Accelerated Life Testing –

Second Edition
by James A. McLinn
Multiple Stress Level Test - Common Analysis

99.00
Arrh/Weib
Data 1
90.00
0.87
F=10 | S=0
0.99
F=10 | S=0
1.09
50.00 F=9 | S=0
1.18
F=10 | S=0
Unreliability

10.00

5.00

1.00
0.01 0.10 1.00 10.00 100.00
Time
Beta=1.2110, B=13.9768, C=1.7997E-6

Published by
The Reliability Division of ASQ - May 2010
ISBN 0-9701923-0-4
PRACTICAL ACCELERATED LIFE TESTING

James A. McLinn, CRE, Fellow ASQ

Library of Congress Cataloging-in-Publication Data

McLinn, James A.

Practical Approaches to Accelerated Life Testing, James A. McLinn.

p em.

Includes bibliographical references related readings references and index.

ISBN 0

1. Engineering-Management. 2. Testing-Management 3. Quality Engineering. I.

Title

HF5415.15R45 2010

@ 2010 by ASQ

All rights reserved. No part of this book may be reproduced in any form
or by any means, electronic, mechanical, photocopying, recording, or
otherwise, without prior written permission of the publisher

ISBN 0-9701923-0-4

Publication Manager and Editor: James A. McLinn Cover Design: James A. McLinn
Attention: Bookstores, Wholesalers, Schools and Corporations:
Reliability Division, ASQ books and Reliability Review Journal subscriptions are
available at quantity discounts with bulk purchases for business, educational, or
instructional use. Please contact the Reliability Division Publisher at 763-498-8814,
or write the Reliability Division Publications at, 10644 Ginseng Lane, Hanover,
Mn 55341. A copy of our Publication Order Form is available on the Reliability
Division web site. Visit http://www.ASQ-rd.org

Published in the United States of America

 Reliability Division ASQ
The Reliability Division of ASQ has sponsored a Monograph Series in response to requests for
articles on a variety of diverse topics of reliability. Copies of any of the RD Monographs may be
ordered from:

James McLinn - Monograph Publisher

10644 Ginseng Lane
Hanover, Minnesota 55341 USA

1 763 498-8814

JMReL2@Aol.com
 Practical Approaches for Accelerated Life
Testing
by James A. McLinn, CRE, Fellow ASQ
10644 Ginseng Lane, Hanover, Minnesota 55341
JMREL @ Aol.com

In Cooperation with the Reliability Division of ASQ

Accelerated Life Testing (ALT) is one of the most important topics for any
reliability engineer. Few reliability engineers know more than the basics of a few
techniques. Most test people do not know the wide range of testing applications
that do exist. ALT theory is often covered by journal articles showing a few
isolated applications, techniques or explanations to existing techniques. Creation
of new ALT techniques through development of theory are likewise scattered
through a number of journals.
This short book is an attempt to bring a variety of tools and techniques
together and present them as a coherent package. No attempt was made to cover
all the useful techniques of ALT. Hence, vibration fatigue applications, thermal
cycling, step-stress methods and HALT techniques are missing from this work.
However, the critical and often under emphasized steps of preparation of an ALT
and administration are included. Ground rules and guidelines which are
contained within will help the reader avoid the most common pitfalls. The last
two sections look at the analysis of some ALT life data. The combination should
aid the reader when performing most types of Accelerated Life Tests.
References are at the end of the work with a two-page supplemental
reading list. These go beyond the scope of this work.

i
Table of Contents
Section and Topic Page Number
1.0 A History and Background of ALT 1
1.1 A Short History of ALT Approaches 1
1.2 Applications of ALT Techniques 4
1.3 Reasons for Employing ALT Techniques 5
1.4 Outline of ALT Procedure 7

2.0 Planning for an ALT and Selecting the Type of Test 8

2.1 Selecting the ALT Method 8
2.2 Examples of Detailed Timing Diagrams 10
2.3 Plan for Success as well as Failure 12

3.0 Selecting and Setting Up Test Conditions and Parameters 14

3.1 Selecting the Test Equipment 14
3.2 Selecting the Test Conditions 15
3.3 Selecting the Applied Stress Test Methods 18
3.4 Selecting Sample Sizes for Meaningful Tests 19
3.5 Considering Other Reliability Goals and Measures 28
3.6 Selecting the Number of Test Failures 29
3.7 Selecting the Time to Failure Distribution 29
3.8 Other Distributions for Reliability 31

4.0 Administering the Life Test 33

4.1 Methods for Selecting Test Stresses 33
4.2 Military Standards 35
4.3 The Relationship of Field Failures and Product Development Activities 37
4.4 Effectiveness of Environmental Stress Screens 39
4.5 A Few Stress Application Examples 43
4.6 Suggested Rules for Performing an ALT 47

5.0 Models for Reliability Analysis of ALT 51

5.1 Degradation of the Systems under Test 51
5.2 Degradation Mechanisms 58
5.3 Creation of a Failure Model Versus Time or Stress 62
5.4 Humidity Testing Issues and Non-Linear Behavior 68

6.0 Selecting the Best Distribution and Performing the Data Analysis 71
6.1 Selecting the Best Distribution 71
6.2 The Data Analysis of ALT 73
6.3 Solutions to Accelerated Life Test of Capacitors 81

7.0 Degradation Test Methods 88

7.1 Background 88
7.2 The Concept of Degradation 89
7.3 Models of Degradation Testing 94
References 101

ii
Table of Figures
Description of the Figure Page Number

Figure 2.1 - Timing Diagram of Stress Tests 11

Figure 3.1 - Analysis of Electronics Module 26
Figure 4.1 – The Evolution of a Failure 37
Figure 4.2 - The Marble Bag Model of System Failures 38
Figure 4.3 - The Eight Dynamics of a Project 39
Figure 4.4 – Temperature Cycle 41
Figure 4.5 – Crack Growth in a Plane 45
Figure 5.1 - Simple Linear Degradation 52
Figure 5.2 - Complex Degradation Models 53
Figure 5.3 – Long Term Wear-In Model 54
Figure 5.4 – Long Term or S-N Curves 54
Figure 5.5 – Power Law Models 55
Figure 5.6 - Potential Well Aging Model 59
Figure 5.7 - The Aging of a Crystal 60
Figure 5.8 - The Crystal Aging Model 61
Figure 5.9 - Weibull Graph of Stress Versus Life 62
Figure 5.10 - Log Stress Versus Log Life 63
Figure 5.11 – Log Log Plot of Concentration 65
Figure 5.12 – Weibull Graph of RAM Accelerated Life Data 66
Figure 5.13 – PDF of RAM Accelerated Life Data 67
Figure 6.1 – Rank Regress Analysis of Interval Data 75
Figure 6.2 - MLE Analysis of Interval Data 76
Figure 6.3 – Data Set B MLE Analysis 77
Figure 6.4 – Data Set B Rank Regression Analysis 79
Figure 6.5 – All Capacitor Data with best Fit Slope 81
Figure 6.6 - Relationship of  with Temperature and Voltage 84
Figure 7.1 – The Distribution of a Measured Parameter 91
Figure 7.2 - A Change in Characteristics at 145ºC and 12V 95
Figure 7.3 – Basic Analysis of Oil Data 97
Figure 7.4 – Independent Analysis of Oil Data 98
Figure 7.5 – Three Parameter Analysis of Oil Data 99
Figure 7.6 – Log Stress Versus Log Life of Oil Data 100

iii
List of Tables
Table Number and Title Page Number

Table 2.1 - A Long List of Reasons to Perform ALT 8

Table 2.2 - A Short List of Methods 9
Table 2.3 – Types of Stress Tests 10
Table 2.4 - A List of Undesirable ALT Events or Possible Occurrences 12

Table 3.1 - The Trade-Off Between Sample Size and Test Length 28
Table 3.2 – The Moments of a Distribution 29

Table 4.1 - The Selection of Applied Stresses 33

Table 4.2 - The Value of Various Applied Stresses 33
Table 4.3A – The Value of Various Applied Stresses (US) 39
Table 4.3B – The Value of Various Applied Stresses (France) 40
Table 4.4 - The Critical  Factors for Vibration Test from Kececioglu 46
Table 4.5 – A List of Suggested Ground Rules 48

Table 5.1 - Common Degradation Events 55

Table 5.2 - Degradation Results of an ALT 57
Table 5.3 – Degradation mechanisms 58
Table 5.4 - The Details of the Stress to Failure Chain 58
Table 5.5 – Results of a Corrosion Test 64
Table 5.6 - Additional Data of Best Fit Power Law Models 65
Table 5.7 - RAM Memory 66
Table 5.8 - Shelf Life Test Results 68

Table 6.1 - Field Data 74

Table 6.2 – Capacitor ALT Data 76
Table 6.3 – Models for Reliability 78
Table 6.4 – Analysis of Life Data from Table 6.2 79
Table 6.5 - A Comparison of Data Sets D and E 80
Table 6.6 - Estimates of E a and N Based Upon Definition of Life 83
Table 6.7 - Table of Potential Interaction Models 84

Table 7.1 - Preliminary Test Results of Component 92

Table 7.2 - Test Results of Components 93
Table 7.3 - Life of Insulation Fluid in Minutes 96
Table 7.4 - MLE Weibull Analysis of Insulating Fluid 97

iv
Preface - Second Edition

This edition was updated as a response to the need for a more examples that are
clear, concise and engineering oriented explanations for critical topics of Accelerated
Life Testing. A number of reliability engineers had expressed frustration that the ALT
topics were covered by only two books published in the last 10 years. This book began as
a series of short articles published in the Reliability Review in the late 1990s that
addressed some simple examples and subjects. It has increased in size from the original
125 pages, yet some valuable ALT topics will always remain unaddressed.
In this work, as in the earlier edition, the statistics and complex math was kept to
a minimum. It is assumed that the reader has a basic knowledge of reliability and some
basic knowledge of how to set-up and run an ALT. It was not the intent to turn this work
into a statistical primer, so statistics are kept to a minimum. Canned software packages
are mentioned, as these are the tools that most people employ to solve their own
problems. Weibull tools and other web resources are discussed. New topics including
some that cover HALT, HASS and ESS are included in this edition. Readers are
encouraged to use this book as a learning tool. Be sure to show it to co-workers, and look
at the reading list at the end for further information. Many of these refer to RAMS
proceedings, Quality Engineering, The Journal of Quality Technology and other
conferences.

A number of people who have helped improve the original articles and the first
edition. Definitely the biggest help came from my wife, Connie, who reviewed this
monograph and the first edition. She made numerous suggestions and improvements on
both. With out her help, it would not have turned out so well.

James McLinn
Hanover, Minnesota
March 2010

v
Preface - First Edition

This monograph was created as a response to the need for a clear, concise and
short explanation for some of the topics of Accelerated Life Testing. A number of
reliability engineers had expressed frustration that the ALT topic was covered by only
one book. This began as a series of short articles published in the Reliability Review to
address some simple examples and subjects. It has doubled the size from the original six
articles, yet some of this valuable ALT topic must remain unaddressed.
In this work, the statistics and math were kept to a minimum. It is assumed that
the reader has a basic knowledge of reliability and some knowledge of ALT. It was not
the intent to turn this work into an ALT statistical primer, so statistics are only
occasionally mentioned. Software aids are mentioned, as these are the tools that most
people employ to solve their own problems. Readers are encouraged to use this short
book as a tool, to show it to co-workers, and move on to the reading list at the end for
further information.

I wish to thank a number of people who have helped improve the original articles
and encouraged me during the creation of this monograph. These include Hal Williams of
Reliability Review , William Stoner of Amway Corp., Norb Santoski, Reliability
Consultant, Valter Loll of Nokia, Copenhagen and Patrick O'Connor of Great Britain.

Perhaps my biggest help came from my wife, Connie, who reviewed this
monograph and made numerous suggestions and improvements. With out her, it would
not have turned out so well.

James McLinn
Hanover, Minnesota
April 2000

vi
 Practical Approaches for Accelerated Life Testing
1.0 The History and Background of ALT –

Modern reliability tools and techniques for Accelerated Life Tests (ALT) have been
developed primarily over the last 30 years. The modern concepts and tools actually reach all the
way back to the 1940s and 1950s when reliability was in its infancy. Many approaches were the
outgrowth of the needs of the U.S. military [23] to achieve and demonstrate long-lived, reliable
electronics equipment. In the 1940s, the expectation was that 25 to 50% of electronic equipment
would not work when called upon to do so. The worldwide use of hardware in a variety of
applications and environments represented a major challenge. Separate tools and techniques have
also been developed for mechanical applications and some are detailed in the Mechanical Design
Reliability Handbook [24]. The approaches for mechanical acceleration are usually somewhat
different from electronic applications because of the difficulty of accelerating materials,
mechanical components and assemblies. Software is also subject to accelerated testing, but
software topics will not be discussed in this book.

1.1 A Short History of ALT Approaches –

The original and current purposes of running ALT are typically for one of the following
reasons:
1) Demonstrate that a design was ready for release.
2) Show that a product would last for a minimum amount of time.
3) Verify that a design is robust.
4) Create an estimate of the warranty failures.
5) Demonstrate a minimum time to failure.
6) Identify some of the failure modes to be expected in the field.
7) Demonstrate the robustness of a design for operation in the customer's environment.

An ALT might be performed as a test at the end of manufacturing to show the customer
that a system would work well or exceed some minimum length of time in a harsh environment.
These early life tests were typically called "demonstration tests", "evaluation tests", "customer
acceptance tests", “verification tests”, “validation tests” or even "Wald sequential tests". These
tests were often run at nominal laboratory conditions or occasionally at the typical customer
conditions. Accelerated tests might be run at worst-case customer conditions or at some other
high stress level. The more adventuresome might use two or more stresses to test a design. Since
reliability was in its infancy during the 1950s, the development of accelerated tests was often
documented at that time by statistical journals such as Annals of Mathematical Statistics or the
Journal of the American Statistical Association. In 1955, a conference on electrical contacts and
connectors was started, emphasizing reliability physics and understanding failure mechanisms.
Other conferences began in the 1950s to focus on some of these important reliability topics. That
same year, the Reliability Analysis Design Center, RADC, issued “Reliability Factors for Ground
Electronic Equipment.” This was authored by Joseph Naresky. By 1956, ASQC was offering
papers on reliability as part of their American Quality Congress. The radio engineers, ASME,
ASTM and the Journal of Applied Statistics were contributing research papers. The Institute of
Radio Engineers, IRE, was already holding a conference and publishing proceedings titled
“Transaction on Reliability and Quality Control in Electronics”. This began in 1954 and
continued until this conference merged with an IEEE Reliability conference and became the
Reliability and Maintainability Symposium (RAMS).[23] Today, this conference is one of the
biggest sources of ALT information. Other early sources include the Bell System Technical

1
Journal and the Department of Commerce, which also sponsored research and improvements of
early accelerated life testing. References [1] to [6] provide examples of these early efforts.
The military services each began studying the ALT problem in earnest during and after
World War II because of the poor performance of electronic and electrical devices. A solution
was proposed and a joint military study was initiated in 1952. The primary impetus for the study
was the widespread recognition that modern electronic gear supplied to the US military were not
performing up to desired reliability levels. After a four year period of study, a report was
generated. This report came to be known as the AGREE report, an acronym for the Advisory
Group on Reliability of Electronic Equipment [6]. This publication in 1957 is probably the best
places to begin any modern history of accelerated life testing. The vacuum tube radio systems
studied by AGREE were found to follow a bathtub-type curve. It was easy to develop replaceable
electronic modules, later called Standard Electronic Modules (or SEMs), to quickly restore a
failed system and they emphasized modularity of design. Hence, there was early emphasis on
maintainability topics. Additional recommendations included running formal demonstration tests
with statistical confidence for products. Also recommended was running longer and harsher
environmental tests that included temperature extremes and vibration. This came to be known as
AGREE testing and eventually turned into Military Standard 781. The last item provided by the
AGREE report was the classic definition of reliability. The report stated that the definition is “the
probability of a product performing without failure a specified function under given conditions
for a specified period of time”. Another major report on “Predicting Reliability” in 1957 was that
by Robert Lusser of Redstone Arsenal, where he pointed out that 60% of the failures of one Army
missile system were due to components. He showed that the current method for obtaining quality
and reliability for electronic components were inadequate and that something more was needed.
ARINC set up an improvement process with vacuum tube suppliers and reduced infant mortality
removals by a factor of four. This decade ended with a lot of promise and activity. Papers were
being published at conferences showing the growth of this field. Over the next several decades,
Birnbaum made significant contributions to probabilistic methods, the reliability of complex
systems, cumulative damage models, competing risk, survival distributions and mortality rates.
The 1960s dawned with several significant events. RADC began the Physics of Failure in
Electronics Conference sponsored by Illinois Institute of Technology (IIT). A strong commitment
to space exploration would turn into NASA, a driving force for improved reliability of
components and systems. Richard Nelson of RADC produced the document “Quality and
Reliability Assurance Procedures for Monolithic Microcircuits,” which eventually became Mil-
Std 883 and Mil-M 38510. Semiconductors came into more common use as small portable
transistor radios appeared. Next, the automobile alternator became possible with the use of low
cost germanium diode later replaced by better silicon diodes and able to meet the under-the-hood
stress and environment requirements. Dr Frank M Gryna published a Reliability Training Text
through the IRE. The nuclear power industry was also growing by leaps and bounds at that point
in history. The demands of the military ranging from missiles to airplanes, helicopters and
submarine applications drove a variety of technologies. During this decade, a number of people
began to use, and contribute to the growth and development of, the Weibull analysis methods and
applications which are closely tied to the analysis of ALT. Professor Gumbel demonstrated that
the Weibull distribution is a Type III Smallest Extreme Value distribution. This is the distribution
that describes a weakest link situation. Dr. Robert Abernethy was an early adaptor at Pratt and
Whitney, and he developed a number of applications and analysis methods.
During the decade of the 1970s, reliability work progressed across a variety of fronts. In
this decade, the use and variety of ICs increased. Bipolar, NMOS and CMOS all developed at an
amazing rate. In the middle of the decade, ESD and EOS were covered by several papers and
eventually evolved into a conference by the decade end. Likewise, passive components which
were once covered by International Reliability Physics Symposium, IRPS, moved to a Capacitor
and Resistor Technology Symposium (CARTS) for continued reliability advancement for discrete

2
components. A few highlights of the decade were the first papers on gold-aluminum inter-
metallic products, accelerated testing, the use of Scanning Electron Microscopes for analysis and
loose particle detection testing (PIND). Perhaps the two most memorable reliability papers from
this decade were one on soft error rates caused by alpha particles (Woods and May) and on
accelerated testing of ICs with activation energies calculated for a variety of failure mechanisms
by D.S. Peck. By the end of the decade, commercial field data were being collected by Bellcore
as they strived to achieve no more than 2 hours of downtime over 40 years. This data became the
basis of the Bellcore reliability prediction methodology.
The Navy Material Command brought in Willis Willoughby from NASA to help improve
military reliability across a variety of platforms. During the Apollo space program, Willoughby
had been responsible for significantly improving spacecraft reliability. He insisted that all
contracts contain specifications for reliability and maintainability instead of just performance
requirements. Willoughby's efforts were successful because he attacked the basics and worked
upon a broad front. Wayne Tustin credits Willoughby with first emphasizing temperature cycling
and random vibration, which later became ESS testing.
The 1980s was a decade of continued great change. Televisions had become all
semiconductor. Automobiles rapidly increased their use of semiconductors with a variety of
microcomputers under the hood and in the dash. Large air conditioning systems developed
electronic controllers, as had microwave ovens and a variety of other appliances.
Communications systems began to adopt electronics to replace older mechanical switching
systems. Kam Wong published a paper at RAMS questioning the bathtub curve. Developments in
statistics made an impact on reliability. Contributions by William Meeker, Gerald Hahn, Richard
Barlow and Frank Proschan developed models for wear, degradation and system reliability and so
fit nicely with ALT methods. The Air Force issued the R&M 2000 which was aimed at making
R&M tasks normal business practice. Altogether, the1980s demonstrated progress in reliability
across a number of fronts from military to automotive and telecommunications to biomedical.
RADC published their first Reliability Tool Kit and later updated this in the 1990s for COTS
applications. ASQ issued a number of handbooks detailing statistical analysis of accelerated life
tests and in 1990 Wayne Nelson issued his value book Accelerated Testing [7].
By the 1990s, the pace of IC development was picking up. New companies built more
specialized circuits and Gallium Arsenide emerged as a rival to silicon in some applications.
These changes drove IC manufacturers to perform more ALTs. It quickly became clear that high
volume commercial components often exceeded the quality and reliability of the small batch
specially screened military versions. Early in the decade, the move toward Commercial Off the
Shelf (COTS) components gained momentum. The Army started the Electronic Equipment
Physics of Failure Project and engaged the University of Maryland CALCE center, under Dr.
Michael Pecht, as part of the process. RAC issued a six set Blueprint for Establishing Effective
Reliability Programs in 1996 including one to discuss ALT (RBPR-4).
While there is much to credit to the early roots of reliability testing, the ALT field itself
has rapidly changed primarily over the last 30 years. This may be credited to the increasingly
sophisticated number of software tools developed and the increasing customer needs for more
sophisticated, yet more reliable products. The need to develop new and different materials for
NASA and biomedical applications as well as finding ways to improve all manufacturing
processes have helped drive interest in ALT. Few comprehensive books on the accelerated life
topic have been produced. More often than not, a small chapter of a general reliability book has
attempted to cover this broad and important topic. Dr. Wayne Nelson, in his classic work,
Accelerated Testing, provided a prophetic statement. He stated in that 1990 publication that
"Statistical methodology is improving rapidly. Thus, books over 5 years old lack important
developments, and books over ten years old are seriously out of date." He predicted such would
be the fate for his own book and updated in it 2004 with Accelerated Testing: Statistical Models,
Test Plans, and Data Analysis The breadth of applications for accelerated life testing included in

3
Nelson's 1990 work is amazing. His list of references and additional information runs the full
gamut, running several pages long and citing industries citations as diverse as concrete
technology, metals and ceramics, pavement and soils, the Plastic and Rubber Institute, the U.S.
FDA Center for Drugs and Biologics, the National Lubricating Grease Institute and the National
Nuclear Data Center. This length of this initial list was increased by Dr. Dimitri Kececioglu and
Dr. Feng-Bin Sun when they provided additional references at the end of chapter one in a more
recent related work, Environmental Stress Screening, 1996. It is now 2010, twenty years having
passed since he wrote those words and new books have come onto the market. In 1998, Reliasoft
issued the Accelerated Life Testing Reference with their new ALTA™ software. Gregg Hobbs
added to this with his Accelerated Reliability Testing: HALT and HASS book in 2000. The
Accelerated Stress Testing Handbook was produced by Anthony Chan and Paul Englert in 2001.
Harry McLean published a HALT, HASS & HASA Explained that same year and updated it in
2009. A last book of note is titled Accelerated Testing: A Practioner’s Guide to Accelerated and
Reliability Testing by Bryan Dodson and Harry Schwab.
By 2010, journals and conference proceedings continue to document the advancements
made in the last 10 years. Some new and different approaches to ALT have come into more
common during the last 20 years. Often, these special approaches were developed internally at
companies wishing to improve their products. For the most part the ALT tests have evolved from
steady state operating conditions at customer limits to rapidly changing environment conditions at
or beyond the specification limits. These extreme test conditions serve a variety of purposes.
Some of these special approaches will be covered in later sections. I document the use of newer
test software, test equipment and guidelines for improved methods such as HALT, (Highly
Accelerated Life Test), HASS (Highly Accelerated Stress Screen) and step-stress techniques.
The application of accelerated techniques has spread far beyond the few original
traditional military arenas to almost every aspect of modern life. The need to quickly evaluate
new products and technologies now extends to computers and computer products, such as hard
drives, appliances of all types, farm equipment - from tractors to hay bailers, recreational vehicles
of all types, implantable biomedical devices - from pacemakers to nerve stimulators, miniature
electromechanical modules or MEMs, automobiles - from engines to air conditioners, airplanes -
from small consumer to large commercial units, telephones and telecommunications equipment,
satellite applications and almost all building structures.

