713 9.9288 0.7765 0.7605
c = Censored.
FIGURE 4.4 Normal probability plot.

Nonparametric confidence limits for reliability can be added to a normal prob-

ability plot using 5% and 95% ranks, as explained in the Exponential Probability
Plotting section of this chapter. The plotting position for the standard normal inverse
of the 5% and 95% ranks is
x = m + sF -1 éë F ( x ) ùû (4.44)
Note that F(x) is the median rank.
Lognormal Distribution
The lognormal probability density function is
1 é 1 æ ln x - m ö2 ù
f (x) = exp ê - ç ÷ ú, x > 0 (4.45)
sx 2p êë 2 è s ø úû
where
μ is the location parameter
σ is the shape parameter
If x is a lognormal distributed random variable, then y = ln(x) is a normally
distributed random variable. The location parameter is equal to the mean of the
logarithm of the data points, and the shape parameter is equal to the standard devia-
tion of the logarithm of the data points. Thus the lognormal distribution does not
have to be dealt with as a separate distribution. By taking the logarithm of the data
points, the techniques developed for the normal distribution discussed in the previous
section can be used to estimate the parameters of the lognormal distribution.
The Weibull probability density function is
b( x - d)
b -1 b
æ x -dö
f (x) = exp ç ÷ , x ³ d (4.46)
qb è q ø
where
θ is the scale parameter
δ is the location parameter
In some cases, a three-parameter Weibull distribution provides a better fit than
the two-parameter Weibull distribution. The difference in the two distributions is
the location parameter δ, which shifts the distribution along the x-axis. By definition,
there is a zero probability of failure for x < δ. Although unusual, the location can
be negative; this implies that items were failed prior to testing.
Three methods for estimating the parameters of the Weibull distribution are
presented in this section:
1. Maximum likelihood estimation
2. Hazard plotting
3. Probability plotting

The following techniques are applicable for the two-parameter and three-parameter
Weibull distributions. When using the three-parameter Weibull distribution, replace
x with x − δ. The maximum likelihood equations for the Weibull distribution are
-1
1 r é n ùé n ù 1
å
r i =1
ln ( x i ) = ê åx ib ln ( x i ) ú ê åx ib ú - (4.47)
ë i =1 û ë i =1 û b
1
é1 n ùb
q̂ = ê åx ib ú (4.48)
ë r i =1 û
where
Iterative techniques are required for this expression. The estimated parameters
are asymptotically normal. The variances of the estimates can be found by inverting
the local information matrix. The local information matrix is
é ¶2L ¶2L ù
ê- 2 - ú
¶b ¶b¶q ú
F =ê (4.49)
ê ¶2L ¶2L ú
ê- - 2 ú
ë ¶b¶q ¶q û
The second partial derivatives of the log-likelihood equation are
¶2L é 1 æ x i öb 2 æ x i ö ù é æ x i öb 2 æ x i ö ù

¶b2
= år b çè q ÷ø çè q ÷ø åk ê- çè q ÷ø lnn çè q ÷øú (4.50)
ê - - ln ú +
êë úû êë úû
¶2L é b æ x i öb æ b ö ù é æ x i öb æ b ö ù

¶q
= å -
ê 2 ç ÷ ç 2÷ ( b + 1) ú å ê - ç ÷ ç 2 ÷ ( b + 1) ú (4.51)
+
ëq è q ø è q ø úû k êë è q ø è q ø
2
r ê úû
¶2L ìï 1 æ x öb æ 1 ö é æx ö ù üï
= å í- + ç i ÷ ç ÷ êb ln ç i ÷ + 1ú ý
¶b¶q r îï q è q ø è q ø ë èq ø û þï
(4.52)
ïìæ x ö æ 1 ö é æ x ö ù ïü
b
+ å íç i ÷ ç ÷ êb ln ç i ÷ + 1ú ý
k ïîè q ø è q ø ë è q ø û ïþ
where
å represents summation over all failures
r
å represents summation over all censored points

k
The variances of the estimated parameters are
é var ( b ) cov ( b,q ) ù

F -1 = ê ú (4.53)
ëcov ( b,q ) var ( q ) û
Approximate (1 − α)100% confidence intervals for the estimated parameters are

æ K a var ( b ) ö
b ç ÷
£ b £ b exp ç 2 ÷ (4.54)
æKa var ( b ) ö è b ø
ç ÷
exp ç 2 ÷
è b ø
æ K a var ( q ) ö
q ç ÷
£ q £ q exp ç 2 ÷ (4.55)
è q ø
æKa var ( q ) ö
ç ÷
exp ç 2 ÷
è q ø
where K a is the inverse of the standard normal probability density function.

2
These confidence intervals are approximate but approach exactness as the
sample size increases. Confidence intervals for reliability can be found using
the expressions
é æ öù é æ öù
exp ê - exp ç u + K a var ( u ) ÷ ú £ R ( x ) £ exp ê - exp ç u - K a var ( u ) ÷ ú (4.56)
êë è 2 ø úû êë è 2 ø úû
u = b éë ln ( x ) - ln ( q ) ùû (4.57)
éæ var ( q ) ö æ u2 var ( b ) ö æ 2u cov ( b,q ) ö ù

var ( u ) » b2 êç ÷ + çç ÷÷ - ç ÷ ú (4.58)
êëè q ø è b4 b2 q
2
ø è ø úû
Confidence intervals for percentiles can be found using the expressions
e y L £ x £ e y U (4.59)
1
x = q éë - ln (1 - p ) ùû b (4.60)
ln éë - ln (1 - p ) ùû
y L = ln q + - K a var ( y ) (4.61)
b
ln éë - ln (1 - p ) ùû
yU = ln q + + K a var ( y ) (4.62)
b
{ln éë- ln (1 - p )ùû} { }

2
var ( q ) var ( b ) 2 ln éë - ln (1 - p ) ùû cov ( q,b )
var ( y ) = + - (4.63)
q 2
b 4
b2 q
Hazard Plotting
The Weibull cumulative hazard function is
H ( x ) = - ln éë1 - F ( x ) ùû (4.64)
Replacing F(x) and rearranging gives a linear expression
ln H ( x ) = b ln x - b ln q (4.65)
By plotting ln H(x) versus ln x, the resulting slope (censored points are not plotted)
provides an estimate of β. The y-intercept of this plot is an estimate of β ln θ. Thus, θ
is estimated from the expression
æ y ö
qˆ = exp çç - 0 ÷÷ (4.66)
ˆ
è b ø
where y 0 is the y-intercept of the hazard plot.

The hazard function h(x) is estimated from the inverse of the reverse rank of the
ordered failures; the cumulative hazard function, H(x), is the cumulative of the values
of h(x). An alternative to plotting ln H(x) versus ln x is to directly plot H(x) versus x
on specialized Weibull hazard paper. The advantage of a hazard paper is that loga-
rithmic transformations do not have to be computed. Computers have made this
technique obsolete.
Example 4.11: Determine the parameters of the Weibull distribution using the

multiple-censored data in the following table. A “c” following an entry indicates
censoring.
Time to fail
309 c 229
386 104 c
180 217 c
167 c 168
122 138

Now the final two columns of Table 4.5 can be plotted, as shown in Figure 4.5. The
slope of the best-fit straight line through the data (found by linear regression) is
equal to 2.34 and provides an estimate of β. The y-intercept of the best-fit straight
line through the data is −13.004. The estimated scale parameter for the Weibull
distribution is
ˆ = exp æ - -13.004 ö = 259.1

q ç 2.34 ÷ø
è
TABLE 4.5 Tabulations for a hazard plot for the Weibull distribution.
Time to fail Reverse rank h(t) H(t) In H(t) In t

104 c 10
122 9 0.1111 0.1111 −2.1972 4.8040
138 8 0.1250 0.2361 −1.4435 4.9273
167 c 7
168 6 0.1667 0.4028 −0.9094 5.1240
180 5 0.2000 0.6028 −0.5062 5.1930
217 c 4
229 3 0.3333 0.9361 −0.0660 5.4337
309 c 2
386 1 1.0000 1.9361 0.6607 5.9558
c = Censored.
FIGURE 4.5 Weibull distribution hazard plot.
By taking the logarithm of the Weibull cumulative distribution function twice
and rearranging,
é æ 1 öù
ln ê ln çç ÷÷ ú = b ln x - b ln q (4.67)
êë è 1 - F ( x ) ø úû
é æ 1 öù
By plotting ln ê ln çç ÷÷ ú versus ln x and fitting a straight line to the points,
êë è 1 - F ( x ) ø úû
the parameters of the Weibull distribution can be estimated. The slope of the plot
provides an estimate of β, and the y-intercept can be used to estimate θ,
æ y ö
qˆ = exp çç - 0 ÷÷ (4.68)
ˆ
è b ø
The cumulative distribution function, F(x), is usually estimated from the median
rank, but other estimates such as the mean rank and the Kaplan-Meier product limit
estimator are also used. Median ranks are estimated using techniques shown in the
Exponential Probability Plotting section of this chapter. A specialized probability
paper is available for probability plotting. Using a probability paper eliminates the
need to transform the data prior to plotting.
Example 4.12: Determine the parameters of the Weibull distribution using prob-

ability plotting for the data given in the hazard plotting example, Example 4.11.
TABLE 4.6 Tabulations for a probability plot for the Weibull distribution.
é æ 1 öù
Median rank, ln êln ç ÷ú
êë è 1 - F ( x ) ÷ø úû
ç
Time to fail Ii Oi F(x) In t
104 c
122 1.1000 1.1000 0.0769 −2.5252 4.8040
138 1.1000 2.2000 0.1827 −1.6008 4.9273
167 c
168 1.2571 3.4571 0.3036 −1.0167 5.1240
180 1.2571 4.7143 0.4245 −0.5934 5.1930
217 c
229 1.5714 6.2857 0.5755 −0.1544 5.4337
309 c
386 2.3571 8.6429 0.8022 0.4827 5.9558
The last two columns of Table 4.6 are plotted in Figure 4.6. The slope of the best-fit
straight line through the data (found using linear regression) is 2.41, which is the
estimated value of β. The y-intercept of the best-fit straight line through the data is
−13.55. The estimated shape parameter is
æ -13.55 ö
qˆ = exp ç - ÷ = 276.6
è 2.41 ø
Confidence limits can be added to this plot using 5% and 95% ranks as described
in the Exponential Probability Plotting section of this chapter. Plotting position for
the 5% and 95% ranks are found from the expression
1
é æ 1 öù b
xi = q ê ln çç ÷÷ ú (4.69)
êë è 1 - F ( x ) ø úû
FIGURE 4.6 Weibull distribution probability plot.
Example 4.13: Repeat Example 4.12, and add 5% and 95% confidence limits.
Solution: The 5% and 95% ranks are computed using the following expression.
Because the failure order is not an integer, the 5% and 95% ranks must be interpo-
lated. Using a failure order of 1, from Equation 4.15 the 5% rank is
1
w0.05 = 10 - 1 + 1
1
19.4457 +
10 - 1 + 1
Using a failure order of 2, the 5% rank is
2
w0.05 = 10 - 2 +1
2
5.8211 +
10 - 2 + 1
Interpolating to a failure order of 1.1 gives a 5% rank of 0.0083. All calculations are
shown in Table 4.7.
Figure 4.7 shows the probability plot with 5% and 95% confidence limits.
TABLE 4.7 Probability plotting computations.

Interpo
Time Order 5% 95% Interpo lated Rank
to incre Median F- F- 95% lated 95% plotting
fail Status ment Order rank critical critical 5% rank rank 5% rank rank position
104 Censored
1 19.4457 0.2863 0.00512 0.25887
122 Failed 1.1000 1.1000 0.0769 0.0083 0.2724 96.27
2 5.8211 0.3416 0.03677 0.39416
138 Failed 1.1000 2.2000 0.1827 0.0469 0.4167 141.17
3 3.9223 0.3648 0.08726 0.50690
167 Censored
168 Failed 1.2571 3.4571 0.3036 0.1160 0.5525 179.80
4 3.2374 0.3706 0.15003 0.60662
180 Failed 1.2571 4.7143 0.4245 0.2018 0.6708 214.24
5 2.9130 0.3632 0.22244 0.69646
217 Censored
6 2.7534 0.3433 0.30354 0.77756
229 Failed 1.5714 6.2857 0.5755 0.3292 0.7982 256.96
7 2.6987 0.3089 0.39338 0.84997
309 Censored 8 2.7413 0.2550 0.49310 0.91274
386 Failed 2.3571 8.6429 0.8022 0.5656 0.9452 334.54
9 2.9277 0.1718 0.60584 0.96323
FIGURE 4.7 Weibull probability plot with confidence limits.

Nonparametric Confidence Intervals

The duration for a bogey test is equal to the reliability requirements being demon-
strated. For example, if a test is designed to demonstrate that a component has specific
reliability at 100 hours, and the test duration is 100 hours, then the test is designated
as a bogey test. Pass-fail tests also are considered bogey tests. A 100(1 − α) % nonpara-
metric confidence interval for reliability when bogey testing is
( n - r + 1) Fa ,2(n -r +1),2r
n-r
£R£ 2
(4.70)
n - r + ( r + 1) Fa r + ( n - r + 1) Fa
, 2 ( r +1) , 2 ( n - r ) , 2 ( n - r +1) , 2 r
2 2
where
n is the sample size
Fα, v1, v2 is the critical value of the F-distribution
Example 4.14: Twenty items are tested, with 3 items failing and 17 items passing.
What is the 90% confidence interval (two sided) for reliability?
Solution: The total sample size, n, is 20; the number of failures, r, is 3. To compute
the confidence limits, two critical values from the F-distribution are needed:

Fa /2,2(n-r +1),2r = F0.05,36,6 = 3.786
Fa /2,2(r +1),2(n-r ) = F0.05,8,34 = 2.225
The confidence interval for reliability is
20 - 3 (20 - 3 + 1)3.786
£R £
20 - 3 + (3 + 1)2.225 3 + (20 - 3 + 1)3.786
0.656 £ R £ 0.958
Summary
Several methods are available to estimate distribution parameters. It is recommended
to use a graphical method, such as hazard plotting or probability plotting, to obtain
a visual goodness-of-fit test. Once the goodness of fit is satisfied, maximum likelihood
estimation should be used because it is more accurate than other methods.
5
Accelerated Test Plans
R
eliability testing is often expensive; expensive prototypes, lengthy tests, and
exotic test equipment are not unusual. To deal with time and cost pressures,
as well as the mathematical complications of censored data, reliability testing
statistics have become specialized. Zero-failure testing, sequential testing, accelerated
testing, and burn-in are discussed in this chapter.
Mean Time to Fail

Manufacturers often quote the mean time to fail as a method of conveying product
reliability. Reliability comparisons should never be made based on the mean time to
fail. Consider two airplanes. Airplane A has a mean time to fail of 30 hours, and
Airplane B has a mean time to fail of 20 hours. You are required to choose one of the
airplanes for a 12-hour flight; which do you prefer? The normal reaction is to choose
Airplane A. Now consider that the time to fail for both airplanes is normally distrib-
uted; the standard deviation for Airplane A is 10 hours, and the standard deviation
for Airplane B is 2.5 hours. Do you still prefer Airplane A? The distributions for the
time to fail for both airplanes are given in Figure 5.1. The reliability for Airplane A at
12 hours is 96.41%. The reliability for Airplane B at 12 hours is 99.93%. The mean
time to fail did not accurately select the airplane with the best reliability at 12 hours.

FIGURE 5.1 Example of mean time to fail.
0.16
0.14
Airplane B
0.12
0.1
Probability density
0.08
0.06
0.04
Airplane A
0.02
0
0 10 20 30 40 50 60
Time to fail
Test Plan Problems

A reliability test plan requires all of the following:
1. A sample size
2. A test duration
3. Criteria for acceptable performance
4. The number of failures allowed
For example, given a Weibull time-to-fail distribution with a shape parameter

of 2, a 95% reliability can be demonstrated at 100 hours with a confidence level of
90% by any of the test plans shown in Table 5.1.
The most efficient reliability test plans assume there will be no failures. Of course,
this is an unrealistic assumption, but if failure occurs, the item under consideration
TABLE 5.1 Weibull reliability demonstration test plans.
Number of units
Test plan tested Number of failures Test duration (hours)
A 45 0 100
B 76 1 100
C 192 5 100
D 318 10 100
E 45 1 131
has not necessarily failed to meet the reliability requirements. The test may be continued
until the reliability requirements are met, or until it is evident that the reliability require-
ments will not be met. This defies traditional statistical thinking, which requires test
parameters and pass-fail requirements to be determined prior to testing.
Consider a unit with a Weibull time-to-fail distribution having a shape parameter
of 2 and a scale parameter of 570 hours. The reliability at 100 hours is
100 2

R 100 e
570
0.9697
Now suppose 45 of these units are tested for 100 hours. To demonstrate 95%
reliability at 100 hours of operation with a confidence of 90%, all 45 units must survive
for 100 hours. The probability of all 45 units successfully completing the test is
= P 0= .9697 45 0.2503
Although the unit has a reliability greater than 95% at 100 hours, the unit is unable
to demonstrate 95% reliability at 100 hours only in slightly more than 25% of the
attempts using the designed test plan. Figure 5.2 shows the probability of successfully
completing this test plan as a function of the true reliability at 100 hours. As the true
reliability increases, the probability of successfully demonstrating 95% reliability
increases; as the sample size increases, the probability of successfully demonstrating
95% reliability increases. From Figure 5.2, it can be seen that this item has a 62.89%
chance of successfully demonstrating the desired reliability if 318 units are tested.
Why does a unit that exceeds the required reliability fail the reliability demonstration
test? Reliability is not demonstrated because the sample size is not large enough.
FIGURE 5.2 Probability of successful completion of the test plan.

FIGURE 5.3 Type I error using a single drive test.
Consider another example: A golfer has a driving distance that is normally distributed

with a mean of 250 meters and a standard deviation of 15 meters. This golfer wants to join
a league that requires all members to have an average drive of at least 240 meters. This is
tested by allowing the golfer to select a number of drives, and after all drives have been
hit, if the average drive distance is greater than 240, the golfer is allowed to join the league.
Recall that Type I error is the risk of incorrectly rejecting the null hypothesis
and that Type II error is the risk of incorrectly accepting the null hypothesis. In this
example, Type I error is the risk of incorrectly determining that the golfer’s average
driving distance is less than 240 meters. For this example, the Type II risk is 0; there
is no chance of incorrectly determining that the golfer’s average driving distance is
greater than 240 meters because the average driving distance is greater than 250
meters. The only reason this golfer will not be allowed to join the league is because
the golfer’s average driving distance is incorrectly evaluated. Figure 5.3 demonstrates
the probability of the golfer being allowed to join the league if the driving average is
based on a single drive. The probability the golfer is not allowed to join the league is
25.3%. This is the Type I error; the golfer’s true average driving distance exceeds 240
meters, but the demonstration test incorrectly determined the true average driving
distance was less than 240 meters.
The distribution of the average n drives has the following properties:
250
15

n
FIGURE 5.4 Type I error using a drive test with a sample size of 5.
Figures 5.4 and 5.5 show the effect of sample size on a Type I error. By increasing
the sample size, the standard deviation of the distribution of the average driving
distance is decreased, resulting in a more accurate estimate of the true driving
distance. If the golfer is evaluated using an average of five drives, there is a 6.8% of
failing to meet the requirements to join the league. This error drops to 1.75% when
the golfer is evaluated using 10 drives.
Figure 5.6 shows the probability of failing to pass the drive test as a function of
the number of drives. This figure shows that there is a substantial risk of incorrectly
disqualifying the golfer if a small sample size is used and that, as the sample size
increases, this error is reduced.
Theoretically, this situation is not a problem. If at least 25 drives are required to
evaluate each golfer, the error is near 0. However, reliability engineers often do not
have this option. Sample sizes of 25 can be too expensive or require too much time.
Using a sample size of 2, more than 17% of the evaluations when the golfer exceeds
the required mean by 10 meters are incorrectly rejected. If this were a test to evaluate
a design or process, the incorrect test result leads to increased cost and delays while
the product or process is redesigned. It may be better to place test more samples,
knowing that the item under consideration may have been rejected incorrectly, based
on a small sample size. This is the concept of sequential testing.
Now consider a golfer who has an average driving distance of 230 meters with a
standard deviation of 15 meters. Type I error—the risk of incorrectly determining that
the golfer’s average driving distance is less than 240 meters—is 0 because the golfer’s true
driving distance is 230 meters. Type II risk—the probability of incorrectly determining
FIGURE 5.5 Type I error using a drive test with a sample size of 10.
FIGURE 5.6 Type I error as a function of sample size.

that the golfer’s average driving distance is greater than 240—is the same as the Type
I risk for the golfer with an average driving range of 250 meters. The only reason this
golfer will be allowed to join the league is because the golfer’s average driving distance
is incorrectly evaluated. Statistical parameters are included in sequential testing to guard
against Type I and Type II errors. A spreadsheet titled “TestDesignInteractive.xls” has
been included on the accompanying CD to allow further investigation of these situations.
Zero-Failure Testing
There are two types of zero-failure tests: (1) bogey testing and (2) Bayesian testing.
Bogey Testing
Bogey testing requires the testing duration to be equal to the required life. For
example, if a 95% reliability is required at 200,000 kilometers of service, then the
units being tested will be removed from testing when they fail or when they complete
the equivalent of 200,000 kilometers of testing. The sample size required to demon-
strate reliability of r with a confidence level of c is
ln 1 c
n (5.1)
ln r
Example 5.1: A windshield wiper motor must demonstrate 99% reliability with 90%
confidence at 2 million cycles of operation. How many motors must be tested to
two million cycles with no failures to meet these requirements?
Solution: Two hundred and thirty motors must function for two million cycles to
demonstrate the desired level of reliability, as shown:
ln ( 1 − 0.9 )
= 229.1
ln0.99
Bayesian Testing
Bogey testing is inefficient. By extending the test duration beyond the required life,
the total time on the test can often be reduced. When the test duration is extended,
it is necessary to make assumptions concerning the shape of the distribution of the
time to fail. This is done by assuming a Weibull distribution for time to fail and by
assuming a shape parameter. Recall from Chapter 2, Probability Fundamentals, that
the Weibull distribution can approximate many other distributions by changing the
value of the shape parameter.
Effect of the Shape Parameter

The shape of the distribution has a tremendous effect on the amount of testing required
to demonstrate reliability. Consider the histogram of time to fail for a population
FIGURE 5.7 Weibull distributed population with 95% reliability at 1 bogey and a
shape parameter of 3.6.
with a Weibull time-to-fail distribution with a shape parameter of 3.6, as shown in

Figure 5.7. Ninety-five percent of the items survived for more than 1 bogey; thus, the
reliability at 1 bogey is 95%. This histogram appears to show a normally distributed
population because the Weibull distribution is nearly normal when the shape param-
eter is 3.6. Also, notice that when the shape parameter is 3.6, to demonstrate a 95%
reliability at 1 bogey, the mean must be at least 2.06 bogeys.
Figure 5.8 shows a population having a Weibull time-to-fail distribution with a
shape parameter of 1. This figure shows a reliability of 95% at 1 bogey and a mean of
19.5 bogeys—nearly 10 times greater than the Weibull distribution with a shape
parameter of 3.6. The variance of the distribution increases as the shape parameter
decreases. Variance is the equivalent of uncertainty, and the amount of testing
required to demonstrate reliability is dependent on the amount of variance in
the population.
shape parameter of 1.8. This figure shows a reliability of 95% at 1 bogey and a mean
of 5.2 bogeys.
shape parameter of 8.0. This figure shows a reliability of 95% at 1 bogey and a mean
of 1.37 bogeys.
Estimating the Shape Parameter

From Figures 5.7 through 5.10 inclusive, it can be seen that the assumed shape greatly
affects the testing requirements. How is the shape parameter estimated? The shape
parameter is governed by the physics of failure. Some phenomena are skewed right,
while others are skewed left or have no skewness. In general, the more the failure
mode is a result of mechanical wear, the larger the shape parameter will be. The shape
parameter is usually stable and tends to remain constant. For example, the shape
parameter for master cylinders will be similar for master cylinders for small cars and
large trucks and for different designs because the physics of how the master cylinder
fails is similar.
The best way to estimate the shape parameter is through prior testing. Many
automotive companies require some testing to failure to allow the shape parameter
to be determined. Keep detailed records of all tests, and build a database of shape
parameters. It is recommended to use the lower 90% confidence limit for the shape
parameter because of the magnitude the shape parameter has on test requirements.
With the lack of any prior knowledge, data sources are available on the Internet, or
the shape parameter can be estimated based on the knowledge of the physics
of failure.
Be careful when estimating the shape parameter for electronics. Many sources
state the shape parameter for electronics is 1.0 because there is no mechanical wear
in electronics. Electronic modules located in environmentally harsh conditions, such
as under the hood of an automobile or in an airplane, fail as a result of mechanical
wear. The extreme vibration, temperature cycling, and, in some cases, contaminants
cause mechanical failures. It is not uncommon to have shape parameters greater than
8.0 for electronic components.
Determining Test Parameters

The statistical properties for Bayesian testing with the Weibull distribution are based
on the exponential distribution. If the parameter t follows a Weibull distribution,
then the parameter tβ is exponentially distributed. The lower (1 − α) confidence limit
for reliability is
t

L ,
RL , t e

(5.2)
where
1
2 n t i
L , i 1 (5.3)
2 ,d

n is the number of units tested, both failed and surviving, and X α,2d is the critical value
of the chi-square distribution with a significance of α (0.05 for a confidence level of
95%) and d degrees of freedom. For failure truncated testing, d is equal to 2r, where
r is the number of failed units. For time truncated testing, d is equal to 2r + 2.
Example 5.2: Fourteen units are placed on a test stand. The first unit fails after
324 hours of testing, the second unit fails after 487 hours of testing, the third unit
fails after 528 hours of testing, and the remaining 11 units are removed from testing.
Given a Weibull time-to-fail distribution with a shape parameter of 2.2, what is the
lower 90% confidence limit for reliability at 400 hours?
Solution: Because testing was suspended at the time of the last failure, this is failure
truncated testing. With failure truncated testing and three failures, the degrees of
freedom for the chi-square distribution are 2(3) = 6. The critical value of the chi-
square distribution given 6 degrees of freedom and a significance of 10% is 10.64.
This value can be found from Appendix A or using the Microsoft Excel expression
= CHISQ.INV ( 1 − 0.9,6 )
The lower 90% confidence limit for the mean of the transformed data is
1
 2 ( 12,872,524.77 )  2.2
=θ L,α =  797.3
 10.64 
The lower 90% confidence limit for the reliability at 400 hours is
  400 2.2 
 −  
RL,0.10
= ( 400) e= 0.80311
  797.3  
Example 5.3: Fourteen units are placed on a test stand. The first unit fails after
324 hours of testing, the second unit fails after 487 hours of testing, the third unit
fails after 528 hours of testing, and the remaining 11 units are removed from testing
after 550 hours without failing. Given a Weibull time-to-fail distribution with a shape
parameter of 2.2, what is the lower 90% confidence limit for reliability at 400 hours?
Solution: Because testing was not suspended at the time of the last failure, this is
time truncated testing. With time truncated testing and three failures, the degrees
of freedom for the chi-square distribution are 2(3) + 2 = 8. The critical value of the
chi-square distribution given 8 degrees of freedom and a significance of 10% is 13.36.
This value can be found from Appendix A or using the Microsoft Excel expression
= CHISQ.INV ( 1 − 0.9,8 )
The lower 90% confidence limit for the mean of the transformed data is
1
 2 ( 13,882,130.48 )  2.2
=θL,α  =  744.1
 13.36 
The lower 90% confidence limit for the reliability at 400 hours is
  400 2.2 
 −  
RL,0.10
= ( 400) e= 0.7747
  744.1  
Reliability tests are often designed in a two-step procedure: (1) “How many test
stands are available?” and (2) “What is the test duration, given the number of test
fixtures?” For a sample of n units, the required test duration to demonstrate a reli-
ability of R at time t with a confidence level of 1 − α assuming no failures is
  1
t   − ln α  β
T =  1   n  (5.4)
 [ − ln R ]β 
 
Example 5.4: How long must 20 units be tested to demonstrate a 99% reliability
with 80% confidence at 200 hours of operation, given a Weibull time-to-fail distri-
bution with a shape parameter of 3.5 and assuming no failures?
Solution: The required test duration is
  1
 200   − ln ( 0.2 )  3.5

=T = 1    362hours
  − ln ( 0.99 )  3.5   20 
  

For a test duration of T, the required sample size to demonstrate a reliability of R

at time t with a confidence level of 1 − α assuming no failures is
− ln α
n=
  1
 
β
  ( − ln R ) β   (5.5)
T  t  
   
 

Example 5.5: How many units must be tested for 300 hours to demonstrate a 99%
reliability with 80% confidence at 200 hours of operation, given a Weibull time-to-
fail distribution with a shape parameter of 3.5 and assuming no failures?
Solution: The required sample size is
− ln ( 1 − 0.80 )
=n = 3.5
38.7 ⇒ 39
  1

300  ( − ln0.99 ) 3.5 
  200  
   
Guidelines for Zero-Failure Test Planning

When designing reliability demonstration tests, a balance must be maintained
between the test duration and sample size. Figure 5.11 shows the effect of sample size
on the test duration for several shape parameters of the Weibull distribution.
Figure 5.12 shows the effect of sample size on the total test time (sample size ×
test duration) for several values of the Weibull shape parameter.
FIGURE 5.11 The effect of sample size on test duration.
FIGURE 5.12 The effect of sample size on the total test time.
From Figure 5.12, it can be seen that when the shape parameter is less than 1.0,
the total test time decreases as the sample size increases. If the shape parameter is
equal to 1.0, the total test time is unaffected by the sample size. If the shape parameter
is greater than 1.0, the total test time decreases as the sample size decreases, with the
total test time at a minimum when only one sample is tested. Testing one sample is
not recommended; if possible, a minimum of four samples should be used to gain an
estimate of the variance in the population.
Example 5.6: A major Tier 1 automotive supplier produces DC motors for windshield
wiper systems. Depending on the size of the vehicle, the windshield configuration, and
other factors, the size and configuration of the motor changes, resulting in a unique
. no i
motor for each windshield program. Because each program has a unique motor, each
motor design must be validated for reliability. The shape parameter results for the
five most recent motor development programs are given in Table 5.2. For the next
windshield wiper motor program, how many units must be tested and for what test
duration to demonstrate 95% reliability with 90% confidence at 1.5 million cycles?
Solution: The most efficient test design will assume no failures. This is not a real-
istic assumption, but it provides the starting point for an efficient test. What value
should be used for the shape parameter? A smaller shape parameter requires more
testing; thus, the conservative approach is to err on the low side when determining
the value of the shape parameter. When data from only one test are available, a
safe approach is to use the lower 95% confidence limit for the estimated shape
parameter. Another approach is to use the smallest estimated shape parameter—in
this case, 7.04. Estimating the shape parameter to use from previous data is not an
exact science. Caution should be used, and the use of a conservative approach is
recommended. In this case, the Tier 1 supplier and the automotive company agreed
to use a shape parameter of 7.04 to design the test. Eight test stands were available
for testing, which results in a test duration of
  1
   − ln 0.1  7.04
T =  1.5  ( ) 1.92million cycles
 
 1
   8 
 − ln ( 0.95 ) 7.04

 
One of the 8 units failed after 1.82 million cycles of testing. This did not result in the
design failing to demonstrate the required reliability. There were two options at this
point: (1) redesign the motor and retest or (2) continue testing the remaining seven
motors until the required reliability is demonstrated. The second choice can be risky; the
motor may not be able to demonstrate the required reliability with additional testing.
TABLE 5.2 Motor test shape parameter results.
Program β Lower 95% confidence limit for β

Chrysler—Compact car 7.41 4.78
Chrysler—Large car 8.63 5.64
Chrysler—Sport utility vehicle 7.04 4.22
Chrysler—Large car 8.93 6.20
General motors—Midsize car 7.99 5.31
To demonstrate, the additional seven motors must survive for a duration of

1

7.04
7.04
2 7.04
2 1.82
1.5
X 1
ln 0.95 7.04
T 2.088 million cycles
2 7
where X is the critical value of the chi-square distribution with a significance of 10%
2
(1 − 90% confidence) and 4 degrees of freedom (2r + 2), which is 7.7794. The Tier 1
supplier and the automotive company agreed to continue testing the remaining seven
motors rather than abandon the motor design. The remaining 7 motors all survived
for 2.088 million cycles.
Sequential Testing
Example 5.6 is an example of sequential testing. Reliability tests often are too lengthy
or expensive to begin testing with a sample size large enough to make a decision with
a high degree of confidence. Conventional testing requires determining a sample size,
null hypothesis, and alternative hypothesis prior to testing. An alternative is to begin
with a small sample size (often only a single sample is used) and to make decisions
concerning the test hypothesis as test data are collected. L
With sequential testing, decisions are based on the likelihood ratio 1,n , where
L0,n
n
L k ,n f x i \ k (5.6)
i 1
where
f(x) is the probability density function of the variable x
θ is the parameter being tested
The hypotheses for a sequential test are
H 0 : 0

H1 : 1
Sequential test plans have a Type I error (i.e., the probability of rejecting the null
hypothesis when it is true) of α when θ = θ0 and a Type II error (i.e., the probability of
accepting the null hypothesis when it is not true) that is less than or equal to β when θ1 < θ.
Pass-Fail Testing
In many instances, the outcome of a reliability test is pass or fail for each of the indi-
vidual items being tested. This situation can be modeled by the binomial distribution,
and for an acceptance test, the hypotheses are
H 0 : p p0

H1 : p p0
where
p is the probability of failure for an item
p0 is the probability of failure; the probability of accepting H0 is 1 − α
where α is the Type I error or the producer’s risk.
As each item is tested, the decision to continue testing, accept the null hypothesis,
or reject the null hypothesis is determined from the expression
y n y
p 1 p1
z 1 (5.7)
p0 1 p0
where
p1 is the probability of failure; the probability of accepting H0 is β
β is the Type II error or the consumer’s risk
n is the total number of trials
y is the total number of failures
The test is continued if A < z < B, where

A (5.8)
1
1
B (5.9)

The null hypothesis is accepted if z < A and rejected if z > B. If neither of these
conditions is met, testing continues.
Example 5.7: A sequential test was conducted to test the hypotheses
H0 : p ≤ 0.02

H1 : p > 0.02
The level of producer’s risk for the test was 0.05, the level of consumer’s risk was
0.1, and p1 = 0.1. The first three items tested were successful, and the fourth and
fifth items tested failed. Determine the status of the test after the results of each
trial were known.
Solution: The values of A and B are
0.1
=A = 0.10526
1 − 0.05
1 − 0.1
=B = 18.0
0.05
The values of z after each trial, and the resulting decisions, are as follows:
Trial z Decision
1 0.9184 Continue
2 0.8434 Continue
3 0.7746 Continue
4 3.873 Continue
5 19.364 Reject H0
It is often useful to represent acceptance tests graphically. The test described in

Example 5.7 can be defined as
Accept H 0 if c < y .
Reject H 0 if d > y .
Continue testing if c < y < d.
where
1 p0 1
n ln ln
c 1 p1
p 1 p0 (5.10)
p1 1 p0
ln 1 ln
p0 1 p1 p0 1 p1
1 p0 1
n ln ln
d 1 p1
p 1 p0 p1 1 p0 (5.11)
ln 1 ln
p0 1 p1 p0 1 p1
Example 5.8: For the test described in Example 5.7, suppose 81 items were tested
with the 4th, 18th, 37th, and 73rd items failing, while all other items were successful.
Show this test graphically.
Solution: Figure 5.13 shows this sequential test.
From Figure 5.13, it can be seen that the fewest number of trials required to reject
the null hypothesis is 2 (i.e., 2 consecutive failures with no successes). It also can
be seen from Figure 5.13 that the fewest number of trials (all trials being successful)
required to accept the null hypothesis is 27.
FIGURE 5.13 Sequential testing with the binomial distribution.

Mathematically, the fewest number of trials required to reject the null hypothesis is
ln
fR
p1 1 p0 1 p0 (5.12)
ln ln
p
0 1 p1 1 p1
The fewest number of trials required to accept the null hypothesis is
ln A
fA
1 p1 (5.13)
ln
1 p0
The discriminatory power of a sampling plan is shown by its operating charac-

teristic curve (OC curve). An OC curve displays the probability of accepting the null
hypothesis versus the fraction defective. For binomial sequential sampling, the prob-
ability of accepting the null hypothesis is
Bh 1
Pa p (5.14)
Bh Ah
where
p is the true fraction defective
h is an arbitrary term used to compute p (h usually ranges from −3 to 3) using
the expression
h
1 p1
1
p 1 p0
h h (5.15)
p1 1 p1

p0 1 p0
Example 5.9: Construct the OC curve for the sampling plan described in
Example 5.8.
Solution: First, a value of −3 is arbitrarily assigned to h, and p is computed. The
resulting value of p is then used to compute Pa. The value of h is then increased, and
the procedure is repeated until enough values are obtained to construct a graph.
Table 5.3 contains the information required to construct an OC curve. Figure 5.14
shows this OC curve.
TABLE 5.3 OC curve computations.
h p Pa
−3.0 0.227 0.001
−2.6 0.201 0.003
−2.2 0.175 0.007
−1.8 0.149 0.017
−1.4 0.124 0.042
−1.0 0.100 0.100
−0.6 0.078 0.224
−0.2 0.059 0.436
0.2 0.043 0.683
0.6 0.030 0.863
1.0 0.020 0.950
1.4 0.013 0.983
1.8 0.008 0.995
2.2 0.005 0.998
2.6 0.003 0.999
3.0 0.002 1.000
FIGURE 5.14 Sequential sampling OC curve.

Another characteristic that is valuable when planning a sequential test is the

expected number of trials required to reach a decision. This value is a function of the
true fraction defective and is
Pa p ln A 1 Pa p ln B
E p (5.16)
p 1 p1
p ln 1 1 p ln
p0 1 p0
This function reaches a maximum between p0 and p1 and is relatively flat in

this region.
Example 5.10: Determine the expected number of trials required to reach a decision
for the sampling plan described in Example 5.8.
Solution: Figure 5.15 shows the expected number of trials required to reach a
decision for the sampling plan described in Example 5.8.
FIGURE 5.15 Expected number of samples required to make a decision.
Testing with continuous data, such as time to fail or pressure to burst, provides more
information than a simple pass or fail designation and reduces the testing time.
Assuming an exponential time-to-fail distribution, a sequential test can be conducted

on the cumulative time to fail,
n
T t i
i 1
where
ti is the test time accumulated for the ith unit
n is the total number of units tested
The test hypotheses are
H 0 : 0
H1 : 0
where λ is the failure rate of the exponential distribution.

The null hypothesis is rejected if

r ln 1 ln
T 0 1
(5.17)
1 0
where
λ0 is the failure rate such that the probability of accepting H0 is 1 − α
λ1 is the failure rate such that the probability of accepting H0 is β
The null hypothesis is accepted if

n ln 1 ln

0 1
T (5.18)
1 0
Testing continues if

r ln 1 ln r ln 1 ln

0 1
0 1
T (5.19)
1 0 1 0
Example 5.11: An accelerated test is needed to demonstrate a failure rate of 0.001

with a producer’s risk of 10%. The test should have a 20% chance of passing if the
true failure rate is greater than 0.005.
Solution: The parameters for this test are
λ0 =0.001
λ 1 =0.005
α =10%
β =20%
Figure 5.16 shows this sequential test.
FIGURE 5.16 Exponential sequential test.
The Bayesian testing described for zero-failure test plans can be expanded to sequen-
tial test plans. Again, this assumes a Weibull distribution for the variable of interest
with a known or assumed shape parameter. The test hypotheses for the Bayesian
Weibull sequential test are
H 0 : 0

H 1 : 0
where
θ is the scale parameter of the Weibull distribution
θ0 is the scale parameter such that the probability of accepting H0 is 1 − α
The desired reliability has been demonstrated if
n
t
i 1

i nb d (5.20)
and the item has failed to meet reliability requirements if
n
t
i 1

i nb c (5.21)
where
1 1
a (5.22)
1 0

1n 0
b 1 (5.23)
a
1
ln

c (5.24)
a
1
ln

d (5.25)
a
and θ1 is the value of the scale parameter that yields a τ probability of accepting H0.
The term τ is used to represent the consumer’s risk because the standard term, β, is
used to represent the shape parameter of the Weibull distribution.
Example 5.12: A component is required to be 95% reliable after 12.96 weeks.

Create a sequential test with a 5% chance of rejecting a component that meets the
reliability requirement, while having a 10% chance of accepting a component with
reliability less than 89.4% after 12.96 weeks. The time to fail for the component being
evaluated is known to follow a Weibull distribution with a shape parameter of 2.2.
Solution: The requirements for reliability must be transformed into the required
scale parameters for the Weibull distribution. This is done by algebraic manipulation
of the Weibull reliability function
β
t 
−  t
(x ) e
R= θ
⇒
= θ 1
(5.26)
{ }
− ln R (t )  β
The scale parameter for a Weibull distribution with a shape parameter of 2.2 having
a 95% reliability at 12.96 million cycles is
12.96
=θ0 = 1
50 weeks
 − ln ( 0.95 )  2.2
The scale parameter for a Weibull distribution with a shape parameter of 2.2 having
a 95% reliability at 12.96 million cycles is
12.96
=θ1 = 1
35 weeks
 − ln ( 0.894 )  2.2
The remaining calculations required to construct the test plan are
1 1
a= − = 0.00021799
352.2 502.2
50 
 ( 2.2)ln
=b =  35  13,259
0.00021799
 1 − 0.01 
ln 
 0.05 
=c = 13,259
0.00021799
 1 − 0.05 
ln  
=d =  0.01  10,327
0.00021799
FIGURE 5.17 Weibull sequential testing example.
Figure 5.17 describes this test plan.
Example 5.13: What is the status of the test described in Example 5.12 if the testing
shown in Table 5.4 has been completed?
Solution: The decision parameter is weeks to fail raised to the shape parameter of
the Weibull distribution (tβ). With no failures, the null hypothesis cannot be rejected,
and from Figure 5.17, the component must accumulate a tβ value greater than 10,327
to demonstrate the required reliability. Table 5.5 shows the computations for the
cumulative value of tβ .
With the cumulative value of tβ equal to 2,384, the test is in the continue region.
TABLE 5.4 Weibull sequential test example.
Component Test time (weeks) Status

A 14.8 Suspended
B 15.1 Suspended
C 7.7 Suspended
D 28.0 Suspended
TABLE 5.5 Weibull sequential testing example computations.
Component Test time (weeks) Status tβ ∑tβ

A 14.8 Suspended 375 375
B 15.1 Suspended 392 768
C 7.7 Suspended 89 857
D 28.0 Testing 1,527 2,384
Example 5.14: For the test described in Examples 5.12 and 5.13, how long must
Component D survive to demonstrate the desired reliability?
Solution: The required cumulative value of tβ is 10,327. Components A, B, and C
accumulated a total of 857 for tβ . This means that tβ for Component D must be
t β= 10,327 − 857= 9,470
Thus the required test time for Component D is

1 1
=t (t=
)
β β
=
9,470 2.2
64.2 weeks
Randomization of Load Cycles

Environmental and stress conditions occur randomly, or randomly within seasons,
when used by consumers. This is sometimes overlooked in laboratory testing where
testing may begin with low stress and increase the level of stress as testing progresses,
completing the test with the highest stress. Because the level of stress randomly occurs
when the product is in the field, the test data cannot accurately predict the field reli-
ability. To accurately model field reliability, the stress levels applied in the laboratory
must accurately match the application rate in the field. In most cases, this requires a
random application of stress levels.
Distribution-specific random numbers can be generated by setting the cumulative
distribution function equal to a unit, uniform random number and taking the inverse
of the function. A unit, uniform random number follows a uniform distribution between
0 and 1. The function =rand() in Microsoft Excel generates unit, uniform random
numbers, and the function RND in Basic generates unit, uniform random numbers.
Example 5.15: Derive a random number generator for the Weibull distribution.
Solution: For the Weibull distribution,
β
x
− 
r = 1−e θ
β
x
− 
1− r =e θ
β
x
− 
ln ( 1 − r ) = θ
1
x
 − ln ( 1 − r )  β =
θ
1
x =θ  − ln ( 1 − r )  β
1
x =θ  − ln ( r )  β
The last step of the solution is possible because the term 1 - r is a unit, uniform
distribution that is identical to the distribution characteristics of r.
Table 5.6 gives random number generators for several distributions.
TABLE 5.6 Distribution-specific random number generators.
Distribution Random number generator

Exponential 1
x = − ln r
λ
Normal x=  −2ln r1 cos ( 2Πr2 )  σ + µ*
 
Lognormal =x exp  −2ln r1 cos ( 2Πr2 )  σ + µ *
 
Weibull 1
x =θ ( − ln r ) β
*
Two unit, uniform random numbers are required to generate one
distribution-
specific random variable.
Example 5.16: Figure 5.18 shows a histogram of the pressure recorded in the fluid
transportation system on several vehicles over a one-year time period; a total of
10,624 pressure readings were recorded.1 A function is needed for the test controller
to randomly generate pressure cycles that follow this histogram.
FIGURE 5.18 Fluid transportation system pressure histogram.
1
The data used to create this histogram are given in the file Chapter5.xls on the accompanying CD.
Solution: Figure 5.19 shows a Weibull probability plot for this data. The Weibull
parameters computed using this plot are
β =1.8
θ =1,505
The function used to randomly generate pressure cycles is

1
=p 1,505 ( − ln r ) 1.8
FIGURE 5.19 Weibull probability plot for pressure.
For some environmental factors, such as temperature, it is impossible or prohibi-

tively expensive to completely randomize the stress levels. For these cases, it is recom-
mended to cycle the stress level as often as feasible during the test.
Reliability Growth2
Reliability growth is the improvement in reliability over a period of time due to
changes in product design or the manufacturing process. It occurs by surfacing failure
modes and implementing effective corrective actions. Reliability growth management
is the systematic planning for reliability achievement as a function of time and other
resources, and controlling the ongoing rate of achievement by reallocation of these
resources based on comparisons between planned and assessed reliability values.
2
Portions of this section were taken from the 1999 AMSAA Reliability Growth Handbook. (This handbook
is available on the accompanying CD.)
The following benefits can be realized by the utilization of reliability

growth management:
•• Finding Unforeseen Deficiencies—The initial prototypes for a complex
system with major technological advances will invariably have significant
reliability and performance deficiencies that could not be foreseen in the early
design stage. This is also true of prototypes that are simply the integration of
existing systems. Unforeseen problems are the norm in achieving seamless
interoperation and interfacing among systems that have already been
developed. Reliability growth testing will surface these deficiencies.
•• Designing in Improvement through Surfaced Problems—Even if some
potential problems can be foreseen, their significance might not. Prototypes
are subjected to a development-testing program to surface those problems that
drive the failure rate so that the necessary improvements in the system design
can be made. The ensuing system reliability and performance characteristics
will depend on the number and effectiveness of these fixes. The ultimate goal
of the development-testing program is to meet the system reliability and
performance requirements.
•• Reducing the Risk of Final Demonstration—Experience has shown that
programs that rely simply on a final demonstration alone to determine
compliance with the reliability requirements do not, in many cases, achieve
the reliability objectives within the allocated resources. Emphasis on reliability
performance prior to the final demonstration using quantitative reliability
growth could substantially increase the chance of passing or even replace
the final demonstration.
•• Increasing the Probability of Meeting Objectives—This can be achieved
by setting interim reliability goals to be met during the development-testing
program and the necessary allocation and reallocation of resources to attain
these goals. A comprehensive approach to reliability growth management
throughout the development program organizes this process.
Reliability Growth Process

Reliability growth is the result of an iterative design process. As the design matures,
it is investigated to identify actual or potential sources of failures. Further design
effort is then spent on these problem areas. The design effort can be applied to either
product design or manufacturing process design. The iterative process can be visual-
ized as the feedback loop shown in Figure 5.20.
Figure 5.20 illustrates the three essential elements involved in achieving the
reliability growth:
•• Detection of failure modes
•• Feedback of problems identified
•• A redesign effort based on problems identified
Figure 5.20 illustrates the growth process and associated management processes
in a skeleton form. This type of illustration is used so that the universal features of
FIGURE 5.20 Reliability growth feedback model.
these processes may be addressed. The representation of an actual program or program

phase may be considerably more detailed. This detailing may include specific inputs
to and outputs from the growth process, additional activity blocks, and more explicit
decision logic blocks.
Reliability Growth Models

Reliability growth models can be used to estimate the amount of time required to
obtain a reliability goal or to predict reliability. These growth models apply to software,
hardware, and processes. There are several reliability growth models; the two discussed
in this section are the Duane model and the Army Materiel Systems Analysis Activity
(AMSAA) model.
Duane Model
J.T. Duane was one of the first to introduce a reliability growth model. He empirically
derived the expression shown in Equation 5.27, which assumes a nonhomogeneous
Poisson process,
c t (5.27)
where
θc is the cumulative mean time between failures
β is the slope parameter
α is the initial mean time to fail
By taking the logarithm of Equation (4.57), the following linear expression

is found:
ln C ln ln t (5.28)
The parameters of the Duane model can be found by plotting Equation (5.28).
The current or instantaneous mean time to fail is
t
i t (5.29)
1
Example 5.17: Ten electronic throttle bodies were each tested for 250,000 cycles.
After each failure, the root cause was determined, and design changes were made
on all 10 throttle bodies. Table 5.7 gives the failure data. Determine the Duane
model parameters. Predict the testing duration to obtain the mean cycles to fail
of 75,000 cycles.
Solution: The parameters of the Duane growth model are determined by plotting
the logarithm of the cumulative failures versus the logarithm of cumulative test
cycles. Table 5.8 contains the information required for the graph.
Figure 5.21 shows a graph of the last two columns of Table 5.8.
The slope of the straight line in Figure 5.21 is β = 0.353. The y-intercept of the graph
is 1.05. The initial mean time to fail is
=α e y=
−intercept
2.864(thousands of cycles)
The Duane reliability growth model for the instantaneous mean cycles to fail is
2.864t 0.353
θi ( t ) =
( 1 − 0.353)
Solving the above expression for time,
2.864t 0.353
75 =
( 1 − 0.353)
t = 3,054 thousand cycles
TABLE 5.7 Reliability growth example data.
Cumulative cycles Cumulative mean cycles to

(thousands) Cumulative failures fail (thousands)
250 12 20.83
500 20 25.00
750 27 27.78
1,000 31 32.26
1,250 33 37.88
TABLE 5.8 Computations for Duane growth model example.
Logarithm of Logarithm of
Cumulative Cumulative mean cumulative cumulative mean
cycles Cumulative cycles to fail cycles cycles to fail
(thousands) failures (thousands) (thousands) (thousands)
250 12 20.83 5.5215 3.0364
500 20 25.00 6.2146 3.2189
750 27 27.78 6.6201 3.3243
1,000 31 32.26 6.9078 3.4738
1,250 33 37.88 7.1309 3.6344
FIGURE 5.21 Duane reliability growth plot.
AMSAA Model
The AMSAA model also assumes a nonhomogeneous Poisson process. The instan-
taneous mean time to fail is
1
t
i t (5.30)

where
α is the scale parameter
These parameters can be estimated graphically using the expression
ln N t ln t ln (5.31)
where [N(t)] is the cumulative number of failures at time = t.

By plotting ln[N(t)] on the y-axis and ln t on the x-axis, the slope of the resulting
best-fit straight line through the data is the estimated shape parameter β. The scale
parameter is determined using the y-intercept
y intercept

e
(5.32)
Example 5.18: For the data given in Example 5.17, determine the AMSAA growth
model parameters and the number of cycles of testing required to achieve a mean
time to fail of 75,000 cycles.
Solution: Table 5.9 contains the data required for plotting.
A plot of the data in Table 5.9 is shown in Figure 5.22 .
The slope of the best-fit line in Figure 5.22 is 0.647, which is equal to the shape
parameter β. The y-intercept of the best-fit line in Figure 5.22 is −1.053. The scale
parameter is
−1.053
−
=α e=
0.647
5.087thousand cycles
The instantaneous mean time to fail is

1−0.647
5.087  t 
θi ( t ) = 
0.647  5.087 
Solving for θi(t) = 75, 000 gives
t = 3,056 thousand cycles
TABLE 5.9 AMSAA growth plot computations.
Logarithm of
Cumulative cycles cumulative cycles Logarithm of
(thousands) Cumulative failures (thousands) cumulative failures
250 12 5.5215 2.4849
500 20 6.2146 2.9957
750 27 6.6201 3.2958
1,000 31 6.9078 3.434
1,250 33 7.1309 3.4965
FIGURE 5.22 AMSAA reliability growth plot.
Summary
There are many approaches to planning an accelerated test. Bayesian and sequential
tests can provide efficiency over conventional testing methods. There is no best method
for planning an accelerated test. Several types of test plans should be investigated for
each situation.
6
Accelerated Testing Models
A
ccelerated testing models are useful for translating reliability performance
across differing stress levels. There are models for single sources of stress and
multiple sources of stress. Proper understanding and use of these models can
greatly improve the efficiency of accelerated testing.
Linear Acceleration
Although any transformation function could be used to model acceleration, a linear
transformation of the timescale is almost always used. Under this assumption, the
time to fail under normal operating conditions is
t 0 = εt ε (6.1)
where
ε is the acceleration factor
tε is the time to fail under increased stress conditions
If f(t) represents the probability density function under accelerated conditions,
then the probability density function under normal operating conditions is
1 t 
f0 (t ) = f   (6.2)
ε ε
The reliability function under normal operating conditions is
t 
R0 ( t ) = R   (6.3)
ε
TABLE 6.1 Reliability functions with an acceleration factor of e.
Distribution Reliability function under normal operating conditions

Weibull  β

(t ) exp −  εθt  
R=
   
Lognormal  ln t − ln ε − µ 
R (t ) = 1 − Φ  
 σ 
Normal  t − εµ 
R (t ) = 1 − Φ  
 εσ 
Exponential −
λt
R (t ) = e ε
The hazard function under normal operating conditions is
1 t 
h0 ( t ) = h   (6.4)
ε ε
Table 6.1 gives the transformed reliability functions for several distributions using
an acceleration factor of ε. It can be seen from the reliability functions in this table
that there is no acceleration when ε = 1.
Example 6.1: When tested with an axial loading of 200 lb, the time to fail for a
component was found to follow a Weibull distribution with a shape parameter of
and a scale parameter of 44.2. The normal axial loading is 38 lb, and the acceleration
factor between these two load levels is known to be 12. Determine the reliability of
the component under normal operating conditions at time = 100.
Solution: The expression for reliability under normal operating conditions is
  t  
2.8
()
R=
t exp  −   
  18 ( 44.2 )  
 
At time = 100, the component reliability is
 1002.8 
R ( 100 ) =
exp  − =0.997
 18 ( 44.2 ) 
Field Correlation
Two conditions are required for an acceleration model to be valid:
1. The failure mode(s) must be the same for all stress levels
2. The shape of the distribution must be the same for all stress levels
One method of comparing the shape of the distributions is to produce probability

plots at each stress level. If failure data collected under different levels of stress are plotted
on the same probability plot, the lines should be parallel. (The lines will never be exactly
parallel because of sampling error.) If parallel lines do not result, either the acceleration
is not linear or the chosen probability density function does not adequately model the data.
Another method for verifying that the shape of the distribution has not changed
due to increased stress is to compare the confidence intervals for distribution shape
parameters. This method is usually applied when the failure data are modeled with
the Weibull distribution. If the confidence intervals for the shape parameters do not
overlap, then the distribution shapes are not equal. Beware that overlapping confidence
intervals do not guarantee that the distribution shapes are equal. With small sample
sizes, confidence intervals are large—often too large to discriminate.
Example 6.2: A component was tested at two pressure levels, and Table 6.2 gives
the results. The failure modes were identical; therefore, determine if increasing the
pressure is a valid method of acceleration.
Solution: Table 6.3 shows the computations required for a Weibull probability plot.
Figure 6.1 shows a Weibull probability plot for both sets of data. The lines appear to
be parallel; thus, increasing the pressure is a valid method of accelerating the test.
TABLE 6.2 Time to fail data for pressure testing.

Time to fail Time to fail
(250 psi) (400 psi)
359 29
207 28
225 28
342 22
282 34
326 44
406 28
483 35
403 21
178 34
326 24
252 28
299 35
351 29
202 15
196 37
459 16
470 33
221 28
340 19
241 38
430 11
22
31
39
27
19
TABLE 6.3 Weibull probability plot computations.
134
  1    1 
ln ln    ln ln   
Time to fail Time to fail Count Count F(t) F(t) ln(t) ln(t)   R    R 
(250 psi) (400 psi) (250 psi) (400 psi) (250 psi) (400 psi) (250 psi) (400 psi) (250 psi) (400 psi)
178 11 1 1 0.0313 0.0255 5.1818 2.3979 −3.4499 −3.6543
196 15 2 2 0.0759 0.0620 5.2781 2.7081 −2.5392 −2.7481
202 16 3 3 0.1205 0.0985 5.3083 2.7726 −2.0523 −2.2659
207 19 4 4 0.1652 0.1350 5.3327 2.9444 −1.7118 −1.9306
Accelerated Testing
221 19 5 5 0.2098 0.1715 5.3982 2.9444 −1.4461 −1.6704

225 21 6 6 0.2545 0.2080 5.4161 3.0445 −1.2254 −1.4557
241 22 7 7 0.2991 0.2445 5.4848 3.0910 −1.0345 −1.2715
252 22 8 8 0.3438 0.2810 5.5294 3.0910 −0.8646 −1.1089
282 24 9 9 0.3884 0.3175 5.6419 3.1781 −0.7100 −0.9623
299 27 10 10 0.4330 0.3540 5.7004 3.2958 −0.5666 −0.8279
326 28 11 11 0.4777 0.3905 5.7869 3.3322 −0.4316 −0.7029
326 28 12 12 0.5223 0.4270 5.7869 3.3322 −0.3027 −0.5854
340 28 13 13 0.5670 0.4635 5.8289 3.3322 −0.1780 −0.4737
342 28 14 14 0.6116 0.5000 5.8348 3.3322 −0.0558 −0.3665
351 28 15 15 0.6563 0.5365 5.8608 3.3322 0.0656 −0.2627
359 29 16 16 0.7009 0.5730 5.8833 3.3673 0.1881 −0.1614
403 29 17 17 0.7455 0.6095 5.9989 3.3673 0.3138 −0.0616
406 31 18 18 0.7902 0.6460 6.0064 3.4340 0.4456 0.0377
430 33 19 19 0.8348 0.6825 6.0638 3.4965 0.5882 0.1373
459 34 20 20 0.8795 0.7190 6.1291 3.5264 0.7494 0.2385
470 34 21 21 0.9241 0.7555 6.1527 3.5264 0.9472 0.3425
483 35 22 22 0.9688 0.7920 6.1800 3.5553 1.2429 0.4511
35 23 0.8285 3.5553 0.5670
37 24 0.8650 3.6109 0.6943
38 25 0.9015 3.6376 0.8404
39 26 0.9380 3.6636 1.0224
44 27 0.9745 3.7842 1.2994
FIGURE 6.1 Weibull probability plot for pressure test data.
Example 6.3: Determine the acceleration factor for increasing the test pressure
from 250 psi to 400 psi for the test described in Example 6.2.
Solution: Using the data from Table 6.3, the scale parameters for the Weibull
distribution are
250psi − θ =352.9
400psi − θ =31.5
The acceleration factor for increasing pressure from 250 psi to 400 psi is
352.9
=ε = 11.2
31.0
In many cases, it is not possible to obtain field data for correlation to laboratory
results. When the expected field service of an item is more than 5 years, it is often
not practical to wait until the field data are available for correlation to the laboratory
data. Even correlating an accelerated version of a laboratory test to a non-accelerated
version is often not feasible due to time constraints. For example, although the equiva-
lent of 10 years’ worth of cycles may be applied to an electronic controller in 2 months,
with affordable sample sizes (usually fewer than 20), it may take more than a year to
obtain enough failures for correlation.
A technique for correlation in these situations is to correlate the life remaining
in a unit that has been in the field. For example, if a unit has been in the field for 18
months and 1 hour of the accelerated test is the equivalent of 1 month of field use,
then it is expected that the unit would survive 18 hours less on the accelerated test
than a new unit. One disadvantage of this approach is it requires a large sample size
because of the variation in time-to-fail data.
Example 6.4: Given the data in Table 6.4, how many miles in the field does 1 hour
of testing simulate?
Solution: Figure 6.2 shows a scatterplot of the hours to failure on an accelerated
test versus the miles accumulated in field service. The slope of the regression line
is −124.7. This means each test hour is the equivalent of 124.7 hours of field service.
The poor correlation seen in Figure 6.2 is typical of a remaining life test. For this
example, the correlation coefficient is −0.51, which is typical for this type of test.
TABLE 6.4 Equivalent life test data.
Miles in service Accelerated testing (hours to fail)

4,483 162.9
6,977 253.4
8,392 147.7
8,659 210.0
9,474 266.6
10,243 137.2
10,268 202.9
11,255 219.4
15,206 66.0
15,634 248.1
16,306 116.0
16,681 199.1
18,019 192.8
18,974 148.9
20,908 160.8
21,159 242.8
23,002 191.6
23,009 186.9
23,539 135.4
24,678 147.7
25,026 105.4
25,151 205.2
25,533 242.1
26,918 175.3
28,094 171.2
30,634 84.4
31,201 91.3
31,236 129.7
32,502 153.7
33,076 160.4
33,197 125.9
TABLE 6.4 (Continued) Equivalent life test data.
Miles in service Accelerated testing (hours to fail)

34,618 134.8
35,412 76.9
35,877 116.8
37,978 151.5
38,003 106.6
40,803 145.4
41,187 163.6
41,204 226.3
41,818 117.6
41,951 77.0
42,639 48.6
44,891 110.9
45,929 197.7
46,052 77.9
46,472 136.8
49,592 74.4
50,079 85.5
51,168 82.0
51,666 213.8
52,697 67.6
FIGURE 6.2 Scatterplot of field miles versus laboratory time to fail.