1.2 Applications of ALT Techniques -

The application of ALT might be divided into a small number of areas for clarity. These
areas include the following range of parts and materials.

1) New materials test and evaluation. This includes lubricants, oil, paints, concrete, new
metals, ceramics, fibers, MEMs, rubber products, memory metals, plastics, adhesives,
nanotubes, nano technology and adhesive products and tapes.

2) Component test and evaluation. Included in this category include Silicon

semiconductors, Gallium Arsenide devices, Indium Gallium Phosphide transistors,
a variety of digital and linear microcircuits, many different formulations of resistors,
the wide range of capacitors, switches, crystals that now exist as well as all batteries,
ball and roller bearings, sliding mechanisms of made with Teflon and a variety of
other materials, inductors and magnetic components, motors, actuators, light bulbs, hand
tools, power tools, and building materials.

3) Hardware systems, with test and evaluation of software controlled hardware. These
types of products include gas powered cars, electrical cars, bicycles, motorcycles, boats

4
 and boat motors, computers, hand held devices ranging from cell phones to personal
assistants, all household appliances and white goods, recreational vehicles, including
motorcycles, snowmobiles and sailboats, farm equipment of all types, airplanes, military
hardware of all types, space-rated systems, submarines and underwater vehicles, nuclear
and coal-fired power plants, medical instruments and sensor devices and all types of
power generation systems.

4) Software packages with test and evaluation. This includes all software packages, both
standard and custom software, all firmware, software-controlled systems, including
warning and safety systems are included. These cover all modern applications and
aspects of software.

It is easy to identify additional products, materials or applications for each of these four
categories. Thus, we can get a better understanding of why so many companies are now trying to
apply Accelerated Life Test techniques for common products of everyday life. The need for such
tests is closely tied to product improvement, fewer warranty failures and market share.

1.3 Reasons for Employing ALT Techniques -

Incentives for employing ALT techniques during modern product development projects
or to enhance any released product include the following activities:

1) Provide a way to more quickly evaluate an early product design.

2) Evaluate or estimate the projected life of a product early in the design phase.
3) Examine the influences of maintenance activities on subsequent system failures.
4) Reduce the post release product support costs and warranty costs.
5) Coordinates with other similar problem prevention tools.
6) Selecting from competing suppliers or materials during design. Determine which
material or suppliers are estimated to be best.
7) Identify some major failure modes early in the design cycle in order to start corrective
actions before the release to production of the system.
8) Assure that no major unexpected failures will occur after production release.
9) Identify the influences of any manufacturing process steps on the performance or
longevity of the design.
10) Early ALT activities will aid the design team in making important decisions.
11) Reduce the time for qualification tests or durability tests by creating and employing
shorter and more effective types of tests.
12) Identifies the need for further screening or burn-in as required.
13) Reduce the corporate risk when business decisions are required.
14) Identify the potential for hazard and safety issues before release to production.

All of these reasons for the use of ALT require more detailed explanations. These are covered in
the following short paragraphs.

1) ALT provides a way to more quickly evaluate an early product design. This
includes an identification of major failure modes, any weak points of a design, identifies critical
supplier components and assemblies and may identify the most critical stresses. For any company
wishing to get to market faster with better products, early ALT is critical. It is through such early
activities, especially when performed by development engineers, that large performance and time
improvements can be made. Reliability people should be part of this early activity and are often

5
key in setting up teams, and designing the ALT tests and stress limits for more effective
development.

2) Evaluate or estimate the projected life of a product early in the design phase. ALT
is the key method for determining or estimating the projected life of a product when an ALT is
operated in a customer-type environment. Information about how long a typical product might
last as a measure of design life is important at this phase. Some suggest operating this ALT test to
double the projected field life in order to achieve reliable operation. In addition, the time to first
failure should be estimated in applications where this measure is critical of customer success or to
reduce filed support costs.

3) ALT examines the influence of maintenance activities on subsequent system

failures. The frequency and quality of maintenance activities may influence the time to
subsequent system failures. Implicitly, this also speaks to the need for improved maintenance
activities to prolong the life of a product. ALT may address two possibilities by looking at proper
maintenance and any degradation mechanisms that result from the maintenance. The object is to
create a set of rational maintenance policies that identify ways that maintenance activities will
prevent system failures.

4) ALT reduces the product support costs and warranty costs. Knowing the time to
first failure, some of the failure modes and the root causes of failure early in the development
history all lead to cost reductions during development, manufacturing cost reductions and reduced
warranty costs in the field. Manufacturing costs are reduced by fewer in-house failures and
repairs during manufacturing and fewer escapes of supplier problems. Reduced field failures is a
direct saved cost, but also leads to improved customer loyalty and ensures future repeat business.

5) It coordinates with other similar problem prevention tools. Tools such as Failure
Mode Effects Analysis, reliability predictions, hazard predictions and Design of Experiments.
The ALT results often provide failure rate and failure mode information to feed Design and
Process FMEAs and other reliability tools. The early ALT results also support reliability
predictions for future projects through similarity of components and assemblies. ALT results can
also influence the size and type of design of experiments activities that improve the ultimate
performance of a product.
Past projects with ALT information always influence the activities for new projects
through the data, test set-up and lessons learned. The past project information is very important to
driving the new FMEAs, new reliability predictions and improved test methods. Alt also become
an element of the management decision to move ahead while selecting performance measures and
product features for new products.

6) Selecting from competing suppliers or materials during design. Material test,

evaluation and supplier selection are all important for success of many projects and processes.
Early in the design process, the evaluation of suppliers and materials leads to improved designs.
Trade-off studies can be time consuming and sometimes costly. Having accelerated approaches
for evaluation can easily reduce the time and costs.

7) Identify some major failure modes early in the design cycle. Knowing the system
or component failure modes early in the design is the start of corrective action. One usually
doesn’t know all the modes, but finding modes early helps reduce the cost and time involved with
correcting the causes. Sample size may be important to observing variability and enough modes
to be valuable.

6
 11) Reduces the time for qualification tests or durability tests by shorter and more
effective types of tests. Reduced time during development is a critical feature in modern product
development. Often this may be as much as 3 to 6 months saved on a 2 year development. The
savings primarily comes from not reworking a design for undesirable failure modes found late in
the development process.

12) Identifies the need for further screening or burn-in as required. The advantages
cover screening, burn-in and all similar types of "manufacturing screening tests" that enhance the
apparent performance of a product as viewed by a customer. The need for additional HALT or
HASS work may be determined from these results.

13) Reduce the corporate risk when business decisions are required. Occasionally
near the end of a development process, if a problem is discovered, a difficult decision may need
to be made. Often the choices are to hold up production until the problem is corrected, screen a
less than perfect design or release the design under controlled conditions. All three choices have
financial impacts. The ALT may help determine which provides the lowest impact on time, the
quickest resolution and lowest financial drain. This tool provides the project manager with critical
and timely information.

14) Identify the potential for hazard and safety issues before release to production.
Avoiding hazards and safety concerns prevents serious embarrassment and recalls. Think of
recent automobile recalls. About eight million cars had to be reworked for brake and accelerator
problems. The total costs would exceed five billion dollars.

1.4 Outline of Accelerated Life Test Procedure –

There are some who believe that ALT is achieved solely through the introduction of
environmental stresses on operating hardware. Others believe effective ALT is accomplished
mainly through the statistical analysis of after-the-fact test results. No time-consuming and costly
activity such as ALT should be started without a thorough outline of all the activities planned and
the estimation of expected outcomes. The plans should include consideration of equipment,
calibration, testing to identify degradation mechanisms, costs, sample size for statistical measures
and data sensitivity and collection. This permits a series of planned activities to be performed in a
proper sequence that enhances the probability of successful completion of all the steps. ALT
should build upon every previous ALT run. The lessons learned and results should lead to
continued product and process improvement. The following short outline of the rest of the book
forms the basis for the sections covering all of the various aspects of ALT. These key sections
include:

Section 2.0 - Planning for an ALT and selecting the type of test.

Section 3.0 - Selecting and set-up the test equipment, conditions and parameters.

Section 4.0 - Administer the accelerated test.

Section 5.0 - Bring the test to a successful conclusion.

Section 6.0 - Analyze the data.

7
2.0 Planning for an Accelerated Life Test and Selecting the Type of Test -

This critical initial activity often starts with a statement of purpose of the ALT. This is a
clear statement of why the ALT is to be performed and the expected outcomes. The following list
covers additional reasons for running Accelerated Life Tests. The expected results noted are part
of why the ALT is performed. This list covers most situations that will be encountered by
reliability engineers, quality engineers, development engineers, manufacturing engineers and
project managers. It covers the needs of companies for rapidly producing high reliability and
high quality products for customers. The list also reflects the fact that there are eight project
attributes - ranging from product performance, product features, quality in the field , reliability,
development schedule (time-to-market), direct development costs, warranty costs and
maintenance (see Section 4.2). These need to be simultaneously maximized for overall corporate
success. Attempting to enhance one can only occur at the expense of the others.

This additional list supplements the reasons of section 1.3 and shows the range of applications of
ALT to products and customers.

Table 2.1 - A Long List of Reasons to Perform an ALT

A. Confirm a minimum life estimate or time to first failure of a large population.

B. Look at variability or robustness of the manufacturing process.
C. Determine whether screening will be required for production to meet customer
requirements or market goals.
D. Look at the impact of various customer environments, such as duty cycle, on the
performance or ultimate life of a system.
E. Determine if customer rough handling, customer abuse or customer misuse is a
significant factor in the expected life of a product in the field.
F. Identify the impact of software redundancy to hardware and hardware redundancy
to the proper and continuous system operation.
G. Determine if little-used, emergency or "one-shot products" will operate properly
when called upon to do so. This is especially true of warning systems or emergency
systems such as fire alarms or extinguishers.
H. Show that no dangerous situations exist for a product. This covers all aspects of
liability, hardware and software warnings and system safe operating conditions and
fail-safe modes.
I. Estimate the acceleration factor for a component or system with respect to a specific
set of customer stresses.
J. Demonstrate successful customer operation and maintenance across the many
divergent customer environments and customer use conditions.

2.1 Selecting the Accelerated Life Test Method –

There is an importance to the selected test method and the limits that the test method
imposes. Each purpose or reason shown in Table 2.1 may have associated with it a different set
of conditions for an ALT. The ALT method is tied not only to the purpose of the test, but also to
the applied stresses, the expected test conditions, the test length, the types of test equipment
available and even limited by the sample size. The test method may include a decision on testing
to run to complete failure versus stopping with some systems still operating. A second choice

8
may be to select a degradation parameter to observe during test versus running samples to failure.
The third choice may be to test for specific failure modes as the primary purpose. A fourth choice
may be to create an ALT test that demonstrates a minimum acceptable MTBF. The example of a
fifth choice will be to set up a HALT test as a means of identifying weak points of a design.
Early in the planning stage for an ALT, the test methods or choices available should be
reviewed and considered. The best test approach could then be selected from the short list of
methods shown in Table 2.2 and Table 2.3. Next, a minimum sample size would be determined
and then equipment availability ascertained to support the sample and method. The test methods
available include:

Table 2.2 - A Short List of Methods

A. Steady state operating conditions with one or more stresses in place. The
stresses may range from nominal customer to worst-case customer conditions to
stresses above the worst-case customer conditions. Stresses remain fixed while systems
are operated to failure. (Figure 2.1A)

B. Changing stress conditions for one or more stresses, cycling from one
limit to another in some coordinated fashion if more than one stress is used. This is a
common test condition when one or more stresses vary. The stress limits and the
rate of change are both important. Less important is the dwell time at each
extreme. (Figure 2.1 C or Figure 2.1D)

C. Rapidly changing stress is also called "rapid rate tests". These usually cycle
quickly between extremes at a very high rate of change for one or more stresses.
High rate temperature changes are sometimes labeled temperature shocks
when the rate of change is greater than 30ºC per minute. (Figure 2.1C)

D. A progressively increasing stress state for one or more stresses. These are
known as ramp stresses and are employed for specific reasons. The stresses could be
in a continuous ramp or as a series of small steps. (Figure 2.1B)

E. Irregular or random application of one or more stresses. The envelope of the

stresses may or may not describe a distribution. There are a limited number of stress
states or conditions available. This is common with mechanical shock or vibration.
(Figure 2.1E)

F. Full-reversing stresses, typically occurring only in some types of material

testing or mechanical testing. This is a special type of test to limits. (Figure 2.1F)

Some of the stress conditions have familiar names as tests. A short list of some of these tests
types are shown in Table 2.3. These names do vary from industry to industry and over time.

9
 Table 2.3 – Types of Stress Tests

Example or Common Term

Run-in or burn-in test

Multiple Steady Environments
Accelerated Life Test
Steady state Short term reliability screens
operation tests "Break something" tests
High Temperature Life - HTL
AGREE Tests
Customer acceptance tests

Worst case conditions test

Maximum or minimum power tests
Maximum or minimum temperature
Maximum or minimum airflow
Multiple worst case conditions
Extremes tests Environmental Stress Screening (ESS)
Overstress Conditions Tests
Multiple Environmental Overstress (MEOST)
STRIFE (Stress-Life) Testing
Step-Stress Tests
Multiple Step-Stress Tests

Power temperature or stress cycling

Multiple Extreme Cycling Tests
Slow rate cycling ( < 30C/min.)
High rate cycling or shock tests ( > 30C/min.)
Design Margin Evaluation (DME)
Cycling tests HASS - Highly Accelerated Stress Screen
HALT - Highly Accelerated Life Test
Mechanical stress loading (On-Off)
Full-reverse loading
Irregular or pulsed tests (ESD, speed bump, etc.)
Spectrum of irregular tests (power line transients,
Miners Rule applications, etc.)

Note: these terms are not universally accepted nor do they mean the same thing at different
companies. There is no standard nomenclature in this reliability area.

2.2 Examples of Detailed Timing Diagrams –

The following timing diagrams, shown in Figure 2.1, apply for many of ALT test
methods listed in Table 2.2 and Table 2.3. They represent the common generic applications of
stresses for a variety of accelerated test. Applied stress conditions may vary slightly from the
timing diagrams because of equipment limits and test requirements. Not every type of test is
covered in these timing examples.

10
 Figure 2.1 - Timing Diagrams of Stress Tests

Max. Rated
Stress Step-Stress
Condition

Steady
Operating
Ramp
Stress

Time Time
Figure 2.1A Figure 2.1B

High Rate + Stress

Stress
Low Rate

Time

Time - Stress
Figure 2.1C Figure 2.1D

+ Stress
Stress Periodic
Irregular
Stresses

Time
Periodic
- Stress Full
Reverse
Time Stresses
Figure 2.1E Figure 2.1F

11
 These few timing diagrams represent just some of the possible examples of how
environmental or customer use stresses can be applied to a system. It is easily possible to
combine several of these single stresses into a multi-stress combined environment ALT. The
complications are limited by the equipment, not the imagination. HALT is a good example of
this, when random six axis vibration is combined with high rate temperature cycling and voltage
margins.

2.3 Plan for Success as well as Failure - before the test begins

One often underutilized area of ALT is the "plan for failure" portion of a test. Often it is
assumed that the systems under test will fail at some convenient time and/or in some convenient
way and so the test results will be easily gathered, data collected and coherent. Some time should
be devoted to thinking about what can happen in the real world. The goal is to prevent the many
types of undesirable situations that can occur during real life tests. Table 2.4 is a short list of
examples of common ALT problems. This list is not exhaustive and ranges from possible
undesirable test outcomes to total failures of the ALT test itself.

Table 2.4 - A List of Undesirable ALT Events or Possible Occurrences

A. About two hours after the systems are placed on test, some to most of the systems fail
suddenly and catastrophically. The systems were not watched or monitored continuously after
about 20 minutes, so this undesirable outcome was discovered until the next day. The time of
failure would be uncertain unless continuous monitoring was in place.

B. During the first over-night of test, the systems begin to fail when no one is around.
Same issues as A above about continuous monitoring early in the test.

C. During the first weekend after the test began, the systems begin to fail in when no one
is around. Same issues as A above about continuous monitoring early in the test.

D. After about 3 weeks on test, the systems begin to fail in large quantities for no
apparent reason and with no warning. This is unexpected, so no one is prepared for the sudden
event. Continuous monitoring during the test is key to catch this problem.

E. After the start of test, the systems begin to fail catastrophically and burn-up, leaving
little left to analyze for cause. A corollary of this is the failed systems also damage and/or destroy
the test chamber and supporting test equipment. Either case represents a double catastrophe as
both precious samples and test equipment are lost.

F. No failures occur over a six week period of accelerated test time and the test needs to
be stopped without failure having occurred. What should one do now? What conclusions can be
drawn from the no-failure test?

G. No failures occur in the first three steps of a step-stress test and many of the systems
fail early in the fourth step. What should one do now? What conclusions can be drawn from the
no-failure test?

12
 H. Halfway into an accelerated life test, it is discovered that a piece of monitoring
equipment is not calibrated and appears to have been drifting over time. It is unclear whether
this was a recent event or if the equipment was drifting before the test began. What should one do
now? What conclusions can be drawn from the test?

I. A piece of monitoring equipment fails in test and no one notices for some time.
Several failures are found to have occurred during the time that the monitor was not properly
operating. The times of failure are now uncertain for these units. There was no warning to
indicate that the monitoring equipment would fail.

J. A piece of recording equipment fails during test so no one is able to say exactly what
stresses were present during the changing stress test. Thus, it is not possible to say what
conditions appeared to cause the failures or to estimate the accumulation of stress.

K. A test is continually interrupted (stopped) to fix a piece of support equipment,

causing questions about the trustworthiness of the equipment and the test results. What should
one do now to finish the test successfully? What conclusions can be drawn from the interrupted
test?

L. There are three stresses present in an accelerated life test and you suspect at some
point that at least two stresses might be interacting. Only two failures have occurred during the
accumulated test time so there is not enough information for any reasonable calculations. What
should one do now to finish this test successfully? What conclusions can be drawn from the
current test?

M. The test is stopped periodically to measure the performance of the systems under
test before placing them back into test. This typically takes about 24 hours to complete the
samples. You wonder if these long and frequent interruptions have any impact on the test results.
What should one do now to finish this test successfully? What conclusions can be drawn from the
current test?

The examples in Table 2.4 raised a number of interesting questions on some of these
tests. Many represent common test outcomes that may be undesirable. None are answered or
directly addressed in this section. However, in a later section, there will be some discussions
covering how to handle of these undesirable results. Some answers will be provided, but it is best
to avoid the negative results whenever possible by proper selection of test equipment, changing
test conditions, improving monitoring equipment or more carefully watch the accelerated test for
unusual behavior.

13
3.0 Selecting and Setting Up Test Conditions and Parameters -

Once the decision is made to run an Accelerated Life Test, planning for success becomes
very important. The critical elements of planning include the selection and validation of the test
equipment and any test loads that are required. This test equipment may also include thermal
chambers for high and low temperature, all vibration tables any sources of humidity, dust or
ultraviolet light. In addition to the required environmental equipment, there is often additional
equipment needed. This might be equipment for electrical or mechanical stimulation of the
samples in order to simulate the customer conditions. These stresses might range from voltage
sources to a dynamic actuator for placing mechanical stress on the samples. One of the most
neglected areas of ALT testing is selecting the best loads for the samples under test. These loads
are meant to simulate the outside world under a variety of dynamic conditions. Too little attention
is paid to the loads as they are often generated as passive load devices such as resistors. In the real
world loads are usually dynamic and changing. Examples of this might be a real system with a
typical load that is a mixture of 1000 Ohms resistance in parallel with 2.0 micro Henries of
inductance and 10.8 micro Farads of capacitance. This is the nominal load and it changes by
+30% depending upon the operating conditions. Such a dynamic load could not be modeled by a
simple resistor alone treating the inductance and capacitance as if they were negligible. With the
wrong load, the ALT can’t reflect the real world.

3.1 Selecting the Test Equipment –

Begin by asking what type of ALT test equipment is readily available and how it would
be employed in the test. Consider the choice of a thermal cycle chamber (temperature only), a
thermal shock chamber, a multiple stress chamber (temperature, one axis vibration, humidity) or
other specialized chambers such as a HALT system (high rate temperature cycle, high levels of
multi-axis vibration, and humidity). Often the available equipment sets fundamental limits on
what stresses can be used and how they will be applied. A few details of the differences between
the four chambers are described in the following.

Thermal Cycle Chamber – This chamber typically operates over the range of -40ºC to
150ºC. The rate of change of interior air temperature is usually limited to about 10ºC to 15ºC per
minute for any small loads. The test samples in the chamber will change temperature at an even
slower rate because of the limited thermal capacity of the chamber. No other source of system
stimulation is present in this type of simple chamber. These chambers may vary in size from a
desk top model of 3 ft. 3 to 4 ft. 3 to as large as an 800 ft. 3 walk-in chamber for testing an
automobile.

Thermal Shock Chamber – This chamber typically operates from -60ºC to 200ºC. The
rate of change of interior temperature around the systems is limited to about an effective rate as
low as 50ºC per minute to a modern change rate of 100ºC per minute. Typically, two chambers
exist in this approach. One chamber is held at the hot extreme and the other chamber held at the
cold extreme. The test systems move back and forth quickly from one chamber to the other to
obtain the high rate air change. The systems respond more quickly to the high rate stresses in this
type of chamber, but the limits are set by the thermal capacity of the chamber itself. No other
source of system stimulation is present in this type of chamber. No vibration, no humidity, no
voltage margins are present. These chambers are typically limited in size from about 3 ft. 3 to
about 25 ft. 3 .

14
 Multiple Stress Chamber - This chamber has a temperature rate of change of 10ºC to
20ºC per minute with extreme limits about -50ºC to 150ºC. It is typically capable of one axis
vibration that may operate to 10 times the force of gravity, or 10 g’s. Additional stresses such as
humidity may be present in the chamber. The three stresses (temperature, humidity, vibration) are
usually tightly controlled so that they may be applied independently and in a coordinated fashion.
Humidity can be set and controlled from about 20% RH to 95% RH over most of the temperature
range from 3ºC to 150ºC. The thermal capacity of the chamber usually limits the temperature rate
of change to the low numbers noted and the one vibration axis table has many limits. It is possible
to add a two or three axis vibration capability to the $60,000 one-axis chamber, but the cost then
will usually exceed $250,000. The chamber size may be from 8 ft. 3 to as large as 120 ft. 3 .

HALT Chamber - This type of chamber was designed with multiple-stress capability to
push the limits of temperature cycling. It typically has a temperature rate of change of 30ºC to
50ºC per minute with chamber extreme limits about -60ºC to 200ºC. A high capacity cooling unit
coupled with a very high rate of air flow achieves such rapid air temperature changes. It has
typically six axis vibration capability, three translation axes of X, Y and Z motion and three
rotational axes about X, Y and Z. The applied g forces may be 50 times the force of gravity for
small parts and as low as about 25 gs for larger parts and assemblies. Humidity is typically
present in the chamber. The three stresses noted are usually tightly controlled so that they may be
applied independently and be coordinated. Humidity can be set and controlled from about 20%
RH to 95% RH over most of the temperature range from 3ºC to 150ºC. The thermal capacity of
the chamber usually limits the temperature rate of change to the numbers noted, though very large
masses in the chamber sometimes further reduces the change rate. Vibration is limited to 3 g’s at
the low end in this large chamber, though small table top units will operate at lower values. These
chambers are limited in size to about 80 ft. 3 with present technology.

Other Specialty Chambers - These specialty chambers may meet the needs of a special
test system or unusual life test requirements. Most of the time these are built as customer
chambers, as modified from some existing model. They may be very large walk in units over 100
ft. 3 or very small table top units of about 2 ft. 3 .

3.2 Selecting the Test Conditions –

Pushing the limit of extreme testing often drives the need for improved capable test
equipment capability. This is both in pushing for absolute extremes such as -60ºC to 250ºC as
well as rapid rate of change of the stresses. It is the combination of system size, end use customer
environment and chamber capability defines the levels of stress that can be applied. Other
considerations include how stresses may be applied for as long the ALT may operate. Thermal
cycle rate is usually limited by the equipment capability. As a general rule it is the cooling
capability (e.g. 20 horsepower cooling at 747 Watts per horsepower) that usually sets the limit on
the rate of cool thermal change. Thus, a 25 horsepower system running at 50% efficiency delivers
only 12.5 horsepower cooling (about 9.3 KW) while consuming almost 19.6 KW from the wall.
This is 82 Amps for a 220 VAC system or 41 Amps for a 440VAC system. This level of power
requires a special set up for power handling in most buildings.
The interior dimensions of the test chamber, the total mass within the chamber, the
heat capacity of the chamber walls and the heat capacity of any internal chamber masses all
impact the speed at which the chamber is able to go from the hot extreme to the cold extreme.
The sum of all these heat capacities combined with the cooling capability of the equipment
normally set the time limit of the chamber for cool temperature change. It is much easier to
generate 20 KW to heat a chamber than to generate the same amount of cooling. The total mass

15
within the test chamber should be kept to a minimum when high rate of temperature changes are
desired. Therefore, it is often best to place system electrical or mechanical loads outside the test
chamber as well as all support equipment that is not required in the chamber. These loads may
still need to be cycled up and down in temperature in a separate chamber in order to simulate the
real-world load fluctuations for the system.
In many situations an elaborate fixture may need to be created to hold the test systems
in the thermal chamber. This fixture, in addition to negatively impacting the thermal cooling rate
of the chamber, will also have a negative impact on any planned vibration. It is common that the
mass of the fixtures often exceed the mass of the systems under test. Special light alloys of
aluminum or magnesium are therefore employed to reduce the mass of any necessary fixtures.
Ultimately, after the chamber limitations are noted, the choices of test conditions may be
further limited by the ability of the applied voltages and/or mechanical actuators to simulate a
high stress version of the customer environment. These stresses may include UV light, the
circulation of dust, humidity or the inclusion of a salt spray. Most of the time corrosive chemicals
are kept out of an expensive test chamber. Usually only humidity may be present and only in a
stainless steel chamber.
One last concern is that these stresses may interact. That is, when two stresses such as
temperature and humidity are combined, the effects are greater than would be expected for simply
adding the effects. An example of this is a degradation test in which a certain temperature over a
certain time causes the systems to degrade by 25%. When humidity alone is applied to the same
systems for the same time, they degrade by 8%. When both stresses are applied simultaneously
for the same length of time, the systems degrade by 42%. This number is larger than the simple
sum of 33% (25% plus 8%). Add a third stress such as vibration and the results may get rapidly
more complex. These possibilities will be covered in detail by modeling in a later section. A
familiarity with Design of Experiments is a great help for this analysis.
What are the optimal test conditions to select for an ALT? Select conditions that simulate
the customer use conditions, but go beyond the worst case customer conditions. This ensures that
a system will be robust when released to the field. Another rule of thumb is to select as many of
the field stresses as possible and combine as many as possible in one test chamber. This approach
allows a variety of field possibilities to be covered with a minimum of equipment. This works
well for getting an overall estimate of the life of the system or the main failure modes of the
system. It has limits, however. This "as many stresses as possible" approach does not facilitate
learning the root cause of the failure modes. It may be difficult to associate specific failure modes
with specific single stresses. This doesn’t have to be a hindrance, but it may be helpful in the
corrective action process. Treat the situation as a Design of Experiments for several test
conditions and stresses. This can help reduce the number of test chambers or time required. It
permits the identification of failure modes with stresses and even the identification of interacting
stresses. Such an approach is much more difficult and can be time consuming, sample intensive
and it often requires many chambers. Ultimately, schedule, equipment availability and resources
help to determine what test conditions will be employed. At some point compromises on test size,
length, conditions and support will arise.

3.2.1 Thermal Capacity of a Chamber –

The cooling capacity of a chamber often determines how quickly the interior of a
chamber can transition from hot extreme to cold extreme. A sample calculation will help clarify
how to “size a chamber”. This is to determine how large a chamber is required? This suggests the
cooling characteristics and capacity that should be purchased to assure a minimum high rate of
temperature change. If it is desirable to be able to go from 100ºC to -50ºC within 20 minutes,
what size of cooling in the chamber is required for a fixed chamber load of 30 pounds?

16
 First, estimate the interior size of the chamber required to support the systems, fixtures
and loads that may be inside the chamber. Look at the volume of the systems, fixtures and loads
and quadruple this volume. This number is the minimum interior size required. Next look at the
masses involved. Here, we will use a 30 pound fixture in the chamber that supports twenty
pounds of systems during thermal cycling. The fixture is 12 inches high by 24 inches wide by 6
inches deep and is an open structure that surrounds the systems and loads. This fixture size is 1
ft. 3 , so the chamber has to be at least 4 ft. 3 The suggested chamber size would be at least 18
inches high by 36 inches wide by 18 inches deep, giving at least 6 inches on each side for
sufficient air flow. Then, the desired chamber volume is 1.5 ft. x 3 ft. x 1.5 ft. or 6.75 ft. 3 . This
makes the interior chamber surface area equal to 2[(1.5 x 3) + (1.5 x 1.5) + (3 x 1.5)] or 22.5
ft. 2 . The thermal load (for cooling) presented by the air volume inside the chamber and surface
area of the chamber can be approximated with the quick calculations.

Volume thermal load = (6.75 ft. 3 )(0.16 lbs./ ft. 3 ) = 1.08 equivalent “pounds of air" (3.1)

Surface area thermal load = (22.5 ft. 2 )(1.3 lbs./ ft. 2 ) = 29.25 equivalent “pounds of air" (3.2)

These calculations are based upon some very simple assumptions about the surface
material (stainless), finish of the chamber (nice smooth polish), air heat capacity, the chamber
materials (mainly stainless with rubber gaskets and seals on door and ports) and geometry (cube
with small windows). The volume of the systems, the equipment and the loads has been ignored
up to this point. The calculation is therefore only approximate, but looks at a worst case.

The total equivalent chamber thermal load is equal to 1.08 + 29.25 = 30.33 pounds of air
at 70ºF and 40% humidity, based upon the volume and the surfaces to be heated and cooled. This
gives us an approximation for later calculations of cooling capacity required.

Example 3.1 - What is the energy required to cool this 30 pound load from 100ºC to -
50ºC or to traverse a 150ºC total range? Note, the heat capacity of air varies a little over this
temperature range, but it averages close to 1 Cal./gram/ºC and we will use this for the whole
range.

Energy Required to cool = (30.33 lbs. of air)(~1 Cal./gram/ºC for air)(150ºC)

= (13,770 grams)(1 Cal./gr./ ºC))(150ºC) =

= 2,065,500 Calories required. = 2,066 Kilo Calories

= (2,066 KCal.)(0.0015586 Horsepower-Hrs./KCal.)

Energy Required = 3.22 Horsepower-Hrs. (3.3)

Thus, to transition a total of 150ºC in only 20 minutes requires (3)(3.22) = 9.66 Horsepower just
to cool the chamber and the air inside. This is a minimum of 7.2 Kilo Watts of cooling capability.
If cooling inefficiencies are included, this requirement will double to about 14.4 Kilo Watts or
just about 20 Horsepower. To transition the 150ºC in only 10 minutes requires almost 40
Horsepower of cooling or 28.8 KW of energy.

17
 Place a 30 pound "mixed metal" load in the chamber. This consists of the test systems,
fixtures and test load. There is an additional cooling energy requirement for the chamber that is
about:

Energy = 30 pounds (~6 Cal./gr./ºC)(150ºC) = (13620 gr.)(6 Cal./gr./ºC)(150ºC)

Energy required = 12,258 KCal. (3.4)

This is an additional (12,258)(0.0015586) = 19.1 Horsepower-Hrs. or an additional total

Horsepower of 54.9 Horsepower for a desired 20 minute temperature swing.

Based upon this simple calculation, a total of 28.8 + 54.9 = 83.7 Horsepower or 62.5 KW
would be required to meet the 20 minute temperature transition requirement. At the 50%
efficiency for a mechanical cooling system already included, this requires 200 Amp service at
440V. This cooling requirement pushes the limits of many large capacity chambers, so select a
chamber with sufficient cooling capacity or accept the temperature transition time that comes
with the chamber and the load present. One option for achieving high cooling capacity is to add a
liquid nitrogen tank (LN2) for extra cooling. This reduces the mechanical cooling required
typically by about a factor of 2. This is exactly what chamber manufacturers do to achieve rapid
rate of cooling. Changing the air flow and reducing the heat capacity of the chamber itself will
further reduce the cooling required. This has been the approach taken by manufacturers of high
capacity, high rate of change chambers. Thus, a chamber with 30 Horsepower cooling and a LN2
boost can make this extreme transition with only 100 Amp service at 440VAC. The only
additional cost is the additional capability for the LN2 tanks nearby. This often includes
installation and maintenance costs if a large external LN2 tank is used.

3.3 Selecting Applied Stress Test Methods –

A variety of applied stress test methods are available as discussed in section 2.0. Stresses
may be applied as steady state stresses, ramp-type stresses, step stress, or by stresses that vary
periodically over time. These include sinusoidal applied stresses or other cyclic stresses. Even
randomly applied stress conditions may be considered as often occurs with vibration. The
envelope of all the random stresses may be defined or approximated by a statistical distribution
in these cases. There are also situations in which the random stresses may not correspond to a
statistical distribution, but can be described and reproduced by equipment. Figure 2.1 provides
some examples of the timing of applied stresses for common ALT situations. All of these
methods have some advantages as well as some disadvantages. It is up to the reliability engineer
to determine which method represents the best option for the samples, test chamber and support
equipment that is available.
Options for multiple stress levels also exist during test. Multiple stresses are sometimes
required if a projection to another stress level (usually lower) is desired. One example is a system
operated at stresses of 120ºC, 100ºC and 75ºC with the results used to estimate the customer
operating life at 40ºC. Only this stress was accelerated and other customer stresses that might be
present should also be considered. If the customer operates in a humid environment then multiple
humidity conditions might be included in an ALT. Multiple test chambers may be required to
perform a matrix of conditions.
Caution, it is sometimes believed that a single high stress level with an assumed
reliability life model is sufficient to project life at another set of lower stress conditions. This
approach sometimes provides misleading estimates and should be approached with caution.
Unless good history has already been obtained, the assumption of a stress-life model and the
unknowns contained within the model may lead to incorrect estimates. This will be discussed in
detail in the section on data analysis.

18
 The most common methods of applied stresses tend to be steady stress, cyclic stresses or
step stress. Only one or two stresses are typically employed in any of these examples, but analysis
methods exist for three or more levels of stress [25]. A variety of test plans have been generated
in the past with these more complex approaches. A later section will show some typical
calculations of acceleration factors from a variety of test conditions. That section will address the
theory about these issues of interacting stresses and stress-life models. The later section will show
specific examples.

3.4 Selecting Sample Sizes for Meaningful Tests –

This area is often treated as the most important feature of an ALT. After all, the number
of test samples is closely tied to other reliability goals, test equipment capability, the test stress
severity and even test length. Sample sizes are often determined either by practical means or by
calculation of statistical sample sizes. Either approach may fill your particular ALT need,
depending upon the original purpose of the ALT, the product development time allowed and the
expected outcomes. Sample size can drive the conditions of an ALT, but remember with
acceleration and prior knowledge sample size changes. Be careful about the method selected for
sample size, some are better than others.

3.4.1 Statistical Sample Sizes -

A "statistical sample size" is always preferred when they are available. A statistical
sample can be established through several different common formulas. Each of the different
approaches to sample size determination does not take into account applied stresses, test
conditions or even test time. It is wise to avoid any statistical approach not closely related to
fundamental items test stresses or conditions such as those mentioned. Most of the time, stress
level, test length and test conditions and definition of failure are related to sample size and
reliability goals. Related measures such as MTBF or minimum percent failure might also be tied
to picking a sample size. The following is a short summary of some of the possible approaches to
sample size.

3.4.1.1 The Distribution-Free Approach –

This is an example of one approach that should be avoided whenever possible. The
simple distribution-free always leads to the biggest sample size and longest test. This formula for
sample size is based upon a reliability goal and confidence limit only. This sample size
calculation method may actually be derived from binomial confidence limits [9]. Since this
approach is not directly related to any of the critical test factors, test conditions or test length, the
large sample size determined by this method should be suspect. Unfortunately, this method
appears in many older reliability text books without sufficient explanation, qualifications,
warnings and limits suggested. Few books published in the last 10 years still carry this approach
in a few examples. This approach is appropriate only when little information is available about
the product to be tested and only pass-fail tests can be determined. This approach is simply an
attribute test of pass-fail. There are several variations of this method and all may be labeled
distribution-free. The following two short examples are shown for clarity. Both make no
assumptions about the underlying distributions of time to failure or stress versus life, since neither
appear in these simple estimates. This might be better labeled information-free approach.

19
Example 3.1 - Let the probability of being successful over a certain period of time at a given
nominal operating stress level be P. If there are N independent samples, what is the reliability in a
failure-free situation?

Here we have:

Survivors
Reliability =
N  Samples
If we require all samples to operate successfully we have

Probability = (R) N = 1 – Confidence

so

Ln(1  Confidence)
N= (3.5)
Ln( R)

Example 3.2 - How many items need to be tested for the full life with no failures to demonstrate
a reliability of 0.95 at 90% confidence?

Ln(1  Confidence) Ln(1  0.90)  2.30259

N= = = = 44.89 ~ 45 samples
Ln( R) Ln(0.95)  0.05129

All of the samples must operate failure-free over the expected life time and operating conditions
specified.

Example 3.3 - What is the sample size required to demonstrate a reliability of 0.92 at 95%
confidence using the success-run approach.
An approach labeled "success-run" and is similar to the above distribution-free method.
The following example is based upon O'Connor, page 307 [10]. He recognized this as an
approximation method and showed an example where failures also occurred during test.

Ln(1  Confidence)
N+1 = (3.6)
Ln( R )
Ln(1  0.95)  2.99573
N+1= = = 35.93 ~ 36 samples
Ln(0.92)  0.08338

All of the samples must operate failure-free over the time and stress conditions specified.

A third approach is possible and that will be described in the next example.

3.4.1.2 The Constant Failure Rate Model –

This approach to sample size selection is based upon the exponential model for
reliability. This model is very simple and does not correspond to electronic or mechanical
systems. We treat it as a rough approximation. We recognize that this MTBF calculation is very
familiar. In addition we have added the confidence estimate to MTBF. The equation is:

20
 2 Nt ( A.F .)
MTBF = (3.7)
12C , 2 f  2

This equation might be treated as an estimate of a real-life situation when the hazard rate
is known by independent information to be close to constant. The equation relates the desired
sample size, N, as a function of five other independent variables. These variables are:

1) An MTBF goal
2) A desired confidence, C, which is one sided here
3) The test length, t, or the time all units on test accrue
4) The acceleration factor, A.F., based upon any stresses
5) The permitted number of failures, f, that occur during the test.

A hidden assumption may be present. It is that items on test are either repaired or replaced upon
failure. Hence, the combination "Nt" represents the total test time of the samples.
The acceleration factor may be identified from prior history of similar tests or from a
well-formed current multi-level stress test. This factor is often estimated or calculated. The test
time and the sample size may not be independent variables since the devices under test might be
repairable. The combination "Nt" represents the total time on test at a given level of stress.

Example 3.4 - Let a non-repairable module be placed on an Accelerated Life Test at constant
stress equal to the field operating stress. The test has a goal of demonstrating a reliability of at
least 0.97 at 500 operating hours. A 95% confidence is desired. What sample size is required?

Solution - Fix the number of failures at zero, the minimum time to establish the
reliability goal. We are asking in essence the time to the first failure for this sample. The
following two formulas exist to help solve the problem.

t'

MTBF
R=e or

 t'
MTBF = (3.8)
LnR
and
2 Nt ' ( A.F .)
MTBF =
 12C , 2 f  2

with A.F. = 1.0,

 02.05, 2 = 5.99147, for zero failures or two degrees of freedom at 95% confidence

' '
Since A.F. = 1.0, then t = t, there is no difference between the time on test, t and the clock
time, t.

Now solving these two equations yields:

21
 t 2 Nt ( A.F .)
= (3.9)
LnR  02.05, 2

or after eliminating t and rearranging, we have:

 5.99147  5.99147
2N =  = 196.6 samples
ln R  0.030459

thus, N = 98.3 samples minimum, or 99 samples are required

The conditions are no failures and all samples operate 500 hours to meet the requirement.

Example 3.5 - Decide the ALT will no longer be a failure-free test in Example 3.4, but rather run
to at least 3 failures in order to obtain some failure mode information. Past history suggests the
acceleration factor we select for test to be at least 2.5 at this level of stress.

Then,  02.05,8 = 15.5073 for three failures and the new MTBF goal becomes 16,415.4 hrs.

This is calculated as follows from equation 3.8.

 t' 500
MTBF = = = 16,415.4 hours
LnR  0.030459

And with equation 3.7

2 Nt (2.5)
MTBF =
 12C , 2 f  2

so we have
2(2.5) Nt 5[( N  3)t  t1  t 2  t 3 ]
16,415.4 = =
15.5073 15.5073

with t 1 , t 2 , and t 3 representing the times of the first three failures. These failures are not replaced
or repaired, so they appear overtly in the equation.

(N-3)t+ t 1 + t 2 + t 3 = 50911.7 hours (3.10)

These first three times to failure are unknown at the start of the test, so N must be selected first
since time cannot be estimated. Let N = 12, this is large enough to be meaningful, but may not be
statistically significant. Equation 3.10 becomes:

9t + t 1 + t 2 + t 3 = 50911.7 (3.11)

22
Assume all 3 failures occur very near the end of the test, say within about 500 hours of the end,
just so we have a rough estimate. Then we have the approximate solution to Equation 3.10
becomes:

~9t + (t-300) + (t-200) + (t-100) = 50911.7

12t -600 = 50911.7

so the test must run at least t = 4292.6 test hours minimum

In real life the failures did not occur based upon our assumption, but are usually more spread out
in time.
With the statement that reliability is 0.97 in example 3.4 and MTBF is estimated at
16,415.4 hours we could have estimated the times to failure as 1428 hours, 2993 hours and 4722
hours. This last estimate is still rough and based upon assumptions.

Example 3.6 – An Alternate solution Example 3.5 - We will let the total time to failure of the
first three failures total to 5250 hours, not as shown in equation 3.11. Thus:

t 1 + t 2 + t 3 = 5250 hours

then equation (3.10) above becomes:

9t = 45661.7
solving yields:
t = 5073.5 hours

This is the minimum test length to accumulate 3 failures, but we do not know when the three
failures occurred. The answer to Example 3.5 does depend upon when the failures actually occur,
which is information not available at the start of the test. Thus, the ALT is run and the test
analyzed along the way to see if the necessary conditions have been met. Along the way, an
estimate of the total test length is automatically provided. Said another way we run to three
failures and look at the results.

Example 3.7 - Let the acceleration factor for a life test in Example 3.5 be determined by the
following stress-life equation.

Life = B (Stress) 3 (3.12)

The stress is increased for the test by 60% over the normal conditions. If the test is stopped at the
first failure, what is the minimum test time?

We have  02.05, 2 = 5.99147

and so MTBF = 16,415.4 hours from Example 3.5.

Calculating the new A.F. yields:

23
 Life1 (1.0) 3.0
A.F. = = 3.0
= (1.6) 3 = 4.1 (3.13)
Life 2 (1.6)

Filling in Equation 3.9 provides:

2 Nt ( A.F .)
16,415.4 =
5.99147

and so
Nt = 11994.2 hours

With t = 500 hours as a minimum test time to first failure, the minimum required sample size is

N = 23.988 ~ 24 samples

If the test is planned for 500 hours duration as the minimum time to first failure, then N = 24
samples required. If the test were planned for 1000 hours duration to the first failure, then only 12
samples are required, in theory. This equation suggests that we may trade off sample size and test
length with no loss of information. This is clearly not true as the next section will show.

3.4.1.3. - The Weibull Model for Selecting Sample Sizes

The Weibull Approach – Non-Constant Failure Rate Let a sample of repairable or

non-repairable systems be placed into a life test. What sample size is required if zero failures are
desired in this test. The following formula covers this situation when a Weibull Model of time to
failure is employed.

 [(t )( A.F .)] 

N
 i 0
 = (3.14)
 ln(1  C )

with MTBF = ( 1+ 1/) (3.15)

Where:

The sample size is N.

The individual test time for each sample is t i .
The unknown, C, represents the confidence level desired.
The term, A.F., is the acceleration factor.

This formula requires an estimate of , which is the shape parameter of the Weibull Distribution,
while  is the characteristic life of that distribution.

Equation 3.14 is for zero failures only. The term, -Ln(1-C), may be replaced by 0.5
times the standard Chi-Square term for failure situations as has been suggested by Dr. Wayne
Nelson. in a 1985 paper [11].

24
 In the situation of a test that operated to multiple failures, the value of  could actually be
determined through the test itself. That is, the Weibull analysis of the first 3, 4 or 5 failures may
provide a reasonable estimate of  for a one-level stress test. The acceleration factor will remain
unknown in a single stress level test. Since this unknown is needed in all situations except under
normal operating conditions, at least a second stress level would be required. The following is a
simple example of the use of this equation.

Example 3.8 - Let R be estimated as 0.97 at 500 hours with an A.F. = 2.0, at C = 95% and  =
2.5 based upon past history for a no failure test. What is the sample size required?

The basic Weibull two-parameter formula is:

t
( ) 

R=e (3.16)

Filling in the equation with the information from above we have:

500
( ) 2 .5

0.97 = e

solving this equation gives:

 = 2020.6 hours

Next, this number will be filled into Equation 3.14, and when all samples are run the same length
of time we have:

 [(t )( A.F .)] 

N

 = i 0

 ln(1  C )

N [(t )(2.0)]2.5 N [(2)(500)]2.5

(2020.6) 2.5 = =
 ln(1  C ) 2.9957

5.49806 x 10 8 = N(1000) 2.5 = N (3.1623 x 10 7 )

and
N = 17.38 samples ~ 18 samples

Thus, the test should run 500 hours with no failures with a sample size of 18 to demonstrate the
requirement of 0.97 at 500 hours with 95% confidence. This is a different size of samples because
the value of  is 2.5. When the failure rate is constant the value of  is 1.0

Example 3.9 - Let R = 0.97 at 500 hours, A.F. = 1, C = 95%,  = 2.5 from past history. The test
is planned to stop at first failure test. If the characteristic life is  = 2020.6, as before, what is the
test length required?