Arrhenius Model
In 1889, the Swedish chemist Svante Arrhenius developed a mathematical model for
the influence of temperature on chemical reactions. Because chemical reactions (e.g.,
lubricant breakdown, corrosion, diffusion of semiconductor material) are responsible
for many failures, the Arrhenius equation has been adapted to model the acceleration
of reliability testing.
The Arrhenius equation states that the rate of reaction, K, is a function of the
absolute temperature T,
B
−
K = Ae T (6.5)
where A and B are constants specific to the reaction conditions.
This equation is often written in the form

 E 
− 
K = K 0e  BT 
(6.6)
where
K0 is the reaction rate at a known temperature
E is the activation energy of the reaction
B is Boltzmann’s constant
The United States (MIL-HDBK-217), British (HRD5), and French (CNET) govern-

ments have developed standards to predict the reliability of electronic equipment
using the Arrhenius equation and assuming an exponential time-to-fail distribution.
Lall et al. [15], Walker [30], and O’Connor [20] have shown that these standards are
of little value in predicting the reliability of electronics. Lall et al. [15] state that these
standards “have been proven inaccurate, misleading, and damaging to cost-effective
and reliable design, manufacture, testing, and support.”
This does not mean that the Arrhenius model cannot be used for accelerated
testing; simply beware of standards based on steady-state temperatures and expo-
nential time-to-fail distributions. A more general form of the Arrhenius model that
allows any desired accelerating stress is
c
t = ke s (6.7)
where
t is the time at which a specified portion of the population fails
k and c are constants specific to the reaction conditions
S is the level of applied stress
The term S in Equation (6.7) could be any type of acceleration factor, including
•• Temperature
•• Temperature cycling rate
•• Vibration level (GRMS)
•• Humidity
•• Load (pressure or pound-force)
•• Salt concentration
•• Voltage cycling rate
By making a logarithmic transformation, Equation (6.7) can be written as
c
ln t = ln k + (6.8)
S
By plotting this expression, the values of the constants in the Arrhenius model
can be determined. Kielpinski and Nelson [13] have developed optimum test plans
for the Arrhenius model. These plans estimate the 10th percentile of the time-to-
failure distribution at the normal operating stress with the greatest possible large-
sample precision. These plans involve testing at only two temperatures, and Meeker
and Hahn [18] caution their use because the plans do not allow any degrees of freedom
for assessing the goodness of fit. Meeker and Hahn make the following recommenda-
tions for developing an accelerated test plan using the Arrhenius model:
•• Restrict testing to a range of temperatures over which there is a good chance
that the Arrhenius model adequately represents the data.
•• Select a second temperature reasonably removed from the highest temperature.
•• Select a low temperature that is as close as possible to the design temperature.
•• Apportion more of the available test units to the lower levels of stress.
Example 6.5: A component was tested at four levels of axial loading, and Table 6.5
shows the results. (A “+” after a data point indicates the item was removed from
testing without failing.) Develop the Arrhenius model for modeling reliability as a
function of axial loading.
Solution: The time variable in the Arrhenius model is the time for a specified
percentage of the population to fail. A percentage of the population (percentile)
could be chosen and determined nonparametrically using the median rank or
another nonparametric method, or a distribution could be used to model the data.

If the Weibull distribution is used, the scale parameter θ is the point in time where
63.2% of the population fails. Figure 6.3 shows a Weibull probability plot for the
hours to fail at the four levels of axial loading.
TABLE 6.5 Arrhenius model sample data.
Hours to fail
2 lb 5 lb 10 lb 20 lb
171.5 39.8 9.1 9.3
178.3 47.1 9.8 9.7
257.2 56.0 11.5 11.4
271.1 57.1 14.4 11.7
333.7 65.9 16.1 13.5
362.9 77.4 20.3 13.8
366.9 79.6 29.4 15.3
400.0+ 85.0 30.8 17.0
400.0+ 88.2 32.6 23.3
400.0+ 89.2 36.1 23.5
400.0+ 100.0+ 36.5 24.7
400.0+ 100.0+ 36.8 24.7
400.0+ 100.0+ 39.3 25.0
400.0+ 100.0+ 40.4 25.1
400.0+ 100.0+ 42.9 26.6
400.0+ 100.0+ 45.2 27.5
400.0+ 100.0+ 47.6 29.4
400.0+ 100.0+ 55.2 30.0
FIGURE 6.3 Weibull probability plot of axial load data.

The Weibull distribution adequately models all data sets, and the four plots
(although not perfectly parallel) are close enough to continue with the analysis.
Remember that the slope of the probability plot is a sample of the Weibull shape
parameter, and the estimation method contains an error.
The plots shown in Figure 6.3 are typical for data from Weibull distributions with
equal shape parameters. (The data were generated using a Weibull random number
generator with a shape parameter equal to 2.8 for all stress levels.) One important consid-
eration when reviewing the Weibull plots is to ensure that the slope is not changing with
the stress level. In Figure 6.3, it can be seen that the slope does change from one data set
to another, but there is movement in both directions as the stress is increased. Note that
the slope for the data at 2 lb is nearly identical to the slope for the data at 20 lb.
For further confirmation of the equality of the four slopes, the results of a
maximum likelihood analysis1 for the data are shown in Table 6.6.
As seen in Table 6.6, the confidence intervals for β overlap considerably for all
stress levels, confirming that the increase in stress had little or no effect on the
distribution shape.
The constants in the Arrhenius model are determined by regressing the inverse of
the stress against the logarithm of the scale parameter (θ—the 63.2 percentile), as shown
in Figure 6.4.
TABLE 6.6 Weibull maximum likelihood parameter estimation.
Weibull parameter 2 lb 5 lb 10 lb 20 lb
β 2.79 3.02 2.46 3.33
Lower 90% Confidence Limit for β 1.77 2.10 1.91 2.58
Upper 90% Confidence Limit for β 4.39 4.36 3.17 4.29
θ 512.70 105.50 34.75 22.49
FIGURE 6.4 Arrhenius model regression.
1
For details on performing a maximum likelihood analysis, consult Chapter 4 “Parameter Estimation.”
The slope of the regression equation, which provides the estimate for the constant
c in the Arrhenius model, is 6.8. The constant k is estimated from the y-intercept,
which is 2.942,
=k e=
2.942
18.9
The Arrhenius model for modeling reliability as a function of axial loading is
6.8
θ = 18.9e pounds
Example 6.6: For the model developed in Example 6.5, determine the reliability
at 200 hours given an axial load of 3.5 lb.
Solution: The estimated scale parameter for an axial load of 3.5 lb is
6.8 6.8
=ˆ 18.9
θ = =
e lb 18.9e 3.5 131.9hours
There were four shape parameters estimated when determining the Arrhenius
model, and there was not enough statistical evidence to prove they were not
equal. What shape parameter should be used for this analysis? There are several
approaches, but a common and simple approach is to use the average of the four
estimated shape parameters. The four estimates for the shape parameter were
2.79, 3.02, 2.46, and 3.33. The average of these four values is 2.9. Using a shape
parameter of 2.9, the reliability at 200 hours, with an axial load of 3.5 lb, is
2.9
 200 
− 
( 200) e=
R=  131.9 
0.0353
Example 6.7: For the data given in Example 6.5, use a nonparametric model to
estimate the time that will yield a 90% reliability with an axial loading of 4 lb.
Solution: The cumulative distribution function (i.e., the percentage of the popula-
tion failing) can be estimated using the median rank,
i − 0.3
MR = (6.9)
n + 0.4
where
i is the order of failure
The median ranks are the same at all load levels, with the exception of the censored
data points. For the second failure, the median rank estimate of F(t) is
2 − 0.3
=
MR = 0.092
18 + 0.4
TABLE 6.7 Median rank computations.
Hours to fail Estimated F(t)

Order 2 lb 5 lb 10 lb 20 lb 2 lb 5 lb 10 lb 20 lb
1 171.5 39.8 9.1 9.3 0.038 0.038 0.038 0.038
2 178.3 47.1 9.8 9.7 0.092 0.092 0.092 0.092
3 257.2 56.0 11.5 11.4 0.147 0.147 0.147 0.147
4 271.1 57.1 14.4 11.7 0.201 0.201 0.201 0.201
5 333.7 65.9 16.1 13.5 0.255 0.255 0.255 0.255
6 362.9 77.4 20.3 13.8 0.310 0.310 0.310 0.310
7 366.9 79.6 29.4 15.3 0.364 0.364 0.364 0.364
8 85.0 30.8 17.0 0.418 0.418 0.418
9 88.2 32.6 23.3 0.473 0.473 0.473
10 89.2 36.1 23.5 0.527 0.527 0.527
11 36.5 24.7 0.582 0.582
12 36.8 24.7 0.636 0.636
13 39.3 25.0 0.690 0.690
14 40.4 25.1 0.745 0.745
15 42.9 26.6 0.799 0.799
16 45.2 27.5 0.853 0.853
17 47.6 29.4 0.908 0.908
18 55.2 30.0 0.962 0.962
Table 6.7 shows all median rank computations.

Because the requirement is to determine a model for time to fail with 90% reli-
ability, the time for 90% reliability (10% median rank) must be determined. Table 6.8
shows that the time of the second failure occurs at F(t) = 9.2%, and the third failure
occurs at F(t) = 14.7%. Table 6.8 shows the linear interpolation to a median rank of
10% for all levels of stress.
The constants in the Arrhenius model are determined by regressing the inverse
of the stress against the logarithm of the 10th percentile, as shown in Figure 6.5.
The slope of the regression equation, which provides the estimate for the constant
c in the Arrhenius model, is 6.8. The constant k is estimated from the y-intercept,
which is 1.99,
=
k e=
1.99
7.3
TABLE 6.8 Interpolation for 10% median rank.
Hours to fail
Percent fail 2 lb 5 lb 10 lb 20 lb
9.2% 178.3 47.1 9.8 9.7
14.7% 257.2 56.0 11.5 11.4
10.0% 189.8 48.4 10.0 9.9
FIGURE 6.5 Nonparametric Arrhenius model regression.
The Arrhenius model for modeling the 10th percentile (i.e., the time at which
90% reliability is achieved) as a function of axial loading is
6.8
t 10% = 7.3e pounds

The estimated 10th percentile for an axial load of 4 lb is
6.8
=
t 10% 7=
.3e 4 40.0 hours
Example 6.8: Determine the Arrhenius model for the characteristic life of the
Weibull distribution for the test described in Example 6.2.
Solution: The parameters of the Weibull distribution estimated from probability
plotting are shown in Table 6.9.
TABLE 6.9 Estimated Weibull parameters.
250 psi 400 psi

Shape parameter 3.68 3.53
Scale parameter 352.9 31.5
With only two data points, a regression cannot be used to determine the slope and
y-intercept of the best-fit line for the Arrhenius model. The slope of the Arrhenius model is
ln ( 352.9 ) − ln ( 31.5 )
c= = 1, 612
 1   1 
  − 
 250   400 
The y-intercept and the Arrhenius constant k are
 1 
y − intercept = ln ( 352.9 ) −   (1, 612 ) = −0.581
 250 
k = e −0.581 = 0.559
The Arrhenius model for modeling the characteristic life of the Weibull distribu-
tion as a function of pressure is
1, 612
θ = 0.559e psi

Eyring Model
The Arrhenius model is limited to one type of stress, which is temperature. When
more than one type of stress acts on a component, a more complex model is required.
Henry Eyring derived a model based on chemical reaction rates and quantum
mechanics. If a chemical process (e.g., chemical reaction, diffusion, corrosion, migra-
tion) is causing degradation leading to failure, the Eyring model describes how the
rate of degradation varies with stress or, equivalently, how time to failure varies with
stress. The Eyring model is
b   d 
c +  T   S1
t = aT α e T e   
(6.10)
where
T is the absolute temperature
S1 is the level of a second stress factor
α, a, b, c, and d are constants
Although the Eyring model was developed for absolute temperature, the absolute
temperature T is sometimes replaced by another stress factor.
The Eyring model can be expanded to include an infinite number of stresses. For
every additional stress factor, the model requires two additional constants. The Eyring
model with three stress factors is
b c +  d   S e +  f   S
  1    2
α t   T  
t = aT e e   T 
e (6.11)
The general Eyring model includes terms that have interactions, that is, the effect
of changing one source of stress varies, depending on the levels of other stresses. In
models with no interaction, acceleration factors can be computed for each stress, and
then multiply them together. This would not be true if the physical mechanism
required interaction terms. In practice, the Eyring model usually is too complicated
to use in its most general form and must be modified or simplified for any particular
failure mechanism. With five constants to be estimated for two stress factors in the
Eyring model, testing must be conducted under at least five distinct stress combina-
tions. Ideally, more than five testing conditions should be used to demonstrate the
goodness of fit. Determining the constants is not straightforward and is best done
with mathematical software, such as Mathcad®.
A modified version of the Eyring model that eliminates the interaction terms is
b
t = ae T S2c (6.12)
This model originally was developed for accelerating chemical reactions with absolute
temperature and voltage, with S2 being replaced by V (voltage). The interactions are elimi-
nated from the model by setting the constants for the interaction terms, α and d, to 0.
With no interaction terms, only three constants require estimation, and this can
be done with multiple linear regression by algebraically manipulating the model into
a linear form. The linear form of the Eyring model with no interaction terms is
b
ln t = ln a + + c ( ln S2 ) (6.13)
T
The simplified Eyring model can also be extended to include an infinite number
of stresses. The general form of this model is
n
b
ln t = ln a + + ∑ [c i ln Si ] (6.14)
T i =1
For example, an Eyring model with four stress sources and no interactions
would be
b
ln t = ln a + + c1 ( ln S1 ) + c 2 ( ln S2 ) + c 3 ( ln S3 ) (6.15)
T
Example 6.9: Given the failure time, temperature (degrees Celsius), and salt spray
concentration data shown in Table 6.10 (a “+” after a data point indicates the item
was removed from testing without failing), develop the Eyring acceleration model
for the characteristic life of the Weibull distribution.
Solution: When two levels of stress are used for the Eyring model, assuming no interac-
tions, three constants must be determined, which means an experiment with three experi-
mental settings is required. Ideally, more than three experimental trials would be used to
allow for the fit of the Eyring model to be checked. Because there are four experimental
trials for this example, there is one degree of freedom to vary the fit of the model.
The time variable in the Eyring model is the time for a specified percentage of the
population to fail. The characteristic life of the Weibull distribution, θ, is the 63.2
percentile. Figure 6.6 shows a Weibull probability plot for the hours to fail at the
four levels of salt concentration.
TABLE 6.10 Eyring model example data.
Hours to fail
Temp. = 20 Temp. = 20 Temp. = 50 Temp. = 50
Salt = 5% Salt = 10% Salt = 5% Salt = 10%
171.5 39.8 9.1 9.3
178.3 47.1 9.8 9.7
257.2 56.0 11.5 11.4
271.1 57.1 14.4 11.7
333.7 65.9 16.1 13.5
362.9 77.4 20.3 13.8
366.9 79.6 29.4 15.3
400+ 85.0 30.8 17.0
400+ 88.2 32.6 23.3
400+ 89.2 36.1 23.5
400+ 100+ 36.5 24.7
400+ 100+ 36.8 24.7
400+ 100+ 39.3 25.0
400+ 100+ 40.4 25.1
400+ 100+ 42.9 26.6
400+ 100+ 45.2 27.5
400+ 100+ 47.6 29.4
400+ 100+ 55.2 30.0
FIGURE 6.6 Weibull probability plot of salt concentration data.

The Weibull distribution adequately models all data sets, and the four plots
(although not perfectly parallel) are close enough to continue with the analysis.
Remember that the slope of the probability plot is a sample of the Weibull shape
parameter, and the estimation method contains an error. The plots shown in Figure
6.6 are typical for data from Weibull distributions with equal shape parameters.
(These data were generated using a Weibull random number generator with a shape
parameter equal to 2.8 for all stress levels.) One important consideration when
reviewing the Weibull plots is to ensure that the slope is not changing with the stress
level. In Figure 6.6, it can be seen that the slope does change from one data set to
another, but there is movement in both directions as the stress is increased.
For further confirmation of the equality of the four slopes, the results of a
maximum likelihood analysis* for the data are shown in Table 6.11.
As seen in Table 6.11, the confidence intervals for β overlap considerably for all
stress levels, confirming that the increase in stress had little or no effect on the
distribution shape.
The original Eyring model uses the absolute temperature (T + 273 for Celsius,
or T + 460 for Fahrenheit). The constants in the Eyring model are determined using
multiple linear regression. Table 6.12 shows the data for this regression.
There are a multitude of options for performing multiple regression2. The multiple
regression is easily performed using Microsoft Excel. The following steps describe
the process3.
TABLE 6.11 Weibull maximum likelihood parameter estimation.

Weibull parameter Salt = 5% Salt = 10% Salt = 5% Salt = 10%
β 2.79 3.02 2.46 3.33
Lower 90% confidence limit for 1.77 2.10 1.91 2.58
Β
Upper 90% confidence limit for 4.39 4.36 3.17 4.29
Β
θ 512.70 105.50 34.75 22.49
TABLE 6.12 Eyring model data for multiple linear regression.
1
θ In(θ) Temp. (°C) Temp. (°K) Salt (%) Temp. ln(Salt)
512.7 6.2397 20 293 5 0.0034 −2.9957
105.5 4.6587 20 293 10 0.0034 −2.3026
34.8 3.5482 50 323 5 0.0031 −2.9957
22.5 3.1131 50 323 10 0.0031 −2.3026
2
In addition to all the available commercial statistical software, JASP is an open source statistical package
built on R with a simple to use interface.
3
Multiple regression can also be performed in Excel using the array based formula LINEST, but the proce-
dure is more difficult.
FIGURE 6.7 Analysis tool pack.
1. Select Data Analysis from the Tools menu.

2. If Data Analysis does not appear on the Tools menu, select Add-Ins from the
Tools Menu and then check the Analysis Tool Pack box, as shown in Figure 6.7.
3. Select Regression from the Analysis Tools list.
4. Select the column of data containing the natural logarithm of the
characteristic life for the Input Y Range field. (Select the column heading
along with the data.)
5. Select the two columns of data containing the inverse of absolute
temperature and the natural logarithm of the salt concentration for the Input
X Range field. (Select the column headings along with the data.)
6. Check the Labels box.
7. Select a cell for the Output Range. This cell will be the upper left corner of the
results of the regression.
Figure 6.8 shows Steps 4 through 7, and Figure 6.9 shows the results of the regression.
Before determining the constants of the Eyring model, the regression results
should be analyzed. The Adjusted R Square measures the amount of variation in the
data explained by the model. In this case, the regression output indicates that 83.12%
of all the variation in the characteristic life is explained by the Eyring model.
The term Significance F in the ANOVA section is a measure of Type I error for
the regression. As a rule of thumb, if the significance is less than 5%, the model
adequately represents the data; if the significance is greater than 5%, then the model
may not be adequate. In this case, a significance of 23.72% indicates that the Eyring
model may not be valid, although it appears that there is an 83.21% of variance in the
data set.
The P-values for coefficients in the regression provide the Type I error (or signifi-
cance) for each of the coefficients estimated in the regression model. If the P-value is
FIGURE 6.8 Analysis tool pack multiple linear regression.
FIGURE 6.9 Eyring model regression results.

less than 5%, then the slope for the stress factor is not statistically different than 0,
which means the stress factor has no proven impact on time to fail. For this example,
the Type I error for the regression slope for the inverse of absolute temperature is
16.8%, and the Type I error for the regression slope for the natural logarithm of salt
concentration is 32.9%. Based on the Type I error, neither of these factors impacts
the time to fail. This is consistent with the lack of significance for the overall
regression model.
If neither of the stress factors affects the time to fail, how can the change in the
characteristic life be explained? Recall that the characteristic life was 512.7 hours
when the temperature was 20°C and the salt concentration was 5%. Likewise, the
characteristic life was 22.49 hours when the temperature was 50°C and the salt concen-
tration was 10%. To determine statistical significance, there must be a method for
estimating the error of the estimate. The most precise method for estimating this
error is to repeat experimental trials. In this example, although each set of experi-
mental conditions (e.g., temperature = 20°C, and salt concentration = 5%) had 18
samples, there was only one result for each set of experimental conditions used in
the regression analysis. The regression analysis does use the individual 18 data points.
If another 18 samples were tested, and the temperature was 20°C and the salt concen-
tration was 5%, would the resulting characteristic life be 512.7? Without providing
repeat trials, the regression assumes that interactions have no statistical significance
and uses the degrees of freedom for the interaction to estimate experimental error.
The degrees of freedom for this regression are defined as follows:
Term Degrees of freedom

Overall mean 1
1 1
Temperature
ln ( Salt concentration ) 1
1 1
Interaction between and
Temperature
ln ( Salt concentration )
If the interaction has no impact on the time to fail, then the regression is valid;
however, if the interaction is present, then the experimental error is over estimated,
causing the remaining terms in the model to appear less significant than they
really are.
In this example, the conclusion is that the Eyring model with no interaction does
not adequately model that data. It appears that interaction is present. To provide
enough degrees of freedom to model the interaction, at least one of the experimental
trials should be repeated. If the Eyring model did adequately model the data, the
constant a is estimated from the y-intercept,
a = e y −intercept = e −22.4 = 1.8439 (10−10 )

TABLE 6.13 Residuals analysis of the Eyring model.
Temperature (°C) 20 20 50 50
Salt 5% 10% 5% 10%
concentration
Actual 0 512.7 105.5 34.75 22.49
Predicted 0 421.94 153.97 44.98 16.41
Residual −90.76 48.47 10.23 −6.08
Percent error −21.5% 31.5% 22.7% −37.0%
The constant b is estimated by the slope for the inverse of absolute temperature,
which is 7,061.8. The constant c is estimated by the slope for the natural logarithm of
the salt concentration, which is −1.454. The Eyring model for the characteristic life
of the Weibull distribution is
7 , 061.8
θ = 1.8439 (10−10 ) e T
S −1.454
A second method for checking the fit of the Eyring model to this data is an
analysis of residuals, as shown in Table 6.13. There appears to be little or no bias in
the residuals, but the error is in the range of 20% to 40% of the predicted value. If
this error is sufficient, then the model can be used.
Example 6.10: For the model developed in Example 6.9, determine the reliability
with a temperature of 35°C and a salt concentration of 8% at 80 hours in service.
Solution: The estimated scale parameter is
7,061.8
(
θ =1.8439 10−10 e ) T
S −1.454
7,061.8
=
=
1.8439 10−10 e ( )
(
0.08−1.454 65.9hours ) 35+273
( )
Four shape parameters were estimated when determining the Eyring model, and
there was not enough statistical evidence to prove they were not equal. What shape
parameter should be used for this analysis? There are several approaches, but a
common and simple approach is to use the average of the four estimated shape
parameters. The four estimates for the shape parameter were 2.79, 3.02, 2.46, and
3.33. The average of these four values is 2.9. Using a shape parameter of 2.9, the
reliability at 80 hours, with a temperature of 35°C and a salt concentration of 8%, is
2.9
 80 
− 
( 200) e=
R=  65.9 
0.1724
Example 6.11: For the data given in Example 6.9, use a nonparametric model to
estimate the time that will yield a 90% reliability with a temperature of 35°C and a
salt concentration of 8%.
Solution: The cumulative distribution function (i.e., the percentage of the popula-
tion failing) can be estimated using the median rank,
i − 0.3
MR = (6.16)
n + 0.4
where
i is the order of failure
The median ranks are the same at all load levels, with the exception of the censored
data points. For the second failure, the median rank estimate of F(t) is
2 − 0.3
=
MR = 0.092
18 + 0.4
Table 6.14 shows all median rank computations.

Because the requirement is to determine a model for time to fail with a 90%
reliability, the time for 90% reliability (10% median rank) must be determined. Table
6.15 shows that the time of the second failure occurs at F(t) = 9.2%, and the third
failure occurs at F(t) = 14.7%. Table 6.15 shows the linear interpolation to a median
rank of 10% for all levels of stress.
The original Eyring model uses the absolute temperature (T + 273 for Celsius,
or T + 460 for Fahrenheit). The constants in the Eyring model are determined using
multiple linear regression. Table 6.16 shows the data for this regression.
The constant a is estimated from the y-intercept,
a = e y −intercept = e 23.68 = 5.17522 (1011 )

The constant b is estimated by the slope for the inverse of absolute temperature,
which is 7,534.3. The constant c is estimated by the slope for the natural logarithm of
the salt concentration, which is −0.993. The Eyring model for the 10th percentile (90%
reliability) is
7 , 534.3
t = 5.17522 (10−11 ) e T
S −0.993
For 35°C and 8% salt concentration, the time at which a 90% reliability is
achieved is
7 , 534.3
t = 5.17522 (10−11 ) e T
S −0.993
7 , 534.3
= 5.17522 (10−11 ) e 35 + 273 ( 0.08−0.993 ) = 26.7hours
TABLE 6.14 Median rank computations.
Hours to fail Estimated F(t)

Temp. = Temp. = Temp. = Temp. = Temp. = Temp. = Temp. = Temp. =
20 20 50 50 20 20 50 50
Salt = Salt = Salt = Salt = Salt = Salt = Salt = Salt =
Order 5% 10% 5% 10% 5% 10% 5% 10%
1 171.5 39.8 9.1 9.3 0.038 0.038 0.038 0.038
2 178.3 47.1 9.8 9.7 0.092 0.092 0.092 0.092
3 257.2 56.0 11.5 11.4 0.147 0.147 0.147 0.147
4 271.1 57.1 14.4 11.7 0.201 0.201 0.201 0.201
5 333.7 65.9 16.1 13.5 0.255 0.255 0.255 0.255
6 362.9 77.4 20.3 13.8 0.310 0.310 0.310 0.310
7 366.9 79.6 29.4 15.3 0.364 0.364 0.364 0.364
8 85.0 30.8 17.0 0.418 0.418 0.418
9 88.2 32.6 23.3 0.473 0.473 0.473
10 89.2 36.1 23.5 0.527 0.527 0.527
11 36.5 24.7 0.582 0.582
12 36.8 24.7 0.636 0.636
13 39.3 25.0 0.690 0.690
14 40.4 25.1 0.745 0.745
15 42.9 26.6 0.799 0.799
16 45.2 27.5 0.853 0.853
17 47.6 29.4 0.908 0.908
18 55.2 30.0 0.962 0.962
TABLE 6.15 Interpolation for 10% median rank.
Hours to fail
Percent fail Salt = 5% Salt = 10% Salt = 5% Salt = 10%
9.2% 178.3 47.1 9.8 9.7
14.7% 257.2 56.0 11.5 11.4
10.0% 189.8 48.4 10.0 9.9
TABLE 6.16 Eyring model data for multiple linear regression.
1
10th ln(10th
percentile percentile) Temp. (°C) Temp. (°K) Salt Temp. ln(Salt)
189.8 5.2458 20 293 5% 0.0034 −2.9957
48.4 3.8794 20 293 10% 0.0034 −2.3026
10.0 2.3073 50 323 5% 0.0031 −2.9957
9.9 2.2973 50 323 10% 0.0031 −2.3026
Voltage Models
There are many empirically derived acceleration models for voltage, humidity, vibra-
tion, and other acceleration factors. The following empirical relationship has been
shown to hold true for accelerating failure by changing voltage levels:
3
t  Vε 
=   (6.17)
tε  V 
where
t is the time at which a specified portion of the population fails under normal
operating conditions
tε is the time at which a specified portion of the population fails under accelerated
operating conditions
V is the normal voltage
Vε is the accelerated level of voltage
The inverse voltage power model, shown in Equation (6.18), is a simplified Eyring
model that has been shown to be useful in predicting the life of capacitors:
t = aV −b (6.18)
where
V is voltage
a and b are constants to be determined
Lifetime as a function of voltage has been shown to follow the exponential model
shown in Equation (6.19):
t = ae bV (6.19)
where
V is voltage
a and b are constants to be determined
A modified form of the Eyring model has been proven useful for modeling the
acceleration of electromigration failures in semiconductors. The ionic movement is
accelerated by temperatures and current density. The modified Eyring model for
semiconductor electromigration is
c
t = aJ −b e T (6.20)
where
J is the current density
T is the absolute temperature
a, b, and c are constants to be determined
Mechanical Crack Growth

The Coffin-Manson model was developed to model material cracking, fatigue, and
deformation as a function of temperature cycles. This number of cycles to fail is
modeled by the expression
t = aF bT c
where
t is the number of cycles at which a specified portion of the population fails
F is the cycling frequency
T is the temperature range during a cycle
a, b, and c are constants to be determined
Example 6.12: For the data given in Table 6.17, determine the parameters of the
Coffin-Manson acceleration model, and determine the number of cycles that gives
a 90% reliability for 2 temperature cycles per hour, with a minimum temperature
of −20°C and a maximum temperature of 100°C. Note that the 10th percentile
represents the time at which 10% of the population fails.
Solution: A linear version of the Coffin-Manson model is
lnt =
ln a + b ln F + c lnT
The coefficients of this equation can be found with multiple linear regression, using
the natural logarithm of the 10th percentile as the independent variable and the
natural logarithm of F and the natural logarithm of T as the dependent variables.
(Example 6.9 shows the procedure for using Microsoft Excel for multiple linear
regression.)
TABLE 6.17 Coffin-Manson example data.
F T 10th percentile
1 80 487.1
5 80 257.3
1 150 301.9
5 150 82.4
Using multiple linear regression,
a = 161, 078
b = −0.602
c = −1.286
The number of cycles that gives a 90% reliability is
t = 161, 078 F −0.602T −1.286
The number of cycles that gives a 90% reliability for 2 temperature cycles per hour,
with a minimum temperature of −20°C and a maximum temperature of 100°C, is
t = 161, 078 ( 2−0.602 )(120−1.286 ) = 224.7
Degradation Testing
Modeling performance degradation can dramatically reduce the testing duration
and sample sizes. Two features of degradation testing allow efficiency in test duration
and sample size:
1. The time to fail is predicted with a model of how performance degrades
over time.
2. The statistical methods use continuous data from the performance
parameters that degrade, rather than binomial, pass-fail statistics.
The ability to predict performance allows a test to be abandoned early if the model
predicts that the reliability requirements will not be achieved, saving testing resources.
Testing also may be stopped early if the model predicts a successful test, but extrapo-
lating for a successful test is much riskier than extrapolating to a prediction of failure.
The disadvantage of degradation testing is that a model is required for the degra-
dation. The acceleration models described in previous sections can be used as degra-
dation models. Models for specific phenomena are also available.
Degradation testing is commonly used for metal creep, crack initiation, crack
propagation, tensile strength, flexural strength, corrosion, rust, elongation, breakdown
voltage, resistance, and many others.
Consider a simple case of linear degradation, such as material wear. (Examples
of this are brake pads and tire tread.) A component fails as the material is removed
by use. When a component is new, the material thickness is 40 millimeters. The
component is considered to be in a failed state when the material thickness is less
than 8 millimeters. Table 6.18 gives a test data for four components, and Figure 6.10
shows a plot of these data.
TABLE 6.18 Linear degradation example data.
Material thickness (millimeters)