Filling in Equation 3.14 as before, selecting the appropriate Chi-Square term for 2
degrees of freedom and 95% confidence, we have:

25
 2i 0 [(t )( A.F .)] 
N
2.5 2 N [(1)(500)]2.5
(2020.6) = =
 02.03, 2 5.99147
so

5.99147(1.8353x10 8 )
N(t) 2.5 =
2

N(t) 2.5 = 5.498 x 10 8

If we select a sample size of 12 for this test and plan to stop at the first failure, we have:

(t) 2.5 = 45,817,201

t = 1159.9 test hours to first failure

Thus, the test conditions meet the requirements if no failure occurs until this point in time.

1 2
In this example we replaced the term -Ln(1-C) with the standard  term per the
2
suggestion of Dr. Nelson and then filled in the equation for the first failure and a 95% confidence
level. A sample size was selected in order to finish the calculation of the test length. A test length
could have been selected and the resultant sample size calculated. The following table shows the
limits of this calculation for Example 3.9.

Table 3.1 - The Trade-Off Between Sample Size and Test Length

Sample Size Test Length Total Cumulative Test Hours

1 sample 3133.9 hours 3133.9 hours
5 samples 1646.2 " 8231 "
10 samples 1247.6 " 12,476 "
15 samples 1060.8 " 15,912 "
20 samples 945.5 " 18,910 "
25 samples 864.8 " 21,620 "
50 samples 655.4 " 32,770 "
100 samples 496.7 " 49,670 "
200 samples 376.4 " 75,280 "
500 samples 260.9 " 130,450 "
1000 samples 197.7 " 197,700 "

We can see that, as the sample size increases, the test length decreases, but not in proportion so
that the total test hours required actually rapidly increases. Treat the increased cumulative test
hours as increased information. If the number of failures were increased, then the value of 
could be determined independently. The Weibull formulas get rapidly complex for non-repairable
and multi-failure situations, as was the case in Example 3.9. In the two-parameter Weibull,
Example 3.9, the A.F. was assumed. The more realistic test situation is when the Acceleration
Factor is not known before the start of the test. A second major purpose of the ALT is to identify
an appropriate stress-life relationship and to look at the failure modes. Thus, in this situation, the

26
ALT is analyzed "on the fly", and the relationships are calculated at or near the end of the test.
These situations, presented in some reliability texts, will be covered in later sections.
The next example covers the situation of having several failures in test using the Weibull
Model. Earlier, we used the Weibull formula, Equation 3.14, to estimate reliability in a no-failure
case. The following is the formula to be used for this failure situation. The unknowns are the
same as before.

2i 0 [(t )( A.F .)] 

N

 = (3.17)
 12C , 2 f  2

Example 3.10 - The following data set of an electronics test consists of 12 systems on test
operated for 1000 hours. Three units failed and were not replaced. Failure times are 458 hours,
677 hours and 855 hours. Ninety percent confidence is desired and  is estimated at 0.8 which is
typical for many electronic systems early in life showing a decreasing failure rate. What is the
characteristic life when the acceleration factor is 1.0?

Solution - Plotting the data on the Weibull graph (following page) showed the first three failures
consistent with a shape parameter of 2.46, not 0.8. This suggests that something unusual may
actually be happening with the electronic modules, since = 2.46 suggests an increasing failure
rate situation. Failure analysis is required to determine the causes of the failures and why the
unusual behavior at this point. We will ignore this analysis situation. Filling in the Equation 3.17
gives:

2i 0 [(t )( A.F .)]2.46

N

 2.46
=
 12C , 2 f  2

2[9(1000) 2.46  (458) 2.46  (677) 2.46  (855) 2.46

 2.46 = =
13.362

4.89828 x10 8
 2.46 = = 3.6658 x 10 7
13.362

 = 1188 hours

This estimate is different from the one derived by Weibull analysis (see Figure 3.1). The
reason for this is the different methods used. Here, we employed the Weibull formula and
performed a hand calculation at 90% confidence. The Weibull analysis, Figure 3.1, was
performed with a software package and the Maximum Likelihood Estimator, MLE at 50%
confidence. A different formula is employed by the software package to evaluate the 12 data
points. Note also that the best fit MLE line does not go through the three data points. That is
because the MLE weighs the suspensions (unfailed systems on test) as heavily as the failures
[12]. A 90% confidence line on this graph would bring the estimate of the characteristic life close
to the calculated one. These confidence limits are shown as well in Figure 3.1, but the hand
analysis remains different from the canned program.

27
ReliaSoft's Weibull++ 6.0 - www.Weibull.c om
Example of Failed Data
99.00
Weibull
Data 1

W 2 MLE - SRM MED

90.00
F=3 / S=9

50.00
Reliability

10.00

5.00

James Mclinn
Rel-Tech
1.00 3/1/2010 10:09
100.00 1000.00
Hours

   

Figure 3.1 - Weibull Graph of time to Failure

3.5 Considering Other Reliability Goals or Measures –

This area is often expressed either as a reliability number at a specified time or as a

MTBF number. Either of these numbers may be employed and both depend upon assumed models
for the relationship between reliability and time or the relationship of MTBF and reliability. Other
measures that complicate this description of reliability include the life-versus-stress relationship.
Measures such as the median and mode will not be discussed in this monograph as they are
primarily distribution measures. This area is most often covered by the detailed models associated
with ALT and reliability. A multi-level stress example in Section 7 will show such an example.
One may also use the moments of the distribution to describe some of the aspects of reliability.
Table 3.2 shows a short summary of the first 4 moments of a distribution for several common
distributions.

28
 Table 3.2 – The Moments of a Distribution

Moment Description Formula – Normal Distribution


X i
= 
i 1
First Average or Mean xf ( x)dx =
 N
N


(X i  X )2
2 = 
i 1
Second Variance x 2 f ( x)dx =
 N 1


Third Skew Skew = 

x 3 f ( x)dx


Fourth Kurtosis Kurtosis = 

x 4 f ( x)dx

3.6 Selecting the Number of Test Failures –

There is no single answer to selecting the number of test failures that should be
encouraged or permitted in the ALT plan. This item may be selected to fit other requirements. For
example, the quote "Run the life test until at least four failures occur" is one simple philosophy.
This failure number is sufficient for testing data and trending failure modes.
The alternative to this is a zero failure approach. In this situation, life tests may provide
only a crude estimate. Failure mode information, which is as valuable as knowing when failures
occurs, is also not available. Most of the time, this zero failure approach is not desired so long as
sufficient time is available. A physics of failure approach may be desired as a means of obtaining
more information about the ALT.

3.7 Selecting the Time to Failure Distribution -

This relationship or distribution is commonly assumed before the start of an ALT. It can
not be shown or proved until after the completion of the ALT, if at all. The four most common
reliability distribution choices are Exponential, Weibull, Normal and Lognormal. Once in a while,
another distribution such as Extreme Value, may be selected. For complex ALT results a
combination of distributions may appear during the analysis of the ALT data. When multiple
stresses are present, interaction between the stresses may add to the difficulty of any analysis.
A number of common distributions of time to failure are possible. Only a few of the
continuous distributions will be mentioned here. These cover over 90% of the distributions,
applications and situations likely to be encountered in ALT analysis.

29
3.7.1 The Exponential Distribution –

This distribution may be formed with up to 2 unknowns to describe this distribution. A

failure rate and position locator such as a time-offset are the two unknowns for the exponential
distribution. Not all software analysis programs have the two-parameter Exponential version
available, so this is rarely used in favor of the one-parameter Exponential. See references 7 and
12 for examples and derivations. The two-parameter Probability Density Function, PDF, and
Reliability may be written for the exponential as:

f(t) = e   ( t  ) PDF, where  > 0 and t >  (3.18)

R(t) = e   ( t  ) Reliability equation

Here,  corresponds to the failure rate and  to the time offset. Complete statistical solutions are
possible for both one and two-parameter situations for both Rank Regression and Maximum
Likelihood approaches. Most of the time the math is rather complex and so is performed by a
custom reliability statistical package. It is beyond the capability of most people to take a standard
package, such as Excel™, and then write the necessary equations to create reliability solutions.
The Exponential distribution appears when the hazard rate (for this distribution only it is
the same as the failure rate) is constant over time. There are a small number of situations for
systems when this may occur. The traditional middle portion of the bathtub curve is sometimes
shown as constant. This constant hazard rate seldom happens for components, as they tend to be
dominated by wear for mechanical components or declining failure rate for electronic
components. This distribution does not appear very frequently in the analysis of accelerated life
data.
Example 3.11 - Consider the following time to failure data set of a dishwasher. Eight
systems were run under accelerated conditions with times to failure of 23 hours, 48 hours, 67
hours, 100 hours and 134 hours. The remaining three units were unfailed at 140 hours. What is
the failure rate?
Solution - Past experience has suggested that the dishwashers followed an exponential
distribution in time to failure. A simple hand calculation of hazard rate shows.

5 _ Failures 5
= = = 0.0063 fails/hour
23hrs.  48  67  100  134  3(140) 792

A check for hazard rate, , and time offset, , was made with ReliaSoft Weibull ++ 6.0 and the
two-parameter analysis with MLE showed best fit values of:

 = 0.0082 and  = 23.00

These are probably a better fit to the estimates calculated by hand. As with any accelerated life
test, the failure modes should be reviewed after failure analysis and the information compared to
past history.

3.7.2 The Weibull Distribution - This distribution has up to 3 unknowns to describe the various
possibilities. These are defined as: a scale factor, , a distribution shape parameter, , and
location parameter, , also called a time offset. The three-parameter Weibull Probability Density
Function, PDF, and Reliability may be written as:

30
 t 
 t    1 (  )
f(t) = ( )( ) e PDF, where t > ,  > 0 and  > 0 (3.19)
 
t 
( )

R(t) = e Reliability

Complete statistical solutions are possible for both two and three-parameter Weibull models for
both the Rank Regression and Maximum Likelihood approaches. It is the two-parameter Weibull
that we typically graph and employ with Weibull analysis. Examples of the analysis of two and
three parameter solutions will be provided in Section 7 when multilevel ALT data is considered.

3.7.3 The Normal Distribution - This distribution may be described by up to 3 unknowns. These
are defined as: a mean, , a standard deviation, , and location parameter,  not shown in
Equation 3.20. The two-parameter Probability Density Function, PDF, is written as:

1 t 
 1  ( 2 )( 
)2
f(t) =  e PDF, where  > 0 (3.20)
 2 

The reliability function is not easily described and usually tables are employed. The Normal
Distribution often appears in ALT through a component or system with many failure modes.
Examples include batteries and incandescent light bulbs, both of which often have a normal
distribution in time to failure.

3.7.4 The Log-Normal Distribution - This distribution may be described by up to 3 unknowns.

These are designated as a Ln mean,  ' , a Ln standard deviation,  ' , and location parameter,
labeled  ' and not shown in Equation 3.20. The two-parameter probability density function is:

1 t  '
 1  ( 2 )( '
)2
f(t) =  '  e where  ' > 0, t > 0 (3.21)
 t 2 

Where
( ' ) 2
' 2
Mean = e e

The Lognormal distribution often appears in ALT through a component or system with many
failure modes subject to slow degradation. The Lognormal distribution may appear in the
analysis of bearings, semiconductors or some materials.

3.8 Other Distributions for Reliability - Distributions, such as the Binomial, Poisson, Extreme
Value or Logistic Distribution, may be employed to describe a specific reliability situation. These
are less common and will not be discussed in this monograph. Reference 12 has examples.

31
3.8.1 Spliced Solutions or Multiple Straight Lines - Some ALT problems have unusual
behavior and are best described by the "multi-Weibull" or "mixed-Weibull" models. These are
often tied to failure modes or time-dependent failure mechanisms uncovered in the test. A few
examples of these will appear in the analysis of ALT data section. Other complex solutions exist
based upon the three-parameter Weibull. These appear as combinations of two and three-
parameter distributions from the list above and can cause interesting analysis problems.
Additional complexity may exist, when multiple stresses are present in a ALT. The
presence of two or more stresses may lead to an interaction between the stresses which is
reflected in the behavior of the data. At times it can be very difficult to separate the effects of
interaction when few failures have occurred or when it is difficult to separate the impact of each
stress. A few examples will be shown in Section 7. The reader is referred to references 7, 13, 14
and 22 for examples and details.

32
4.0 Administering the Life Test

Actually running the ALT is important to getting data and results. Performing an
accelerated life test and administering the details of the test have been planned, is an important
task. Selecting the best stresses and environmental operating conditions is the next most
important task to sample size. The list of likely or critical stresses may include any of those listed
in Section 4.1. Other, more exotic stresses, do exist for some applications. The selection and use
of these are similar to the ones shown here. The problem is how to determine what the "best
stresses" should be for any particular situation. Consider the following methods for selecting
these best stresses. Best is a relative term, and often people are limited by the equipment at hand.

4.1 Methods for Selecting Test Stresses

Start with the selection of the stresses that are standard for the particular customer,
environment or market. Consider also the relevant specifications from government organizations
or industrial specifications. Keep in mind that these stresses have typically been established
typically more than 10 years prior may not be as relevant today. Even though, a review of the
historical practices is often the best place to begin. Examples of the standard type stresses with
associated failure modes are included in Table 4.1. These are closely tied to Physics of Failure.
Table 4.2 shows the process for selecting a stress for a given failure mode or mechanism.

Table 4.1 – Failure Mechanisms and Stresses

Failure Mechanism Applied Stress
Electromigration Current Density or High Temperature
Thermal Cracks Dissipated Power or High Temperature
Corrosion Humidity or Temperature
Mechanical Fatigue Repeated cycle of vibration
Thermal Fatigue Repeated Temperature cycles

Table 4.2 - The Selection of Applied Stresses

A) Select standard environmental stresses, including items such as high temperature,

low temperature, random vibration, dust or humidity.

B) Select cyclic stress conditions, such as temperature cycling, humidity cycling, stress
cycling or sine wave vibration.

C) Select stresses that are frequently called out in the Military Standards and have
been found useful to demonstrate conformance to stringent military customer
environments.

D) Select stresses based upon tradition for the market or business. These may be unique
to an industry or situation.

E) Select stresses based upon the expected customer environment or industry. This
includes consideration of worst-case customers, abusive customers and customer who fail
to perform suggested maintenance.

33
 F) Select stresses based upon known or anticipated failure modes or physics of failure
for critical components or system functions. Such stresses are thought to dominate the
operation of a system

G) Select stresses based upon customer safety or liability considerations.

H) Select stresses based upon long term degradation modes such as corrosion or
material degradation.

Standard stresses may be very useful and are often employed in some industries and
companies. These standard stresses then serve to be a test of robustness of the design and
manufacturing processes. However, they may or may not correlate well with the current customer
use environment, the market needs or reveal potential field failure modes. It takes time to develop
the important tests and to demonstrate that they work well to identify customer problems. It may
be a few years before the tests become a specification that the industry will accept and most
people begin to use. About 10 to 15 years after the initial use of any stress to improve products in
the customer markets, there is sufficient information to create a meaningful specification. But
time may have passed for technology to change, the customers' environment to change or
knowledge about the usefulness of the stress to change. The specification may already be obsolete
in a quickly-changing situation. The following example, based upon personal experience, echoes
this possibility of slow development.

Example 4.1 – The following three example show what can happen with ALT. The first example
describes an implantable biomedical product that is subjected to a centrifuge (i.e. spin) test at
high g force level. This test had been employed for years as a "proof of robustness of all new
designs" at one company. No one there remembered or knew what this test did as a stress, or
whether or not it was relevant. It did not reveal any failure modes of interest or any new
information over the period of 30 years. It was once a standard industry test, when products first
were marketed. In more than 15 years no product had ever failed the test or appeared to be
degraded by the stress conditions. Every new product had to go through this stress as part of the
design qualification. The company decided that it was easier to continue this meaningless test
then to try to explain to the regulatory agency why it should be dropped. Tradition is not always a
good thing.
Most standard tests have much more value than the one just noted. A second example is a
standard test condition for a high temperature bake of 85ºC for 168 hours. This test was
developed to look at some of the stability characteristics of an integrated circuit. This test
occasionally produced a few failures and showed some long term degradation. These potential
degradation failure modes could be correlated with lot-to-lot performance because some of the
failure modes were sensitive to high temperature. Thus, this standard test produced some useful
information on future failures in-house and possible correlation to the field. It also provided quick
feedback for future process changes and supports improvements.
A last real-life example is a vacuum cleaner that was tested to an industry standard of
Accelerated Life Test conditions. Ten samples were operated and their performances observed
and measured over time. At the end of the test, any design that passed without showing
degradation would be acceptable. Any design that showed problems or too much performance
degradation would need to be improved before it could be released. This test was correlated with
eventual field performance and was considered an essential test by the company.

4.2 Military Standards

34
 At one time, there was a plethora of military standards that described the suggested
testing of components, assemblies and systems for the military markets. These tests were often
developed over time as a solution to specific military hardware problems. Sometimes these
problems were failure to perform in the field at temperature or under vibration. At other times,
these standards were ways to identify a weak design or a flawed manufacturing process. In any
case, military standards issued in the 1960s and 1970s became well-accepted by military suppliers
in the 1970s. Eventually, these standards spread to other industries such as NASA, automotive
and biomedical by the 1980s and continue to be used today. The need for highly reliable
components and systems drove the use of these standards in a variety of other markets and
applications. These standards, through their recommended stresses and test conditions, tended to
correlate with specific failure modes when they were developed. The customer environments and
expectations of the current decade were two driving forces for selecting stress types, stress levels
and stress duration. Yet, time did show the stress combination to be harsh enough to identify
some design weaknesses and even some systematic manufacturing or supplier component
problems. Eventually, the consumer, commercial, and automotive markets, through design
improvements and improved technology, made these original stress conditions, effectively
obsolete by 1990. Some of these original military standards proved to be less meaningful and
effective for the product development cycles of the 1980s and 1990s. It is no surprise that
knowledgeable people began looking for something better. In some cases, the older military
standards were updated and made more relevant. That was the ultimate point of the R&M 2000
process when it began in the early 1990s.
In some cases, industries had developed their own reliability methods and standards
based upon earlier military standards. Examples include FMEA and ISO type standards in
industries such as automotive, telecommunications, gas industry, commercial aviation and even
the semiconductor industry. The abundance of standards were developed to aid work with
component suppliers, improve the internal development process, create more robust products and
deliver long-lived, low life-cycle cost products for the customer. This process has been one of
learning, combined with trial and error for over 40 years now. Market acceptance, reduced
development costs, reduced development time, international competition and the applications of
new technology have driven most markets. Yet in the late 1990s, the unimaginable happened -
some military specifications were recognized as obsolete and abandoned without being replaced.
The plan at that time was to improve the procurement side for components and assemblies for
military application by issuing updated specifications for Commercial Off The Shelf components,
or COTS, using equivalent standards. The older specifications and standards disappeared without
being replaced in a timely fashion by the COTS specifications as originally planned. Thus, once
again reliability professionals are on their own for suggestions and help when developing
Accelerated Life Tests. The following are real examples of the use of past test standards; some
are good and some show the all too human element.

Example 4.2 - At one large military supplier in the late 1970s, a system was required to
go through a 24 hour acceptance test which consisted of 6 gs of three-axes vibration combined
with temperature cycling between -40ºC to 100ºC while operating. No electrical performance
failures were permitted during the 24 hour test period. The levels of vibration were so high that an
occasional integrated circuit would fly off a printed wiring board during the vibration and high
temperature portion of the 24 hour test. This was usually as a result of the accumulation of fatigue
from the vibration and heat. This acceptance test was thought to be a good measure of the design
robustness and a measure of any residual manufacturing defects. Some correlation was obtained
between the occurrence of a failure during the acceptance test and the eventual field performance.
That is, a system failing the acceptance test would be repaired and run back through the
acceptance test. Any system requiring three passes through this acceptance test before shipping to
the field were also discovered to be problem systems in the field. No one looked at the units

35
passing the acceptance test to see if they failed more often than might be expected. It could be
argued that this 100% test left the "good systems" passing, though with some degree of latent
damage. Three passes through this 24 hour test with repairs certainly could cause additional
damage. The stress level of the acceptance test probably shortened the life of the good systems. It
is wise to look at the optimum stress screen time and conditions as many of the authors of ESS
and Burn-In books and articles later suggested [8, 15, and 17]. Here, the reliability knowledge
and theory required to understand and avoid such problems followed the practice by about 10
years.
Often the application of military standards can be an effective approach to Accelerated
Life Testing if the "best stresses" are applied. Rugged customer environments and those requiring
high reliability are often the best application for these standard military stresses. It is easy to
overuse strong stresses such as those used in military standards. All stresses should be applied
with some care and good engineering judgment. A cause-effect relationship between stress and
customer environment should apply [15]. Be sure to correlate the stresses with field performance,
both positive and negative when running life tests and stress tests.

Tradition is an all-too-frequent reason to select specific Accelerated Life Test stresses.

That is, doing the same thing as has always been done, even though the designs, materials or
technology may have changed. Be careful, this argument in favor of keeping to tradition may be
subtle. As materials, designs or technologies evolve; some thought should be given to evolving
Accelerated Life Tests as well. Example 4.1 provided evidence of problems with maintaining
tradition and not knowing why.

Customer Environments or Industrial requirements often provide the biggest reasons

to change or add to Accelerated Life Tests to product development. These requirements may be
characterized by new market applications, new designs or new technical applications of old
designs. When this happens, greater reliability numbers and more thorough tests are often
required.

Example 4.3 - The following is another real-life example of a new requirement driving
reliability improvement in new markets. An existing piece of scientific equipment was employed
to support a production line part time. New market applications of the products being
manufactured called for "in-line" 100% use of the instrument. The line support application was
recognized by customers. This instrument would be typically used eight to nine hours per day,
five days a week with an availability of 0.75. This is, the equipment should be able to be used at
least 75% of the time (of 9 hours per day) when desired. In essence, the desired information could
be actually obtained within 24 hours with little overall impact on design or manufacturing. This
situation actually dropped the required Availability to about 0.38. The new “in-line” system
requirement would be operation of 24 hours, 7 days a week with the Availability requirement of
about 0.95. This is a much more difficult target. The equipment would be employed more than
three times as often in this role than in a line-support function. There would be little time for
planned maintenance and low tolerance for equipment being in a failed state. The achievement of
the new market requirements could be easily accomplished through fewer system failures. That
is, improve reliability or establish longer Mean Time Between Failures, MTBF, for the
equipment. Less importantly, shorter down-time or Mean Time to Repair, MTTR, for any failure
reason would also increase the availability. This MTTR approach could become more important
than the reliability improvements as it might be more easily obtained through equipment design
changes and other design improvements. Items such as MTTR and MTBF measures do not
normally appear as part of an ALT measure. Thus, any Accelerated Life Test would need to focus
upon the specific applications for repair as well as the Mean Time Between Failure. The older,
now obsolete military standards did have test conditions for the measurement of MTTR. The

36
newer Semi-E10 standard which, governs the use of similar production equipment, hardly
mentions or identifies the MTTR measure as important to achieving reliability goals. In some
applications when MTTR can be short and the economic consequences of failure small, then
measures such as Availability can be valuable.

4.3 The Relationship of Field Failures and Product Development Activities

Field failure mode information is very important to the selection of any stresses for an
Accelerated Life Test [15]. This is because the reliability is often dominated by a small number of
failure modes which are often tied to just a few environmental stresses. Identifying these critical
few stresses and their potential impact on the system becomes an important early step for setting
up any meaningful Accelerated Life Test. This is one aspect that the traditional Physics of
Failure approach employs when determining the relationship between the applied stress and
eventual system failure. The applied stress is coupled first through a latent defect and then a
failure mechanism that propagates the flaw until it becomes large enough to become a system
failure. Once the failure is recognized by the customer, it may then be labeled with a failure mode
(see Figure 4.1) . This sequence is commonly documented by a Failure Mode Effects Analysis
(FMEA) or Fault Tree Analysis, FTA. The order is different in an FMEA. This document starts
with the failure mode and works backward to a root cause. Thus, the selection of Accelerated Life
Test stresses should be closely related to expected field failures and failure modes.

Failure Flaw System

Flaw
Mechanism Growth Failure
Failure Failure
Mode
Analysis
Defined

Figure 4.1 – The evolution of a Failure

One model of a system could be that of a marble bag. Each marble represents a failure mode
of the system. The shading of the marbles relates to the applied stresses. So each marble represents a
failure mode as activated by a stress, either a normal operating condition or a high stress condition.
Only some failure modes are evident as a result of a given single stress. Multiple applied stresses
may lead to only a few important failure modes. One of these failure modes could be the result of
interaction between two applied stresses. For any given system, there are about 20 failure modes total,
with only four or five dominating as a result of a single applied stress. Overstress conditions and
acceleration may make some of the marbles evident more quickly.

37
 Figure 4.2 - The Marble Bag

The development process of any complex hardware or software system is really a series
of difficult choices [15]. This might be described as a balancing act between the "Eight Project
Factors" which are system performance, schedule of development, cost to develop, system
quality, system reliability, features, long term field support (or life cycle cost) and field warranty
failures. The object is to optimize the combination of all eight factors at the same time. No single
factor can be neglected or else the others must bear increased expense or problems. No single
factor can be improved by itself, except at the expense of some or all of the others. This model
shows a complex juggling act during the development process. The object is to maximize the
combination of all factors in order to achieve low cost and reliability. Accelerated Life Testing
directly relates to the quality, reliability and performance issues as part of the development. Items
such as schedule, features, field failures and life cycle costs are indirectly involved with the ALT.

38
 Schedule Project
Cost
Quality

Reliability

Field
Support

Features
Warranty
Performance

Figure 4.3 – The Eight Dynamics of a Development Project

Often, the corporate development philosophy has a strong impact on how and when ALT
is performed. In the real world, a reliability professional must work within these constraints while
trying to provide some objective evidence of why ALT should be increased in time, sample size
or be changed in some way. Economic reasons may go a long way to help this discussion. The
following section provides some information about the relative value of various applied stresses.

4.4 Effectiveness of Environmental Stress Screens

The use of environmental screens as part of the manufacturing process provides

information about ALT stresses that could be used to improve the overall product development
process. Think of the information in Table 4.3 as evidence of areas to improve. This information
was derived from a large number of sources and so reflects the thinking across the years [8, 16,
26]. It still has some use for testing today and in the future. The table shows that there is a residue
of problems that often remain in systems and that it is the stresses that reveal the hidden flaws.

Table 4.3A - The Value of Various Applied Stresses (US Study)

Applied Stress Ranking

Temperature Cycle 360
Random Vibration 240
High Temperature 200
Electrical 150
Thermal Shock 140
Sine, fixed freq. vibration 130
Low Temperature 120
Sine, swept 110
Combined environment 100

39
 Mechanical shock 80
Humidity 60
Acceleration 40
Altitude 20

The higher the ranking number; the more important the stress.

Table 4.3B - The Value of Various Applied Stresses (French Study)

Applied Stress Ranking

Temperature Cycle 150
High Temperature 100
Room Temperature operation 80
Electrical 75
Thermal Shock 70
Low Temperature 65
Sine, swept 60
Random Vibration 55
Humidity 50
Sine, fixed frequency vibration 40
Mechanical shock 35
Acceleration 30
Combined environment 25
Altitude 20

The higher the ranking number; the more important the stress.

4.4.1 - Discussion of the Stress Effectivity Identified in Table 4.3

Temperature Cycle - This stress has been identified by a series of studies as the best
single stress [8, 16, 26] to select when trying to identify the main failure modes of a system. This
stress works well for both mechanical systems, electronic systems and most components. It may
be because this stress "stresses" the materials as a result of the temperature gradients and causes
sufficient internal stress to initiate and propagate both cracks and other defects in structures. This
stress tends to follow a simple power law relationship in total temperature change, T. The stress-
life relationship can be described as:

Life = B(T)  N (4.1)

This simple model has a scale factor B, were T is the total temperature change, and N is a
physical constant associated with materials. The model can be expanded to allow for more
complicated geometry, various failure mechanisms, and include additional effects such a rate of
temperature change. In the end, despite the increased possible complexity; the formula always
looks something like Equation 4.1.
The Coffin-Manson relationship describes mechanical crack growth driven by thermal
cycles. The equation for Coffin-Manson shows a similar relationship to Equation 4.1. The thermal
cycles to failure are labeled N TC , the cycle frequency is F and T represents the temperature
difference. Thus Coffin-Manson can be written as:

40
 N TC = A(F) b (T) c (4.2)

Where A, b and c are constants that depend upon material and how the thermal cycle is applied.

Max.
Temp.
Temp.

T

Min.
Temp.

Figure 4.4 – Temperature Cycle

The temperature stress need not be uniform across a material, component or system.
Depending upon mass and geometry, there may be stress concentrations in a system and these
will dominate the failure modes.

Vibration - This is one of the best stresses for many systems because the vibration leads
to the accumulation of fatigue usually in some weak aspect of a component or system. It works
well for mechanical and electronic systems for which structure or mechanical integrity is
important for proper system operation. When applied with temperature or temperature cycling,
we need to be concerned about interaction of vibration with temperature. At high temperature,
when many materials approach a Type I phase change, such as melting, the result of the
combined stresses rapidly exhibits non-linear behavior. Some metals are subject to creep at a
temperature above half their melting temperature. This is often just above room temperature.
Thus the failure mode would be a deformation associated with temperature. Vibration at high
temperature exacerbates creep, basic crack growth and the accumulation of fatigue in general. At
low temperatures, vibration may become important especially when Type II phase changes are
present. Type II changes occur when atomic, crystalline or structural rearrangement is present.
Solder is a good example of such a material because it is not obvious that there are internal
structure changes on going.
Vibration may be applied as a random event, as a swept sine stress or even a fixed
frequency sine stress. All of these stresses may produce different results for a system. As a
complicating factor, both narrow-band and wide-band frequency processes may be described by a
simple power-law relationship as shown in Equation 4.2. Examples in a later section will show
the use of these processes.

N
Life = B(Stress)

The stress may be measured by a power spectral density, a sine frequency, or the envelop of
stresses described by a g force level.

41
 High Temperature - This has long been tied to burn-in and shelf life tests for
components and systems. This stress has been most commonly employed for ALT for a variety of
products ranging from electronics to mechanical structures, most materials and even drugs. When
temperatures are static, the life usually follows the Arrhenius relationship in temperature. This is
simply:

Ea
K bTk
Life = A e (4.3)

where A is a scale factor, E a is the activation energy in electron volts or eV, K b is Boltzmann's
constant or 8.62 x 10 5 eV/ºK and T k is the temperature expressed in ºKelvin = 273 + ºC. This
formula works for many electronic and mechanical components as well as diffusions processes.

Electrical or Operating Voltage - The electrical stress may appear as either an absolute
operating voltage stress for some components, as a current density or sometimes as an on-off
stress. For incandescent light bulbs, both the operating voltage and the on-off cycle may be
important in determining the absolute life and many of the failure modes. This stress typically
follows a power law relationship. It may be expressed as:

Life = B(Voltage)  M (4.4)

or Life = B (J)  P

where B is a scale factor, M and P are physical constants that are typical of the materials
involved, J is a current density, typically amps/cm 2 or amps/cm 3 . The negative signs of P and
M indicate that the life gets shorter as the stress increases.

Thermal Shock - This stress is similar to the thermal cycle one. The main difference is
that the effective rate of temperature change is greater than 30ºC/minute for the components or
systems involved. The equations are often similar to those observed with temperature cycling.
Temperature shock is more effective than temperature cycles at discovering latent failures. This
may be accomplished by a rapid rate of temperature change of the air, by a movement from a hot
condition to a cold condition or by moving from a hot liquid to a cold liquid. Each of these three
conditions will produce different results.

Sine or Fixed Frequency Vibration - This stress is considered less important then
random vibration because of the limited stress frequency range involved. Typically there are
limited applied g forces present with this type of stress. This tends to follow a power law
relationship. The application of this stress is typically performed at one or more system
resonances. Most of the time, the energy from the single stress frequency is poorly coupled into
the system, so little fatigue accumulates. Fixed frequency vibration can be effective when it is at
one of the resonances of the system.

Low Temperature - This is often treated as an important stress for some systems and
components. Many systems react to this stress because a low temperature limit crosses a phase
transition of one of the materials in use in the system. Solder is a good example of such a
material. Low temperature can cause failures because the thermal change causes high stress at

42
some point of the design. A second reason for failure is that some materials become brittle at low
temperature and easily fail from ordinary operating stresses at this temperature.

Swept Sine Vibration - This is considered less effective than fixed frequency sine
because most of the time the applied frequency is not close to a system resonance and so little
damage may be accumulated as a result of the stress. Swept sine is one way to find the system
resonances for a future dwell.

Speed, Frequency or Load - These stresses are common, especially for mechanical
systems. Although not appearing in Table 4.3, any one of the stresses may be tied to a specific
failure mode and have a direct impact on life. These typically follow a power law relationship.

Combined Stresses - Combined stresses are considered by some past studies to be less
effective than other single stresses. This may be because the applied stress levels in the 1980s
have traditionally not been high enough to see failures modes quickly. By the 1990s, high
combined stresses became more common, hence the development of HALT, HASS and ESS as
techniques. Future surveys will find these approaches to ALT rated higher on the scale of
effective test methods.

Mechanical Shock - This stress is not commonly used by many reliability people except
when testing packaging material. It is employed primarily with shipping tests due to concerns
about damage from a drop. Vibration tends to be the mechanical stress of choice.

Humidity - This stress is the least commonly used stress and typically follows a simple
power law. Humidity often appears in material evaluations and ALT for simple components. For
plastic integrated circuits and many material corrosion situations, humidity versus life follows a
simple power law. Systems tend to be too complex to describe a simple relationship. When used
with high temperature, non-linear effects must be considered. Many ASTMs tests are based upon
temperature and humidity as a combination. Here RH stands for relative humidity.

Life = B(RH)  M

Ultraviolet Light or Chemical Degradation - These stresses, while not in Table 4.3, are
often strongly associated with materials and products that are exposed to the outdoors. A Physics
of Failure approach may be desired with these stresses. Corrosion, chemical aging or internal
structural changes may all occur. These tend to follow a power law relationship between life and
stress.

Biological Stresses - These stresses are not usually considered and are sometimes
required when a product must work in a high-humidity or condensing environment. Sterility and
shelf life environments are also important. These may measure resistance to biological attack or
the growth of simple biological products. Occasionally, a simple attack, such as by cockroaches,
mice, squirrels may need to be considered.

Other Stresses - Special stresses tied to markets or environments are sometimes

employed. Examples are altitude, beta radiation, dust, ESD or air flow. Each needs to be
developed for a specific product or market. These tend to follow a simple power law.

4.5 - A Few Stress Application Examples

43
 A few examples will help show how some of these stresses may be applied. Not every
important stress is showed here. No interacting stresses are included in this section, but examples
of interaction are in Section 7.

Example 4.4 - Let the humidity stress-life relationship be determined by the following:

M
 RH 
Life = B   (4.5)
 85% 
If the applied stress or humidity is raised from 30% to 60%, and M = 3.0, how much is the life
shortened? We can write the equation for the Acceleration Factor, A.F. as the following:

30% 3.0
) B(
Life1 85 %
A.F. = = = (2) 3 = 8.0 (4.6)
Life 2 60% 3.0
B( )
85%

This equation reflects the fact that, for the two different life conditions based upon Equation 4.5,
we can write them separately and then divide the two equations to get Equation 4.6.
The life would be shortened by a factor of eight when the humidity is doubled as shown by
Equation 4.6.

Example 4.5 - The breakdown of oxide in a CMOS integrated circuit has been suggested to
follow the formula:

C
(V1 V2 )
t ox .
AV = e (4.7)

where A V is the Acceleration Factor for operating voltages, V 1 and V 2 . The term t ox. measures
the oxide thickness in Å, and where the constant, C, is 300 Å ,and V 1 is the accelerated voltage
with V 2 being the standard operating voltage. Let t ox. be 400 Å , with V 1 be 8 volts and V 2 be
5.0 volts. Completing Equation 4.7 with the information provided yields:

A V = e ( 0.75(( 3.0 ) = 9.488

That is, each operating hour at 8 volts is equivalent to 9.488 operating hours at 5 volts. Clearly,
there is an unidentified upper limit to the application voltage for this type of test. Likewise, there
may be a temperature dependence or effect as well.

Example 4.6 - Crack growth is an important failure mechanism in many components and
systems. There are several models available to describe different aspects, but for short cracks the
following model from Carter [18] works well for stress driven crack growth.

da
= A ( p )  (d - a) (4.8)
dn

44
where A and  are material constants,  p measures the plastic shear strain, a is the initial crack
length and d is the shortest crack distance that leads to system failure. The term, n, measures the
number of applications of some external load. This load may be thermal cycles, mechanical
stress, internal heating or even the adsorption of humidity, causing swelling. All of these stresses
may lead to crack growth and development and a "common failure mode". When two or more
stresses are present, interaction between stresses may occur and confuse the results of the ALT.
Statistical analysis at the end of the test may need to look at this carefully and take this into
account during the analysis.
Let A = 2.5,  = -4.3, a = 0.007 inch, d = 0.050 inch,  p = 8.5, what is the crack
da
growth per stress cycle, or ?
dn

da
= A (  p )  (d - a) = 2.5 (8.5) 4.3 (0.043) = 1.084 x 10 5 inch/stress cycle
dn

High Stress
Area
a

Figure 4.5 Crack Growth in a Plane

Example 4.7 - Vibration with intrinsic damping may be a common situation for many
mechanical systems. The following vibration equation from Kececioglu [8] is a simplification of
a more complex equation that works in many situations when light damping is present in a
vibrating system. The equation describes the ratio of the standard deviation of the damage
accumulation,  d , to the total accumulated damage as measured through the Damage
Coefficient, D. This ratio is given by:

d 1 (b)
= (4.9)
D ( )(nY )

where  1 is a constant from Table 4.4,  measures the level of damping and n Y measures the
cycles accumulated during a specific period of time. Let the light damping be,  = 0.05, let n Y =
1909.85 stress cycles accumulated in 500 hours. The value of D = 0.9097 must be derived from
the Equation 4.10 below, while  1 = 3.11 was taken from the Table 4.4.
The value of the Damage coefficient, D, depends upon the Palmgren-Miner model of
cumulative damage. This depends upon the following factors:

45
 1) Narrow versus wide-band stressing during accelerated vibration.
2) Static versus dynamic loads during vibration.
3) Unidirectional versus full-reversing application of stresses.
4) Compression versus shear in stresses.
5) Type of material, ferrous versus non-ferrous or plastic materials.
6) Finish of the material under stress.
7) Geometry or stress concentration factors.

For a narrow-band, single peak situation, we can write the equation in the following example. It is
expressed as:

n T b
D=( )( ) ( y 2 ) b ( + 1) (4.10)
2 A 2

where:
T is the duration of the stress, here 500 hours
 n is the natural frequency of the assembly, here 510 Hz.
 y is the Raleigh scale factor, here 150 MPa.
b is a Palmgren-Miner exponent, here 6.5
A is a material constant, here 1.767 x 10 24
 is the Gamma Distribution

n T b 510 500
D=( )( ) ( y 2 ) b ( +1) = ( )( ) (150 2 ) 6.5 (4.25)
2 A 2 2 1.767 x10 24

D = 0.9097

Table 4.4 - The Critical  Factors for Vibration Testing, from Kececioglu [8]

b  1 (b)  2 (b)  3 (b)

1 0.0414 0.00323 0.0796

3 0.369 0.0290 0.2120
5 1.280 0.0904 0.6790
7 3.720 0.2230 2.3300
9 10.700 0.5180 8.280
11 31.50 1.230 30.00
13 96.70 3.060 111.2
15 308.00 8.110 415.0

Filling in this information in Equation 4.9, we have:

d 1 (b)
=
D ( )(nY )

46
 3.11
 d = (0.9097) = 0.16417
(0.05)(1909.85)

We can use the relationship for the Normal Distribution, Equation 4.11, since both D and  d are
treated as if they are normally distributed. We can use the normal distribution to estimate
reliability under these conditions at any point in time. With the values of  d = 0.16417 and D =
0.9097, we have:

1  0.9097
so R(t) 500 =  [ ] =  [ 0.5499 ] = 0.7088 (4.11)
0.1642

Thus, we have an estimate of the reliability of a product under certain conditions of stress. It is
helpful to have such an estimate before the start of a test because it suggests how long the test
might last and may help with estimating results. All that was needed was a description of the
natural frequencies of the product, material constants and an idea of the internal damping.

Example 4.8 – Now let  = 0.03, let n Y = 1200 stress cycles accumulated in T = 600 hours, with
b = 7.5 and  n is the natural frequency of the assembly, here 620 Hz.,  y is the Raleigh scale
factor, here 160 MPa and A = 1.61 x 10 23 . Calculating the value of D first gives:

n T b 620 600
D=( )( ) ( y 2 ) b ( + 1) = ( )( ) (160 2 ) 7.5 (4.75)
2 A 2 2 1.61x10 23

D = (98.676)(3.727 x 10 21 )(4.568 x 10 17 )(9.478) = 15922.6 x 10 4

D = 1.5923

5.47
And  d = (1.5923) = 0.6204
(0.03)(1200)

1  1.5923
We have R(t) 350 =  [ ] =  [- 0.9547 ] = 0.518
0.6204

This situation is less reliable than the previous example.

4.6 - Suggested Rules for Performing an Accelerated Life Test

Many accelerated tests are defeated before they really begin. This is often due to the way
they were set up and conducted. The following Table 4.5 list of suggested ground rules will help
avoid some of the most common problems associated with ALTs. They are good generic advice,
but as always may need to be adapted for specific situations.
Table 4.5 – List of Suggested Ground Rules for ALT

47
 1) Verify the upper limit of the stresses by running at least one part through the full
stress conditions before committing the whole sample. This also provides an opportunity to verify
the equipment, calibration and all fixtures before committing all of the samples.

2) Verify all samples are working correctly at the start of the test as well as in the
fixture in the environmental conditions in which they will operate. That is, verify that they work
initially at room conditions and at the start of any environmental stresses.

3) Closely monitor the samples through the first few accelerated test cycles. Make sure
they are operating correctly. This usually means starting the test early in a work day.

4) Never start a life test on a Friday or a day before a holiday, if no one will be
available to monitor the test during the second day. If it is necessary to start a test on a Friday, the
test must be monitored on Saturday to verify the samples are still operating correctly.

5) Read important performance measures for the test samples during the test. If
measurements cannot be taken continuously, then make measurements at intervals. That is, take
parametric measurements so the drift can be monitored during test. Do this within 24 hours after
the start of the test and a few days later. The following times to check are suggested if continuous
monitoring is not possible: initial reading at nominal conditions, initial reading at stress limits,
and read at 24 hrs, read at 72 hrs. These would be at about 10%, 25%, 50%, 75% of the expected
total test time and then at the end. These are suggestions. The timing can be modified to fit
schedules, technology and materials. The issue is that periodic measures should be made to
monitor drift well before any failure occurs.

6) Verify the measuring equipment is capable of the precision required. Do this by

capability studies before the life test begins. Be sure they are calibrated. Check the equipment at
least half way through the ALT and at the end to confirm the same capability is still present. Use
known good samples, sometimes called "golden units" to verify measurement stability. Be sure to
look at repeatability of the measures from a statistical perspective.

7) Prepare for sudden and unexpected failure. Assume that the test equipment, test
chamber or a large fraction of the samples will suddenly and unexpectedly fail during test.
Determine the back up plan for this possibility before the ALT begins.

8) Verify the test conditions for stress before, several times during the ALT and at the
end of test. That is, use independent measures such as thermocouples, voltmeters or humidity
gauges to confirm the chamber conditions, voltages and samples are as they should be. Look into
any closed chambers periodically.

9) Place some known good and marginal samples in with the life test units if such
items are available. These special units are not part of the life test, but serve as a comparison to
the units being tested. Sometimes this is best done by combining Design of Experiments with the
standard life test. This latter condition may apply only if sufficient knowledge and samples are
present.

10) Assume a fire or similar catastrophe will occur during test. What will you do
before the start of the test to protect the samples (with fuses, thermal cut-outs, etc.), the test
equipment and the laboratory (have fire alarms, sprinklers, etc.)? Remember, these unpleasant
events seem to occur when no one is around and usually take some time to develop. Even test
chambers and test loads eventually fail from continued use by many ALTs.

48
 11) Consider the possible results before the test begins. Create hypothetical best and
worst case scenarios and look at times to failure under these circumstances. This helps visualize
when significant events may occur. This is especially important in a multiple-level stress test.
Here, the highest stress condition should typically have the first major change and first failure.

12) Start any data analysis after the second failure of each of the various stress
conditions. This may seem a little premature, but permits an initial consistency check to occur
while the test is still in operation. The first failure does not provide more than a crude estimate of
statistical characteristics. The second failure permits the start of trending and provides better
estimates. If a problem with test samples or test conditions has occurred, then this early analysis
will help identify it.

13) Verify the failure modes by performing analysis of the failed units as they occur.
Don't wait until the end of the test to look at failed samples. Be sure to check the failure modes
between the different stress levels and verify they are as expected.

14) Perform analysis on some of the non-failed units from each stress level at the end
of test. They often provide useful information about the propagation of the failure modes as well
and help with identifying mechanisms and degradation information.

15) Never assume that the failure modes are caused by "bad parts" or explain away
failures as "can never occur in the field because...". Both of these are forms of denial that is
sometimes used as a reason for not improving a product, not making changes or a desire to stick
to a tight schedule. All failure modes derived from an ALT are meaningful and bring value. Be
sure to perform the analysis on all failed samples and don't assume failure causes based upon
failure modes or failure symptoms.

16) Retest any design improvements and confirm they have the desired effects. This
should be done even if the product is in production. After all, it is better to find out about a design
flaw in production by an Accelerated Life Test than by waiting for customer complaints to arrive
from the field.

17) Perform a detailed and thorough analysis of all of the data at the end of test. Make
use of all the data available, including history of similar products and tests.

4.6.1 - The most important rules

Five of the proposed ALT rules are more important than the others and will be repeated
here since they are so essential for successful completion of an ALT and start the data analysis
phase. The original rule is underlined.

Rule 3) Closely monitor the samples through the first few accelerated test cycles. Make
sure they are operating correctly. This usually means starting the test early in the work day. This
rule helps ensure the ability to make parametric measurements, and even, later in the test, to be
able to follow a trend in measurements that occur. The data collected on the fly during test may
suggest one of three possibilities before the end of test. These include:

a) The test may be ended early because there are significant results that suggest
conclusions. Unfortunately, this happens most often with a failed test. This is, a test in
which unexpected and "bad" results are the quick outcome. Remember the fundamental

49
 law of reliability - "Bad news happens fast, but good news usually takes a long time".
Occasionally, we are able to end a test early because a positive result has been reached to
some degree of confidence.

b) The test will run as planned because results are within the acceptable range for
numbers of failures, size of parametric changes or failure modes. Most tests fit in this
group.

c) The length of the test is extended because the anticipated test results or failures
have not yet occurred. This possibility happens with robust designs and with specialty
testing such as Accelerated Shelf Life Tests. These latter tests are often open-ended, it
is necessary to run the test until there is evidence of degradation and/or failures.