Test duration (hours) Sample A Sample B Sample C Sample D
10 36.8 36.5 36.9 37.1
20 33.4 35.3 36.0 33.8
30 29.7 32.2 33.6 31.7
40 25.8 28.8 30.3 29.3
50 25.2 27.1 27.4 26.8
FIGURE 6.10 Linear degradation example data.
From Figure 6.10, it can be seen that the assumption of linear degradation is valid
through 50 hours of use. Without prior experience, extrapolating beyond 50 hours
could cause errors. Material thickness as a function of test duration is modeled with
the expression
y = a + bx + e
where
y is the material thickness
x is the test duration
e is the error in the model
The parameters in this model can be estimated using least squares regression:
b=
n ∑ i =1xi yi −
n
( ∑ x )( ∑ y ) (6.21)
n
i =1 i
n
i =1 i
n ∑ i =1 x − (∑ x )
n 2 n
i i =1 i
∑ y − b ∑ i =1xi
n n
i =1 i
a= (6.22)
n
where n is the sample size.
A prediction interval that contains a single response with a confidence of (1 − α) is
1 (x − x )
2
ŷ 0 ± t α s 1 + + 0 (6.23)
2 n Sxx
where t the critical value of the t-distribution with n − 2 degrees of freedom, and
S yy − bSxy
s= (6.24)
n−2
( ∑ x ) (6.25)
n 2
n
i =1 i
S xx = ∑x i2 −
i =1 n
( ∑ y ) (6.26)
n 2
n i
S yy = ∑y i2
i =1
−
i =1 n

n
S xy = ∑x i y i −
( ∑ x )( ∑ y ) (6.27)
n
i =1 i
n
i =1 i
i =1 n
A confidence interval that contains the mean response with a confidence of

(1 − α) is
1 ( x0 − x )
2
ŷ 0 ± t α S + (6.28)
2 n Sxx
Solving Equations (6.21) and (6.22), the predicted material thickness is
y = 39.63 − 0.265 x
Recall that the component is considered to be in a failed state when the material
thickness is less than 8 millimeters. If the component is required to function for 200
hours, the expected material thickness would be
y = 39.63 − 0.265 ( 200 ) = −13.3

Thus there is no need to continue this test for 200 hours. Failure can be predicted
at 50 hours.
Now consider the same component in the preceding example, with a change in
requirements. When a component is new, the material thickness is 40 millimeters.
The component is considered to be in a failed state when the material thickness is
less than 8 millimeters, and the required life of the component is 105 hours. The
expected material thickness after 105 hours of use would be
y = 39.63 − 0.265 (105 ) = 11.8
The expected material thickness meets the requirement. Confidence limits and
prediction limits for the expected material thickness can be computed using the
method of maximum likelihood [31] or by using the error of the regression model.
Using the regression model, a 90% confidence interval for the mean material thick-
ness after 105 hours of operation is
9.2 ″ y ″ 14.5
A 90% prediction interval for the material thickness after 105 hours of operation is
8.4 ″ y ″ 15.2
After 105 hours of operation, there is a 95% chance (one side of the 90% confidence
interval) that the mean value of the material thickness is greater than 9.2, and it is
expected that 90% of the population will have a material thickness greater than 8.4.
Testing could be suspended at this point because predictions indicate success. Caution
should be taken when accelerated degradation testing is used to cease testing early
in cases of success. Extrapolating accelerated degradation test models could cause
invalid test results.
Example 6.13: Table 6.19 gives degradation data for five prototype switches. The
designed life of the switch is the equivalent of 155 hours of testing.
a. Projecting to 155 hours of testing, 90% of the population will have a resistance
less than what value?
b. If the maximum design resistance is 550, what decision should be made after
70 hours of testing?
Solution: Figures 6.11 and 6.12 show linear and logarithmic models fitted to the
switch degradation data in Table 6.19.
TABLE 6.19 Switch degradation data.
Test duration Resistance

(hours) Switch A Switch B Switch C Switch D Switch E
10 240.7 255.5 293.3 268.7 295.5
20 316.2 297.9 316.1 305.0 348.1
30 380.7 392.4 343.7 338.2 362.0
40 409.1 436.0 397.5 359.5 400.9
50 419.2 459.8 429.2 442.5 445.3
60 455.8 472.8 444.0 464.6 455.3
70 475.8 479.8 449.3 468.9 470.2
FIGURE 6.11 Linear model of switch degradation data.
FIGURE 6.12 Logarithmic model of switch degradation data.

As seen from Figures 6.11 and 6.12, the logarithmic model fits the degradation
data better than the linear model. Linear regression can be used to project degrada-
tion by transforming the logarithm of x in place of x. (This example is contained in
Microsoft Excel on the accompanying CD.) Using the logarithmic model, the expected
switch resistance after 155 hours is 553.1, and the upper 95% prediction limit for
resistance is 600. Given a design resistance requirement of 550, testing should
be stopped at this point; this switch does not meet reliability requirements.
Qualitative Tests
Because of the difficulty of predicting field performance from accelerated tests, several
types of tests are designed to reveal failure modes and performance limits with little
or no regard for field correlation. These types of tests have three purposes:
a. To define the operating envelope
b. To discover failure modes that may or may not be addressed
c. To compare
Step-Stress Testing
The purpose of step-stress testing is to define an operating envelope. A component is
subjected to successively higher levels of stress until the component fails or the limits
of the test equipment are reached. Figure 6.13 shows a typical step-stress test plan.
FIGURE 6.13 Typical step-stress test.

TABLE 6.20 Step-stress temperature test plan.
Time (am) Test temperature (°C)

8:00 20
8:05 25
8:10 30
8:15 35
8:20 40
8:30 45
8:40 50
8:50 60
9:00 70
9:10 80
9:20 90C
9:30 100
9:40 110
9:50 120
10:00 End test
Step-stress testing may involve one or more levels of stress. In some cases, all but
one type of stress are held constant; in some cases, the level changes for multiple levels
of stress. For example, the temperature may be held constant while the vibration level
is increased, or both the temperature and vibration may be increased during a single
step-stress test.
An area of concern with step-stress testing is that the failures are a result of
accumulated damage rather than the latest level of stress. For example, consider a
component that fails after 55 minutes of operation due to a design flaw. This compo-
nent is exposed to a step-stress test to determine the maximum operating temperature.
Table 6.20 gives the test plan. The step-stress plan would show failures at 60°C and
incorrectly assume the maximum operating temperature is 50°C.
This problem can be avoided by testing multiple samples at different beginning
levels of stress. For example, a second sample could start testing at 30°C, and a third
sample could begin testing at 40°C.
Elephant Tests
Elephant testing is named because the test “steps on the product with the force of an
elephant.” Elephant tests do not attempt to correlate to field life; they merely try to
precipitate failures. When failures are found, those failures may be
•• Ignored
•• Corrected to improve reliability
•• Used for comparison to another supplier, another design, or another
production lot
Elephant tests may contain one or more types of stress, and each type of stress
may be at one or more levels. Products are commonly subjected to more than one
elephant test, with each test designed for a particular type of failure mode.
HALT and HASS4

Highly accelerated life testing (HALT) is designed to discover failure modes. The
temperature and vibration levels used during a HALT test may purposely be different
than the actual operating conditions. During the HALT process, a product is subjected
to increasing stress levels of temperature and vibration (independently and in combi-
nation), rapid thermal transitions, and other stresses specifically related to the opera-
tion of the product.
The information goals of HALT are as follows:
•• To determine multiple failure modes and root causes.
•• To determine functional operating limits.
•• To determine functional destruct limits.
HALT was developed specifically for solid-state electronics but can be applied
to any product, with the correct equipment. In general, the information goal of HALT
is most effectively met by testing at the lowest possible subassembly. Consider a balance
between a “functional assembly” and determine the technological limit of a product
feature. Consider removing structure that dampens vibration or blocks the airflow,
reducing the applied stresses to subassemblies. Consider that loss of critical circuitry
or connections may affect the subassembly tested.
A complaint with HALT is the inability to reproduce failure modes. This some-
times occurs because of the random nature of HALT tests. Especially with the use of
air hammers, vibration is random and uncontrolled across a frequency range. If a
test is repeated, the vibration exposure will likely be different.
The results of a HALT test often are used to design a highly accelerated stress
screen (HASS). A typical HASS applies all stresses simultaneously. During vibration,
the temperature is continuously ramped between brief dwells at extremes. Some
typical HASS properties are as follows:
•• Includes all critical product stresses, not only temperature and vibration. For
example, a HASS test may include power cycling, voltage spikes, and other
environmental factors.
•• A HASS test may screen and evaluate the product beyond operating limits
and near destruct limits.
•• A HASS test precipitates failures in the product due to latent defects.
4
HALT and HASS material is used with permission of Intertek, Inc., www.intertek.com.
Summary
There are many types of accelerated tests. Some of the tests are quantitative, and some
are qualitative. It is recommended to consider the limitations of the quantitative
models when making business decisions. For example, how well can a test with a
duration of weeks (or less) mimic the field usage of a product for 10 years? Also the
statistical confidence computed from quantitative tests requires a random sample
representative of the population that will be placed in the field. These tests are usually
prototypes and do not account for all sources of variation, such as
•• Changing operators
•• Multiple batches of raw materials
•• Changing suppliers
•• Equipment maintenance
•• Shift changes
Before deciding on a testing plan, review the limits of each test when assessing
the alternatives.
7
Environmental
Stress Screening
A
ccelerated testing models are often used to create screening or burn-in
programs for use in production. This chapter presents the details of envi-
ronmental stress screening (ESS) and a model to financially optimize the
return on investment (ROI).
Stress Screening Theory

Electronic hardware product reliability requires both an understanding of why
products fail and the advantages of finding, forcing, and correcting these failures. Of
critical importance is being able to address the types of product failure modes that
may not be discovered through inspection or testing, such as the following:
•• Parameter drift
•• Printed circuit board shorts and opens
•• Incorrect installation of a part
•• Installation of the wrong part
•• Contaminated part
•• Hermetic seal failure
•• Foreign material contamination
•• Cold solder joints
•• Defective parts
Product Reliability Equation

In the efforts to qualify components, assemblies, and finished systems, electronics
manufacturers have placed a concerted emphasis on reliability to the extent that
reliability is now generally considered a factor in product performance.
The reason for this is that even well-designed products, manufactured through
an accepted or mature process environment, can fail. The “how” and “why” of many
of these failures can be traced to a number of factors, although most fall within
three categories:
1. Defective components
2. Workmanship error
3. Process stress
These are typically defined as latent defects, which are flaws that are not readily
apparent—or identifiable through conventional testing—but that, with time, can
precipitate product failure.
Cumulative failure data indicate that latent defects generally lead to failures early
in the life of a product. Once a product passes through this stage, it is usually reliable,
as shown in a typical life-cycle curve in Figure 7.1, often referred to as a bathtub curve,
or in the trimodal distribution of failure density plotted over the life of a product.
From this basic model, life-cycle patterns can be plotted, as shown in Figure 7.2.
The first break, where most failures can be traced, is referred to as infancy. The center
section of the curve is the useful life of a product.
FIGURE 7.1 Bathtub curve.

FIGURE 7.2 Life-cycle failure patterns.
The final upsweep represents the normal wear-out period of a product. The critical
factor to controlling product reliability is to ensure that products have successfully
passed through infancy prior to delivery to a customer. Put another way, it requires
establishing a means of precipitating infancy failures in-house.
Customer satisfaction is only one reason why an intensified focus on reliability
is necessary. Others include the following:
•• Mandated Qualification—Both the consumer and military industries have
stepped up emphasis on reliability certification. Some even include a penalty
clause for failure to prove reliability as part of a contract.
•• Short Market Life Cycles—Rapidly changing technologies and increased
competition diminish the amount of attention or interest a product can
generate. A defective introductory product can virtually kill a company’s
opportunity in that market.
•• High Rework Costs—Repairing or replacing components at later phases of
production is both time and cost intensive.
•• High Warranty Costs—The expense of having to disassemble a product in
the field, or administer a product recall, can have both a hidden and needless
impact on unit cost.
•• Lost Market Share—Continued customer dissatisfaction can have a long-term
impact that not only affects profitability but may require increased marketing
and public relations expenditure to counter.
Resolving these concerns has led to a reexamination of qualification theory. More

conventional forms of testing, in which failure-free product operation is desired, are
being augmented by ESS, in which failures are both expected and necessary to assure
parts and workmanship reliability.
What Is ESS?
By basic definition, ESS consists of five screening products, ideally at the most
cost-effective point of assembly to expose defects that cannot be detected by visual
inspection or electrical testing. These defects typically are related to defective parts,
workmanship, or process and are major contributors to early product field failure.
Through an ESS program, 100% of a group of products is subjected to an envi-
ronmental stimulus, or a set of stimuli, for a predetermined time for the purpose of
forcing failures to occur in-house.
Critical to an understanding of ESS is that, within the program, failures are
expected, normal, and unavoidable. In this sense, ESS is radically different from conven-
tional certification testing, which requires failure-free operation as proof of reliability.
It is equally important to realize that ESS is not simply an amplified certification test
forcing greater, and possibly damaging, stress on the product. In fact, it is not a test at
all, but a program used to upgrade product reliability in ways testing cannot.
Toward this end, two important factors are key to the proper implementation of ESS:
•• An optimum level of stress must be applied to the product to force defects
into failure.
•• The stress environment must not exceed the electrical or mechanical limits of
the product, forcing needless failure or reducing the useful life of the product.
When these ideas are properly applied, ESS becomes a dynamic product reliability
tool, with these attendant benefits:
•• Reduced field repair expense
•• Fewer defects and waste
•• Elimination of less effective screening procedures
•• Early detection of design and manufacturing flaws
•• Improved production efficiency
•• Lower unit costs
•• Increased product value
•• Improved customer satisfaction
•• A better ROI
Evolution of ESS
ESS has its origins in environmental testing methodology—basically, a means of
exposing samples of product assemblies to one or more simulated field conditions.
Driven by the military, environmental testing became prominent during World War
II when the proliferation of sophisticated weaponry, aircraft, and communications
systems demanded a less cost- and time-intensive method of proving reliability than
actual use testing. Laboratory experimentation with small-scale hardware was
conducted. An insulated chamber, equipped with the technology to simulate environ-
ments such as temperature, humidity, altitude, and others, was used. These tests were
performed during developmental stages to verify the design and, on a manufacturing
audit basis, to measure design compliance.
After the war, the new electronic technologies became available to the consumer
market, a situation that created a different need. As products were downsized, their
complexity increased, involving unique processing and assembly procedures. The
accumulation of field failure data showed that design compliance was no longer
sufficient evidence of reliability. It also led to the discovery that a number of unpre-
dictable factors involved in parts design, product workmanship, and manufacturing
processes were contributing to failure rates. As many of these were occurring during
the infancy of the product, it was determined that the testing methodology that could
mimic the infancy stage of the product would provide the cure.
In an attempt to pass products through infancy, a process of powering products
for an extended length of time was introduced. Referred to as burn-in, the process
also generated a high degree of heat, a stress that many believed would have an added
impact in precipitating early product failures.
Burn-in testing did succeed in forcing a small number of infancy failures, but
not enough to significantly alter field failure rates. In addition, the process was time
consuming, a factor that slowed production, increased unit cost, and delayed delivery.
An answer came, again from the military, with the introduction of its newly
devised approach of environmental simulation testing in the form of military stan-
dards. These standards required that products be operated at specific environmental
extremes. The logic was that field operation would demand that products be used in
a variety of environments, as well as be exposed to cyclical changes within these
environments. For example, a jet climbing from a desert runway to an altitude would
be exposed to a significant temperature change in a relatively short period. A static
testing process, such as burn-in, could not provide the proper environmental simula-
tion. To address this problem, a combination form of environmental simulation was
introduced. This early application involved temperature cycling between hot and cold
extremes paired with low-frequency vibration in a mechanical shaker.
These basic concepts led to the development of mission profile testing, in which
products were exposed to environments that simulated actual use conditions. Mission
profiling produced some important results. First, it was discovered that a number of
products could operate successfully at the temperature extremes but would fail while
being taken through multiple temperature cycles. This process, known as thermal
cycling, was the basis for ESS theory. Second, a significantly higher number of failures
were forced than with burn-in, particularly the types of failures that were occurring
in the field. Finally, thermal cycling precipitated failures in a much shorter time. It
also became evident that testing of this type produced a stress—through changing
temperatures and the resultant expansion and contraction this caused—that would
force previously undetectable latent defects into product failures.
From these discoveries, the concept of ESS was born, and the distinctions between
this discipline and conventional test methodology became more defined. Where
environmental testing was used primarily for design validation, environmental
screening could be used to qualify product materials and workmanship.
As interest in ESS developed, a number of companies became involved in experi-
mentation with different forms of stress and made attempts to establish specific
guidelines for each type of program. As a result, some of the early distinctions drawn
between ESS and conventional testing were blurred, creating confusion in industries
where ESS had gained only preliminary interest. In an attempt to clarify the situation,
the Institute of Environmental Sciences (IES) undertook a comprehensive study of
ESS. Its book Environmental Stress Screening of Electronic Hardware (ESSEH) sets
out to demystify the process and provide some general guidelines for ESS applications.
This effort, and follow-up studies conducted by the IES and other sources,
provided solid evidence that ESS is a highly successful method of enhancing
product reliability.
Misconceptions about ESS

Like any specialized technology, ESS is subject to misunderstanding. The following
list includes the more common misconceptions:
•• ESS Is a Test—ESS is not a test. It is not intended to validate the design, which
requires a failure-free simulation process. Rather ESS is a screening process
that requires stimulation to expose latent defects in products that would
otherwise fail in the field.
•• ESS Is the Same as Burn-In or Aging—ESS evolved from burn-in techniques,
but it is a considerably advanced process. Burn-in is a generally lengthy process
of powering a product at a specified constant temperature. ESS is an accelerated
process of stressing a product in continuous cycles between predetermined
environmental extremes.
•• ESS Involves Random Sampling—ESS requires that all products be exposed.
Because latent product defects are random by nature, screening all products
is the only way to assure the effectiveness of the program.
•• ESS Can Be Used to Validate Design—Although ESS may occasionally
expose a design inconsistency, its intent and methodology are different.
Design validation uses environmental simulation and laboratory analysis to
determine whether a product is viable and whether it will work within the
environment for which it was designed, usually called its mission profile. ESS,
on the other hand, applies maximum permissible stimulation to products for
the purpose of exposing latent defects related to parts, manufacture, assembly
technique, or processing. ESS produces data that can be analyzed (e.g., for
pattern part failures, process defects, or design errors), but its principal intent
is to force failures. The mission profile, in this sense, is irrelevant. What is
important is that the maximum degree of stress is applied to precipitate failure
without detracting from the useful life of the product.
•• ESS Forces Needless Failures—ESS is designed to force only those failures
that would normally occur. ESS programs are based on a product profile that
determines the maximum allowable stress the product can absorb without
affecting its useful life. This determination is made prior to implementing the
program by analyzing product function, tolerances, and failure rates. Ideally,
there is never a loss of useful life. If there is, it will be negligible, particularly
in light of the failures precipitated in-house.
•• All ESS Systems Are Alike—ESS is as different as each product that is made.
It is, by design, a product-specific program that must be tailored according
to predetermined product variables. The type of screen, profile variations,
fixturing, power supplies, inputs, outputs, and other considerations must
be resolved before the program is implemented.
•• ESS Is Expensive—ESS saves money and can deliver a fast payback. An
ESS program is a cost that can and should be measured for its return-on-
investment potential. By implementing the program properly, dramatic
cost savings can be realized through reduced field warranty repair expense,
lower unit costs, improved product value perception in the marketplace, and,
ultimately, increased profitability. ESS is ideally implemented at a point in
production where the cost to repair is lowest.
Types of Environmental Stress

ESS consists of exposing a product to one or more stressful environments, such as
thermal cycling, vibration, high temperature, electrical, thermal shock, and others.
Profile characteristics, such as stress extremes, rates of change, amplitude, frequency,
and duration, must be tailored for each product. The goal in determining equipment
requirements and profile characteristics is to achieve maximum screening strength
within the realm of product limitations and overall affordability. Useful screening
strength information is available in the U.S. Department of Defense document
DOD-HDBK-344 (USAF). A description of some of the more commonly used
environments follows.
Temperature Cycling
Temperature cycling consists of multiple cycles of changing temperature between
predetermined extremes. Figure 7.3 shows a typical temperature profile.
Because all variables in a screen are product dependent, temperature extremes
must stop short of damaging the product but must be far enough apart to allow for
optimum stressful lengths of temperature change between the extremes. Cycle times
must follow this same rule of thumb because the constant rate of change provides
the expansion and contraction necessary to sufficiently stress the product.
The product temperature change rate is dependent on specific heat properties of
the product, the difference between the product and air temperatures, and surface
conduction factors involving air velocity and direction. Assuming proper airflow,
the temperature of a typical circuit board would resemble that of Figure 7.4 when
subjected to a typical temperature profile.
Because the dwell period at the temperature extremes does not significantly
contribute to the stress, many manufacturers allow the product to remain at tempera-
ture extremes only long enough to allow a functional test of the product, as shown
in Figure 7.5.
FIGURE 7.3 Typical ESS temperature profile.
FIGURE 7.4 Chamber and product temperature profile.
Another step some manufacturers have taken to maximize the stress during thermal
cycling is to adjust the chamber air temperature so that the high and the low tempera-
tures are close to the extremes the product can withstand. However, in researching this
stress profile, it must be determined that multiple cycles between these two extremes
do not erode the useful life of the product. Figure 7.6 illustrates such a profile.
FIGURE 7.5 Temperature profile with functional test at temperature extremes.
FIGURE 7.6 Temperature profile near product destruct limits.
The airflow in a thermal cycling chamber is critical because it directly affects

product temperature change rates. Figure 7.7 illustrates the product temperature
during four separate runs of a profile that had identical air temperature change rates.
The only difference between the runs is the air velocity. Clearly, the chart shows
that the lower the air velocity the slower the product temperature increases and
FIGURE 7.7 Temperature profile with air velocity.
decreases, and less stress is brought to bear on the product. The higher air velocities
cause a more rapid product temperature change; thus, these higher air velocities
subject the product to a higher degree of stress.
The optimum airflow depends on the product. There is an air velocity at which
maximum heat transfer is obtained. Exceeding that air velocity can be counterpro-
ductive. The air velocities listed in Figure 7.7 may or may not be the most appropriate
for the product. The correct air velocity rates and air direction for a particular product
can be determined through experimentation.
The general consensus is that thermal cycling is the most effective screen. Similar
to any screen, it must be properly implemented. Failure rates must be analyzed to
determine which latent defects are causing failures, and experimentation must
be performed to determine the screen profile best suited to trigger those particular
latent defects into failures.
Random Vibration
Considered the most effective of three principal types of vibration, random vibration
involves the excitation of a product with a predetermined profile over a wide frequency
range, usually from 20 Hz to 2,000 Hz. Product stress is created through simultaneous
excitation of all resonant frequencies within the profile range.
Random vibration is applied by mounting or attaching a product to a unit called
an electrodynamic shaker, controlled by a closed-loop digital system and dedicated
fixturing. Fixturing must be extremely rigid to ensure that stress is transmitted
directly to the product and that the process is repeatable with reasonable accuracy.
Products may be vibrated on a single axis or on multiple axes concurrently or
consecutively. Currently, opinions vary as to which provides the more

stressful environment.
Random vibration screens generally require less time to run than other ESS
programs and are considered particularly effective in exposing mechanical defects,
such as loose solder, improper bonds, and printed circuit board shorts.
The primary drawbacks are equipment cost and lack of screen uniformity.
Regarding the former, electrodynamic shakers may be expensive to install, control,
and maintain. Random vibration also is, by nature, less effective than temperature
cycling in providing a uniform stress environment. To avoid overstress, random
vibration screens are tailored to exert maximum stress on parts at joints located at
the center of a product and are gradually reduced for parts closer to the edge, a factor
that may result in failure to expose all potential defects.
Nonetheless, random vibration is considered superior to other forms of vibration
and is often used, where financial objectives allow, in conjunction with thermal
cycling screens.
High-Temperature Burn-In
This process is generally a static one in which a product is subjected to elevated
temperatures over a predetermined period. This screen evolved from the idea that
continuous operation of a product would force infancy failures. It also was believed
that providing additional heat would escalate the rate of failure. It is more likely,
however, that increases in screen effectiveness are related to changing temperatures
during heat up and final cooldown, rather than powering at a constant high temperature.
Electrical Stress
This process is used to exercise circuitry and simulate junction temperatures on
semiconductors. There are two basic types:
1. Power cycling, which consists of turning the product’s power on and off at
specified intervals.
2. Voltage margining, which involves varying the input power above and below
the nominal product power requirements
Electrical stress does not, according to research, expose the number of defects
commonly found through thermal or random vibration screens; consequently, it is
considered much less effective. It can, however, be relatively inexpensive to implement
electrical stress with another screen to increase overall screen effectiveness. It also
may be necessary to power products in order to find soft failures.
Thermal Shock
Thermal shock is a process that exposes products to severe temperature extremes,
usually in rapid fashion, that is, a product is continually transferred—either mechani-
cally or manually—from an extremely hot environment to an extremely cold environ-
ment and back. Thermal shock is generally considered a cost-effective way to screen
for defects at the component level, particularly integrated circuit (IC) devices that
require a high degree of stress to experience the rates of change needed to force latent
defects into failure.
Thermal shock also may be useful at other levels of assembly, as long as the
severity of its rates of change do not cause needless damage. This is a particular risk
with more complex assemblies, such as those containing components other than ICs.
Of equal consideration is the cost efficiency of a thermal shock screen. Generally, the
equipment needed to provide an effective screen is expensive. In addition, because
products must be shifted from one environment to another, there will always be an
unused chamber. If manual transfer is involved, the risk of accidental product damage
increases. Finally, a thermal shock screen is difficult or, in some cases, impractical to
power or monitor. This limits opportunities for accumulating field failure analysis data.
Sine Vibration, Fixed Frequency

This is a form of vibration that operates on a fixed-sine or single operational frequency.
This method usually requires a mechanical shaker with a frequency range of up to
60 Hz. Although less expensive and easier to control than a random vibration screen,
fixed frequency vibration is generally viewed as not providing as effective a level
of stress.
Low-Temperature Screen
Similar in concept to a high-temperature screen, a low-temperature screen is based
on the principle that failures will be forced by the contrast between the heat generated
by powering a product and the cold environment.
Sine Vibration, Swept Frequency

This is a form of vibration that operates on a swept-sine or multiple operational
frequencies. It usually requires a hydraulic shaker with a frequency range of up to
500 Hz and is viewed similarly to a fixed frequency in terms of overall effectiveness.
Combined Environment
Depending on product complexity, cost, and reliability specifications, environmental
screens may be used in concert with each other. For example, thermal cycling and
random vibration are often combined in an ESS program. The primary considerations
should be whether an additional stress—applied simultaneously or consecutively—
will expose a significant number of additional defects and whether the cost of the
additional stress is justifiable.
In general, the types of stress used and how the stress is applied are dependent
entirely on what will precipitate the greatest number of failures in a specific product
in the shortest period of time. As with all screens, the most effective profile is product
dependent. However, some recommendations are available in the IES Environmental
Stress Screening of Electronic Hardware (ESSEH) guidelines, the Rome Air Development
Center (RADC) studies, and the U.S. Navy Manufacturing Screening Program,
NAVMAT P-9492.
Advantages of Temperature Cycling

A number of independent studies have been conducted that analyze the effectiveness
of various types of ESS processes. Organizations such as the IES and the RADC have
FIGURE 7.8 ESS screen types.
compiled in-use data that conclude thermal cycling is the most effective type of screen,
as shown in Figure 7.8. A comparison in the weighted rank of effectiveness among
various environments is shown in this table published by the IES in its Environmental
Stress Screening of Electronic Hardware (ESSEH) guidelines.
Although it has been shown that the most effective screening program may
involve more than one screen, thermal cycling is considered the single most effective
screen in terms of identifying latent defects over a period of time. According to the
IES Environmental Stress Screening of Electronic Hardware (ESSEH) guidelines
published in 1981, 1984, and 1988, thermal cycling, when compared with random
vibration (ranked as second most effective), regularly detected an average of two-
thirds more latent product defects.
Thermal cycling provides the additional advantage of a more uniform stress
environment—one in which all areas of the product are subjected to a relatively more
equal amount of stress throughout the screen profile. Some other environmental
stresses cannot, by their nature, provide uniformity and have a higher inherent
risk factor.
Stress uniformity also makes thermal cycling easier to control, permitting greater
flexibility in implementation, revision, or refinement.
Finally, thermal cycling takes less time to perform than most other types of stress.
(It is less time consuming than any form in terms of its ability to force the greatest
number of defects into failure. Random vibration, which has a shorter cycle time,
does not provide the same overall efficiency.) The net result is a better product
throughout with minimal impact on production.
FIGURE 7.9 Effect of temperature on screen effectiveness.
The basic concept of thermal cycling involves three uncomplicated principles.

First, the temperature extremes must be as far apart as possible—regardless of
the intended operational limits of the product—as long as they do not cause
needless damage. Second, the rate of change between the extremes must be as
rapid as possible—usually greater than or equal to 5°C per minute—to create the
optimum level of stress, again without damaging the product. Figure 7.9 shows a
comparison of various temperature change rates and their impact on
screen effectiveness.
Finally, the temperature of the product must follow, as closely as possible, the
temperature inside the chamber. This is achieved by assuring that high-velocity air
is properly routed over the product. Figure 7.7 illustrates how rapidly the product
temperature changes at various air velocities. Clearly, the temperature of the product
changes faster as the air velocity across it is increased. Faster temperature change
creates greater stress on the product; consequently, it can more quickly force latent
defects into failure.
A good starting point to determine an effective thermal cycling stress profile
would be to start with the following:
•• Temperature extremes equal to or greater than a 100°C temperature difference
between hot and cold
•• Product temperature rate of change equal to or greater than 5°C per minute
•• Air velocity of approximately 750 fpm (feet per minute) at the product
•• Enough cycles to effectively cause latent defects to produce failures
Levels of Screen Complexity

Making ESS work to your best advantage requires consideration of whether it is either
necessary or justifiable to augment the primary stress environment. Certain products
may or may not benefit from electrical stress, such as powering or signal generation,
during the screen. Others may require it.
To make this determination, some basic considerations are necessary. The envi-
ronment in which the product must operate is important. A device that will be used
in a controlled environment, such as a computer, will most likely not require special
treatment. Shipboard instrumentation, on the other hand, may be exposed to a volatile
range of environments and most likely would require more than a basic approach.
Beyond the environment, different screening levels can provide information
relating to types of failures. It may be necessary to know whether a failure is hard
(i.e., a point at which the product will stop functioning) or soft (i.e., a situation in
which a product will cease functioning under certain conditions but will operate
under others). Hard failures are always defined as critical because the product is
unable to operate under the conditions in which it is expected to operate. Soft failures
can be critical if they occur within the intended operational specifications of a product.
Amplified levels also can provide failure analysis data that indicate when and where
a failure occurred. Finally, various levels of power cycling create additional stress that
can be beneficial in forcing latent defects to cause failure.
The overriding consideration is to provide as much stress as is cost justifiable to gain
the desired result. Once this need has been established, the appropriate level of screen
complexity can be factored into the program. The basic types are described as follows:
•• Static—This means subjecting an unpowered product to a single stressful
environment. The product is typically tested before and after the screen to
verify operation.
•• Dynamic—Power is applied to the product during the screen. Another type of
stress may be added to the product at the same time the power is being cycled
on and off. Also power levels can vary above and below nominal values. Power
is usually turned off during hot-to-cold transitions so that heat generated by
the product does not affect descending (i.e., pulldown) product and chamber
temperature rates, which reduces stress during that part of the screen.
•• Exercised—Power and the additional stresses of supplying inputs and
monitoring or loading outputs are applied to the product. In some cases,
signals or loads must be applied to a powered product to avoid damage.
•• Full Functional—The product is operated as though it were in actual use.
This can include power and all necessary signals and loads, which may or
may not be cycled on and off.
•• Monitored—This is a description of any powered product screen. It consists
of monitoring the product for failures and is used in indicating, for example,
products that fail and then resume operation (i.e., soft failures), the conditions at
the time of failure, and the time of failure. This type of information is critical in
cases where even soft failures cannot be tolerated. It also can be used to identify
and correct product or process deficiencies, or to fine-tune the screen profile.
Failure Analysis
Once an ESS program is implemented, its long-term success depends on monitoring
the results. This, in fact, can often be the single most critical factor in determining
whether ESS delivers the proper ROI. New suppliers, a change in process technology,
new assembly technology, or a sudden drop in in-house failure rates without a corre-
sponding drop in the field failure rates are all reasons for reviewing a program.
Data acquisition is essential to the process because it helps establish the learning
curve required to keep personnel integrated with program objectives by providing
the following:
•• Failure symptoms
•• Environment at the time of failure
•• Cycle during which the failure occurred
•• Time into the cycle during which the failure occurred
•• The cause of failure, based on the analysis of the failed part or assembly
Analysis of this data should be grouped into major failure categories and should
be plotted against the screening cycle and the time at which the defect was found. As
data accumulate, problems can be more readily identified and resolved, and screen
profiles can be altered accordingly.
Observing sound failure analysis practice goes beyond establishing the initial
ESS profile; in effect, it creates a closed-loop system that builds accountability into
the program and allows for change based on hard data, rather than supposition.
Case Histories
The IES, RADC, and other groups have researched ESS in detail, compiling studies
and case histories relating to methodology, combination screening, screen complexity,
cost analyses, and other factors.
In addition, a number of companies have independently written about their
success with ESS programs. A few examples are detailed in the following paragraphs.
Analogic Corporation
Analogic Corporation found an excessive number of early-life or infancy failures
in one of its products, even after steady burn-in screening. Detecting no inherent
design problems, company executives applied an ESS program of thermal cycling
to the product under power in an attempt to detect a greater number of
defective products.
The thermal cycling reduced the final-assembly-level reject rate to less than half
the rate achieved on the same line in the preceding year. This reject rate included a
significant increase in the number of defects identified in ICs and displays—defects
that did not appear in the previous steady-state burn-in screening. In addition,
thermal cycling shortened the necessary aging time and reduced production time.
As a result of these findings in the laboratory, Analogic installed six thermal cycling
systems on its production lines.
Bendix Corporation
Bendix Corporation, makers of electronic fuel injector systems from 1975 to 1980,
published some significant findings of its experience with ESS technology. In designing
production lines for its injector systems, Bendix wanted to ensure product reliability
because an injector system field failure could disable a car and result in expensive
field repairs. Because the product was extremely complex (involving 420 components
and 1,700 soldered joints), the conventional screening methodology was considered
inadequate. Bendix also wanted screening to be more effective in identifying defects
that caused intermittent, or soft, failures.
Bendix replaced the traditional burn-in screen on the production line with thermal
cycling. The ESS system proved more effective in spotting the solder, workmanship,
and semiconductor defects that contributed to soft failures. As a result of using thermal
cycling, field failures caused by these defects were reduced from 23.5% to 8%.
Hewlett-Packard (HP)
HP introduced its 9826A desktop computer line in 1983. To be competitive, the
company needed to spend less on production and still produce an extremely
reliable product.
The company installed an ESS program consisting of a thermal cycling screen.
The screen enabled the company to cut back production-line screening time from
the two to five days needed for a typical burn-in screen to a few hours. The thermal
cycling screen prevented a production-line bottleneck, increased productivity, and
reduced the unit production cost.
Because the ESS screen was applied very early in the production phase on rela-
tively simple subassemblies, the company also spent less time on troubleshooting and
repairs. There was also a substantial reduction in the field warranty rate. After six
months, thermal cycling was exposing two to three times more latent defects than
similar production lines. According to a six-month study, ESS directly contributed
to a 50% drop in the number of field repairs made under warranty. Research shows
that total savings to the company will amount to more than $1.5 million over five
years, after the cost of installing the ESS program.
In addition, ESS has become a powerful tool in the automotive industry, where
“the Big Three” automakers have embraced the concept as readily as SPC. In aerospace,
companies such as Motorola, Inc.; Hughes, Inc.; and Boeing have successfully imple-
mented programs. Many consumer electronics companies also have used ESS to
reduce unit costs and to gain a competitive edge.
Implementing an ESS Program

To be successful, ESS must be treated as a highly technical discipline. It must involve
careful research, a sound set of objectives, and a system for monitoring ongoing
effectiveness. The basic steps are described as follows:
1. Establish an Objective—The reasons for undertaking any form of ESS must
be clear for the program to work. Some principal considerations are as follows:
•• Achieve and maintain high field reliability by finding latent defects in-house.
•• Reduce production costs by exposing and correcting defects early in the
manufacturing process.
•• Meet a contractual reliability demonstration requirement.

•• Assure cost-effectiveness in a reliability improvement warranty (RIW)
contract.
•• Reduce field costs of operations and maintenance.
•• Upgrade in-house reliability methodology.
2. Involve Key Personnel—Essential to program planning is that technical

personnel from all applicable departments must be involved. This in-house
team may consist of, but not be limited to, representatives from engineering,
production, testing, and quality. For example, quality can examine field
failure histories and report the type of latent defects that cause field failures.
These are the defects that must also precipitate failures in-house for ESS to
be effective. Engineering can provide information on product tolerances,
handling limitations, and other factors. Production can point out areas of
process complexity. Testing and field services can assist in selecting the most
appropriate screen profiles to expose latent defects, as well as develop field
failure expectancy rates for new products.
3. Involve a Supplier—Selecting the right supplier can make an important
difference in both the cost and effectiveness of an ESS program. This may
include being able to provide assistance in clarifying program goals, as well
as the equipment information necessary to develop cost analyses. The latter
point is of particular importance because the need for, and the cost of, future
system integration will be determined by how complete a program the
supplier can initially deliver. A credible supplier should also be able to
provide appropriate agency approvals for equipment and be able to support
equipment, both in terms of warranty and service.
4. Analyze the Business Impact—Based on information from all applicable
departments, management can then determine whether an ESS program is
justifiable. This can and should be evaluated on an ROI basis, involving
considerations such as the following:
•• Will the reduced field maintenance cost justify the program cost?
•• Is ESS necessary to eliminate excessive latent defects?
•• Is ESS necessary to meet a reliability requirement?
•• Will ESS save money in production?
•• Is an improved production schedule worth the cost of the program?
•• Will the goodwill and increased sales potential from delivering latent-defect-
free products balance the cost of the program?
5. Determine the Best Level for the Program—Implicit in the cost advantages
of ESS is the capability to force failures at the most cost-effective point in the
production process prior to shipment. Figure 7.10 illustrates the basic
production levels, including cost-to-repair values taken from a typical
industry example and a level of complexity ratio that is fairly common
throughout the industry.
FIGURE 7.10 ESS failure versus production level.
Although logic might suggest that screening at the relatively inexpensive

component level is best, also note that screening at such a low level of complexity
may not produce meaningful results. Screened components, for example, can
still be damaged during assembly. On the other hand, looking at the board level,
there is a level of complexity very close to the final assembly, but at a significantly
lower cost to repair. Depending on the levels of complexity involved at each
stage of production, it may be justifiable to screen at more than one level.
In the case of the company cited in this example, the decision to screen
at the board level was made. The early identification of defects greatly reduced
field repair costs, a savings that quickly paid for the ESS program and later
translated into increased profitability.
In each case, it is important to realize that a certain number of trade-offs
are implicit at each level of screening and that careful analysis of these vari-
ables will help determine where a program will be most effective.
•• Tailor ESS to the Product Requirements—ESS is a product-specific program,
requiring a screen or set of screens that will provide the highest level of stress
without diminishing product life. To achieve this, each product must be treated
differently in establishing a screening profile. Its operating requirements, in-
use environment, and failure history must be used to develop the proper
program.
•• Establish a Method for Monitoring the Program—Follow-up in ESS is not
only good practice—it is essential. Because each program is based on field
failure analyses, it is only logical that this would continue throughout the
program. This assures the program is performing as specified. In addition,
monitoring provides a basis for assessing change—from new parts and

processes to a falloff in screen effectiveness—and responding with the
appropriate action.
Equipment
Because an ESS program must, by nature, be tailored to meet a specific set of objec-
tives, it is important that equipment be considered in relation to the primary intent
of the ESS: to provide optimum stress in the shortest amount of time without
damaging the useful life of the product.
Keeping the product always as the focal point, it is possible to take a logical
approach to equipment specification, as illustrated in Figure 7.11.
1. Profile—This will determine which types of equipment and accessories are
required, based on an examination of screen types, screen complexity, and
research and analysis of field failures.
2. Chamber—The major considerations associated with selecting a chamber are
as follows:
•• The type and size of the product to be screened
•• The quantities to be screened at any one time
•• The number of interconnects required to and from the product, if it is to
be powered
•• The temperature range and change rates of the screens to be performed
•• The data transfer rates of the product, if applicable
•• Airflow, velocity, and direction to maximize product heat transfer
•• Product handling that can affect throughput and product reliability
Chamber construction is usually contoured to the types of stress and

amount of heat that will be generated. This may or may not include complex
fixturing, signal lines, and different types of carriers to hold or load products.
Of consideration is that a minimum of product handling should be involved
FIGURE 7.11 ESS program equipment defined by the product.

because this can increase the risk of product damage outside screening and
minimize the effectiveness of the program.
3. Airflow—To ensure proper screen effectiveness, equipment should
be designed to provide adequate air velocity and uniform air distribution. As
a function of both chamber design and equipment placement, the objective is
to assure that the same temperature change rates are occurring on all
products so that the full impact of the thermal cycle is realized.
4. Fixturing—Fixtures are necessary to hold products in the proper orientation
to benefit from the screening process. Fixtures may also be designed to allow
for easier handling and for compatibility with power supplies and interface
electronics in situations where different levels of screen complexity
are involved.
5. Power Supplies—Power supplies are necessary to apply power cycling.
6. Product Electronics—Electronics are used to create an interface with a
product when it must be exercised or monitored and may include input and
output (I/O) lines, sensor devices, and instrumentation connections.
7. Interconnects—These bring signal and power lines outside the chamber and
must be compatible with available hookups and power sources.
8. Instrumentation—Various types of instrumentation are available for system
and product safety and operation. Consideration must be based on need, as
well as the potential for upgrading the data acquisition capability of a system.
9. Computer Control—Computer control can be used to control total program
operation and monitoring. Off-the-shelf software programs are available, but
it is still a buyer-beware situation. Many programs are too general for specific
applications or are too complicated for personnel to use. A customized, user-
friendly package should be sought.
10. Material Handling—This is an occasionally overlooked but decidedly
important aspect of a program, particularly with large products involving a
high level of screen complexity where cost-effective devices that
automatically load or unload product or that facilitate manual handling of
product should be used. This reduces the possibility of causing needless
product damage outside the screening process.
Burn-In Optimization
The use of environmental stress tests for burn-in can be expensive. In some cases,
more factory floor space is dedicated to burn-in than for assembly. Burn-in costs can
be minimized by optimizing the burn-in duration. There are three components to
burn-in costs:
1. Laboratory costs
2. Laboratory failure costs
3. Field failure costs
FIGURE 7.12 Failure rate characteristics.
Burn-in is effective only if the failure rate is decreasing. Referring to Figure 7.12,
if a product is burned-in for the period from time = 0 to time = x, the product will
be impacted in one of three possible manners:
1. If the failure rate is increasing, the failure rate at time = x is greater than the
failure rate at time = 0. Burn-in has reduced the reliability of the product.
2. If the failure rate is constant, the failure rate at time = x is equal to the failure
rate at time = 0. Burn-in has had no effect on the reliability of the product.
3. If the failure rate is decreasing, the failure rate at time = x is less than the
failure rate at time = 0. Burn-in has improved the reliability of the product.
The total cost to burn in a component or assembly is
 R(x + w)
C = Ax + L 1 − R ( x )  + F 1 −  (7.1)
 R(x) 
where
A is the cost of burn-in per unit time
L is the cost of a failure during burn-in
F is the cost of a failure in the field
w is the warranty period (beyond this time, there is no field failure cost)
Using the Weibull reliability function, the total cost to burn in a component or
assembly is
 
β
 x +w 
−
 
β 
x
−   e  θ  
C = Ax + L 1 − e  θ   + F 1 − β  (7.2)
  
x
− 
θ 
 e 
Example 7.1: An electronic component has a Weibull time-to-fail distribution with

a shape parameter of 0.6 and a scale parameter of 140,000. The warranty period
for this component is 8,760 operating hours. A field failure costs $900, and a failure
during burn-in costs $25 to repair. Burn-in costs $0.037 per hour per unit. What is
the optimum burn-in period?
Solution: The optimum burn-in duration is found by minimizing the following
equation in terms of x:
 
0.6
 8,760 
−
 
0.6 
 x 
−   e  140,000  
C= 0.037 x + 25 1 − e  140,000   + 900 1 − 0.6 
  

−
x 

  
 140,000 
 e 
From Figure 7.13, it can be seen that the optimum burn-in duration is 163 hours.
FIGURE 7.13 Graphical solution for burn-in duration.
Summary
ESS is an important consideration for any manufacturer with an unacceptable field
failure rate caused by production process problems, random defective components,
or other problems not related to product design.
ESS is a comprehensive program for improving product field reliability because
it involves 100% of the product and is structured to force failures of latent defects
in-house, instead of in the field. It will achieve this end at a much more effective and
faster rate than conventional aging methods.
The benefits of ESS include increased profitability by lowering the number of

warranty repairs, reduced costs, and improved product and corporate market
value perception.
Before implementing an ESS program, a careful analysis involving key technical
personnel is necessary to ensure that the program is both performance effective and
cost effective. Continued in-house failure analysis is vital to maintaining
program effectiveness.
An ESS program should be tailored to meet specific product requirements and
to subject the product to stresses that will force the maximum number of failures
without altering the life expectancy of the product. Equipment specification and
implementation should follow this logic and be installed by a supplier able to explain,
justify, and support the equipment.
Finally, as with any investment, ESS must pay for itself. A properly considered
and applied program will generate a measurable ROI—through improved product
quality, reduced failure rate, and increased profitability.
8
Test Equipment Methods
and Applications
T
he basis of all environmental testing is to evaluate a product based on its
mechanical failures. Either the product survives a defined test without failures
or failures precipitated by the test are evaluated for product improvement. The
field of environmental testing is very broad, covering many different disciplines.
Vibration and thermal cycling are the two most commonly used test disciplines;
however, the operating requirements of a product frequently necessitate the use of
other environments such as humidity, salt fog, sand and dust, and electromagnetic
interference (EMI). Often the best test method is to simultaneously combine the
applicable environments.
The testing process is begun by determining the environments to which the
product will be subjected and then evaluating which of these environments are critical
to the survivability of the product in its real-world usage. Next, data are acquired
from these environments so that the environments can be simulated in the laboratory.
This is then followed by the development of the laboratory test and its application to
the product. Every detail of each and every step of this process is critical for success.
Any incorrect assumptions about the consistency of the real world, any improperly
acquired or reduced data, or a poorly applied test will culminate in less- than-reliable
test results.
Obviously, the best test for any product is to subject it to the real-world environ-
ment for its lifetime. This is usually impractical and excessively expensive. For
example, an electronic box that is mounted in an automobile would best be tested by
driving the automobile for many thousands of miles over different road conditions
in cold and hot climates and all weather conditions. The resultant test would take

months and destroy an entire automobile in the process. The goal of laboratory testing
is to achieve results that are identical to those of the real world but at a lower cost and
often with a reduced time.
The first step is to define the real world for the product. Where, when, and how
will the product be used? If it is an electronic entertainment center designed for use
in the home, its operating environments will be considerably less severe than those
of the electronic box in an automobile. The entire life cycle of the product must
be considered—from the time its construction is completed until its desired opera-
tional lifetime ends. This includes shelf storage and shipping before final usage. It is
possible that shocks and thermal cycles that can be induced by handling and long-
term storage of the product in its shipping container are critical to the subsequent
reliability of the product.
Serious consideration must be paid to determining what is the real world for
the product. Temperature extremes, numbers of thermal cycles, humidity levels,
and salt fog conditions can be evaluated from the meteorological conditions in the
geographic regions where the product will be located. Electrical environments are
based on operating conditions. Shock loads can be developed based on handling
and operation. However, vibration is a much more difficult environment to deter-
mine, evaluate, and test. The size of the product, its orientation to the input levels,
the applied frequency range, the critical frequencies of the product, and the available
testing equipment all must be considered in the development of the laboratory
vibration test.
With a product located in an automobile, there is a great range of environments.
The vehicle must be able to operate in both Arctic and tropical conditions, at altitudes
from sea level to those in high mountain ranges, in humidity levels from the desert
to the salt fog at the seashore, and over an infinite variety of road surface and speed
conditions. All of these must be considered, evaluated, and included in the laboratory
testing environments.
Three classifications of testing are commonly used:
1. Defect exposure
2. Acceptance
3. Qualification
The purpose of defect exposure tests is to evaluate the product during its assembly
to ensure that it has been manufactured properly. The most commonly used type of
defect exposure test is ESS (see Chapter 7 “Environmental Stress Screening”), which
uses vibration and/or thermal cycling, with the test environments being based on
standard procedures rather than on the intended environments of the product or on
correlation with the real world. Acceptance tests are used to evaluate production
hardware and ensure that it meets the desired requirements and is ready for instal-
lation and use in the final product. Qualification tests are used for evaluating devel-
opment hardware and initial production hardware, and usually simulate the entire
lifetime of the product.
All of these tests are critical to developing and producing cost-effective, reliable
products. Considerable effort must be placed on developing and performing these
tests so that they are accurate and effective.
Simulation versus Stimulation

Simulation is an attempt to reproduce usage conditions, whereas stimulation testing
is conducted in the absence of real-world data or to rapidly precipitate failures.
Simulation
Simulation is an attempt to match the real-world environment in the laboratory. The
real-world environment must be defined and then data acquired while the product
is being subjected to it. By nature, the laboratory environment approximates the real
world as closely as possible, within the limitations of data acquisition, data reduction,
laboratory equipment capabilities, and practical limitations (e.g., schedule).
Application
Simulation testing is applicable to every product that must withstand cycling envi-
ronments during its usage. This includes thermal, humidity, and electrical cycling,
as well as shock and vibration and other environmental conditions.
Typical Test Types

Because of the broad spectrum of physical environments, there are many typical
types of simulation tests. For large items such as full vehicles (e.g., cars and trucks),
a laboratory vibration test would consist of multiple axes inputs at the wheel loca-
tions, using real-time methods. The electronics in the vehicle can better be tested
using single-axis electrodynamic vibration where their critical frequencies can
be applied.
The optimum laboratory test matches the real world as closely as possible. Because
the real world includes the simultaneous application of multiple environments, these
should be included in the laboratory test whenever possible. To this end, environ-
mental chambers can be used over shakers to simultaneously apply thermal cycling
and humidity along with vibration and shock.
Defining the Real World

The first critical step in developing a laboratory simulation environment is to define
the real world that is being simulated. By necessity, the following can give only general
guidelines for defining the real world because the environments experienced by
hardware vary greatly. These guidelines give an indication of what must be considered.
Defining the real world is a difficult process and can require considerable research.
It includes everything that happens to the hardware after it has been manufactured.
Table 8.1 can be used as a guide in considering the requirements for defining the real
world. This table lists the possibility of a specific environment occurring during a
phase of the life of the product. The three phases are as follows:
1. Storage
2. Shipping and handling
3. Customer usage
TABLE 8.1 Real-world environments.
Product life phase

Shipping and
Environment Storage handling Customer use
Thermal:
High temperature Occasionally Usually Usually
Low temperature Occasionally Usually Usually
Cycling Rarely Usually Usually
Shock Rarely Occasionally Occasionally
Mechanical:
Shock Rarely Usually Usually
Vibration Rarely Usually Usually
Electrical:
Voltage Rarely Rarely Usually
Frequency Rarely Rarely Usually
EMI Rarely Rarely Usually
Other:
Humidity Occasionally Occasionally Usually
Dew cycle Rarely Rarely Occasionally
Altitude Rarely Rarely Occasionally
Submersion Rarely Rarely Occasionally
Dust Rarely Rarely Occasionally
Wind/rain Rarely Occasionally Occasionally
Ultraviolet Rarely Rarely Occasionally
Salt spray Rarely Rarely Occasionally
Salt fog Rarely Rarely Usually
Ozone Rarely Rarely Occasionally
The following are the three categories of applicability:

1. Rarely
2. Occasionally
3. Usually
These are relative terms to assist with determining a real-world definition. They
refer to how commonly the environments are experienced by a general spectrum of
products. “Usually” indicates that a specific environment is included as a phase of
environmental testing for most products. “Occasionally” indicates that a specific
environment is included as a phase of environmental testing for approximately 25%
products, and “Rarely” indicates that a specific environment is included as a phase
of environmental testing for approximately 1% products.
The actual environment that a product must survive varies greatly. For example,
although vibration is rarely a concern for the storage phase, it is possible that an
electronic box could be “stored” as a replacement unit on an ocean-going vessel, where
it would be subjected to the vibration of heavy machinery near the storage area. Is the
product to be stored in a shed without environmental control in the Arctic or the

tropics? All aspects of the life of the product must be considered in detail so that the
definition of the real world that will be used for developing the laboratory test envi-
ronments is accurate.
Products are usually stored in environmentally controlled buildings and within
boxes or containers. Thus they will rarely see thermal cycling or shock. However,
when they are shipped (in their containers), they can be moved quickly from a ther-
mally controlled environment to an extreme thermal environment, inducing thermal
shock and cycling.
After the real world is defined, the next step in the process is to instrument the
product when it is being subjected to the real world. For an environment such as
thermal soak during storage, this probably is unnecessary. For an environment such
as vibration under customer usage, a considerable amount of data must be acquired
and reduced for developing the laboratory environments. The instrumentation, data
acquisition, and data reduction processes require extensive efforts. Furthermore, for
a given industry, the data must be reacquired on a regular basis to continually update
the subsequent laboratory testing to ensure that it is current and representative.
Some environments, such as thermal shock, occur only occasionally in real-world
situations but are commonly used for defect exposure testing. Specialized tests such
as altitude and ozone depend on the anticipated application of the product by the
customer. The severity of the test levels should be considered in developing a test
program based on the definition of the real world. If the customer usage vibration
levels greatly exceed those of the transportation environment, then it may be possible
to exclude the transportation vibration environments from the test program.
The sequence of applied environments to be used in the testing program must
be considered, based on the real world. The tests should follow the real-world sequence
as closely as possible. If the facilities are available, then combined environmental
testing should be performed. For example, vibration can often be combined with
thermal cycling.
Stimulation
Stimulation testing is conducted in the absence of real-world data or to rapidly precipi-
tate failures. The environment is applied based on predefined limitations for qualifi-
cation testing. To precipitate failures, the applied levels are increased. Because it is
conducted with arbitrary test levels, stimulation testing must be used with great care.
Time and money are saved because the efforts of real-world definition, data acquisi-
tion, and data reduction are eliminated. However, the test results can be unreliable
because they lack a solid basis in reality.
Application
Stimulation testing is performed to develop a product or to precipitate failures.
A prototype product can be subjected to an environment either to survive a prede-
termined level or to determine its weaknesses by precipitating failures, which are
then corrected. Also field failures can possibly be duplicated rapidly in the absence
of real-world data by applying laboratory environments.
Typical Test Types

The most commonly applied stimulation vibration test is a sinusoidal input. This can
be a sweep of a frequency range or a dwell at a specific frequency. One procedure is
to sweep a frequency range to determine the fundamental resonant frequency of the
product and then dwell at this frequency for a specified time at a defined level
(g input). Thermal cycling and thermal shock are also stimulation tests that are
commonly used. Other typical tests include a multiple shock-induced vibration
combined with thermal cycling.
Types of Stresses
This section discusses the various types of stresses used in accelerated testing.
Temperature
Stresses—and failures—can be induced in hardware by thermal changes. These are
usually due to differential thermal expansion. Assemblies are made at room tempera-
ture and then—in the real world—are subjected to environments from the Arctic to
the tropics. These assemblies are made of many different materials, each with a
different coefficient of thermal expansion. When the temperature is changed, the
different parts expand or contract different amounts. For electrical assemblies, breaks
can occur in the circuitry. For structural assemblies, stresses are induced as two parts,
which are attached to each other, to change the two different lengths.
Another cause of thermal-induced failures is the transition temperature of metals.
Some metals and alloys exhibit the property that they become brittle and lose ductility
when their temperature decreases to a certain value. Mechanical stresses applied to
the material when it is below the transition temperature can induce catastrophic
failures. A historic example of this is the Liberty ships that were produced during
World War II. Many of these ships were made of steel, which had a marked transition
temperature that was not known. The high seas, combined with cold winter tempera-
tures, in the North Atlantic caused several of these ships to break in two and disappear
without a trace.
Steady State
Steady-state thermal testing is used to evaluate the operating characteristics of
hardware in environmental extremes.
•• Low Temperature—Low-temperature steady-state testing is used for evaluating
hardware in Arctic conditions. The cycling of mechanical components
(e.g., latches) at low temperatures evaluates their durability characteristics.
Cyclic load testing at steady-state low temperature exposes the brittleness of
materials below their transition temperatures. Mechanical operation at low
temperatures is used for evaluating the effectiveness of lubricants, bushings,
and bearings. Differential thermal expansion can cause the seizing of moving
parts. Elastomeric components can become brittle and crack. Bonding agents
may become brittle and lose their adhesive capabilities.
•• High Temperature—High-temperature steady-state testing is used for

evaluating hardware in extreme tropical conditions. Mechanical operation
at high temperatures is used for evaluating the effectiveness of lubricants,
bushings, and bearings. Elastomeric components can lose pliability and
become excessively soft, failing to perform their intended functions. Bonding
agents may lose their adhesive capabilities.
Temperature Cycling
Temperature cycling is used to evaluate the hardware during exposure to thermal
extremes. Due to differential expansion over several thermal cycles, it is possible to
induce cracking. In electronic assemblies, this can be evidenced by open circuitry.
Thermal testing is performed within a chamber that has capabilities for extreme
temperatures as well as rapid changes in temperature. Depending on the test being
performed, there are two different approaches to fixturing. The test article can be set
on an open framework so that all sides are exposed to the chamber ambient condi-
tions. This maximizes the effects of thermal cycling and induces maximum rate
thermal changes within the test article. If the test article is attached to a large heat
sink and the effects of the heat sink are critical to the operation of the unit at extreme
temperature, then the heat sink should be simulated in the test. This is usually accom-
plished by mounting the test article on a large metallic block with the appropriate
thermal mass.
The rate of thermal change in the chamber is often critical to inducing stresses
that are potentially failure inducing. Various methods can be incorporated into
chambers to create rapid thermal changes. Standard heat-exchanger coolers can
be augmented by the use of liquid nitrogen. Humidity also can be a problem in a
thermal chamber. Condensation will form and freeze if ambient air is used in the
chamber. The condensation problem can be handled by using a nitrogen atmosphere
in the chamber.
During thermal cycle testing, it is possible to decrease the required soak time by
overshooting the ambient temperature in the chamber. By exceeding the required
soak temperature temporarily, the point of maximum thermal mass on the test article
can reach the required soak temperature more quickly. When the point of maximum
thermal mass reaches the required temperature, then the chamber ambient is brought
back to this temperature for the soak. Care must be taken to ensure that the remainder
of the test article can withstand the excessive temperature extreme without
inducing failures.
Thermal Shock
Thermal shock is a test where the ambient temperature surrounding a test article is
rapidly changed from one extreme to another. This is usually accomplished by ther-
mally soaking the test article in a chamber at one temperature extreme and then
quickly moving it to a second chamber at the other temperature extreme.
Dual thermal chambers are available, which can be used for thermal shock
testing. The test article is placed in one chamber at one temperature extreme. After
stabilizing at temperature, the test article is then automatically moved through a
door in the wall between the two chambers and exposed to the opposite
temperature extreme. The two chambers may contain various types of environ-
ments, such as the following:
•• Air to air
•• Air to liquid
•• Liquid to liquid
Vibration
Vibration is possibly the most complex environment that is used for testing. In the
early days of mass production and testing, only two options were available for testing:
(1) expose the test article to the entire real-world environment or (2) simulate it in
the laboratory with a sinusoidal input. Although frequency and amplitude could
be varied for the input, the real world is rarely sinusoidal, and these tests were at best
a poor simulation. It required the advent of the computer so that vibration data could
be reduced, understood, and reproduced in the laboratory.
Laboratory vibration testing falls into three primary categories:
1. Sinusoidal sweeps and dwells
2. Real-time simulation
3. Random vibration
Sinusoidal sweeps and dwells are used for evaluating resonant frequencies, mode
shapes, and damping. They also can be used for quick hardware evaluation in the
absence of any real-world data. Real-time simulation measures the output from the
real world and uses it as an input for the laboratory test. Both sinusoidal vibration
and real-time simulation are analyzed and reproduced in the “time domain.” This
means that the data are interpreted as a graph of amplitude versus time. Random
vibration is analyzed and reproduced in the “frequency domain,” meaning that the
data are interpreted as a graph of amplitude versus frequency. This is an abstract
concept that must be understood in order to effectively apply this means of testing.
Random vibration can accurately simulate the complex vibration input experienced
by hardware that is simultaneously exposed to several environments.
Other testing methods include mechanical vibration inputs, vibration induced
by repetitive shocks, and shock input. Mechanical vibration is generally used for
simulating unsecured cargo being transported in a rough environment. Vibration
induced by repetitive shocks from multiple sources can be used to develop a random
vibration input. Shock inputs simulate short-time-duration, high-frequency, high-
level inputs to the product. All of these methods have advantages and disadvantages
for use in laboratory testing.
The sine wave is the basis of all data analysis and reproduction. Combinations
of sine waves of different amplitudes, durations, and frequencies are used to reproduce
more complex environments.
Analysis Tools
There are two primary categories of analytical tools for working with vibration data:
(1) tools for the time domain and (2) tools for the frequency domain. The time-domain
tools are used for sinusoidal and real-time simulation testing. The primary time-
domain tool is the time history plot, which shows the variation of amplitude with
respect to time. As the name implies, it displays the acceleration, displacement, and
force history of a specific measurement. The frequency-domain tools are used for
random vibration testing. The primary tool in the frequency domain is the power
spectral density (PSD) plot. Fourier transforms are used to define the energy level at
each frequency of the measurement for developing the PSD. Correspondingly, a PSD
plot is a concise method of defining a time history.
Mechanical and Pneumatic