Rule 5) Read important performance measures for the test samples during the test. The
goal is to make periodic measures to monitor drift or degradation. When these are not available,
identify failures at the best time estimates if samples cannot be monitored continuously. It was
initially suggested measurement points at initial, initial at extremes, after 24 hours, 72 hours and
then at about 10%, 25%, 50%, 75% of total test time and near the anticipated end of test. These
times were suggested when continuous monitoring was not a possibility. The review of Rule 11
will suggest why the spread of times proposed is a reasonable choice for most ALT situations.
Special tests, limits of test equipment or unusual circumstances may suggest changes to these
percentages. The greatest problem is having to interrupt the test conditions to take readings. Most
of the time, such temporary interruptions have little effect on the results. Consider both
possibilities carefully!

Rule 11) Consider the possible results before the test begins. Create hypothetical best and
worst case scenarios and look at times to failure under these circumstances... This is especially
important in a multiple-level stress test. The task here is to simply describe what types of test
results are available. Most examples will assume a single stress and consider only three possible
responses, unchanged, degraded or failed. With multiple stresses, the results may become more
complicated. The real world is much more complex, adding noise to all measurements. The
examples in the next section will illustrate some of the basics for data analysis.

Rule 14) At the end of test, perform analysis on some of the non-failed units from each
stress level. They often provide useful information about the failure modes as well. This rule is
always a good idea that promotes learning about the failure mechanisms and modes of non-failed
components. We may then see the progression of the failure mode and mechanism. The
information gathered by this activity is seldom mentioned, except in the Physics of Failure
approach, as it is infrequently done.

Rule 17) At the end of test, perform a detailed and thorough analysis of all of the data. It
is at this step where most books and articles on Accelerated Life Testing actually begin. The
practical side of all ALT efforts is to assume that ALT tests are set up and administered properly.
Sometimes, however, reliability professionals do not fully appreciate what small changes of test
conditions, fixtures or monitoring equipment can do to enhance or impede the learning
experience. Examples in Section 7 will highlight this and some of the problems associated with
analysis of complex or "messy" data.

50
5.0 Models for Reliability Analysis of ALT

The analysis of all ALT data starts with the creation of a model to describe reliability
versus time or a model of life versus stress. Both types of models are important and have been
covered in some detail in the various references [7, 12, 13, 14, 25, and 26]. The general models
for each type of distribution will be covered in this section. The reader is encouraged to pursue
this through the references and recent journal articles listed at the end of the book.

5.1 Degradation Models for Systems under Test

A first model for describing the failure of a component or system is that of degradation.
This model recognizes the possibility that systems and component do not fail suddenly and
unexpectedly. Most of the time, components and systems provide warning through some
performance measure changes indicating that some important aspect of the system function is
degrading. This change leads ultimately to failure if the system is operated a sufficient amount of
time. One example is that of an automobile tire. The performance measure is the tread depth. The
tire tread slowly wears away until there is none left, and the tire may then fail catastrophically.
For now, the focus will be on monitoring the performance measure change as a function of the
applied stress and time on test. This degradation model is important when looking at a variety of
systems and components.

5.1.1 Model One - Linear degradation - This occurs when there is a linear response to the
application of stress to a component or simple system. The linear response is documented through
a performance measure or P.M. A number of components and systems will approximate this
linear degradation for some portion of the life. The following is a simple mathematical approach
to this linear situation:

PM = A - Bt (5.1)

Where A is an initial value, B is the degradation rate and t is time. This equation may also be
expressed as degradation as a function of stress cycle. In that situation t will represent the number
of cycles.
Equation 5.1 describes slow, linear degradation portion of a more complex performance
curve. Often it ignores that fact that, in the real world, there may be transient events at early life
time and non-linear events near the end of life. These are treated as small in comparison to the
overall linear degradation in this model. Figure 5.1 shows this type of linear degradation situation
through the two straight lines. One line is an example of a fast linear degradation while the
other shows a slow linear degradation. The vertical scale (Y axis) of the graph is measured in
the same units as the unknown A from Equation 5.1. The horizontal scale (X axis) is typically
time or stress cycles. The performance measure begins with a reading of A at time zero or very
A
near time zero, while the time ordinate is the ratio . This is, when the P.M. reaches zero, this
B
is the implicit definition of failure in this model. It is possible to change the definition of failure
and use one such as:

A) Failure is defined when a change of 20% from the initial reading occurs.
B) Failure is defined when the P.M. drops to 30% of the original value.

This model, as modified, works as an approximation for many products.

51
 Performance A
Measure A’ Slow Degradation
Rate

Fast Degradation
Rate
Confidence
Limits

A’/B’ A/B Time or cycles

Figure 5.1 – Simple Linear Degradation

5.1.2 Model Two - Log Degradation. This model follows the relationship in the log of the
performance measure. It is described as:

Log (PM) = A - Bt (5.2)

The log model is similar to Model One, but can describe a more complex, real-world situation.
Here, the performance function (A-Bt) is always positive, but may still describe a fast or slow
degradation through the log function. This formula may be rewritten in the more familiar format
of Equation 5.3 as:

A  Bt  Bt
PM = e e = A'e (5.3)

We immediately recognize this as an exponential decay when we use the natural log function to
describe the P.M. versus stress relationship. A variety of situations, systems and components are
described by this Equation 5.3.

5.1.3 Model 3 – Complex, Non-Linear, Degradation - Typical degradation formulas often

neglect the effects of early or late non-linear behavior also known as "wear-in" and "wear-out".
Wear-in occurs early in the history of a component or system and typically lasts only a short time
with rapid change in P.M. Wear-out is more often a long, slow process near the end of life. Both
non-linear effects are shown in Figure 5.2 and correspond to common phenomenon.
Wear-in is a measure of an early-changing characteristic for a simple component,
assembly or system. It is often so short in time it can be neglected for many situations because it
may be eliminated by taking a “second set” of initial readings a little while after the start of test.
This occurs in the real world with batteries, gasoline engines and air conditioning systems. In
those few unique situations wear-in actually occurs over a long period of time rather than quickly.
Here a screening test or burn-in may be employed to remove or reduce the impact of this wear-in
and stabilize the performance measure. Some electronic parts have a very long wear-in period, as
measured in thousands of hours. This early portion, associated with classic infant mortality, tends
to dominate the early useful life in many electronic applications. One good example is the 7000

52
hour infant mortality time that exists for many semiconductors [19]. During this time, a small
number of a production lots rapidly change their characteristics and may become unacceptable for
use after an extended burn-in. This long wear-in may be characteristic of an unstable component
or process. The military specifications for burn-in were originally written with this characteristic
in mind. Such a situation is shown by the width of the confidence limits in Figure 5.2 and by the
example in Figure 5.3. Even at time zero, there is a finite width, scatter or large variability at time
zero. These limits document the small probability that a few of the components or systems will be
far from the mean performance measure and actually close to the definition of failure. We call
these “outliers” and it is these extreme units that usually fail first. Thus, the degradation model
can be correlated with classical reliability measures. This extended degradation situation is less
true for many mechanical parts and systems because of their typically short wear-in times. The
exception is for mechanical systems that have a dual stress life relationship as shown in Figure
5.4 for S-N curves (Stress versus cycle life). We normally treat these mechanical situations
differently.
The term "wear-out" is customarily observed as an indicator of the end of life, hence it
has always been associated with measures of reliability. It is only during Accelerated Life Tests
that we see long-life, wear-out failure modes of products.
To describe the more complex curves shown in Figure 5.2, the degradation mechanism
must be characterized by an inverse power. The wear-out portion can often be included with the
linear degradation shown in the middle of life. The component, system or material in question
will lose strength following a simple power law that covers all but early wear-in effects. This is
shown in Equation 5.4:

dS C
= (5.4)
dt (S ) N

Where S is a measure of the strength parameter, N is an unknown usually related to failure

mechanisms and materials. The constant C is an unknown material constant or degradation rate
measure and the negative sign indicates degradation of the strength S.

So Slow Degradation

Confidence Limit

P.M. or
Strength
Failure
Limit

Rapid
Degradation

Time or Cycles
Figure 5.2 - Complex Degradation Models

53
 Extended wear-in
P.M.
period
Normal Life

Confidence Limits Wear-Out

Failure Limit

Time
Figure 5.3 - Long Term Wear-In

Stress Stress
Dependent Life
Relationships

Cycles of Stress to Failure

Figure 5.4 - Long Term or S-N curves

Rearranging Equation 5.4 leads to:

S N dS = - C dt (5.5)

Integrating we have:

t
S
 ( S ) dS  C  dt
N
S0
0

or S N 1 = S 0N 1 - (N+1) Ct (5.6)

This equation provides the basic relationship between strength and time for a variety of situations.
Figure 5.5 shows three possible relationships of strength (S) degradation versus time (t).
The value for N is shown as an integer in this simple model, but it can be, and is often, a non-
integer. Thus, any component, assembly or system with little wear-in or that can be screened past
the wear-in time may be simply approximated by a power law. Equation 5.6 is a general

54
degradation equation that may reduce to other simpler models through the selection of the value
of N.

1) The linear degradation is the case when N =0

2) Exponential degradation occurs when N = -1
3) Common linear degradation, followed by a wear-out, is modeled by N = 3

These three possibilities look like many of the common degradations we might have
already observed. Consider the following short list in Table 5.1, of common degradation events.

Strength
N=3

N= 0

N= -1

Log Time
Figure 5.5 – Power Law Models

Table 5.1 - Common Degradation Events

1. Corrosion leading to degradation of a material.

2. The loss of a rechargeable battery capacity due to chemical degradation (non-

reversible reactions) from recharging.

3. The loss of strength of an adhesive over time or under applied stress.

4. The loss of gain of a transistor due to time, temperature or high voltage stresses.

5. The aging of a resistor or capacitor (loss of resistance or capacitance) due to time,

temperature or high voltage stress.

6. The loss of breakdown voltage of an electronic component due to transient electrical

stresses such as over voltage or ESD.

Example 5.1 - Consider the following situation as an example of degradation. Past history
suggests that an organic adhesive chemical degrades slowly with time. This degradation follows
a power law relationship as follows:

55
 S 3 = S 30 - 3Ct (5.7)

Where S measures the adhesive bonding strength of the material under study. The adhesive is
required to "bond adequately" for at least three years. The minimum acceptable bond strength for
retention is set at 50% of the initial strength. With this information we can set up an accelerated
test to verify that any new adhesive can meet this requirement. Filling in the information now
known into Equation 5.7, we have:

(0.5S 0 ) 3 = (S 0 ) 3 - 3C(3 years) (5.8)

0.125 (S 0 ) 3 = (S 0 ) 3 - 9C

(0.875)S 30 = 9C

and C = (0.097) S 30

Thus, the one remaining unknown can be expressed in terms of the initial strength of the
adhesive. If we can measure differences as small as 5% in adhesive strength, by the use of
multiple samples and statistics, how long will we have to run a test to verify the adhesive meets
the 3 year requirement? Filling in Equation 5.7:

(0.95S 0 ) 3 = (S 0 ) 3 - 3(0.097) S 30 (t) (5.9)

(0.857)S 30 = S 30 - (0.291)S 30 (t)

0.143 S 30 = 0.291 S 30 (t)

and t = 0.491 years minimum = 25.6 weeks

The expected time to a 5% average decline in the adhesive strength, without acceleration is about
½ year. This assumes the whole life of the degradation follows the formula, that is, there is no
"wear-in" associated with the adhesive strength degradation. Likewise we don’t have measure of
the spread of the distribution of adhesive strength. This would be obtained by actually running a
number of samples through the test and the time to failure adjusted. Any wear-out is covered by
this equation, but we are still on the early life part of the curve and wear-out would probably not
be an issue with this test.
Now, an Accelerated Life Test is performed by raising the temperature of some of the test
samples above room temperature by 23ºC, setting the new test temperatures at 45ºC and at 68ºC
respectively. The results of the higher temperatures are then compared to samples running at
room temperature. Every week, a small number of samples are removed for evaluation of
degradation. The degradation measurements are listed in Table 5.2. Only some of the weeks are
shown even though measurements were made weekly during the test. The three stress levels all
show changes from the initial strength. These are presented as a percentage of the initial value.
Improvements at the start of the test in strength measurement technique, combined with improved
statistical analysis, now permit a change in mean adhesive strength as small as 2% to be
measured. This sets the resolution measurement in Table 5.2. Will this adhesive meet the three-
year minimum strength requirement?

56
 Table 5.2 - Degradation Results of an ALT

Time 1 week 4 weeks 8 weeks 12 weeks 16 weeks

Temperature
Room, at 22ºC 0% 0.1% 0.6% 1.4% 2.1%
at 45ºC 0.1% 0.5% 1.3% 1.9% 2.3%
at 68ºC 0.2% 1.2% 2.2% 3.3% 4.1%

Mean readings of the samples

The data suggests that the noise or variability for these samples is about 0.1 to 0.2% and the
ability to identify measurements is better than expected.

Completing Equation 5.7 with the information in Table 5.2 provides the following three equations
at 16 weeks:

At Room Temperature (0.979) S 30 = S 30 - 3C 0 (16 weeks)

At + 45ºC (0.977) S 30 = S 30 - 3C 1 (16 weeks)
At + 68ºC (0.959) S 30 = S 30 - 3C 2 (16 weeks)

Other equations can be written such as:

and (0.978) S 30 = S 30 - 3C 2 (8 weeks) (0.967) S 30 = S 30 - 3C 2 (12 weeks)

(0.987) S 30 = S 30 - 3C 1 (8 weeks) (0.981) S 30 = S 30 - 3C 1 (12 weeks)
(0.994) S 30 = S 30 - 3C 0 (8 weeks) (0.986) S 30 = S 30 - 3C 0 (12 weeks)

Solving these for the best fit values of C 0 , C 1 , C 2 yields:

C 0 = 0.000359 S 30 C 1 = 0.000516 S 30 C 2 = 0.000896 S 30

The Acceleration Factor, A.F., for these test conditions and times can be estimated by the ratio of
the values of the degradation coefficients, C.

C1 0.000516 C2 0.000896
AF 1 = = = 1.437 AF 2 = = = 2.496
C0 0.000359 C0 0.000359

Employing the factor, AF 2 , shows that this 16-week test at three stress levels estimates that a
4.0% change at room temperature will occur over:

(2.496)(15.5 weeks) = 38.7 weeks

As a minimum estimate, since the data is checked only once a week, and 0.2% variability exists it
is possible the failure may occur just past the 37th week. This is a lower limit estimate on the
room temperature degradation process.

57
Either estimate is longer than the time to 5% change as estimated before the improved resolution
occurred. Any estimates from the results at +45ºC lead to similar results. Table 5.3 shows the
relative relationship of the three curves. Our analysis has compared a single point on three curves
in order to provide the estimate of time to failure at room conditions. With more test data, this
estimate can be improved.

Degradation Mechanisms - The probability of a component failure is often directly related to

one or more of the following reasons. These are all terms for “gets worse over time”. Words such
as wear, fatigue, aging, erosion or degradation are in common use.

Table 5.3 - Degradation Mechanisms

Wear Corrosion Erosion Diffusion

Fatigue Degradation Strain Accumulative fatigue
Aging Chemical reaction Depletion

The fact that any of these mechanisms may be ultimately related to failure is not directly
included in any of the failure life distributions. Rather, we associate or correlate a life distribution
with some known failure mechanisms because of the way the mechanisms propagate. We may
even associate root causes and failure modes with a time-to-failure distribution. The overall
relationship between applied stress and the ultimate system failure may run through latent failure,
cause, mechanism or failure mode as described in Table 5.4.

Table 5.4 The Details of the Stress to Failure Chain

Applied External Stress  Microscopic Strain Latent Defect Accumulation of

Fatigue Generation of Internal Fracture or Start of Damage Development of Crack

Growth Macroscopic Observation of System Changes Observation of Degradation

Continuation of Observed Degradation Ultimate System Failure

This simple connection begins with a stress, moves through a failure mechanism,
propagates, and then results in observable parametric changes or degradations and ultimately
failure. The following short version is what many people present as the stress-versus-failure
connection.

Stress Failure Mechanism Parametric Change System Failure

5.2.1 Activation Model of Aging - The following simple model, from Feinberg and Widon [21],
relates the simple concept of an energy well to the degradation mechanism leading to failure. This
model is based upon the probability of surmounting an energy barrier of height . This model
actually follows the Maxwell-Boltzmann statistic, a common item of statistical physics and
thermodynamics.
Let the probability of escape from a square potential well be p, as shown in Figure 5.6.
This probability is proportional to a reaction taking place, which is the analogy of a propagating
failure mechanism in this case. The energy of the particle is related to , which is classically the

58
vibration frequency of the particle in the potential well. We can write, from the Maxwell-
Boltzmann statistic, the time-dependent probability of escape as:

0
dp ( )
=  0 e KT (5.10)
dt

where K is Boltzmann's constant and T is the absolute temperature,  0 is the mean energy and
 0 is the basic vibration frequency at nominal conditions.

p Escape

 0

 0

Figure 5.6 – Potential Well Aging Model

The degradation can be defined as a change in any related measurable parameter, here
written as R 0 . The change in the parameter is R. Thus, we write an equation that relates the
change in a measurable performance parameter to the thermodynamic model as:


R da ( )
a= with = e KT (5.11)
R0 dt

where  is a function of the temperature, T. Expanding a Maclaurin series about  0 , we can write
the first two terms as:

1 2
 =  0 + (a)Y 10 + (a )Y 20 (5.12)
2

  2
where by definition Y 10 = and Y 20 = (5.13)
a a 0 a 2 a 0

Using Equation 5.11 and the definitions in 5.13, we can rearrange Equation 5.11 and with
Equation 5.14 we get:
0
( )
(T) =  0 e KT
(5.14)

59
Integrating equation 5.11 yields:

 KT    (T )Y10 (t ) 
a=   ln 1   (5.15)
 Y10   KT 

Thus, the degradation can be expressed as a function of physical constants, time and the energy of
f
the initial state. The function, a, will be similar to that shown in Figure 5.8 as .
f

Example 5.3 - The simple aging of a crystal due to defect accumulation on the surface may be
described by the following equation:

1 c
f= (5.16)
2w 

Where f is the basic resonant frequency, c is a bulk elastic wave constant for then material,  is
the density of the material, A is the surface area, w is the total thickness of the crystal itself, made
up of the crystal and a defect film area w m shown on one end in Figure 5.7.

With w = w0 + wm

Wm
A
W0

W
Figure 5.7 – The Aging of a Crystal

Differentiating Equation 5.16 yields:

1 c
df = d wm (5.17)
2w 2 

It is easy to show that creating a Maclaurin series around the nominal conditions will lead to:

60
 
Y 10 =
GKf 0

where  is the chemical potential, N is the number of defects in the crystal, m is the foreign mass
of defects and m = NG, where G is the gram molecular weight of the defective film.

The degradation formula becomes:

f  t 
= -ln 1   (5.18)
f  

2 KTGf 1
with =
 cA 2

f
For a small ratio of the crystal ages slowly at a near-constant level as shown in Figure 5.8.
f
f
As becomes larger, the crystal is forced into a large vibration mode, leading to the rapid
f
accumulation of damage and then quickly to a catastrophic response. The time to such a
catastrophic response is labeled t c in Figure 5.8.

f/f Catastrophic
Change
Large
Change
Wear-in
Linear Degradation
Small
Change

Time Tc

Figure 5.8 – Crystal Aging Model

This aging model can be generalized to cover a variety of similar situations that depend
upon chemical reactions for degradation. This is the basis of the potential well model of
reliability, sometimes called thermodynamic reliability. In fact, it is possible to relate this model
to the familiar Arrhenius relationship [21].
There are a number of typical aging processes that might follow a similar equation with
this type of relationship. These situations include crystal aging, corrosion of thin films, cold
worked metals, gate oxide stresses in semiconductors, battery life degradation, semiconductor

61
aging and rechargeable battery degradation. Note that Figure 5.8 is the complement of Figure 5.2.
This is a reflection of the vertical axis showing degradation as upward versus downward.

5.3 Creation of a Failure Model versus Time or Stress

The aging model in Section 5.2 generated a simple model for relating a performance
measure versus time at a fixed stress. The performance measure can be easily turned into a failure
by drawing a line on the graph such as is shown in Figures 5.4 and 5.5. In both cases one only
need declare a fail zone at some low performance measure. While this may defeat the purpose of
continuous monitoring of a performance parameter, in many engineering cases all we have is a
simple pass/fail measure.
With multiple levels of applied stress, the absolute value of the stress may become very
important as shown in the stress-life relationship of Figure 5.9. The changing relationship of
stress at the highest levels makes it important to select a consistent point on the life curve when
modeling or comparing the relationship between stress and life. This is easily shown by the lines
drawn at 10% cumulative failure and 60% cumulative failure.

Moderate
Cum High
Stress Low Stress
Percent Stress
Failure 60%

Stress-Life
Relationships

10%

Log Time
Figure 5.9 - Weibull Graph of Stress versus Life

The life distribution at various levels of accelerated stress should be similar to the life
distributions at normal stress. On the Weibull graph, this would appear as a series of straight and
parallel lines as in Figure 5.9. This also leads to a single straight line when the stress-life
relationship is plotted on the Log-Stress versus Log-Life graph as shown in Figure 5.10.

62
 Highest
Log-Stress
Stress
S3
S2 Lowest
Stress
S1

Sn Normal Stress

Log-Life Normal Life

Figure 5.10 – Log-Stress Versus Log-Life

The slope of the Log-Stress versus Log-Life graph is proportional to the acceleration
present. There are two main stress-life models when a single stress is applied. These are the
Arrhenius (Equation 5.19) and the Inverse Power Law (Equation 5.20) models. Equation 5.21
shows the situation where two stresses are present, here they are temperature and operating
voltage. In this combined form, we will neglect the possibility of interaction between the two
stresses.

Ea
K bTk
Life = A e Arrhenius (5.19)

Life = B(S)  N Inverse Power Law (5.20)

Combined version would be:

Arrhenius Inverse Power Law Portion

Ea

] [(V)  N ]
K bTk
Life = A [e (5.21)

Example 5.4 - Mechanical Corrosion - In this example, we will estimate the accelerating
influences of a corrosive solution on a material. Chemical concentration will be the accelerating
factor. The performance measure will be the material changes observed when exposed in a
concentrated solution for a specified period of time. The approach to the problem is to first
identify the definition of a failure. This forces the determination of what is important and also
identifies the means and ability for making subjective measurements. The details of the test are:

Control Sample Material Soaked in Solution at Room Temperature

stress level 1 concentration of 1part in 150 parts water- called 1/150
stress level 2 concentration of 1/50
stress level 3 concentration of 1/20
stress level 4 concentration of 1/5

63
 The Definition of Failure - The surface area of the material will be fixed and a predetermined
one inch square piece of material will be inspected. The corrosion will be measured in percent of
surface coverage of this one square inch area on each sample. The four different concentration
levels act as different stresses to drive the corrosion on the surface. Stress concentration points
such a welds, bends or cracks are deliberately avoided. Other life measures that could have been
selected would include the “thickness of corrosive film” or the “time to the beginning” of the
appearance of corrosion. Let the following table represent a set of hypothetical corrosion results
for such a corrosion resistance test.