Mechanical vibration falls into two distinct categories: (1) rotary inputs for loose
cargo testing and (2) cyclic inputs for testing mechanisms. Loose cargo testing is
performed by placing a product in its shipping configuration onto a bed, which is
then excited with a mechanical rotary input. Mechanisms such as latches, pedals,
and door window cranks can be tested with pneumatic actuators to evaluate their
capability to withstand multiple cyclic inputs.
Pneumatics operate using linear actuators. These are powered by “shop air” at
approximately 100 psi. Some rotary actuators may be available. They can be controlled
by computers or simple timed relays, which open and close the valves that duct the
compressed air to the actuators. A primary advantage of pneumatic systems is that
they are inexpensive and easy to use. Most laboratories and factories already have a
supply of compressed air for running various tools. Actuators can be obtained that
apply loads up to and greater than 1,000 lb. The common range employed is usually
less than 100 lb.
Pneumatics are best for cyclic operational loads such as opening and closing
doors and repeatedly operating mechanisms. They have an upper-frequency limit of
approximately 1 Hz. Pneumatic actuators are ineffective in applying vibration loads
because their frequency range is very limited.
Loose Cargo Testing

The purpose of loose cargo testing is to determine if the product in its shipping
container can survive the transportation environments applied to it. The frequency
and amplitude of the input sinusoidal vibration can be varied. The product in its
shipping container is placed onto the bed of a machine. The bed has a wooden floor
and wooden walls. The bed is then activated by a rotary input for a specified length
of time, allowing the product to move freely.
Cyclic Testing
The primary tool used for cyclic testing is pneumatic actuation. Pneumatic actuators
are available in a wide range of displacement and load capabilities, in both linear and
rotary configurations. The fixturing is specific to each setup. However, note that the
fixturing must always be more durable than the hardware being tested. The test will
usually require several sets of hardware to be evaluated, and it is cost effective to use
the same fixturing for all of them.
Pneumatic systems are excellent for performing cyclic tests. Actuators can
be obtained with many different stroke lengths and load ratings. However, pneumatic
actuators usually are limited to frequencies below 1 Hz and correspondingly are
ineffective for applying vibration environments.
Electrodynamic
An electrodynamic shaker is essentially a large loudspeaker coil. It is driven electri-
cally by an amplifier. A computer is used to control it by applying sinusoidal or random
vibration using an acceleration feedback. (A load cell for force feedback also can
be used.) Electrodynamic shakers can be used for applying very-high-frequency
inputs. The inputs usually are defined as random vibration PSD plots or sinusoidal
waves that can be repeated easily on any setup without the requirement of drive
file development.
Small components, engine-mounted components, brackets, fluid operating
systems, and especially electronics can be tested using electrodynamic excitation,
usually in one axis at a time. Electrodynamics is excellent for running modal surveys
of structures, as well as determining the audible resonances for resolving
noise problems.
The lower frequency range of electrodynamic shakers is limited by displacement.
For old shakers, this is 1.0-in. double amplitude; for new shakers, it is 2.0-in. double
amplitude. This usually translates to a lower frequency limit of approximately 5 Hz
to 10 Hz. The upper-frequency limit of electrodynamic shakers is usually above
2000 Hz. Shakers are readily available in a load range from 50 force pounds up to
and exceeding 35,000 force pounds.
Items tested using electrodynamics are mounted to a plate on or attached to
the shaker head. A “slip table” can be attached to the shaker for obtaining the
axes of vibration other than those obtained on the shaker head. The fixture that
attaches the test article to the shaker head or slip table must be rigid and free of
all resonances in the frequency range of the test. The fixture is usually constructed
of an aluminum plate with a minimum thickness of 25 mm (1.0 in.). If more than
a f lat plate is required, then multiple plates are welded together using full
penetration welds.
Electrodynamic shakers are excellent for testing electronics and are used for ESS
of electronics assemblies. They can be used for applying real-world simulations such
as servo-hydraulic systems; however, because of their bulkiness, they usually are
limited to one axis of excitation at a time. Three-axis slip tables are available, but
because of cross-talk problems, the three shakers driving them must be run open
loop, and the bearings present control problems at high frequencies.
Hydraulic
Hydraulics operate using linear actuators. These are powered by hydraulic pumps
and reservoirs at high pressure. Some rotary actuators are available. They are
controlled by computers to apply sinusoidal or real-world simulation inputs using
force, displacement, or acceleration feedbacks as controls, using servo-valves to control
the flow of pressurized hydraulic fluid into the actuators.
Hydraulic systems can be used to apply high loads. Multiple actuators can be used
to simultaneously apply loads in several directions. The primary application of
hydraulics in dynamic testing is to apply real-world simulation of motor vehicles in

multiple directions simultaneously. They are used extensively to simulate test tracks
in the laboratory.
The lower frequency of hydraulic systems is limited by the available displacement
on the actuators that are being used. Practically, this usually is approximately 2 Hz.
The upper limit is approximately 40 Hz. Theoretically, higher frequencies are avail-
able, but they lack adequate controllability. Actuators are readily available that can
apply loads up to approximately 10,000 lb and higher.
Fixturing for servo-hydraulic testing can be of many different forms. For a simple
(i.e., one actuator) test, a linkage is made between the actuator and the test article.
For a “four-poster” test, four hydraulic actuators are mounted vertically with flat
plate “wheel pans” placed on top of them. This is used to apply the vertical inputs that
a vehicle experiences when on a rough surface. More complex setups can apply
vertical, lateral, and torsional (“braking”) loads at each wheel hub simultaneously,
along with a torsional input due to the engine driveshaft.
Servo-hydraulics is good for simulating the loads on major structures. However,
the higher frequencies that are present in the real world cannot be applied. Because
of the 40-Hz limitation, any structures having resonances in excess of this frequency
cannot be adequately tested. This includes all small components, brackets, and elec-
tronics. The “drive files” that are used for controlling multi-axis tests in real-world
simulations are very setup specific. If the setup is changed in any way, a new drive
file must be developed from the original data.
Mechanical Shock
Mechanical shock is defined as either a time history pulse or a shock response spectrum.
The most commonly used time history pulses are defined as half-sine waves or (trian-
gular) sawteeth. Shock response spectra are defined as a plot of response accelerations
(g) versus frequency (Hertz). These accelerations represent the response that a single-
degree-of-freedom system would experience at every frequency in the range when
subjected to the required input. Shock response spectra are useful tools for analysis
because if the resonant frequencies of a product are known, then the loads to which the
product will be subjected are defined. A good reference for shock testing is MIL-STD-
810, Section 516.
Mechanical shock can be induced by either a shock machine or an electrodynamic
shaker. A shock machine has a platen upon which the test article is mounted. The
platen is then raised to a predetermined height and dropped on a crushable object.
The configuration of the crushable object determines the resulting shock pulse. The
shock machine is calibrated for the actual test by first conducting drops using dummy
hardware and varying the object onto which the platen is dropped. When the desired
shock pulse or shock response spectrum is obtained, then the actual hardware is
mounted on the platen for the test. The fast response capability of an electrodynamic
shaker can also be used for inducing a mechanical shock by programming the time
history of the shock pulse into the shaker. Fixturing for shock tests usually is minimal.
The test article must be solidly mounted or clamped to the shock test machine in the
desired orientation.
Electrical Stress*
Electrical stress is used to test a product near or at its electrical limits. Exercising
circuitry and simulating junction temperatures on semiconductors are both good
examples of electrical stress tests. There are two basic types of electrical stress tests:
(1) power cycling and (2) voltage margining.
Power cycling consists of turning the product’s power on and off at specified
levels. Voltage margining involves varying input power above and below nominal
product power requirements. A subset of voltage margining is frequency margining.
Typically, electrical stress alone does not expose the number of defects commonly
found through thermal cycling and vibration stress. However, because it is typically
necessary to supply power to products to find the soft or intermittent failures, it can
be relatively inexpensive to implement electrical stress in combination with other
stress environments to increase overall effectiveness.
Combined Environments
Although each environment (e.g., vibration, thermal cycling) must be separated and
understood individually, in the real world, these environments are combined. A car
or truck traverses rough roads in both the summer and winter, in the tropics and the
Arctic, and at sea level and high in the mountains. The most representative laboratory
tests can combine the vibration, shock, and thermal environments, as well as others.
A common combination of environments is that of vibration and thermal cycling.
Thermal chambers are made for placing over electrodynamic shakers. Using these, it
is possible to simultaneously apply a vibration input combined with thermal extremes
and thermal cycles. For example, it is possible that an electrical connection could
be cracked due to thermal extremes and differential thermal expansion. However,
because the two halves of the joint are in contact, the electrical performance is unaf-
fected. When a vibration environment is applied, the resulting motion causes physical
separation of the halves of their joint, resulting in an intermittent short in the electrical
circuit and exposing the defect. On a larger scale, it is possible to construct a test setup
where an entire car or truck can be mounted for multiple axes of real-time simulation
vibration input in a chamber where thermal extremes can be applied.
Other Types of Stress

In addition to the standard thermal, humidity, vibration, and electrical environments
that are tested, many other real-world environments must be considered. Depending
on the location of a component and its anticipated usage, it could be adversely affected
by many different factors. Salt fog and salt spray can cause corrosion and can accelerate
the effects of dissimilar metals. Ozone and ultraviolet light can cause the deteriora-
tion of elastomers and coatings. Moisture and dust driven by wind can penetrate and
damage many devices. All of these must be considered when developing a
laboratory test.
* This material is used courtesy of Thermotron Industries, www.thermotron.com.

1. Altitude (MIL-STD-810, Method No. 500.3)—The low pressure at high

altitudes can cause pressure loss within volumes that are nominally sealed.
When the unit returns to sea level, it then can ingest air with impurities,
which can adversely affect the unit.
2. Submersion (MIL-STD-810, Method No. 512.3)—Submersion is a test
usually applied to products that are applied to marine environments where
the possibility exists that the products could become completely submerged
in water. The submersion test evaluates the effectiveness of the seals in
the product.
3. Dust (MIL-STD-810, Method No. 510.3)—Dust, sand, and other impurities
can be detrimental to a product. These particles can be extremely abrasive
when driven by wind.
4. Moisture, Wind, and Rain (MIL-STD-810, Method No. 506.3)—Moisture
that is driven by wind can penetrate seals and cause damage to a product.
5. Ultraviolet (MIL-STD-810, Method No. 505.3)—Ultraviolet radiation can
cause heating of the product in a different manner than that of a
thermal soak.
6. Corrosion—Several forms of corrosion can affect the performance of a
product. These include salt spray and salt fog (MIL-STD-810, Method No.
509.3), brake fluid, soft drinks, and lotions.
7. Explosive Atmosphere (MIL-STD-810, Method No. 511.3)—Explosive
atmospheric testing is used to determine if the electronics in a product can
ignite a potentially explosive atmosphere in which it is operated. This
primarily is applicable to fuel system components.
8. Fungus (MIL-STD-810, Method No. 508.4)—Fungus infiltration can
be detrimental to the durability and operation of a product.
9. Icing (MIL-STD-810, Method No. 508.4)—Icing and freezing rain can
adversely affect the operation of components that are exposed to
these environments.
Summary
Stress environments can be used individually or in combination. The duration of the
tests containing these stresses may be measured in seconds or months. In general,
short-duration tests targeted to specific failure modes are used during product devel-
opment. This type of test is likely to use a single type of stress. Qualification tests are
typically longer in duration and use multiple types of stress.
Appendix A:
Statistical Tables
TABLE A.1 The gamma function.
x G(x) x G(x) x G(x) x G(x)

1.00 1.0000 1.25 0.9064 1.50 0.8862 1.75 0.9191
1.01 0.9943 1.26 0.9044 1.51 0.8866 1.76 0.9214
1.02 0.9888 1.27 0.9025 1.52 0.8870 1.77 0.9238
1.03 0.9835 1.28 0.9007 1.53 0.8876 1.78 0.9262
1.04 0.9784 1.29 0.8990 1.54 0.8882 1.79 0.9288
1.05 0.9735 1.30 0.8975 1.55 0.8889 1.80 0.9314
1.06 0.9687 1.31 0.8960 1.56 0.8896 1.81 0.9341
1.07 0.9642 1.32 0.8946 1.57 0.8905 1.82 0.9368
1.08 0.9597 1.33 0.8934 1.58 0.8914 1.83 0.9397
1.09 0.9555 1.34 0.8922 1.59 0.8924 1.84 0.9426
1.10 0.9514 1.35 0.8912 1.60 0.8935 1.85 0.9456
1.11 0.9474 1.36 0.8902 1.61 0.8947 1.86 0.9487
1.12 0.9436 1.37 0.8893 1.62 0.8959 1.87 0.9518
1.13 0.9399 1.38 0.8885 1.63 0.8972 1.88 0.9551
1.14 0.9364 1.39 0.8879 1.64 0.8986 1.89 0.9584
1.15 0.9330 1.40 0.8873 1.65 0.9001 1.90 0.9618
1.16 0.9298 1.41 0.8868 1.66 0.9017 1.91 0.9652
1.17 0.9267 1.42 0.8864 1.67 0.9033 1.92 0.9688
1.18 0.9237 1.43 0.8860 1.68 0.9050 1.93 0.9724
1.19 0.9209 1.44 0.8858 1.69 0.9068 1.94 0.9761
1.20 0.9182 1.45 0.8857 1.70 0.9086 1.95 0.9799
1.21 0.9156 1.46 0.8856 1.71 0.9106 1.96 0.9837
1.22 0.9131 1.47 0.8856 1.72 0.9126 1.97 0.9877
1.23 0.9108 1.48 0.8857 1.73 0.9147 1.98 0.9917
1.24 0.9085 1.49 0.8859 1.74 0.9168 1.99 0.9958
∞
2.00 1.0000
∫ e −τ τ(
x − 1)
Γ(x) = dτ
0
Γ(x + 1) = xΓ(x)

206 Appendix A: Statistical Tables
TABLE A.2 Standard normal cumulative distribution function.
z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
−3.40 0.0003 0.0003 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0005 0.0005
−3.30 0.0005 0.0005 0.0005 0.0005 0.0006 0.0006 0.0006 0.0006 0.0006 0.0007
−3.20 0.0007 0.0007 0.0007 0.0008 0.0008 0.0008 0.0008 0.0009 0.0009 0.0009
−3.10 0.0010 0.0010 0.0010 0.0011 0.0011 0.0011 0.0012 0.0012 0.0013 0.0013
−3.00 0.0013 0.0014 0.0014 0.0015 0.0015 0.0016 0.0016 0.0017 0.0018 0.0018
−2.90 0.0019 0.0019 0.0020 0.0021 0.0021 0.0022 0.0023 0.0023 0.0024 0.0025
−2.80 0.0026 0.0026 0.0027 0.0028 0.0029 0.0030 0.0031 0.0032 0.0033 0.0034
−2.70 0.0035 0.0036 0.0037 0.0038 0.0039 0.0040 0.0041 0.0043 0.0044 0.0045
−2.60 0.0047 0.0048 0.0049 0.0051 0.0052 0.0054 0.0055 0.0057 0.0059 0.0060
−2.50 0.0062 0.0064 0.0066 0.0068 0.0069 0.0071 0.0073 0.0075 0.0078 0.0080
−2.40 0.0082 0.0084 0.0087 0.0089 0.0091 0.0094 0.0096 0.0099 0.0102 0.0104
−2.30 0.0107 0.0110 0.0113 0.0116 0.0119 0.0122 0.0125 0.0129 0.0132 0.0136
−2.20 0.0139 0.0143 0.0146 0.0150 0.0154 0.0158 0.0162 0.0166 0.0170 0.0174
−2.10 0.0179 0.0183 0.0188 0.0192 0.0197 0.0202 0.0207 0.0212 0.0217 0.0222
−2.00 0.0228 0.0233 0.0239 0.0244 0.0250 0.0256 0.0262 0.0268 0.0274 0.0281
−1.90 0.0287 0.0294 0.0301 0.0307 0.0314 0.0322 0.0329 0.0336 0.0344 0.0351
−1.80 0.0359 0.0367 0.0375 0.0384 0.0392 0.0401 0.0409 0.0418 0.0427 0.0436
−1.70 0.0446 0.0455 0.0465 0.0475 0.0485 0.0495 0.0505 0.0516 0.0526 0.0537
−1.60 0.0548 0.0559 0.0571 0.0582 0.0594 0.0606 0.0618 0.0630 0.0643 0.0655
−1.50 0.0668 0.0681 0.0694 0.0708 0.0721 0.0735 0.0749 0.0764 0.0778 0.0793
−1.40 0.0808 0.0823 0.0838 0.0853 0.0869 0.0885 0.0901 0.0918 0.0934 0.0951
−1.30 0.0968 0.0985 0.1003 0.1020 0.1038 0.1056 0.1075 0.1093 0.1112 0.1131
−1.20 0.1151 0.1170 0.1190 0.1210 0.1230 0.1251 0.1271 0.1292 0.1314 0.1335
−1.10 0.1357 0.1379 0.1401 0.1423 0.1446 0.1469 0.1492 0.1515 0.1539 0.1562
−1.00 0.1587 0.1611 0.1635 0.1660 0.1685 0.1711 0.1736 0.1762 0.1788 0.1814
−0.90 0.1841 0.1867 0.1894 0.1922 0.1949 0.1977 0.2005 0.2033 0.2061 0.2090
−0.80 0.2119 0.2148 0.2177 0.2206 0.2236 0.2266 0.2296 0.2327 0.2358 0.2389
−0.70 0.2420 0.2451 0.2483 0.2514 0.2546 0.2578 0.2611 0.2643 0.2676 0.2709
−0.60 0.2743 0.2776 0.2810 0.2843 0.2877 0.2912 0.2946 0.2981 0.3015 0.3050
−0.50 0.3085 0.3121 0.3156 0.3192 0.3228 0.3264 0.3300 0.3336 0.3372 0.3409
−0.40 0.3446 0.3483 0.3520 0.3557 0.3594 0.3632 0.3669 0.3707 0.3745 0.3783
−0.30 0.3821 0.3859 0.3897 0.3936 0.3974 0.4013 0.4052 0.4090 0.4129 0.4168
−0.20 0.4207 0.4247 0.4286 0.4325 0.4364 0.4404 0.4443 0.4483 0.4522 0.4562
−0.10 0.4602 0.4641 0.4681 0.4721 0.4761 0.4801 0.4840 0.4880 0.4920 0.4960
0.00 0.5000 0.5040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359
TABLE A.2 (Continued) Standard normal cumulative distribution function.
TABLE A.6
Poisson cumulative distribution function.
z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.10 0.5398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5753
0.20 0.5793 0.5832 0.5871 0.5910 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141
0.30 0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517
0.40 0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879
0.50 0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224
0.60 0.7257 0.7291 0.7324 0.7357 0.7389 0.7422 0.7454 0.7486 0.7517 0.7549
0.70 0.7580 0.7611 0.7642 0.7673 0.7704 0.7734 0.7764 0.7794 0.7823 0.7852
0.80 0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8078 0.8106 0.8133
0.90 0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389
1.00 0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0.8599 0.8621
1.10 0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830
1.20 0.8849 0.8869 0.8888 0.8907 0.8925 0.8944 0.8962 0.8980 0.8997 0.9015
1.30 0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177
1.40 0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9279 0.9292 0.9306 0.9319
1.50 0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441
1.60 0.9452 0.9463 0.9474 0.9484 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545
1.70 0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.9633
1.80 0.9641 0.9649 0.9656 0.9664 0.9671 0.9678 0.9686 0.9693 0.9699 0.9706
1.90 0.9713 0.9719 0.9726 0.9732 0.9738 0.9744 0.9750 0.9756 0.9761 0.9767
2.00 0.9772 0.9778 0.9783 0.9788 0.9793 0.9798 0.9803 0.9808 0.9812 0.9817
2.10 0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857
2.20 0.9861 0.9864 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890
2.30 0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916
2.40 0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9931 0.9932 0.9934 0.9936
2.50 0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952
2.60 0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964
2.70 0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974
2.80 0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9979 0.9980 0.9981
2.90 0.9981 0.9982 0.9982 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986
3.00 0.9987 0.9987 0.9987 0.9988 0.9988 0.9989 0.9989 0.9989 0.9990 0.9990
3.10 0.9990 0.9991 0.9991 0.9991 0.9992 0.9992 0.9992 0.9992 0.9993 0.9993
3.20 0.9993 0.9993 0.9994 0.9994 0.9994 0.9994 0.9994 0.9995 0.9995 0.9995
3.30 0.9995 0.9995 0.9995 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9997
3.40 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9998
TABLE A.3 Chi-square significance.
α
y 0.005 0.01 0.02 0.025 0.05 0.1 0.2 0.25 0.3
1 7.8794 6.6349 5.4119 5.0239 3.8415 2.7055 1.6424 1.3233 1.0742
2 10.5966 9.2103 7.8240 7.3778 5.9915 4.6052 3.2189 2.7726 2.4079
3 12.8382 11.3449 9.8374 9.3484 7.8147 6.2514 4.6416 4.1083 3.6649
4 14.8603 13.2767 11.6678 11.1433 9.4877 7.7794 5.9886 5.3853 4.8784
5 16.7496 15.0863 13.3882 12.8325 11.0705 9.2364 7.2893 6.6257 6.0644
6 18.5476 16.8119 15.0332 14.4494 12.5916 10.6446 8.5581 7.8408 7.2311
7 20.2777 18.4753 16.6224 16.0128 14.0671 12.0170 9.8032 9.0371 8.3834
8 21.9550 20.0902 18.1682 17.5345 15.5073 13.3616 11.0301 10.2189 9.5245
9 23.5894 21.6660 19.6790 19.0228 16.9190 14.6837 12.2421 11.3888 10.6564
10 25.1882 23.2093 21.1608 20.4832 18.3070 15.9872 13.4420 12.5489 11.7807
11 26.7568 24.7250 22.6179 21.9200 19.6751 17.2750 14.6314 13.7007 12.8987
12 28.2995 26.2170 24.0540 23.3367 21.0261 18.5493 15.8120 14.8454 14.0111
13 29.8195 27.6882 25.4715 24.7356 22.3620 19.8119 16.9848 15.9839 15.1187
14 31.3193 29.1412 26.8728 26.1189 23.6848 21.0641 18.1508 17.1169 16.2221
15 32.8013 30.5779 28.2595 27.4884 24.9958 22.3071 19.3107 18.2451 17.3217
16 34.2672 31.9999 29.6332 28.8454 26.2962 23.5418 20.4651 19.3689 18.4179
17 35.7185 33.4087 30.9950 30.1910 27.5871 24.7690 21.6146 20.4887 19.5110
18 37.1565 34.8053 32.3462 31.5264 28.8693 25.9894 22.7595 21.6049 20.6014
19 38.5823 36.1909 33.6874 32.8523 30.1435 27.2036 23.9004 22.7178 21.6891
20 39.9968 37.5662 35.0196 34.1696 31.4104 28.4120 25.0375 23.8277 22.7745
21 41.4011 38.9322 36.3434 35.4789 32.6706 29.6151 26.1711 24.9348 23.8578
22 42.7957 40.2894 37.6595 36.7807 33.9244 30.8133 27.3015 26.0393 24.9390
23 44.1813 41.6384 38.9683 38.0756 35.1725 32.0069 28.4288 27.1413 26.0184
24 45.5585 42.9798 40.2704 39.3641 36.4150 33.1962 29.5533 28.2412 27.0960
25 46.9279 44.3141 41.5661 40.6465 37.6525 34.3816 30.6752 29.3389 28.1719
26 48.2899 45.6417 42.8558 41.9232 38.8851 35.5632 31.7946 30.4346 29.2463
27 49.6449 46.9629 44.1400 43.1945 40.1133 36.7412 32.9117 31.5284 30.3193
28 50.9934 48.2782 45.4188 44.4608 41.3371 37.9159 34.0266 32.6205 31.3909
29 52.3356 49.5879 46.6927 45.7223 42.5570 39.0875 35.1394 33.7109 32.4612
30 53.6720 50.8922 47.9618 46.9792 43.7730 40.2560 36.2502 34.7997 33.5302
TABLE A.3 (Continued) Chi-square significance.