Table 5.5 - Results of Corrosion Test

Time 1 week 2 weeks 3 weeks 4 weeks 5 weeks 6 weeks

Concentration
1/150 none 0.2% 1.0% by area ~ 2.5% by area 4.5% by area 8%

1/50 none ~ 1% 2.2% 5% 8% 14%

1/20 ~ 1% ~2% 4.0% 8% 12.3% 19%

1/5 ~ 3% 10% 16% 24% 35% 50%

Employing a simple power law to model for analyzing this data, we can write as the stress-life
relationship:

Life = B(S)  N (5.22)

If we use the time to 5% corrosion coverage by area as a measure of failure and we

estimate life as measured in accelerated days of test, the following four equations can be easily
developed from the results in Table 5.5. One equation exists for each concentration. Since the
samples are checked once a day, this is an interval test. We don’t exactly know when the failure
occurred, all we know is that we inspected the samples at 24 hour intervals.

9 days = B (1/5)  N 23 days = B (1/20)  N (5.23)

28 days = B (1/50)  N 36 days = B (1/150)  N

Solving these four equations for best fit values provides:

N = 0.334 and B = 6.85 days

This represents an approximation of this simple power law model as a function of stress. The time
to failure model can be developed from a model based upon the information in the rows of Table
5.5.
The simple analysis does not prove that the test follows the power law. At present, all we
have shown is that it is possible to fit a model to data. We could have used more test information
that was available in Table 5.5 to see if the Inverse Power Law was the best model for the data.
In fact, we can see that at high concentration this model may not be correct. A second test might
be required to make a more definite statement about the model. Figure 5.11 shows a Log-Log plot

64
of the concentration data at 1/5 and 1/20 since this is the most we have. The plot is a fairly
straight line for each data set which suggests that a Power Law may be a good model.

Accelerated Concentration Test

0
0 1 2 3 4
-1
Ln-Concentration

-2

-3

-4

-5
Ln- Time

Figure 5.11 – Ln-Ln Plot of the Concentration Data

A quick analysis of this graph suggests that the simple power law is not sufficient for analysis of
the data set. The value of B in Equation 5.22 seems to vary with concentration. This suggests
some non-linear effect may be present and requires additional attention. Non-linear effects will be
covered in chapter 7. Additional analysis using time to failure of 3% and 10% generated the data
in Table 5.6. This is the same as the stress-life lines sown in Figure 5.9. It adds additional
evidence to the suggestion that as the concentration changes the values of B and N change. This
suggests the Log-Log graph shown in Figure 5.10 is probably not straight.

Table 5.6 – Additional Data of Best Fit Power Law Models

Time to Concentration in Best Fit Value of N at Time Best Fit Value of B at Time
Percent to Failure to Failure
3.0 0.405 4.41 days
5.0 0.334 6.85 days
10.0 0.330 9.76 days

Example 5.5 - Burn-in results of RAM memory - The following data in Table 5.7 shows a
summary of time to failure data for an ALT performed at 125ºC [20]. All failure times are shown
in hours for the Random Access Memory, or RAM.
Figures 5.12 and 5.13 show the Weibull analysis of this single stress data set as
performed with Weibull ++ 6.0 [12]. Figure 5.12 is the Weibull graph of the data while Figure
5.13 presents the probability density plot of the data. Both graphs suggest that two failure modes
may exist, but this idea should be confirmed by failure analysis. The techniques of Weibull
analysis, employed to create the graphs, were discussed in The Weibull Analysis Primer [14].
The early failure population and the normal population are separately identified in Figure 5.12 for
their suggested Weibull parameters. One initial suggestion is that the early failure population,
about 49%, could be screened out. This percentage is so high that another method; one for
preventing the cause of the problem, should be found. In Figure 5.13, the heavy line shows the

65
smoothed plot of the data. The light lines suggest the make up of the two sub-populations that
may be present.

Table 5.7 - RAM Memory

Fail # Time Fail # Time Fail # Time

1 215 hrs. 2 301 hrs. 3 315 hrs.
4 319.5 hrs. 5 328 hrs. 6 347 hrs.
7 425 hrs. 8 452.5 hrs. 9 491 hrs.
10 706 hrs. 11 739 hrs. 12 944.5 hrs.
13 1349 hrs. 14 1482 hrs. 15 1763 hrs.

One unit remained non-failed at 1800 hrs.

RAM Failure Data

99.00
Weibull
Data 1

W5 RRX - SRM M
90.00
F=15 / S=1

50.00
Cum Percent Failure

10.00

5.00

James Mclinn
Rel-Tech
1.00 3/7/2010 10:39
100.00 1000.00 10000.00
Hours


Figure 5.12 – The Weibull Graph of RAM ALT Data

66
 RAM Failure Data
2.50E-3
Weibull
Data 1

W 5 RRX - SRM M
F=15 / S=1

2.00E-3
Cum Percent Failure

1.50E-3

1.00E-3

5.00E-4

James Mclinn
Rel-Tech
0 3/7/2010 10:43
0 600.00 1200.00 1800.00 2400.00 3000.00
Hours


Figure 5.13 – The PDF Plot of RAM Failures

Example 5.6 - A technique to estimate shelf life through ALT - The goal in the following
example is to verify a two-year shelf life for a product. The test that evaluates shelf life is
destructive to the material being tested. How can this best be accomplished? Select 3 or 4 levels
of stress with about 24 samples for each stress level. Fifty samples are also required for the
normal (non-accelerated) stress conditions. After determining a strategy, estimating the
background noise in the measurements and creating definitions for failure, the test may begin.
Table 5.8 represents a hypothetical set of results for this test. The amount of degradation is shown
in the table as a percentage. Therefore, an entry such as 0.96 suggests this product has degraded
from the initial 100% to the level of 96%. This number represents the average of the samples
destroyed to make the measurement. The accuracy of the measurement is treated as 1%. That is,
96% is most probable, but there is a small probability the number is could be as high as 97% or as
low as 95%.

67
 Table 5.8 - Shelf Life Test Results

Non-accelerated Stress Level 1 Stress Level 2 Stress Level 3

25C and 40% 40C, 40% 60C, 60% 60C, 80%

Month Degr. % Month Degr. % Month Degr. % Month Degr. %

3 0.97 1 1.00 1 1.00 1 0.99
6 0.94 2 0.99 2 0.98 2 0.97
9 0.90 3 0.98 3 0.97 3 0.96
12 0.87 4 0.96 4 0.95 4 0.94
15 0.84 5 0.95 5 0.93 5 0.92
18 0.80 6 0.93 6 0.91 6 0.90
21 0.77 All samples were consumed at these stress levels
24 0.74

A simple graph or analysis of this data may lead to an estimate of the acceleration factor
associated with the different levels of stress. The failure modes may be identified through
analysis to assure the same root causes and mechanisms are operating at each level of stress. As
one simple measure of life, consider the time to 10% degradation. For the non-accelerated parts,
it was about 9.0 months, for stress level 1 it was estimated at 8.6 months, for stress level 2 it was
estimated at 7.6 months and for stress level 3 it was estimated at 6.0 months. (This interval data
may be a little noisy because the samples were pulled every 4 weeks). The time to 10%
degradation was interpolated for most of the stress levels.
The first estimation of a model for this two-stress situation, assuming no interaction, can
be written as:

Ea

](Humidity)  N
K bTk
Life = A [e (5.22)

With these numbers, a best fit value for the model unknowns can be made. It was evaluated as:

E a = 0.0544 eV and N = -0.3264

The simple model with these numbers filled in becomes:

0.0544

](Humidity) 0.3264
K bTk
Life = A [e (5.23)

Please note that details of the analysis of this data set have been skipped here since it is similar to
the three prior examples. No test for interaction between the two stresses was made in this case.
An interaction example and details will be shown in Section 7.0. This data set gives us both time
to failure for each stress condition and acceleration between conditions. The acceleration appears
low for both temperature and humidity. Even stress level three took 6 months to get to 10%
degradation. Thus it is possible with a six month test to simulate 24 months. A few comments
about humidity testing are appropriate here.

5.4 - Humidity Testing Issues and Non-Linear behavior

68
Humidity testing can be a mystery in many reliability situations. Most environmental tests can be
divided into two major categories. These are durability tests, where the accumulation of wear
eventually causes a product to fail. Common examples of durability tests are corrosion, vibration,
thermal cycling or mechanical abrasion situations. The second type of test is a capability test.
These often occur as stress tests and sometimes may be called “overstress tests”, but this term can
be misleading since overstress is a relative term. The goal of all stress tests is to determine how
long a product can last under stress conditions such as operating temperature or operating
voltage. When stress are above the normal customer operating conditions, the life becomes short
and may be used to estimate life from a short term, high stress condition.
Humidity tests may be different as humidity can cause a change of failure mode.
Humidity may bring on failure mechanisms such as electro-migration, corrosion, dendritic
growth, or chemical changes to materials. These induced failure mechanisms may follow a wear
process or be impacted by a second stress present. If the second stress is temperature, then failure
mechanisms such as chemical reactions may be increased. When the second stress is operating
voltage, then dendritic growth or corrosion may be accelerated. The combined results may follow
the Eyring relationship which includes interaction between the stresses. D.S. Peck had proposed
[27] that the Eyring relationship for operating bipolar plastic integrated circuit could be described
by

0.79

](Humidity) 2.7
K bTk
Life = A [e (5.24)

Intel, on the other hand in their 1988 handbook has suggested this relationship might best be
described as

Ea

][e  B ( RH ) ]
K bTk
Life = A [e (5.25)

A more complex relationship such as Equation 5.26 might be required to describe the interaction
of temperature and humidity at some levels of stress.

0.79 B ( RH ) Ea
K bTk 2.7 K bTk
Life = A [e ](Humidity) [e ] (5.26)

The last term, which contains both temperature and humidity is the interaction term.
All the models in Equations 5.24 to 5.26 are really for operation and not shelf life tests.
We might start with similar models, but would be surprised if the operation life model was the
same as the storage life model. This is because humidity often causes a change of mechanical and
electrical properties of materials which often changes their susceptibility to failure
mechanisms. The modulus of elasticity of many materials changes after absorbing moisture. This
includes materials as diverse as wood, many plastic and some epoxies. These often become more
prone to developing cracks or soften as a result of humidity exposure. Such failure mechanisms
can not be easily described by any known simple models such as shown in Equations 5.24
through 5.26.

Despite this caution, many reliability engineers still apply such simple equations for life
test calculations. Such simple models really are first order approximations and are easy to
use. More complex models can be derived from real data coming from a thorough test. Most of us
end us using a standard set of test conditions and then compare test results as a measure of
differences. Standard test conditions include 85C and 85% relative humidity. Others use 90%

69
relative humidity and 70C. Any of these test conditions are often based on some historical data
rather than on test rationale or good understanding of failure mechanisms that might be present.
The real questions to ask before starting a temperature and humidity test include:

1) What are the expected humidity-triggered failure mechanisms for my product?

2) Are any forms or electro-migration involved with the product?

3) Will the humidity affect any of the material properties and alter the test results?

4) What type of humidity test is most appropriate for my products? Steady-state humidity, cyclic
exposure, or combined environment conditions?

5) Will any of the possible failure mechanisms be accelerated in a non-linear fashion by the
humidity and temperature combination?
6) Do any of the known acceleration models apply?

7) Can I trust that my test will represent the whole life of the product?

70
 6.0 - Selecting the Best Distribution and Performing the Data Analysis

Most reliability texts and many articles about Accelerated Life Testing begin at the point
of looking at a set of data generated from a test. Completed life data is often presented as a table
or a short list and the task of data analysis commences. Data analysis really begins before the
review of the data. Test conditions, data collection, calibration, resolution of the measurements all
impact the data. Next should be a review of the test results to assure that the data points seem to
be valid. There are two main methods to do this. The first is to look over the times to failure
and failure mode information for basic consistency. Next, review the data from each stress level
and verify consistency across these levels. This second item may be easily done by plotting the
data on the Weibull graph and looking at the fit of the lines relative to the stress levels. Be sure to
verify that all of the failures appear to be related to the test stresses and not an artifact of the test
fixtures or accidents. Then plot the log of mean life versus the log of stress. After these cursory
reviews, the next set of tasks involves more in-depth review.

6.1 Selecting the Best Distribution

With all of the possible statistical distributions to describe the failure data, how can the
best distribution be selected to describe each data set? This distribution is the life distribution. It
describes in detail the time to failure at each level of stress. If the Lognormal distribution is
selected to describe the lowest level of stress, it should also apply at the highest level of stress in
the test. If this does not create a good fit to the data for all stress levels, then the reasons for the
poor fit are important. Try another distribution such as Weibull until a best fit across all stresses is
obtained. A program such as ALTA™ [22] does this automatically and models the distribution
and stress-life choices both quickly and provides metrics of fit.
The second "distribution" is really the stress-life relationship [7]. This need not be a
statistical distribution, but is often described by a simple equation. Two common ones are the
Arrhenius relationship and the Inverse Power Law. Both types of stress-life relationships may be
related to common failure mechanisms such as oxidation or material degradation and so
appear to describe a number of different components and simple systems. Either choice needs to
be explored carefully and related to the failure modes and mechanisms observed as well as
through past history.
Sample size, the number of failures, normal variation in components, incomplete
information about the time of any failures and uncertainty about stress conditions all add to the
difficulty of making sense of the data. Consider the following, more detailed set of guidelines for
selecting the best distribution to describe the data or describe the stress-life relationship.

1) Employ existing knowledge to help identify the best distribution choice. - Ask the
following questions about the time to failure test data.

A) Is there history of a similar situation to guide us? What distribution was employed
in the past for any similar ALT results? Was this prior analysis adequate and
appropriate? Is there history or results from other people based upon similar parts or
systems?

B) Do engineering models or equations of stress versus life help? What do we know

about the underlying structures or relationships between the variables and the
relationships of stress to ultimate life? Many components and materials can be
described by such equations.

71
 C) Does any information on the "Physics of Failure" help select a distribution? Can
the reasons for failure be described mathematically? This often starts with a
microscopic description of suspected failure causes, the failure mechanism(s) and details
of how the underlying defect continues to propagate, transforming a defect or flaw into a
failure.

D) Does the knowledge of the failure mode(s) help in selecting a distribution of life?
Can we estimate how many different failure modes exist in any particular part or
assembly? Is there detailed information about each of the failure modes?

E) Can hard component or system failures be well described? Is there non-failure

data concerning degradation mechanisms that can be modeled? Is there a software
program to answer the question of the best distribution for the data? The following two
questions may be asked at the data analysis point:

1. How can the best fit model for the data be selected?
2. Is this model still consistent with all other relevant data?

2) Select models based upon the best fit of the known data to a statistical distribution.

A) What distribution-fitting software program will best aid the modeling? Since the
test data can come in a number of different ways, the answer may depend upon the
software method selected. These methods include:

1. Complete data - That is, all of the samples are run to failure. The exact
times or cycles to failure are known. There are no uncertainties concerning the
definition of a failure and the times of failures.

2. Suspended data - This is, some of the test samples were stopped before
they were run to failure. An incomplete "time to failure" distribution may be
the result, especially with small sample sizes. The reasons for suspensions
should have nothing to do with the test itself. These suspensions are holes in
the time to failure distribution, as they occur at any time of test. In some cases,
when two failure modes are present, one is treated as a suspension to the other.
We should be very careful that failure modes themselves are exclusive. That
is, failing for one mode excludes the possibility of another failure mode.

3. Censored data - Here, the tests end before all of the samples have failed.
This is called right-censoring and is a very common situation. The most
common example is a simple time-terminated test. For example, ten samples are run to
four failures.

4. Grouped data - This results when multiple failures occur or are found at the
same time. This measure is within the resolution of the time measuring equipment or
observations. Normally this is a very short time interval. For example, a computer may
slowly rotate through a set of test samples and periodically identifies any failed samples.

5. Interval data - This occurs when a periodic check of the test samples is made.
Normally, the time resolution is relatively large - checking once a day or once a week or
some other time interval. Samples will most often fail in between two of the check
points, so the exact times of failure will not be known. This is a further complication to

72
 any statistical modeling package; some packages do not handle this situation well.

6. Small sample size and noisy data - This is a common difficulty. With few
failures in a small sample or large scatter of the times to failure, it makes fitting
to any distribution difficult. Data fitting usually has a poor goodness-of-fit and two or
three distributions may fit equally well.

7. Data with multiple failure modes - This situation occurs with many complex
parts and many systems. Several failure modes and/or competing failure modes
complicate the analysis of ALT data. Failure modes may need to be separated for
meaningful analysis. Three non-interacting failure modes with one having
non-linear behavior may cause great difficulty during the Weibull analysis or ALT
analysis. These modes need to be separated first, then any remaining non-linear behavior
analyzed.

8. Data with interacting failure modes - When multiple stresses are present,
it is a real possibility to have interacting failure modes. The best examples are when the
stresses are temperature and humidity or temperature and voltage. Many failure modes of
components are driven by both stresses. Non-linear behavior may result in these
situations.

9. Data with high levels of stress - There are many failure modes that may
change as the level of stress increases. The best example is with crack generation.
Below certain values of stress, no cracks may be generated and/or propagated.
Above a given threshold, cracks can be generated by the stresses and then propagated. As
the level of stress increases, the generation rate and propagation rate of the cracks
increase non-linearly to the level of sudden system failure. This last possibility skips the
whole earlier crack generation and propagation process.

B) The choices of models for fitting the ALT data to best fit lines include:

1. MLE - The Maximum Likelihood Estimator is often the best choice when
there are a lot of suspended or censored items. Said simply, this model
weighs the non-failed items as importantly as the failed items. [7, 12]

2. Rank Regression - This is often the best choice for samples when complete
data is available (no suspensions present). In this situation, a line is fitted to the failure
data only. It ignores the impact of suspensions (which are few) on the Weibull slope and
includes censored data primarily through the total items exposed to test and estimates of
the characteristic life. This is a common situation with many analytical programs. [7, 12]

3. Least Squares Fit - An alternate fit for "best line" to represent the data points
is the least squares fit. It is generated by minimizing the total of the squares of the data
point distances from the best fit line. This is probably the least-favored choice as it
typically looks only at the failure data. It may ignore the effect of suspensions and
censored items as well as the impact on the Weibull slope and characteristic life. This is
often the choice by programs such as Excel™. It does not have a statistical model basis
and should be considered as the least desired approach to fitting data of the three
mentioned.

6.2 The Data Analysis of Accelerated Life Test

73
 The impact of these prior model choices will be shown by the following several examples
of data analysis of several different types of ALTs. These consider some of the selections
mentioned previously.

Example 6.1 - Interval data with censoring.

The following is an example of real data that was collected from a group of field failures.
It is interval type data with the time intervals expressed in months. That is, the uncertainty about
when failures actually occurred is three months. The number of failed units is five with 1100 non-
failed units still operating in the field. No units are older than 12 months.

Table 6.1 – Field Data

Time interval Number failed or censored

0.1 to 3 months 1 failure and 100 non-failed in this interval

3.01 to 6 months 1 failure and 300 non-failed in this interval
6.01 to 9 months 2 failures and 600 non-failed in this interval
9.01 to 12 months 1 failure and 100 non-failed in this interval

The data in the table says that the units in test were only checked every three months. At
this point a few failures were found with the vast amount of systems still operating failure free.
This interval data is a little unusual, but not unreasonable. Figure 6.1 shows the four data points
plotted at the middle of the four time intervals in question. The analysis is based upon the rank
regression method on the Y axis (hence it is called RRY) of calculating the best fit line for all of
the data. This method uses the Y axis distance of the best fit line from the data points and
minimizes that distance. The best fit line is actually a close fit to the four data points which are
represented by the circles placed in the middle of the time interval. The time interval is
represented by the brackets in this figure. The rank regression correlation factor,  is 0.9903,
which is ordinarily a very good fit of line to data. The Weibull slope, , is 1.224 with an
estimated characteristic life,  of 742.3 months with this method.
Figure 6.2 is a different analysis of the same data set plotted using the Maximum
Likelihood Estimator or MLE. The best-fit line selected by the program goes near the center of
one of the four data points, but not all four. The line is slightly biased to the right of the four data
points and has a steeper slope than the rank regression line. It almost appears that the line drawn
is a “deliberate poor fit” to these four data points. This is a reasonable initial conclusion, though
clearly in error. The MLE model attempts to create a best-fit for all of the data, including
information derived from both the 1100 non-failed units as well as the five failures. That is why
the line appears to be a poor fit to the four data points. The Weibull slope in this MLE case is
2.106 and the characteristic life is 107.4 months. This life is a factor of 6.9 times lower than the
same number calculated by the rank regression method. Most contrasting examples do not
represent such an extreme difference. The Likelihood measure for the MLE data fit is only L = -
35.8 as calculated by this software package. It is a reasonable fit for the data based upon the few
data points and the selection of the model.

74
 Figure 6.1 - RRY Analysis
99.00
Weibull
Data 1
90.00
W 2 RRY - RRM M
F=5 / S=1100

50.00

10.00
Cum Percent Failure

5.00

1.00

0.50

0.10

0.05

James Mclinn
Rel-Tech
0.01 3/11/2010 16:32
1.00 10.00 100.00
Months

Figure 6.1 – A Rank Regression Analysis of Interval Data

After all, what is the measure of the best fit? Is it the model that runs the line through the four
data points representing the five failures out of 1105 samples? Perhaps the model that attempts to
best fit all of the data points may be the better model? Clearly, the latter is the best choice. It is
difficult to believe that the best model doesn't necessarily go through all five of the data points.
The MLE model choice appears to be counter-intuitive, at least until further explanation. Since
the MLE models all the failures and the non-failed samples, it is not surprising that the best fit
line doesn’t go through the 4 data points alone. The data shown in Figure 6.2 is the same as in
Figure 6.1. The main difference is the slope of the line, . The MLE analysis concluded the slope
is about 2.1 which suggests a more peaked distribution and tighter time to failure distribution than
the rank regression slope of 1.2.
Look at the data and be sure to include the consideration of MLE versus rank regression
as a consideration in Accelerated Life Tests. Use MLE when the number of suspensions is more
than a few. When performing a multilevel test, all of the stress analyses must be done the same
way to have a thorough comparison of the levels. The following example shows the importance
of selecting the optimum data analysis method before detailed analysis begins.

75
 Figure 6.2 - MLE Analysis
99.00
Weibull
Data 1
90.00
W 2 MLE - RRM M
F=5 / S=1100

50.00

10.00
CumPercent Failure

5.00

1.00

0.50

0.10

0.05

James Mclinn
Rel-Tech
0.01 3/11/2010 16:30
1.00 10.00 100.00
Months
   

Figure 6.2 – A Maximum Likelihood Analysis of Interval Data

Example 6.2 - Interval Data with Multiple Stresses.

The following data set is based upon an ALT of capacitors operated at 8V, 12V and 16 V
at temperatures of 125ºC, 105ºC and 85ºC in the combinations of stresses shown in the Table 6.2.
There were fifty samples per each test condition and the tests were run for 1000 hours (6 weeks)
and then each test was terminated. All of the test data is right-censored and the test read times
are shown. These indicate that it is interval data based upon irregular intervals. Failures were not
replaced during test.

Table 6.2 - Capacitor ALT Data

Data Set A Set B Set C Set D Set E Set F

Read 125ºC 125ºC 105ºC 105ºC 85ºC 85ºC
Time 8V 12V 12V 16V 12V 16V

24 hrs. 0 fails 0 fails 0 fails 0 fails 0 fails 0 fails

100 " 1 " 4 0 0 0 0
150 6 " 3 1 1 0 0
250 6 " 23 3 10 1 1
500 23 " 13 19 21 1 5
750 9 " 5 15 10 3 15
1000 3 " 1 6 5 5 6

Total fails 48 49 44 47 10 27

76
The complete table of the stress versus life data provides an opportunity to analyze the data for
the two stresses and even look at the possibility of interaction between the two stresses. We know
within the interval when each capacitor failed, but do not know the failure modes of the
capacitors. The time to failure for each of the six data sets shows similar behavior. The highest
stress combination (set B) experienced the most failures while the lowest stress combination (set
E) experienced the fewest failures.