α
0.5 0.6 0.7 0.8 0.85 0.9 0.95 0.975 0.99 0.995
0.4549 0.2750 0.1485 0.0642 0.0358 0.0158 0.0039 0.0010 0.0002 0.0000
1.3863 1.0217 0.7133 0.4463 0.3250 0.2107 0.1026 0.0506 0.0201 0.0100
2.3660 1.8692 1.4237 1.0052 0.7978 0.5844 0.3518 0.2158 0.1148 0.0717
3.3567 2.7528 2.1947 1.6488 1.3665 1.0636 0.7107 0.4844 0.2971 0.2070
4.3515 3.6555 2.9999 2.3425 1.9938 1.6103 1.1455 0.8312 0.5543 0.4117
5.3481 4.5702 3.8276 3.0701 2.6613 2.2041 1.6354 1.2373 0.8721 0.6757
6.3458 5.4932 4.6713 3.8223 3.3583 2.8331 2.1673 1.6899 1.2390 0.9893
7.3441 6.4226 5.5274 4.5936 4.0782 3.4895 2.7326 2.1797 1.6465 1.3444
8.3428 7.3570 6.3933 5.3801 4.8165 4.1682 3.3251 2.7004 2.0879 1.7349
9.3418 8.2955 7.2672 6.1791 5.5701 4.8652 3.9403 3.2470 2.5582 2.1559
10.3410 9.2373 8.1479 6.9887 6.3364 5.5778 4.5748 3.8157 3.0535 2.6032
11.3403 10.1820 9.0343 7.8073 7.1138 6.3038 5.2260 4.4038 3.5706 3.0738
12.3398 11.1291 9.9257 8.6339 7.9008 7.0415 5.8919 5.0088 4.1069 3.5650
13.3393 12.0785 10.8215 9.4673 8.6963 7.7895 6.5706 5.6287 4.6604 4.0747
14.3389 13.0297 11.7212 10.3070 9.4993 8.5468 7.2609 6.2621 5.2293 4.6009
15.3385 13.9827 12.6243 11.1521 10.3090 9.3122 7.9616 6.9077 5.8122 5.1422
16.3382 14.9373 13.5307 12.0023 11.1249 10.0852 8.6718 7.5642 6.4078 5.6972
17.3379 15.8932 14.4399 12.8570 11.9463 10.8649 9.3905 8.2307 7.0149 6.2648
18.3377 16.8504 15.3517 13.7158 12.7727 11.6509 10.1170 8.9065 7.6327 6.8440
19.3374 17.8088 16.2659 14.5784 13.6039 12.4426 10.8508 9.5908 8.2604 7.4338
20.3372 18.7683 17.1823 15.4446 14.4393 13.2396 11.5913 10.2829 8.8972 8.0337
21.3370 19.7288 18.1007 16.3140 15.2788 14.0415 12.3380 10.9823 9.5425 8.6427
22.3369 20.6902 19.0211 17.1865 16.1219 14.8480 13.0905 11.6886 10.1957 9.2604
23.3367 21.6525 19.9432 18.0618 16.9686 15.6587 13.8484 12.4012 10.8564 9.8862
24.3366 22.6156 20.8670 18.9398 17.8184 16.4734 14.6114 13.1197 11.5240 10.5197
25.3365 23.5794 21.7924 19.8202 18.6714 17.2919 15.3792 13.8439 12.1981 11.1602
26.3363 24.5440 22.7192 20.7030 19.5272 18.1139 16.1514 14.5734 12.8785 11.8076
27.3362 25.5093 23.6475 21.5880 20.3857 18.9392 16.9279 15.3079 13.5647 12.4613
28.3361 26.4751 24.5770 22.4751 21.2468 19.7677 17.7084 16.0471 14.2565 13.1211
29.3360 27.4416 25.5078 23.3641 22.1103 20.5992 18.4927 16.7908 14.9535 13.7867
TABLE A.4 F significance.
210
Appendix A: Statistical Tables
TABLE A.4.1 α = 0.01.
Y2
1
Y 1 2 3 4 5 6 7 8 9 10 15 20 25 50 100
1 4052.18 98.50 34.12 21.20 16.26 13.75 12.25 11.26 10.56 10.04 8.68 8.10 7.77 7.17 6.90
2 4999.50 99.00 30.82 18.00 13.27 10.92 9.55 8.65 8.02 7.56 6.36 5.85 5.57 5.06 4.82
3 5403.35 99.17 29.46 16.69 12.06 9.78 8.45 7.59 6.99 6.55 5.42 4.94 4.68 4.20 3.98
4 5624.58 99.25 28.71 15.98 11.39 9.15 7.85 7.01 6.42 5.99 4.89 4.43 4.18 3.72 3.51
5 5763.65 99.30 28.24 15.52 10.97 8.75 7.46 6.63 6.06 5.64 4.56 4.10 3.85 3.41 3.21
6 5858.99 99.33 27.91 15.21 10.67 8.47 7.19 6.37 5.80 5.39 4.32 3.87 3.63 3.19 2.99
7 5928.36 99.36 27.67 14.98 10.46 8.26 6.99 6.18 5.61 5.20 4.14 3.70 3.46 3.02 2.82
8 5981.07 99.37 27.49 14.80 10.29 8.10 6.84 6.03 5.47 5.06 4.00 3.56 3.32 2.89 2.69
9 6022.47 99.39 27.35 14.66 10.16 7.98 6.72 5.91 5.35 4.94 3.89 3.46 3.22 2.78 2.59
10 6055.85 99.40 27.23 14.55 10.05 7.87 6.62 5.81 5.26 4.85 3.80 3.37 3.13 2.70 2.50
15 6157.28 99.43 26.87 14.20 9.72 7.56 6.31 5.52 4.96 4.56 3.52 3.09 2.85 2.42 2.22
20 6208.73 99.45 26.69 14.02 9.55 7.40 6.16 5.36 4.81 4.41 3.37 2.94 2.70 2.27 2.07
25 6239.83 99.46 26.58 13.91 9.45 7.30 6.06 5.26 4.71 4.31 3.28 2.84 2.60 2.17 1.97
50 6302.52 99.48 26.35 13.69 9.24 7.09 5.86 5.07 4.52 4.12 3.08 2.64 2.40 1.95 1.74
100 6334.11 99.49 26.24 13.58 9.13 6.99 5.75 4.96 4.41 4.01 2.98 2.54 2.29 1.82 1.60
TABLE A.4.2 α = 0.025.
Y2
1
Y 1 2 3 4 5 6 7 8 9 10 15 20 25 50 100
1 647.79 38.51 17.44 12.22 10.01 8.81 8.07 7.57 7.21 6.94 6.20 5.87 5.69 5.34 5.18
2 799.50 39.00 16.04 10.65 8.43 7.26 6.54 6.06 5.71 5.46 4.77 4.46 4.29 3.97 3.83
3 864.16 39.17 15.44 9.98 7.76 6.60 5.89 5.42 5.08 4.83 4.15 3.86 3.69 3.39 3.25
4 899.58 39.25 15.10 9.60 7.39 6.23 5.52 5.05 4.72 4.47 3.80 3.51 3.35 3.05 2.92
5 921.85 39.30 14.88 9.36 7.15 5.99 5.29 4.82 4.48 4.24 3.58 3.29 3.13 2.83 2.70
6 937.11 39.33 14.73 9.20 6.98 5.82 5.12 4.65 4.32 4.07 3.41 3.13 2.97 2.67 2.54
7 948.22 39.36 14.62 9.07 6.85 5.70 4.99 4.53 4.20 3.95 3.29 3.01 2.85 2.55 2.42
8 956.66 39.37 14.54 8.98 6.76 5.60 4.90 4.43 4.10 3.85 3.20 2.91 2.75 2.46 2.32
9 963.28 39.39 14.47 8.90 6.68 5.52 4.82 4.36 4.03 3.78 3.12 2.84 2.68 2.38 2.24
10 968.63 39.40 14.42 8.84 6.62 5.46 4.76 4.30 3.96 3.72 3.06 2.77 2.61 2.32 2.18
15 984.87 39.43 14.25 8.66 6.43 5.27 4.57 4.10 3.77 3.52 2.86 2.57 2.41 2.11 1.97
20 993.10 39.45 14.17 8.56 6.33 5.17 4.47 4.00 3.67 3.42 2.76 2.46 2.30 1.99 1.85
25 998.08 39.46 14.12 8.50 6.27 5.11 4.40 3.94 3.60 3.35 2.69 2.40 2.23 1.92 1.77
50 1008.12 39.48 14.01 8.38 6.14 4.98 4.28 3.81 3.47 3.22 2.55 2.25 2.08 1.75 1.59
100 1013.17 39.49 13.96 8.32 6.08 4.92 4.21 3.74 3.40 3.15 2.47 2.17 2.00 1.66 1.48
211
TABLE A.4.3 α = 0.05.
212
Y2
1
Y 1 2 3 4 5 6 7 8 9 10 15 20 25 50 100
1 161.45 18.51 10.13 7.71 6.61 5.99 5.59 5.32 5.12 4.96 4.54 4.35 4.24 4.03 3.94
2 199.50 19.00 9.55 6.94 5.79 5.14 4.74 4.46 4.26 4.10 3.68 3.49 3.39 3.18 3.09
3 215.71 19.16 9.28 6.59 5.41 4.76 4.35 4.07 3.86 3.71 3.29 3.10 2.99 2.79 2.70
4 224.58 19.25 9.12 6.39 5.19 4.53 4.12 3.84 3.63 3.48 3.06 2.87 2.76 2.56 2.46
5 230.16 19.30 9.01 6.26 5.05 4.39 3.97 3.69 3.48 3.33 2.90 2.71 2.60 2.40 2.31
6 233.99 19.33 8.94 6.16 4.95 4.28 3.87 3.58 3.37 3.22 2.79 2.60 2.49 2.29 2.19
7 236.77 19.35 8.89 6.09 4.88 4.21 3.79 3.50 3.29 3.14 2.71 2.51 2.40 2.20 2.10
8 238.88 19.37 8.85 6.04 4.82 4.15 3.73 3.44 3.23 3.07 2.64 2.45 2.34 2.13 2.03
9 240.54 19.38 8.81 6.00 4.77 4.10 3.68 3.39 3.18 3.02 2.59 2.39 2.28 2.07 1.97
10 241.88 19.40 8.79 5.96 4.74 4.06 3.64 3.35 3.14 2.98 2.54 2.35 2.24 2.03 1.93
15 245.95 19.43 8.70 5.86 4.62 3.94 3.51 3.22 3.01 2.85 2.40 2.20 2.09 1.87 1.77
20 248.01 19.45 8.66 5.80 4.56 3.87 3.44 3.15 2.94 2.77 2.33 2.12 2.01 1.78 1.68
25 249.26 19.46 8.63 5.77 4.52 3.83 3.40 3.11 2.89 2.73 2.28 2.07 1.96 1.73 1.62
50 251.77 19.48 8.58 5.70 4.44 3.75 3.32 3.02 2.80 2.64 2.18 1.97 1.84 1.60 1.48
100 253.04 19.49 8.55 5.66 4.41 3.71 3.27 2.97 2.76 2.59 2.12 1.91 1.78 1.52 1.39
TABLE A.4.4 α = 0.1.
Y2
1
Y 1 2 3 4 5 6 7 8 9 10 15 20 25 50 100
1 39.86 8.53 5.54 4.54 4.06 3.78 3.59 3.46 3.36 3.29 3.07 2.97 2.92 2.81 2.76
2 49.50 9.00 5.46 4.32 3.78 3.46 3.26 3.11 3.01 2.92 2.70 2.59 2.53 2.41 2.36
3 53.59 9.16 5.39 4.19 3.62 3.29 3.07 2.92 2.81 2.73 2.49 2.38 2.32 2.20 2.14
4 55.83 9.24 5.34 4.11 3.52 3.18 2.96 2.81 2.69 2.61 2.36 2.25 2.18 2.06 2.00
5 57.24 9.29 5.31 4.05 3.45 3.11 2.88 2.73 2.61 2.52 2.27 2.16 2.09 1.97 1.91
6 58.20 9.33 5.28 4.01 3.40 3.05 2.83 2.67 2.55 2.46 2.21 2.09 2.02 1.90 1.83
7 58.91 9.35 5.27 3.98 3.37 3.01 2.78 2.62 2.51 2.41 2.16 2.04 1.97 1.84 1.78
8 59.44 9.37 5.25 3.95 3.34 2.98 2.75 2.59 2.47 2.38 2.12 2.00 1.93 1.80 1.73
9 59.86 9.38 5.24 3.94 3.32 2.96 2.72 2.56 2.44 2.35 2.09 1.96 1.89 1.76 1.69
10 60.19 9.39 5.23 3.92 3.30 2.94 2.70 2.54 2.42 2.32 2.06 1.94 1.87 1.73 1.66
15 61.22 9.42 5.20 3.87 3.24 2.87 2.63 2.46 2.34 2.24 1.97 1.84 1.77 1.63 1.56
20 61.74 9.44 5.18 3.84 3.21 2.84 2.59 2.42 2.30 2.20 1.92 1.79 1.72 1.57 1.49
25 62.05 9.45 5.17 3.83 3.19 2.81 2.57 2.40 2.27 2.17 1.89 1.76 1.68 1.53 1.45
50 62.69 9.47 5.15 3.80 3.15 2.77 2.52 2.35 2.22 2.12 1.83 1.69 1.61 1.44 1.35
100 63.01 9.48 5.14 3.78 3.13 2.75 2.50 2.32 2.19 2.09 1.79 1.65 1.56 1.39 1.29
213
TABLE A.5 t significance.
Y 0.1 0.05 0.025 0.01 0.005

1 3.0777 6.3138 12.7062 31.8205 63.6567
2 1.8856 2.9200 4.3027 6.9646 9.9248
3 1.6377 2.3534 3.1824 4.5407 5.8409
4 1.5332 2.1318 2.7764 3.7469 4.6041
5 1.4759 2.0150 2.5706 3.3649 4.0321
6 1.4398 1.9432 2.4469 3.1427 3.7074
7 1.4149 1.8946 2.3646 2.9980 3.4995
8 1.3968 1.8595 2.3060 2.8965 3.3554
9 1.3830 1.8331 2.2622 2.8214 3.2498
10 1.3722 1.8125 2.2281 2.7638 3.1693
11 1.3634 1.7959 2.2010 2.7181 3.1058
12 1.3562 1.7823 2.1788 2.6810 3.0545
13 1.3502 1.7709 2.1604 2.6503 3.0123
14 1.3450 1.7613 2.1448 2.6245 2.9768
15 1.3406 1.7531 2.1314 2.6025 2.9467
16 1.3368 1.7459 2.1199 2.5835 2.9208
17 1.3334 1.7396 2.1098 2.5669 2.8982
18 1.3304 1.7341 2.1009 2.5524 2.8784
19 1.3277 1.7291 2.0930 2.5395 2.8609
20 1.3253 1.7247 2.0860 2.5280 2.8453
21 1.3232 1.7207 2.0796 2.5176 2.8314
22 1.3212 1.7171 2.0739 2.5083 2.8188
23 1.3195 1.7139 2.0687 2.4999 2.8073
24 1.3178 1.7109 2.0639 2.4922 2.7969
25 1.3163 1.7081 2.0595 2.4851 2.7874
26 1.3150 1.7056 2.0555 2.4786 2.7787
27 1.3137 1.7033 2.0518 2.4727 2.7707
28 1.3125 1.7011 2.0484 2.4671 2.7633
29 1.3114 1.6991 2.0452 2.4620 2.7564
∞ 1.2816 1.6449 1.9600 2.3263 2.5758
TABLE A.6 Poisson cumulative distribution function.

μ
r 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0 0.9048 0.8187 0.7408 0.6703 0.6065 0.5488 0.4966 0.4493 0.4066
1 0.9953 0.9825 0.9631 0.9384 0.9098 0.8781 0.8442 0.8088 0.7725
2 0.9998 0.9989 0.9964 0.9921 0.9856 0.9769 0.9659 0.9526 0.9371
3 1.0000 0.9999 0.9997 0.9992 0.9982 0.9966 0.9942 0.9909 0.9865
4 1.0000 1.0000 1.0000 0.9999 0.9998 0.9996 0.9992 0.9986 0.9977
5 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9998 0.9997
6 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
μ
r 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
0 0.3679 0.2231 0.1353 0.0821 0.0498 0.0302 0.0183 0.0111 0.0067
1 0.7358 0.5578 0.4060 0.2873 0.1991 0.1359 0.0916 0.0611 0.0404
2 0.9197 0.8088 0.6767 0.5438 0.4232 0.3208 0.2381 0.1736 0.1247
3 0.9810 0.9344 0.8571 0.7576 0.6472 0.5366 0.4335 0.3423 0.2650
4 0.9963 0.9814 0.9473 0.8912 0.8153 0.7254 0.6288 0.5321 0.4405
5 0.9994 0.9955 0.9834 0.9580 0.9161 0.8576 0.7851 0.7029 0.6160
6 0.9999 0.9991 0.9955 0.9858 0.9665 0.9347 0.8893 0.8311 0.7622
7 1.0000 0.9998 0.9989 0.9958 0.9881 0.9733 0.9489 0.9134 0.8666
8 1.0000 1.0000 0.9998 0.9989 0.9962 0.9901 0.9786 0.9597 0.9319
9 1.0000 1.0000 1.0000 0.9997 0.9989 0.9967 0.9919 0.9829 0.9682
10 1.0000 1.0000 1.0000 0.9999 0.9997 0.9990 0.9972 0.9933 0.9863
μ
r 6 7 8 9 10 12 15 20 25
0 0.0025 0.0009 0.0003 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000
1 0.0174 0.0073 0.0030 0.0012 0.0005 0.0001 0.0000 0.0000 0.0000
2 0.0620 0.0296 0.0138 0.0062 0.0028 0.0005 0.0000 0.0000 0.0000
3 0.1512 0.0818 0.0424 0.0212 0.0103 0.0023 0.0002 0.0000 0.0000
4 0.2851 0.1730 0.0996 0.0550 0.0293 0.0076 0.0009 0.0000 0.0000
5 0.4457 0.3007 0.1912 0.1157 0.0671 0.0203 0.0028 0.0001 0.0000
6 0.6063 0.4497 0.3134 0.2068 0.1301 0.0458 0.0076 0.0003 0.0000
7 0.7440 0.5987 0.4530 0.3239 0.2202 0.0895 0.0180 0.0008 0.0000
8 0.8472 0.7291 0.5925 0.4557 0.3328 0.1550 0.0374 0.0021 0.0001
9 0.9161 0.8305 0.7166 0.5874 0.4579 0.2424 0.0699 0.0050 0.0002
10 0.9574 0.9015 0.8159 0.7060 0.5830 0.3472 0.1185 0.0108 0.0006
11 0.9799 0.9467 0.8881 0.8030 0.6968 0.4616 0.1848 0.0214 0.0014
12 0.9912 0.9730 0.9362 0.8758 0.7916 0.5760 0.2676 0.0390 0.0031
13 0.9964 0.9872 0.9658 0.9261 0.8645 0.6815 0.3632 0.0661 0.0065
14 0.9986 0.9943 0.9827 0.9585 0.9165 0.7720 0.4657 0.1049 0.0124
15 0.9995 0.9976 0.9918 0.9780 0.9513 0.8444 0.5681 0.1565 0.0223
16 0.9998 0.9990 0.9963 0.9889 0.9730 0.8987 0.6641 0.2211 0.0377
17 0.9999 0.9996 0.9984 0.9947 0.9857 0.9370 0.7489 0.2970 0.0605
18 1.0000 0.9999 0.9993 0.9976 0.9928 0.9626 0.8195 0.3814 0.0920
19 1.0000 1.0000 0.9997 0.9989 0.9965 0.9787 0.8752 0.4703 0.1336
20 1.0000 1.0000 0.9999 0.9996 0.9984 0.9884 0.9170 0.5591 0.1855
21 1.0000 1.0000 1.0000 0.9998 0.9993 0.9939 0.9469 0.6437 0.2473
22 1.0000 1.0000 1.0000 0.9999 0.9997 0.9970 0.9673 0.7206 0.3175
23 1.0000 1.0000 1.0000 1.0000 0.9999 0.9985 0.9805 0.7875 0.3939
24 1.0000 1.0000 1.0000 1.0000 1.0000 0.9993 0.9888 0.8432 0.4734
Appendix B: Government
Documents
T
he United States government used to publish government documents.
The following documents are available on the accompanying CD. The following
documents are available on the website listed on page xi.
MIL-HDBK-189 Reliability Growth Management

MIL-HDBK-217 Reliability Prediction of Electronic Equipment
MIL-HDBK-251 Reliability/Design Thermal Applications
MIL-HDBK-263 Electrostatic Discharge Control for Protection of Electrical and
Electronic Parts, Assemblies, and Equipment
MIL-HDBK-338 Electronic Reliability Design Handbook
MIL-HDBK-472 Maintainability Prediction
MIL-HDBK-781 Reliability Test Methods, Plans, and Environments for
Engineering Development, Qualification, and Production
MIL-STD-690 Failure Rate Sampling Plans and Procedures
MIL-STD-721 Definition of Terms for Reliability and Maintainability
MIL-STD-785 Reliability Program for Systems and Equipment Development
and Production
MIL-STD-883 Test Methods and Procedures for Microelectronics
MIL-STD-965 Parts Control Program
MIL-STD-1472 Human Engineering Design Criteria for Military Systems,
Equipment, and Facilities
MIL-STD-1629 Procedures for Performing a Failure Mode, Effects, and
Criticality Analysis
MIL-STD-1635 Reliability Growth Testing
MIL-STD-2074 Failure Classification for Reliability Testing

218 Appendix B: Government Documents
Other documents of interest include the following:
DOD-HDBK-344 Environmental Stress Screening of Electronic Equipment, U.S.

(U.S.A.F.) Department of Defense, 1993
Environmental Stress Screening of Electronic Hardware
(ESSEH), Institute of Environmental Sciences and Technology
(IEST), 1990
LC-78-2 Storage Reliability Analysis Report (Vol. 1, Electrical and
Electronic Devices; Vol. 2, Electromechanical Devices; Vol. 3,
Hydraulic and Pneumatic Devices; Vol. 4, Ordnance Devices;
Vol. 5, Optical Devices and Electro-Optical Devices)
MIL-STD-280 Definitions of Item Levels, Item Exchangeability, Models, and
Related Terms, U.S. Department of Defense, 1990
MIL-STD-470 Maintainability Program Requirements for Systems
and Equipment
MIL-STD-756 Reliability Modeling and Prediction
MIL-STD-757 Reliability Evaluation from Demonstration Data
MIL-STD-781 Reliability Testing for Engineering Development,
Qualification, and Production
MIL-STD-790 Reliability Assurance for Electronic Parts
MIL-STD-810 Environmental Test Methods and Engineering Guidelines
MIL-STD-2068 Reliability Development Tests
NAVMAT P-9492 Navy Manufacturing Screening Program, U.S. Department of
Navy, May 1979
NPRD-2 Nonelectronic Parts Reliability Data
RADC-TR-75-22 Nonelectric Reliability Notebook
Rome Air Development Center, various studies, Griffiss AFB, Rome, NY
13441-5700.
Appendix C: Glossary
Accessibility—A measure of the relative ease of Debugging—A process to detect and remedy
admission to the various areas of an item for the inadequacies; not to be confused with terms such
purpose of operation or maintenance. as Burn-In, Fault Isolation, or Screening.
Achieved—Obta i ned as t he resu lt of Degradation—A gradual impairment in the
measurement. ability to perform.
Alignment—Performing the adjustments that Demonstrated—That which has been measured
are necessary to return an item to a specified by the use of objective evidence gathered under
operation. specified conditions.
Availability—A measure of the degree to which Dependability—A measure of the degree
an item is in an operable and committable state at to which an item is operable and capable of
the start of a mission when the mission is called for performing its required function at any (random)
at an unknown (random) time. (Item state at the time during a specified mission profile, given the
start of a mission includes the combined effects of item availability at the start of the mission. (Item
the readiness-related system reliability and main- state during a mission includes the combined
tainability [R&M] parameters but excludes mission effects of the mission-related system R&M
time; see Dependability). parameters but excludes non-mission time; see
Burn-In (Pre-conditioning)—The operation Availability.)
of an item under stress to stabilize its characteris- Derating—(a) Using an item in such a way that
tics; not to be confused with Debugging. applied stresses are below rated values or (b) the
Calibration—A comparison of a measuring lowering of the rating of an item in one stress field.
device with a known standard; not to be confused Direct Maintenance Man-Hours per
with Alignment (see MIL-C-45662). Maintenance Action (DMMH/MA)—A
Chargeable—Within the responsibility of a measure of the maintainability parameter related
given organizational entity (applied to terms such to item demand for maintenance manpower. The
as failures and maintenance time). sum of direct maintenance man-hours divided
by the total number of maintenance actions
Checkout—Tests or observations of an item to (preventive and corrective) during a stated period
determine its condition or status. of time.
Corrective Action—A documented design, Direct Maintenance Man-Hours per
process, procedure, or materials change imple- Maintenance Event (DMMH/ME)—A
mented and validated to correct the cause of failure measure of the maintainability parameter related
or design deficiency. to an item demand for maintenance manpower.
Criticality—A relative measure of the conse- The sum of direct maintenance man-hours divided
quence of a failure mode and its frequency of by the total number of maintenance events (preven-
occurrences. tive and corrective) during a stated period of time.

220 Appendix C: Glossary
Disassemble—Opening an item and removing ability to perform within specified limits without
a number of parts or subassemblies to make the any remedial action.
item that is to be replaced accessible for removal. Failure Mechanism—The physical, chemical,
This does not include the actual removal of the electrical, thermal, or other processes that result
item to be replaced. in failure.
Dormant—The state wherein an item is able to Failure Mode—The consequence of the mecha-
function but is not required to function; not to nism through which the failure occurs (i.e., short,
be confused with downtime operable, which is open, fracture, excessive wear).
the state of being able to perform the intended
function; the same as Not Operating. Failure Modes and Effects Analysis
(FMEA)—A procedure by which each potential
Downing Event—The event that causes an item failure mode in a system is analyzed to determine the
to become unavailable to initiate its mission (the results or effects thereof on the system, and to classify
transition from uptime to downtime). each potential failure mode according to its severity.
Durability—A measure of useful life (a special Failure, Non-chargeable—(a) A nonrel-
case of reliability). evant failure or (b) a relevant failure caused by a
Environment—The aggregate of all external and condition previously specified as not within the
internal conditions (e.g., temperature, humidity, responsibility of a given organizational entity. (All
radiation, magnetic and electric fields, shock relevant failures are chargeable to one organiza-
vibration), whether natural or man-made, or self- tional entity or another.)
induced, that influences the form, performance, Failure, Nonrelevant—(a) A failure verified
reliability, or survival of an item. as having been caused by a condition not present
E nvi ro n m e n t a l S t r e s s S c r e e n i n g in the operational environment or (b) a failure
(ESS)—A series of tests conducted under envi- verified as peculiar to an item design that will not
ronmental stresses to disclose weak parts and enter the operational inventory.
workmanship defects for correction. Failure, Random—Failure whose occurrence
Failure—The event, or inoperable state, in which is predictable only in a probabilistic or statistical
any item or part of an item does not or would not sense. This applies to all distributions.
perform as previously specified. Failure Rate—The total number of failures
Failure Analysis—Subsequent to a failure, the within an item population divided by the total
logical systematic examination of an item and its number of life units expended by that population,
construction, application, and documentation to during a particular measurement interval under
identify the failure mode and to determine the stated conditions.
failure mechanism and its basic course. Fault—Immediate cause of failure (e.g., malad-
Failure Catastrophic—A failure that can justment, misalignment, defect).
cause item loss. Fault Isolation—The process of determining
Failure Critical—A failure, or combination of the location of a fault to the extent necessary to
failures, that prevents an item from performing a effect a repair.
specified mission. Fault Localization—The process of deter-
Failure Dependent—Failure that is caused mining the approximate location of a fault.
by the failure of an associated item(s); not Failure Inherent R&M Value—A measure of RM) that
Independent. includes only the effects of an item design and its
Failure Effect—The consequence(s) a failure application and assumes an ideal operation and
mode has on the operation, function, or status of an support environment.
item. Failure effects are classified as a local effect, Interchange—Removing the item that is to
next higher level, and end effect. be replaced and installing the replacement item.
Failure Independent—Failure that occurs Inventory, Active—The group of items
without being caused by the failure of any other assigned to an operational status.
item; not Failure Dependent.
Inventory, Inactive—The group of items being
Failure Intermittent—Failure for a limited held in reserve for possible future assignments to
period of time, followed by the item’s recovery of its an operational status.
Item—A nonspecific term used to denote any Maintenance Ratio—A measure of the total
product, including systems, materials parts, maintenance manpower burden required to
subassemblies, sets, and accessories. (Source: maintain an item. It is expressed as the cumulative
MIL-STD-280.) number of man-hours of maintenance expended in
Life Profile—A time-phased description of the direct labor during a given period of the life units
events and environments an item experiences from divided by the cumulative number of end-item life
manufacture to final expenditures of removal from units during the same period.
the operational inventory, to include one or more Maintenance, Scheduled—Preventive main-
mission profiles. tenance performed at prescribed points during the
Life Units—A measure of use duration applicable life of an item.
to the item (e.g., operating hours, cycles, distance, Maintenance Time—An element of downtime
rounds fired, attempts to operate). that excludes modification and delay time.
Maintainability—The measure of the ability of Maintenance, Unscheduled—Corrective
an item to be retained in or restored to a speci- maintenance required by item conditions.
fied condition when maintenance is performed
by personnel having specified skill levels, using Malfunction—The event, or inoperable state,
prescribed procedures and resources, at each in which any item or part of an item does not, or
prescribed level of maintenance and repair. would not, perform as previously specified; the
same as Failure.
Maintainability, Mission—The measure of
the ability of an item to be retained in or restored Mean Maintenance Time—The measure
to a specified condition when maintenance is of item maintainability, taking into account the
performed during the course of a specified mission maintenance policy. The sum of preventive and
profile (the mission-related system maintainability corrective maintenance times divided by the sum
parameter). of scheduled and unscheduled maintenance events
during a stated period of time.
Maintenance—All actions necessary for
retaining an item in or restoring it to a specified Mean Time between Demands (MTBD)—
condition. A measure of the system reliability parameter
related to the demand for logistic support. The
Maintenance Action—An element of a
total number of system life units divided by the
maintenance event. One or more tasks (i.e., fault
total number of item demands on the supply
localization, fault isolation, servicing, inspection)
system during a stated period of time, such as shop
necessary to retain an item in or restore it to a
replaceable unit (SRU), weapon replaceable unit
specified condition.
(WRU), line replacement unit (LRU), and shop
Maintenance, Corrective—All actions replaceable assembly (SRA).
performed as a result of failure to restore an item
to a specified condition. Corrective maintenance Mean Time between Downing Events
can include any or all of the following steps: local- (MTBDE)—A measure of the system reliability
ization, isolation, disassembly, interchange, reas- parameter related to availability and readiness.
sembly, alignment, and checkout. The total number of system life units divided by
the total number of events in which the system
Maintenance Event—One or more mainte- becomes unavailable to initiate its mission(s)
nance actions required to effect corrective and during a stated period of time.
preventive maintenance due to any type of failure
or malfunction, false alarm, or scheduled main- Mean Time between Failures (MTBF)—A
tenance plan. basic measure of reliability for repairable items. The
mean number of life units during which all parts
Maintenance Manning Level—The total
of the item perform within their specified limits
authorized or assigned personnel per system at
during a particular measurement interval under
specified levels of organizational maintenance.
stated conditions.
Maintenance, Preventive—All actions
Mean Time between Maintenance
performed in an attempt to retain an item in a
(MTBM)—A measure of reliability, taking into
specified condition by providing systematic
account maintenance policy. The total number of
inspection, detection, and prevention of incipient
life units expended by a given time divided by the
failures.
total number of maintenance events (scheduled Mission Time between Critical Failures
and unscheduled) due to that item. (MTBCF)—A measure of mission reliability. The
total amount of mission time divided by the total
Mean Time between Maintenance
number of critical failures during a stated series
Actions (MTBMA)—A measure of the system
of missions.
reliability parameter related to the demand for
maintenance manpower. The total number of Mission Time to Restore Functions
system life units divided by the total number of (MTTRF)—A measure of mission maintain-
maintenance actions (preventive and corrective) ability. The total corrective critical failure main-
during a stated period of time. tenance time divided by the total number of critical
failures during the course of a specified mission
Mean Time between Removals (MTBR)—
profile.
A measure of the system reliability parameter
related to the demand for logistic support. The Not Operating (Dormant)—The state
total number of system life units divided by the wherein an item is able to function but is not
total number of items removed from that system required to function; not to be confused with Time,
during a stated period of time. This term is Down (Downtime).
defined to exclude removals performed to facili- Operable—The state of being able to perform
tate other maintenance and removals for product the intended function.
improvement.
Operational Readiness—The ability of a
Mean Time to Failure (MTTF)—A basic military unit to respond to its operation plan(s)
measure of reliability for non-repairable items. The upon receipt of an operations order (function of
total number of life units of an item divided by the assigned strength, item availability, status, supply,
total number of failures within that population training, etc.).
during a particular measurement interval under
Operational R&M Value—A measure of R&M
stated conditions.
that includes the combined effects of item design,
Mean Time to Repair (MTTR)—A basic installation, quality, environment, operation,
measure of maintainability. The sum of corrective maintenance, and repair.
maintenance times at any specific level of repair Predicted—That which is expected at some
divided by the total number of failures within future time, postulated on analysis of past experi-
an item repaired at that level during a particular ence and tests.
interval under stated conditions.
Reassembly—Assembling the items that were
M e a n T i m e t o R e s t o r e Sys t e m removed during disassembly and closing the reas-
(MTTRS)—A measure of the system maintain- sembled items.
ability parameter related to availability and readi-
ness. The total corrective maintenance time, asso- Redundancy—The existence of more than one
ciated with downing events, divided by the total means for accomplishing a given function. Each
number of downing events during a stated period means of accomplishing the function need not
of time. Excludes time for off-system maintenance necessarily be identical.
and repair of detached components. Redundancy, Active—That redundancy
wherein all redundant items are operating
Mean Time to Service (MTTS)—A measure
simultaneously.
of an on-system maintainability characteristic
related to servicing that is calculated by dividing Redundancy, Standby—That redundancy
the total scheduled crew, operator, and driver wherein the alternative means of performing the
servicing time by the number of times the item function is not operating until it is activated upon
was serviced. failure of the primary means of performing the
function.
Mission Profile—A time-phased description
of the events and environments an item experi- Reliability—(a) The duration or probability of
ences from initiation to completion of a specified failure-free performance under stated conditions
mission, to include the criteria of mission success or (b) the probability that an item can perform its
or critical failures. intended function for a specified interval under
stated conditions. [For nonredundant items, this system (e.g., data processing subsystem, guidance
is equivalent to definition (a). For redundant items, subsystem). (Source: MIL-STD-280.)
this is equivalent to the definition of mission System, General—A composite of equipment
reliability.] and skills, and techniques capable of performing or
Reliability Growth—The improvement in supporting an operational role, or both. A complete
a reliability parameter caused by the successful system includes all equipment, related facili-
correction of deficiencies in item design or ties, material, software, services, and personnel
manufacture. required for its operation and support to the
degree that it can be considered self-sufficient in
Reliability Mission—The ability of an item to
its intended operational environment.
perform its required functions for the duration of
a specified mission profile. System R&M Parameter—A measure of R&M
in which the units of measurement are directly
Repairable Item—An item that can be restored
related to operational readiness, mission success,
by corrective maintenance to perform all of its
maintenance manpower cost, or logistic support
required functions.
cost.
R&M Accounting—That set of mathematical
Test, Acceptance—A test conducted under
tasks that establish and allocate quantitative R&M
specified conditions by, or on behalf of, the govern-
requirements and predict and measure quantitative
ment, using delivered or deliverable items, to
R&M requirements.
determine the compliance of the item with speci-
R&M Engineering—That set of design, devel- fied requirements (including acceptance of first
opment, and manufacturing tasks by which R&M production units).
are achieved.
Test Measurement and Diagnostic
Screening—A process for inspecting items to Equipment (TMDE)—Any system or device
remove those that are unsatisfactory or those likely used to evaluate the condition of an item to identify
to exhibit early failure. Inspection includes the or isolate any actual or potential failures.
visual examination, physical dimension measure-
Test, Qualification (Design Approval)—
ment, and functional performance measurement
A test conducted under specified conditions, by or
under specified environmental conditions.
on behalf of the government, using items represen-
Servicing—The performance of any action tative of the production configuration, in order to
needed to keep an item in operating condition determine compliance with item design require-
(i.e., lubricating, fueling, oiling, cleaning), but ments as a basis for production approval; also
not including preventive maintenance of parts or known as demonstration.
corrective maintenance tasks.
Testing Development (Growth)—A series
Single-Point Failure—The failure of an item of tests conducted to disclose deficiencies and to
that would result in failure of the system and is verify that corrective actions will prevent recur-
not compensated for by redundancy or alternative rence in the operational inventory. Note that
operational procedure. repair of test items does not constitute correction
Sneak Circuit Analysis—A procedure of deficiencies; also known as test-analyze-and-fix
conducted to identify latent paths that cause the (TAAF) testing.
occurrence of unwanted functions or inhibit Time—The universal measure of duration.
desired functions, assuming all components are The general word “time” will be modified by
functioning properly. an additional term when used in reference to
Storage Life (Shelf Life)—The length of time operating time, mission time, test time, and so
an item can be stored under specified conditions forth. In general expressions such as “mean time
and still meet specified requirements. between failure (MTBF),” time stands for “life
units,” which must be more specifically defined
Subsystem—A combination of sets, groups, whenever the general term refers to a particular
and so forth that perform an operational function item.
within a system and are a major subdivision of the See Figure C.1 for time relationships.
FIGURE C.1 Time relationships.
Time, Active—That time during which an item Time, Reaction—That element of uptime
is in an operational inventory. needed to initiate a mission measured from the
Time, Administrative—That element of delay time that the command is received.
time, not included in the supply. Time, Supply Delay—That element of delay
Time, Alert—That element of uptime during time during which a needed replacement item is
which an item is assumed to be in specified oper- being obtained.
ating condition and is awaiting a command to Time, Turnaround—That element of mainte-
perform its intended mission. nance time needed to replenish consumables and
Time, Checkout—That element of mainte- check out an item for recommitment.
nance time during which performance of an item Time, Up (Uptime)—That element of active
is verified to be a specified condition. time during which an item is in condition to
Time, Delay—That element of downtime during perform its required functions.
which no maintenance is being accomplished on the Uptime Ratio—A composite measure of opera-
item because of either supply or administrative delay. tional availability and dependability that includes
Time, Down (Downtime)—That element of the combined effects of item design, installation,
active time during which an item is not in a condi- quality, environment, operation, maintenance,
tion to perform its required function (reduces repair, and logistic support; the quotient of uptime
Availability and Dependability). divided by uptime plus downtime.
Time, Inactive—That time during which an Useful Life—The number of life units from
item is in reserve. manufacture to when the item has an unrepair-
able failure or unacceptable failure rate.
Time, Mission—That element of uptime
required to perform a stated mission profile. Utilization Rate—The planned or actual
number of life units expended, or missions
Time, Modification—The time necessary to attempted, during a stated interval of calendar time.
introduce any specific change(s) to an item to
improve its characteristics or to add new ones. Wear Out—The process that results in an
increase of the failure rate or probability of failure
Time, Not Operating—That element of with the increasing number of life units.
uptime during which an item is not required to
operate.
Appendix D: List
of Acronyms
The following acronyms are used throughout this book:
AMSAA - Army Materiel Systems Analysis Activity

DMMH/MA - Direct maintenance man-hours per maintenance action
DMMH/ME - Direct maintenance man-hours per maintenance event
EMI - Electromagnetic interference
ESD - Electrostatic discharge
ESS - Environmental stress screening
ESSEH - Environmental Stress Screening of Electronic Hardware
FMEA - Failure modes and effects analysis
FRACAS - Failure rate analysis and corrective action system
GPA - Grade point average
GRMS - g-force root mean square
HALT - Highly accelerated life testing
HASS - Highly accelerated stress screen
HP - Hewlett-Packard
IC - Integrated circuit
IES - Institute of Environmental Sciences
I/O - Input and output
LRU - Line replacement unit
MTBCF - Mission time between critical failures
MTBD - Mean time between demands
MTBDE - Mean time between downing events
MTBF - Mean time between failures
MTBM - Mean time between maintenance
MTBMA - Mean time between maintenance actions
MTBR - Mean time between removals
MTTF - Mean time to failure

226 Appendix D: List of Acronyms
MTTR - Mean time to repair

MTTRF - Mission time to restore functions
MTTRS - Mean time to restore system
MTTS - Mean time to service
OC - Operating characteristic (curve)
PSD - Power spectral density
RADC - Rome Air Development Center
RIW - Reliability improvement warranty (contract)
R&M - Reliability and maintainability
ROI - Return on investment
SPC - Statistical process control (program)
SRA - Shop replaceable assembly
SRU - Shop replaceable unit
TAAF - Test-analyze-and-fix (testing)
TMDE - Test measurement and diagnostic equipment
WRU - Weapon replaceable unit
Bibliography
Accelerated Testing Research: Special Publications (Warrendale, PA: Society of Automotive

1.
Engineers, 2000)
2. Chan, H.A., Accelerated Stress Testing Handbook: Guide for Achieving Quality Products
(Piscataway, NJ: IEEE Press, Institute of Electrical and Electronics Engineers, 2001)
3. Dhillon, B.S., Quality Control, Reliability, and Engineering Design (New York: Marcel Dekker,
1985)
4. Dodson, B.L. and Mulcahy, M.D., Certified Reliability Engineer Examination Study Guide
(Tucson, AZ: Quality Publishing, 1992)
5. Dodson, B., Weibull Analysis (with Software) (Milwaukee, WI: ASQC Quality Press, 1994)
6. Dodson, B.L., Hammett, P.C., and Klerx, R., Probabilistic Design for Optimization and
Robustness for Engineers (London: John Wiley & Sons, 2014)
7. Dovich, R.A., Reliability Statistics (Milwaukee, WI: ASQC Quality Press, 1990)
Environmental Stress Screening Handbook (Holland, MI: Thermotron Ltd., 2004)

8.
Fundamentals of Accelerated Stress Testing (Holland, MI: Thermotron Ltd., 2004)

9.
10. Hobbs, G.K., Accelerated Reliability Engineering: HALT and HASS (New York: John Wiley &
Sons, 2000)
11. Ireson, G.W. and Coombs, C.F., Handbook of Reliability Engineering and Management (New
York: McGraw-Hill, 1996)
12. Kapur, K.C. and Lamberson, L.R., Reliability in Engineering Design (New York: John Wiley &
Sons, 1977)
13. Kielpinski, T.J. and Nelson, W., “Optimum Censored Accelerated Life Tests for Normal and
Lognormal Life Distributions,” IEEE Transactions on Reliability, R-24 (Piscataway, NJ: Institute
of Electrical and Electronics Engineers, 1975), 310-320.
14. Krishnamoorthi, K.S., Reliability Methods for Engineers (Milwaukee, WI: ASQ Quality Press,
1992)
15. Lall, P., Pecht, M.G., and Hakim, E.B., Influence of Temperature on Microelectronics (Boca
Raton, FL: CRC Press, 1997)
16. Lewis, E.E., Introduction to Reliability Engineering (New York: John Wiley & Sons, 1995)
17. Meeker, W. and Nelson, W., “Optimum Accelerated Life Tests for Weibull and Extreme Value
Distributions,” IEEE Transactions on Reliability, R-25 (Piscataway, NJ: Institute of Electrical
and Electronics Engineers, 1976), 20-24.

228 Bibliography
18. Meeker, W.Q. and Hahn, G.J., How to Plan an Accelerated Life Test—Some Practical Guidelines
(Milwaukee, WI: ASQ Quality Press, 1985)
19. Nelson, W., Accelerated Testing: Statistical Models, Test Plans and Data Analysis (New York:
John Wiley & Sons, 1990)
20. NIST/SEMATECH, e-Handbook of Statistical Methods, 2013, http://www.itl.nist.gov/div898/

handbook/.
21. O’Connor, P.D.T., Practical Reliability Engineering (New York: John Wiley & Sons, 2002)
22. Shooman, M.L., Probabilistic Reliability: An Engineering Approach (Malabar, FL: Robert E.
Krieger, 1990)
23. Staudte, R.G. and Sheather, S.J., Robust Estimation and Testing (New York: John Wiley & Sons,
1990)
24. Tobias, P.A. and Trindade, D.C., Applied Reliability (New York: Van Nostrand Reinhold, 1995)
25. U.S. Army Materiel Systems Analysis Activity, Reliability Growth Handbook (Aberdeen
Proving Ground, MD: U.S. Army Materiel Systems Analysis Activity, 1999)
26. U.S. Department of Defense, “MIL-HDBK-189: Reliability Growth Management,” Naval

Publications and Forms Center, Philadelphia. 1981.
27. U.S. Department of Defense, “MIL-HDBK-781: Reliability Test Methods, Plans and
Environments for Engineering Development, Qualification and Production,” Naval
Publications and Forms Center, Philadelphia, 1996.
28. U.S. Department of Defense, “MIL-STD-810: Environmental Test Methods and Engineering
Guidelines,” Naval Publications and Forms Center, Philadelphia, 1995.
29. U.S. Department of Defense, “MIL-STD-1635: Reliability Growth Testing,” Naval Publications
and Forms Center, Philadelphia, 1978.
30. Walker, N.E., The Design Analysis Handbook (Boston, MA: Newnes, 1998)
31. Young, W.R., “Accelerated Temperature Pharmaceutical Product Stability Determinations,”

Drug Development and Industrial Pharmacy 16, no. 4 (1998): 551-569.
About the Authors
Bryan Dodson is Executive Engineer Group Quality

at SKF. Prior to joining SKF, Dr. Dodson was Director
of forward model Quality and Reliability Engineering
at Visteon in Dearborn, Michigan. He formerly was
Manager of Reliability Engineering and Quality
Information Systems for Continental Teves, Brake
and Chassis, North America. Prior to joining
Continental Teves, Dr. Dodson held the positions of
Total Quality Management Leader and Reliability
Leader at Alcoa and the position of Industrial
Engineer at Morton Thiokol.
Dr. Dodson has authored several books and
commercial software programs, including , including
Probabilistic Design for Optimization and Robustness
for Engineers the Reliability Engineering Handbook, and Weibull Analysis: with
Software, and Certified Reliability Engineer Examination Study Guide.
Dr. Dodson holds a BS in petroleum engineering, an MS in industrial engi-
neering, an MBA, and a Doctor of Business Administration. In addition, he is a
Certified Quality Engineer, a Certified Reliability Engineer, a Six Sigma Master Black
Belt, and a licensed Professional Engineer in Michigan and Indiana.

230 About the Authors
Harry Schwab has almost forty years of experience

as an engineer, specializing in structural dynamics
and vibrations. He holds a Bachelor of Aerospace
Engineering from the Georgia Institute of
Technology, an MS in Engineering Mechanics from
the University of Missouri–Rolla, and an MBA from
the Florida Institute of Technology.
Mr. Schwab has worked in the aerospace, auto-
motive, and commercial industries. The products
with which he has been involved include automo-
biles, trucks, railroad freight cars, aircraft, boats,
and guided missiles. He has many years of experi-
ence in working with random vibrations and devel-
oped a method for deriving laboratory testing environments from nonstationary
data. This technique can be used to accurately simulate real-world environments
while greatly decreasing test time.
Mr. Schwab has authored several technical papers for SAE International and the
Institute of Environmental Sciences and Technology (IEST), as well as for other
technical organizations and publications. These cover the fields of vibration testing
and analysis. He developed innovative testing techniques for both the aerospace and
automotive industries, and he is a registered Professional Engineer in Missouri
and Florida.
Index
A motor component Analogic Corporation, 182

Accelerated testing, 1, 16 tolerances, 7 Army Materiel Systems
Arrhenius model, 138–145 number of tolerance Analysis Activity
degradation testing, 157–162 combinations, 6 (AMSAA), 128
design life, 2 simple system, 5, 6 growth plot computations,
brake force application, tolerance stackup, 7, 8 129
distribution of, 3 worst-case tolerance reliability growth plot, 130
durability test, 5 combination, 7 Arrhenius constant, 145
minivan brake voltage models, 155 Arrhenius equation, 138
applications, Accelerated test plans Arrhenius model, 138, 139
percentiles for, 4 mean time to fail, 97–98 linear interpolation, 143
development tests, 1 probability of successful median rank
equivalent life test data, completion, 99 computations, 143
136–137 randomization of load modeling reliability, 142
Eyring model, 145–154 cycles, 122, 124 nonparametric Arrhenius
financial considerations, 8 reliability growth, 124 model regression, 144
engineering, response AMSAA model, 128–130 regression, 141
from, 12 Duane model, 126–128 sample data, 140
finance, response from, 11 feedback model, 126 Axial loading, 144
manufacturing, response management, 125
from, 12–16 process, 125, 126 B
reliability program, 9–11 type I error Bayesian testing
linear acceleration, 131–132 drive test, sample size, determining test
mechanical crack growth, 101, 102 parameters, 106
156–157 as function of sample guidelines for zero-failure
qualitative tests size, 101, 102 test planning, 108–111
elephant tests, 163–164 single drive test, 100 sequential testing, 111
HALT and HASS, 164 Weibull reliability shape parameter
step-stress testing, demonstration test effect of, 103–105
162–163 plans, 98 estimating, 106
quality assurance tests, 1, 2 zero-failure testing Bayes’ theorem, 64–67
reliability demonstration Bayesian testing, 103–106, Bendix Corporation, 183
sample sizes, 2 108–111 Binomial cumulative
scatterplot of field miles vs. bogey testing, 103 distribution function, 53
laboratory time to exponential distribution, Binomial distribution, 23, 56
fail, 137 116–117 binomial cumulative
statistical sample size pass-fail testing, 111–116 distribution
determination, 5 Weibull distribution, function, 53
tolerances, 5 118–122 binomial probability density
best-case tolerance Acceleration factor, 135, 139 function, 53
combination, 7 Acceptance tests, 192 mean, 54
231
232 Index
sequential testing, 113 D E

variance, 54 Defect exposure tests, 192 Electrical stress, 177, 202
Binomial probability density Degradation testing Electrodynamics, 200
function, 53 features, 157 Electromagnetic interference
Bogey testing, 103 linear degradation, 157, 158 (EMI), 191
Burn-in optimization, 187–189 prediction interval, 159 Environmental stress screening
switch degradation data, (ESS), 167
C
160, 161 Analogic Corporation, 182
Censoring, 18
linear model, 161 Bendix Corporation, 183
left, 18
logarithmic model, burn-in optimization,
multiple, 18
161, 162 187–189
right, 18
Discrete modeling combined environment, 178
type I, 18
distributions electrical stress, 177
type II, 18
binomial distribution, 53–54 equipment, 186–187
Central limit theorem, 38
correct discrete ESS failure vs. production
Chi-square distribution, 50, 51
distribution, 58 level, 185
chi-square probability
geometric distribution, 57 ESS program, implementing,
density function, 59, 60
hypergeometric distribution, 183–186, 190
random variable, 59
54–56 evolution of, 170–172
Chi-square probability density
Poisson distribution, 51–52 failure analysis, 182
function, 59, 60
Discrete variables, 19 Hewlett-Packard (HP), 183
Chi-square significance,
Distributions high-temperature
208–209
Bayes’ theorem, 64–67 burn-in, 177
Coffin-Manson model, 156
continuous modeling implementation, 170
Confidence interval, 50, 72–74,
distributions, 31 levels of screen
82, 83, 89, 90, 96,
exponential distribution, complexity, 181
133, 159
47–50 low-temperature screen, 178
Continuous modeling
lognormal distribution, misconceptions, 172–173
distributions, 31
43–47 product reliability equation,
exponential distribution,
normal distribution, 168–170
47–50
37–43 random vibration, 176–177
lognormal distribution,
Weibull distribution, sine vibration, fixed
43–47
32–37 frequency, 178
normal distribution, 37–43
discrete modeling sine vibration, swept
Weibull distribution, 32–37
distributions frequency, 178
Continuous variables, 18
binomial distribution, temperature cycling,
Cumulative distribution
53–54 173–176, 178–180
function, 24–26, 29, 35,
correct discrete thermal shock, 177–178
84, 93, 153
distribution, 58 Estimated Weibull
binomial, 53
geometric distribution, 57 parameters, 144
exponential, 76
hypergeometric Exponential cumulative
F-cumulative distribution
distribution, 54–56 distribution
function, 63
Poisson distribution, function, 76
hypergeometric, 55
51–52 Exponential cumulative hazard
Poisson cumulative
sampling distributions function, 74, 75
distribution function,
chi-square distribution, Exponential distribution, 47,
52, 215
59–60 116–117
Weibull cumulative
F-distribution, 62–63 exponential hazard
distribution
t-distribution, 60–61 function, 48, 49
function, 92
Distribution variance, 28 exponential probability
Cumulative hazard
Duane model, 126 density function, 47, 48
function, 70
computations, 128 hazard plotting, 74–76
exponential, 74, 75
failure data, 127 maximum likelihood
normal, 83
growth plot, 128 estimation, 71
Weibull cumulative hazard
parameters, 127 mean, 49
function, 90
Index 233
probability plotting, 76–79 normal operating Lognormal hazard function,

variance, 47 conditions, 132 46, 47
Exponential hazard function, Weibull hazard function, Lognormal probability density
48, 49 33, 34 function, 44, 45, 87
Exponential probability density Hazard plotting, 70–71 Low-temperature steady-state
function, 47, exponential distribution, testing, 196
48, 71 74–76 M
Exponential sequential normal distribution, 83–85 Maximum likelihood
test, 118 Weibull distribution, 90–92 estimation, 70
Eyring model, 145, 146 Hewlett-Packard (HP), 183 exponential distribution, 71
absolute temperature, 148 Highly accelerated life testing normal distribution, 80–83
analysis tool pack, 149, 150 (HALT), 10, 164 Weibull distribution, 88–90
data, 147 Highly accelerated stress screen Mean time to fail, 97–98
degrees of freedom, 151 (HASS), 164 Mechanical crack growth,
general form, 146 High-temperature burn-in, 177 156–157
linear form, 146 High-temperature steady-state Mechanical vibration, 199
linear interpolation, 153, 154 testing, 197 Multiple-censored data, 18,
median rank computations, Hydraulics, 200–201 19, 73
153, 154 Hypergeometric cumulative Multiple linear regression, 148,
multiple linear regression, distribution 150, 153, 156
148, 154 function, 55
regression results, 150 Hypergeometric N
residuals analysis, 152 distribution, 54 Nonparametric Arrhenius
simplified, 146 hypergeometric cumulative model regression, 144
distribution Nonparametric confidence
F intervals, 96
function, 55
Failure modes and effects
hypergeometric probability Normal cumulative hazard
analysis (FMEA), 9
density function, 55 function, 83
F-cumulative distribution
mean, 55 Normal distribution, 37, 79–80
function, 63
variance, 55 distribution of average
F-distribution, 62
Hypergeometric probability voltage, 41, 42
critical values, 63
density function, 55 hazard plotting, 83–85
F-probability density maximum likelihood
function, 63 K estimation, 80–83
Fluid transportation Kaplan-Meier product limit normal hazard function, 43
system, 123 estimator, 70, 85, 93 normal probability density
F-probability density Kurtosis, 27 function, 38, 39
function, 63 probability of strength, 40
L probability plotting, 84–87
G
Least squares regression, 158
Geometric distribution Normal hazard function, 43
Linear acceleration, 131–132
geometric probability O
Log-likelihood equation, 88
density function, 57 Operating characteristic curve
Lognormal data, 44
mean, 57 (OC curve), 114
Lognormal distribution, 43, 87
variance, 57 computations, 115
location parameter, 46
Geometric probability density sequential sampling, 115
lognormal hazard function,
function, 57
46, 47 P
Government documents,
lognormal probability Parameter estimation
217–218
density function, consistency, 69
Grade point average (GPA),
44, 45 exponential distribution, 71,
12, 13
mean of transformed 74–79
H data, 45 hazard plotting, 70, 71
Hazard function, 26, 29 random variables, 44 lack of bias, 69
exponential, 48, 49 transformation of lognormal lognormal distribution, 87
lognormal, 46, 47 data, 44 maximum likelihood
normal, 43 variance, 46 estimation, 70
234 Index
minimum variance, 69 Q Skewness, 27, 28

nonparametric confidence Qualification tests, 192, 203 Standard normal cumulative
intervals, 96 distribution function,
normal distribution, 79–87 R 44, 46
probability plotting, 70 Random vibration, 176–177 Standard normal probability
sufficiency, 69 Rayleigh distribution, 32 density function, 38
Weibull distribution, 87–95 Reliability, 46 Statistical inferences, 18
Pass-fail testing, 111–116 confidence interval, 74, 82, 89 Statistical methods, 17
Pneumatics, 199 confidence limit, 74, 107 Statistical process control
Poisson cumulative Reliability function, 25–26, 29 (SPC), 7, 10, 11
distribution acceleration factor, 132 Statistical tables
function, 215 exponential, 50 chi-square significance,
Poisson distribution, 47 normal operating 208–209
Poisson cumulative conditions, 131 F significance, 210–213
distribution posterior, 64–66 gamma function, 205
function, 52 prior, 64, 65 standard normal cumulative
Poisson probability density Reliability growth, 124 distribution function,
function, 51, 52 AMSAA model, 128–130 206–207
Poisson probability density Duane model, 126–128 t significance, 214
function, 51 feedback model, 126 Steady-state testing
Poisson process, 126, 128 management, 125 high-temperature, 197
Probability density function, process, 125, 126 low-temperature, 196
19, 24–26, 29, 41, 70 Reliability program Step-stress testing, 162–163
binomial, 53 concept validation, 10 Stimulation
chi-square, 59, 60 customer’s requirements, 9 application, 195
continuous distributions, 21 design validation, 10 typical test types, 196
definition, 23 execution score vs. failure Stresses
discrete distributions, 21 costs, 13 electrical stress, 202
exponential, 47, 48, 71 feedback, 9 temperature
F-probability density intelligent design, 10 steady-state, 196–197
function, 63 manufacturing, 10–11 temperature cycling, 197
geometric, 57 poor reliability program, thermal shock, 197–198
hypergeometric, 55 13, 14 vibration, 198
length, 21–23 scores, 12 analysis tools, 198–199
lognormal, 44, 45, 87 vehicle resale value, 14, 15 cyclic testing, 199–200
normal, 38 Reliability rating electrodynamic, 200
overlaid model, 20 automobile resale value vs. hydraulic, 200–201
Poisson probability density reliability rating, 15 loose cargo testing, 199
function, 51, 52 retail value, 15 mechanical and
posterior, 64–66 pneumatic, 199
prior, 64, 65 S mechanical shock, 201
t-probability density Sampling distributions T
function, 60, 61 chi-square distribution, t-distribution
valid, 23 59–60 mean, 61
Weibull probability density F-distribution, 62–63 t-probability density
function, 32, 87 t-distribution, 60–61 function, 60, 61
Probability of strength, 39, 40 Second moment-generating variance, 61
Probability plotting, 70 function, 27 Temperature
exponential distribution, Sequential test, 111–113, 116, steady-state, 196–197
76–79 118, 121 temperature cycling, 197
normal distribution, Simulation thermal shock, 197–198
84–87 application, 193 Temperature profile, 173, 174
Weibull distribution, 92–95 real-world environments, air velocity, 176
Product reliability equation, 193, 194 chamber and product, 174
168–170 typical test types, 193
Index 235
functional test, temperature W Weibull probability plot

extremes, 175 Weibull cumulative axial load data, 140
near product destruct distribution computations, 133, 134
limits, 175 function, 92 pressure test data, 135
Tensile strength, 39, 40 Weibull cumulative hazard salt concentration data, 147
Test equipment methods and function, 90 Weibull reliability, 35
applications Weibull distribution, 37, 87, demonstration test plans, 98
simulation vs. stimulation, 104, 106, 118, 119, 122, function, 188
193–196 133, 141, 144, 148, 152 Weibull time-to-fail
stresses hazard function, 33, 34 distribution, 98, 99,
electrical stress, 202 hazard plotting, 90–92 104, 105, 107, 108
temperature, 196–198 location parameter, 33, 34 Weibull variance, 35, 36
vibration, 198–201 maximum likelihood Z
Thermal shock, 177–178, 197–198 estimation, 88–90 Zero-failure testing
t-probability density function, mean, 35, 36 Bayesian testing
60, 61 probability plotting, determining test
Type I censoring, 18 92–95 parameters, 106
Type II censoring, 18 scale parameter, 33, 120, 135 guidelines for zero-failure
shape parameter, 32, 120
V test planning, 108–111
variance, 35, 36
Vibration shape parameter, effect of,
Weibull probability density
analysis tools, 198–199 103–105
function, 32
cyclic testing, 199–200 shape parameter,
Weibull sequential
electrodynamic, 200 estimating, 106
testing, 121
hydraulic, 200–201 bogey testing, 103
Weibull maximum likelihood
loose cargo testing, 199 exponential distribution,
parameter estimation,
mechanical and 116–117
141, 148
pneumatic, 199 pass-fail testing, 111–116
Weibull probability density
mechanical shock, 201 Weibull distribution,
function, 32, 87
Voltage models, 155 118–122
A Practitioner’s Guide to
Accelerated and Reliability Testing
2nd Edition
Bryan L. Dodson | Harry Schwab
Cover image used under license from Shutterstock.com.

The scientific theory of accelerated testing is highly developed, but the application
of this theory has proven difficult, especially in the mobility industries. The required
design life for many components exceeds
10 years, and the application environment Combines theory of
is harsh and highly variable. Despite the
statistical testing while
difficulty, the application of accelerated
testing theory can result in considerable
discussing the practical
time and cost savings, as well as increasing limitations of simulating
a product’s useful life. By presenting decades of real-world
accelerated testing from a practical use in a laboratory.
viewpoint, this book is both a valuable
text for practitioner and ideal text for a Amazon Reviewer for
college course. In Accelerated Testing: First Edition of Accelerated Testing
A Practitioner’s Guide to Accelerated
and Reliability Testing, readers are exposed to the latest, most practical knowledge
available in this dynamic and important discipline. Authors Bryan Dodson and Harry
Schwab draw on their considerable experience in the field to present comprehensive,
insightful views in this book.
Readers have access to a wealth of free resources including:
• Microsoft Excel templates: Various types of templates used for calculations in the
book are available for as a free download including templates for statistical tables,
burn-in optimization, and random number generator (Weibull, normal, lognormal,
and exponential distributions).
• Government Documents: PDF files for numerous military standards and
handbooks related to reliability and accelerated testing.
• AMSAA Reliability Growth Handbook
RELATED RESOURCES:
Bosch Automotive Handbook, 10th Edition

Author: Robert Bosch
ISBN: 978-0-7680-9578-8
ISBN: 978-1-4686-0350-7

BRYAN SCHWAB HARRY DODSON - ACCELERATED TESTING - A Practitioner's Guide To Accelerated and Reliability testing.-SAE SOC OF AUTOMOTIVE ENG (2021) PDF

Uploaded by

Copyright:

Available Formats

BRYAN SCHWAB HARRY DODSON - ACCELERATED TESTING - A Practitioner's Guide To Accelerated and Reliability testing.-SAE SOC OF AUTOMOTIVE ENG (2021) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

BRYAN SCHWAB HARRY DODSON - ACCELERATED TESTING - A Practitioner's Guide To Accelerated and Reliability testing.-SAE SOC OF AUTOMOTIVE ENG (2021) PDF

Uploaded by

Copyright:

Available Formats

Accelerated Testing:

Bryan L. Dodson | Harry Schwab

BRYAN L. DODSON AND HARRY SCHWAB

Warrendale, Pennsylvania, USA

To purchase bulk quantities, please contact: SAE Customer Service

Visit the SAE International Bookstore at books.sae.org

For access to free resources to accompany this book, please visit

©2021 SAE International v

Accelerated Test Plans 97

Accelerated Testing Models 131

Eyring Model 145

Environmental Stress Screening 167

Test Equipment Methods and Applications 191

©2021 SAE International xi

•• Random Number Generator—This Microsoft Excel template generates

©2021 SAE International xiii

Purpose of Accelerated Testing

©2021 SAE International 1

TABLE 1.1 Reliability demonstration sample sizes.

Reliability (%) Confidence (%) Sample size

FIGURE 1.1 Distribution of brake force application.

TABLE 1.2 Percentiles for minivan brake applications.

Percentile Number of brake applications

FIGURE 1.2 Percentiles for minivan brake applications.

Statistical Sample Size Determination

Tolerances in Accelerated Testing

FIGURE 1.3 Simple tolerancing example.

maximum diameter and the inner diameter of Component B at a minimum or

TABLE 1.3 Number of tolerance combinations.

Number of characteristics Number of tolerance combinations

TABLE 1.4 Motor component tolerances.

FIGURE 1.4 Tolerance stackup of 3 components with Cpk of 0.50.

FIGURE 1.5 Tolerance stackup of 3 components with Cpk of 0.83.

The following is a consolidated set of responses based on actual experiences.

Response from Finance

Response from Engineering

Response from Manufaturing

TABLE 1.5 Example reliability program scores.

Reliability program item Score (GPA)

TABLE 1.6 Example field reliability performance.

Reliability performance item Cost

FIGURE 1.6 Reliability program execution score versus failure costs.

FIGURE 1.7 Effect of poor reliability on company profits.

Similar to termites damaging a home without the owner’s knowledge, hidden

TABLE 1.7 Vehicle resale value.

Consumer reports reliability

FIGURE 1.8 Automobile resale value versus reliability rating.

•• Explore the impact on manufacturing. What is the impact of material changes,

1. Probability density function

©2021 SAE International 17

TABLE 2.1 Time to fail for light bulbs.

TABLE 2.2 Multiple-censored data.

112 172 220

Probability Density Function

FIGURE 2.1 Example histogram.

FIGURE 2.2 Example histogram with an overlaid model.

A histogram is an approximation of the shape of the distribution. The histo-

and f(x) ≥ 0 for continuous distributions. For discrete distributions,

FIGURE 2.3 Probability density function for the length of fish.