Cap Set B Failure Data

99.00
Weibull
Data 1

W 2 MLE - RRM M
90.00
F=49 / S=1

50.00
Cum Percent Failure

10.00

5.00

James Mclinn
Rel-Tech
1.00 3/14/2010 18:04
10.00 100.00 1000.00
Hours


Figure 6.3 – Data Set B – MLE Method

The analysis of the data in Table 6.2 must deal with the following questions:

1. What statistical distribution(s) do these capacitors follow for the two stresses?
2. What is the temperature acceleration based upon this data?
3. What is the voltage acceleration based upon this data?
4. Is there an interaction between the stresses in this test?

Now question one may be addressed by looking at the six sets of data. These various time-to-
failure distributions were tested for a “best fit” statistical distribution. The following summary is
typical of the all the six data sets. Only the MLE was used for analysis.

77
 Figure 6.3 shows the data from Set B as plotted with the MLE method and Figure 6.4
shows the same data plotted with the Rank Regression method. The basic parameters for MLE are
=1.59 and  = 332.5 This contrasts to the same numbers as calculated with Rank Regression on
Y which are = 1.63 and  = 306.1. Data set B has only one suspension out of 50 samples. It is
difficult to determine the best fit when looking at each set of the lines on the Weibull graph.
Normally, the data from Rank Regression and MLE would yield closer results as was the case
with data set B. It is not certain why these two methods differ so greatly for data sets D and E.
The nature of the interval data undoubtedly contributes to this difference as well as the large
number of suspensions for these two sets.
The two parameter Weibull appears to be the best choice per table 6.3. A three parameter
Weibull doesn’t fit well and a mixed population Weibull has no basis in information. Thus, a two
parameter Weibull was selected for all data sets. Perhaps the mixed population might be a better
fit, but there is no evidence for multiple failure modes, so this option was discarded. The Weibull
is better than the Normal and Log-Normal distribution, based upon the fit parameters, so Weibull
will be used for all additional modeling and calculations of this problem. Since all test conditions
finished with units non-failed, the MLE is the preferred analysis method. Figures 6.3 and 6.4
show how different these two calculations may become. The MLE data set will be the one
employed to complete the analysis of the stress versus life relationship in the later sections. Table
6.5 shows additional calculations for differences between Rank Regression and MLE.

Table 6.3 – Models for Reliability

Data set Model   L Portion

Set B Weibull – 2p 1.59 332.5 -82.1

Weibull – Mixed 2.27 225.1 -316 0.694
2 modes 1.87 541.8 -316 0.306

Mean Std Dev. L

Log-Normal 5.50 0.612 -77.6
Normal 296.5 196.5 -91.7

Model   L

Set D Weibull – 2p 1.92 540.6 -78.6

Weibull – Mixed unable to calculate for MLE
Mean Std Dev. L
Log-Normal 6.02 0.564 -75.2
Normal 477.0 255.9 -83.6

Model   L

Set F Weibull – 2p 2.35 1092.0 -67.3

Weibull – Mixed unable to calculate for MLE

Mean Std Dev. L

Log-Normal 6.83 0.599 -66.1
Normal 928.9 383.5 -68.8

78
 Cap Set B Failure Data
99.00
Weibull
Data 1

W 2 RRY - RRM M
90.00
F=49 / S=1

50.00
Cum Percent Failure

10.00

5.00

James Mclinn
Rel-Tech
1.00 3/14/2010 18:02
10.00 100.00 1000.00
Hours


Figure 6.4 - Data Set B – Rank Regression Method

Table 6.4 - Analysis of Life Data from Table 6.2

Data Set A Set B Set C Set D Set E Set F

At MTBF 418.2 hours 298.2 591.0 479.6 1704.2 967.7

Characteristic Life - 
470.0 hours 332.5 667.3 540.6 1923.7 1092

Time to 10% failures

132.8 hours 80.7 234.3 167.4 723.1 419.1

Time to 20% failures

202.4 hours 129.4 332.2 247.5 1002.1 576.8

Value of 
1.78 1.59 2.15 1.92 2.30 2.35

79
Table 6.4 shows us that there is some variation along the time to failure axis. This is because
when we let each data set be calculated independently. The lines reflect the different values of 
that resulted. This means the best fit lines for each data set are not parallel, so the point for
calculating the stress versus life relationship makes a difference. In the later calculations I have
used the time to 10% failures for all comparisons.
Figure 6.5 shows the situation when a best fit slope was made for all of the data. The
average slope was 2.10 and one can see that the best fit lines don’t always fit the data points very
well in this compromise. Thus one attempts to best fit each time to failure data set or
compromises with one best fit model for all of the data. Using a single best fit model for all six
data sets would miss issues such as interaction between the stresses and changes to the stress-life
relationship as a function of stress. Therefore, I have elected to treat the six data sets as
independent samples and not use the best fit lines depicted in Figure 6.5.

Table 6.5 - A Comparison of Data Sets D and E

Data Set slope 10% Characteristic Method

 Life Life - 

Set B 1.59 80.7 332.5 hours MLE
49 fails 1.63 76.8 306.1 hours Rank Regression

Set D 1.92 167.4 540.6 hours MLE

47 fails 2.66 202.3 472.0 hours Rank Regression

Set E 2.30 723.1 1923.8 hours MLE

10 fails 1.868 641.4 2139.4 hours Rank Regression

A quick look at the data sets D and E is enlightening. Data set D, exhibited 47 total
failures in the sample of 50 at 105ºC and 16 Volts while data set E showed a total of 10 failures at
85ºC and 12 Volts by the same 1000 hours. These are very different results, it seems the data sets
should be a smooth progression from the lowest stress combination (set E) to the highest stress
combination (set B). In order of increasing stress, the sets are Set E, Set F, Set D, Set C, Set A
and Set B. Table 6.2 also suggests that data sets A, B and C reached a peak fall-out during test
while data sets E and F may not have. Data set E is the lowest stress combination, with set F
being the second lowest stress combination and only other date set similar to E. This suggests
something unusual may be happening above 85ºC. Such an observation would suggest a threshold
for different behavior and this question will be pursed in further analysis of section 6.3. Figure
6.5 shows all of the data plotted on a single graph and this figure can’t reflect the data set
differences shown in table 6.4.

80
 Capacitor Life Test
99.00
T-NT/Weib
Data 1
90.00 358.2 / 12
F=50 | S=0
358.2 / 16
F=27 | S=23
378.2 / 12
50.00 F=44 | S=6
378.2 / 16
F=47 | S=3
398.2 / 8
Unreliability

F=48 | S=2
398.2 / 12
F=49 | S=1

10.00

5.00

James McLin
Rel-Tech
3/15/2010
8:44:23 AM
1.00
10.00 100.00 1000.00
Hours

Beta=2.1002, B=3540.6105, C=0.1479, n=0.4006

Figure 6.5 – All Capacitor Data with a Best Fit Slope

6.3 - Solutions to Accelerated Life Test of Capacitors

A check of the raw data will be performed by calculating the Activation Energy and
power law unknowns based upon an Arrhenius model for temperature and a power law model for
voltage. The initial measure of life will employ the time to 10% failures. This measure should be
fairly consistent across the various combinations of stress employed in the test as shown in Table
6.4. The formula for life based upon temperature stress is through the Arrhenius relationship as:

E a (11605 )
273.2  C
Life = Ae (6.1)

The equation set from the data in Table 6.2 can be developed as:

E a (11605 ) E a (11605 )
398.2 358.2
125ºC, 12V 80.7 hours = A e 85ºC, 16V 419.1 hours = A e

E a (11605 ) E a (11605 )
358.2 378.2
85ºC, 12V 723.1 hours = A e 105ºC, 16V 167.4 hours = A e

81
Dividing the two lower equations in each group into the one above eliminates A and gives:

3.2545 Ea 1.7133 Ea
8.9603 = e 2.50358 = e

2.1928 = 3.2545E a 0.91773 = 1.7133 E a

E a = 0.6738 E a = 0.5357

The third pair of points, set B, (at 125ºC, 12V) and set C, (at 105ºC, 12V) yields the highest value
of the activation energy as E a = 0.6916. The combined average from these three data points and
the others that can be calculated is E a = 0.634.
These numbers show scatter, but are considered reasonably consistent if they do not show
more than about 5% scatter from maximum to minimum. We accept the results as a preliminary
estimate for this analysis. Ordinarily, a more detailed analysis should be considered in order to
look at the consistency of the numbers.
Using the voltage versus life information we likewise calculate for an estimate for N. The
power law relationship this stress versus life is taken as:

Life = B(V)  N (6.2)

The three pairs of equations can be formed by the six data sets. These are shown after dividing
one by the other to eliminate the unknown, B:

132.8 8 234.3 12 723.1 12

= ( ) N = ( ) N = ( ) N
80.7 12 167.4 16 419.1 16

1.6456 = (1.50) N 1.39964 = (1.333) N 1.72536 = (1.333) N

0.49811 = N(0.40547) 0.33618 = N(0.28768) 0.54544 = N(0.28768)

N = 1.2285 N = 1.1686 N = 1.8961

More scatter exists in these values of N than is desirable. The average is N = 1.431, but the
scatter is close to + 35 % for this value.
The scatter in the calculations of N and E a suggest there may be a problem with noise in
the data or in the simple models selected. Such scatter may also reflect problems with the
components themselves, with the test method, with the measurement method, with the definition
of failure or even the test equipment itself. Each should be carefully considered and then
eliminated as a source of data scatter. Standard analysis can separate the effects of temperature
and voltage, but may not be able to handle the possibility of stress interactions that may also
complicate this case. Since there are six data points and many different stress conditions, one can
evaluate this possibility and determine if stress or stress interaction has a non-linear effect on
these samples.
Summarizing the same data sets by use of the time to 10% failures and time to 20%
failures presents an additional chance to review the earlier analysis. This data provides:

82
 10% failure 20% failure
N = 1.431 1.349
Ea = 0.634 0.578

Likewise, we can use the characteristic for each of the data samples as a way of estimating the
model unknowns. The results are:

@ 125ºC/8V characteristic life = 470 hrs. @105ºC/16V characteristic life = 540.6 hrs.
@ 125ºC/12V characteristic life = 332.5 hrs. @ 85ºC/12V characteristic life = 1923.7 hrs.
@ 105ºC/12V characteristic life = 667.3 hrs. @ 85ºC/16V characteristic life = 1092 hrs.

With these characteristic life numbers, we calculate values of N = 1.185 and E a = 0.467 .
Table 6.6 shows estimates for N and E a across these different definitions of life. There is
a steady downward trend in values of N and E a as the life definition gets longer.

Table 6.6 - Estimates of E a and N Based Upon Definition of Life.

Method of Life Estimation N Ea

10% time to Fail 1.431 0.634
20% Time to Fail 1.349 0.578
MTBF Fail 1.176 0.463
Characteristic Fail 1.185 0.467

A review of the value of  for the same set of conditions would provide additional
information about the stress dependence of the data. There appears to be some non-linear
dependence between voltage and the value of . Figure 6.6 shows this - at the lowest temperature,
85ºC, there is a different relationship between  then at other combinations of temperature and
voltage. The value of  seems to follow a simple relationship with regard to temperature only, but
not with temperature and voltage. This stress versus life relationship probably does have a strong
impact on the test results and the models generated.
The fact that the voltage seems to show some changes suggests that this is an area to
consider when looking at potential interaction issues. There may be non-linear behavior in the
model of this stress.

83
  
2.5 2.5
85C
X X
X

X 105C X
2.0 2.0
X
X
X
X
1.5 1.5
125C

85C 105C 125C 8V 12V 16V

Figure 6.6 – The Relationship of and Temperature with Voltage

6.3.1 - A Non-Linear Solution showing Interaction between Stresses

The possibility of interaction between stresses is strong in the previous example of

temperature and voltage. One model of interaction will be explored in detail, though many others
may exist. The following short list of interaction models was considered. Those marked "does not
work well" in Table 6.7 may provide unphysical answers when fitted to the data.
Based upon the use of time to 10% failure as a measure of life, the following simple
equation showing interaction was fitted to the data. This is the first entry in Table 6.7.

E a (11605 ) B (V ) Ea (11605 )
273.2  C N 273.2  C
Life = A [e ] [(V) ][ e ] (6.3)

Table 6.7 - Table of Other Potential Interaction Models

Interaction type Life Model Model Status

BVE a
BV TK
e Calculated here
T

BV BV
1+ Calculated here
KT KT

BV + CT 1 + BV + CT Not calculated here

VT 1 + BVT Not calculated here

BEa
B
e VTK Calculated here
VT

84
Three sets of two equations from the six stress levels can be created and these solved for some of
the parameters. These equations are A & B, C & D, E & F. They are:

E a (11605 ) B ( 8 ) E a (11605 )
398.2 N 398.2
A 132.8 hours = A [ e ][(8) ][ e ]

E a (11605 ) B (12 ) Ea (11605 )

398.2 N 398.2
B 80.7 hours = A [ e ][(12) ][ e ]

E a (116050 ( b )(12 ) Ea (116050

C 234.3 hours = A [ e 378.2

][(12)  N ][ e 378.2
]

E a (11605 ) B (16 ) E a (11605 )

378.2 N 378.2
D 167.4 hours = A [ e ][(16) ][ e ]

E a (11605 ) B (12 ) Ea (11605 )

358.2 N 358.2
E 723.1 hours = A [ e ][(12) ][ e ]

E a (11605 ) B (16 ) Ea (11605 )

F 419.1 hours = A [ e 358.2

][(16)  N ][ e 358.2
]

Dividing these two equations A &B gives:

B (8 ) E a (11605 )
398.2
132.8 8 e
= 1.645601 = ( )  N B (12 ) Ea (11605 )
80.7 12 398.2
e

Working out the math and taking the ln of both sides we have:

0.49811 = N(0.405465) – 116.5744 (E a )(B)

Now doing the same with data sets C and D will give:

B (12 ) Ea (11605 )

234.3 12 e 16 e 368.2179( Ea )( B )
378.2
= 1.39964 = ( )  N B (16 ) E (11605) = ( ) N 490.9572 ( E )( B )
167.4 16 a
12 e a

e 378.2
122.7392 ( E a )( B )
1.39964 = (1.333) N e

0.33622 = N(0.287657) – 122.7392(E a )(B)

Now doing the same with data sets E and F will give:

85
 723.1 12 129.5924 ( E a )( B )
= 1.725364 = ( )  N e
419.1 16
129.5924 ( Ea )( B )
1.725364 = (1.333) N e

0.557808 = N(0.287657) – 129.5924(E a )(B)

It is easy to eliminate N and solve for (E a )(B), Taking the equations from C &D and E & F
gives:

0.33622 = N(0.287657) – 122.7392(E a )(B)

0.557808 = N(0.287657) – 129.5924(E a )(B)
---------- ------------- --------------------
-0.221566 = 6.8532 (E a )(B)

or (E a )(B) = -0.032334

Doing this several times and finding an average values yields:

( E a )( B ) = - 0.012088

with the following equations, one can solve for N

0.49811 = N(0.405465) – 116.5744 (E a )(B)

0.49811 = N(0.405465) +1.409151

-0.911041 = N(0.405465)

N = -2.2469

After several other equations we calculate N = - 3.2475 = -3.25

Taking equations for data sets D & E gives:

E a (11605 ) B (12 ) Ea (11605 )

723.1 hours = A [ e 358.2

][(12)  N ][ e 358.2
]

E a (11605 ) B (16 ) E a (11605 )

378.2 N 378.2
167.4 hours = A [ e ][(16) ][ e ]

dividing gives

1.7133 Ea 102.18 ( E a )( B )
4.3196 = e (1.333) N e

(1.333) N e 1.2352
1.7133 Ea
4.3196 = e

86
 1.46316 = 1.7133 E a + N(0.287657) +1.2352

1.46316 = 1.7133 E a -0.934022 +1.2352

1.16198 = 1.7133 E a

or E a = 0.678

Solving the set of equations for the other values leads to:

N = - 3.247 E a = 0.678 B = - 0.012088

Does this result solve the problem with the data? The answer is an emphatic no! All that was
done was to find a model with interaction that seemed to fit the data. This is not proof that it is the
right solution or the best solution. The exercise is one of model fitting only. The fact that N is
negative should be a warning that something may be amiss such as stress dependent results. There
is not enough data to resolve this question.
Note that the values of N and E a are different from that found in section 6.3 where N =
1.431 and E a = 0.634. This is because the suggested interaction including a value for "B"
strongly impacts the other two unknowns. The same change would have been observed with A if
this unknown had been calculated. This suggests that the small value of B is big enough to
significantly change the relationship of the unknowns.
Another model of Table 6.7 was also tried using the exponential interaction model, which
is the last entry of the table.

E a (11605 ) BE a (11605 )
273.2  C N (V )( 273C )
Life = A [e ] [(V) ][ e ]

The results of this exponential model with the same data set was:

N = - 2.695 E a = 0.560 B = + 2.9249

A last model of Table 6.7 was also tried using the linear interaction model, which is in
the middle of the table.

E a (11605 )
BV
Life = A [e 273.2  C
] [(V)  N ][1 + ]
KT

The results of this linear model with the same data set were:

N = + 2.4426 E a = 0.581 B = - 0.27027

This result of the linear model seems the most sensible of the three calculated and is the closest
to the results of section 6.3 calculations that had no correction for interaction.

87
7.0 – Degradation Test Methods
This section will briefly summarize some of the knowledge presented in the prior
sections and then point the direction for future ALT work. Additional examples are shown in this
section. The emphasis is on degradation approaches to ALT conditions. Most often these
conditions are experienced with mechanical components and systems, wear dominated situations
and some electronic assemblies.

7.1 Background

An outline of the expected ALT activities was developed earlier. These represent crucial
activities required to be successful. If we were to estimate the time that should be allotted for each
of the preliminary activities, the following percentages might be typical for many situations and
test plans.

1) Planning the ALT - About 5% of the total test time for a six week test might be placed in this
area for a first-time test. This would be from about 20 to 50 man-hours total. Often less than 10
hours is spent by many engineers for planning purposes. It is easy to neglect preparation time
when tests are repeated or are similar to something that may have been performed before. Ten
hours is too little time, when one considers that equipment should be checked, calibration should
be performed and samples collected and randomized.

2) Setting up the Life Test - The initial tests of the equipment and any calibrations required
should consist of about 3% of the total time for a first-time test. This is about 30 man-hours in a
1000 hour test. This is equipment check-out, set up of the fixtures, verification of the set-up and
initial readings to verify components. Be sure to measure noise and repeatability of the measures.
For a repeat test this could fall to as few as five hours for well-designed, calibrated and proven
equipment.

3) Administering the Test - This stage should consume about 7% of the total test time or up to
70 man hours for a first-time test with first time equipment and fixtures. It includes time for
periodic electrical and mechanical checks of all the samples at several points during the test. Less
than this may be consistent with a short test or one that has automated monitoring. This total time
includes quick calculations, any consistency analysis and checks on the fly. If testing of samples
is long this stage could easily double in length.

4) Successful Conclusion Stage of Test - This stage might consume about 3% of the total test
time or up to 30 man-hours for a first-time test. This assumes that no unusual test sample
behavior, equipment failure, calibration problem or sudden catastrophic sample failures. This
stage includes all final tests with any special tests required to complete the ALT and document
changes. Repeat ALT tests may be accomplished with less time.

5) Analyzing the Data - This stage may consume up to about 4% of the total time of the test or
perhaps 40 man-hours to analyze and summarize all of the test results. One should also produce a
formal report for management and for the future. Less time may be required with repeated tests or
for automated data collection and analysis systems. Be sure to close the loop on any failures with
some failure analysis.

Now that all of the preliminary stages have been totaled, the actual test time for the test samples
might be defined as about 78% of the total test. That is the six weeks or the “Test Time” is:

88
 Test Time = 100% - (5% + 3% + 7% + 3% + 4% ) = 78%

This 78% is the minimum active test time for the test samples. Thus, a 1000 hour test could really
take about 1280 hours or as high as 1425 hours to run from the start to finish for a first-time test.
This is just under 9 weeks and hasn’t allowed all the needed time for failure analysis near the end.
Repeated tests may lower this time to as low as 1150 hours, but most are longer than 1200 hours.
These numbers are not unusual, but it is easy to find situations where some of these percentages
may get larger or smaller. About 3/4 of the planning time, (3/4 of 22%) or 15%, actually went
into a series of proactive activities associated with planning, setting up, calibrating and
administering. The remaining percentage, or 7%, covers the remaining completion activities. The
message is to be proactive for ALT tests, but then to pay close attention to the analysis. Don’t
assume your tests will run as planned and be completed in a minimum time.
The engineering portion of any Accelerated Life Test might be said to be made up of
studying the failure modes, the cause and effect relationship between applied stress and the
resultant response of the system. Items such as time to failure, failure mechanisms and the
Physics of Failure should all be studied. Accelerated Life Test results are often comprised of the
projected device life, any system robustness or resistance to customer abuse and all the
maintenance activities required to keep a system operating in the field. These activities all lead to
more than customer satisfaction. Rather they are directly tied to corporate reputation, corporate
profits and future market expansion or contraction.
Think of these five ALT steps as part of a ballet. A ballet is a planned and choreographed
activity that requires preparation, timing and practice to do it consistently well. A number of
different seemingly chaotic activities are brought together in a carefully timed sequence to
achieve balance and harmony. A hockey game, on the other hand, is a disorganized activity.
While a hockey coach plans plays and has the team practice these set plays before any game, once
the game begins, things sometimes change quickly. The referee starts the game by blowing the
whistle and then "drops the puck". The next thing that happens is people are skating all over the
ice, with each team trying to simultaneously implement both an offensive and defensive game
plans. This usually leads to neither plan working as expected, since opponents have their own
plans. So the players "make it up as they go" and skating may or may not follow game plans.
There have never been two hockey games that were the same because of the need to adjust the
game plans on the fly. The hockey game approach to Accelerated Life Test is to be avoided. The
ballet is preferred.

7.2 The Concept of Degradation

Degradation is the slow change of characteristics of a system that occur while the system
operates under stress with environmental conditions. Sometimes simple un-lubricated operation is
sufficient for degradation. At other times environmental factors such as humidity or salt air may
drive degradation. One needs to look at degradation situations differently. Many times it is hard
to make a meaningful, noise free measurement since the when the characteristic being measured
is slowly changing. The application of statistics may be required to “pick the signal out of the
noise”. A few examples may help enlighten these challenges faced by such an ALT.
Degradation models can reduce the size of samples needed to determine reliability
because the time to failure shows a smooth and progressive change. Statistical methods aid the
measure of performance degradation since it is usually smooth. This continuous situation is
always better than a pass-fail (binomial) model of time to failure. A model is required for
degradation and the parameters of this model may be dependent upon the stresses present. For
more complete understanding it is best to tie the model to failure modes and Physics of Failure.
Degradation testing is commonly used for basic semiconductors, linear integrated circuits, tensile

89
strength of materials, corrosion situations, resistance changes, breakdown voltage changes of
many components, gain of transistors, flexural strength of materials, elongation changes (plastic
behavior) of materials, creep, wear, lubrication and other failure modes.

Example 7.1 - The Accelerated Life Test of an Electro-Mechanical Assembly – This example
highlights degradation through all five steps ranging from planning to set-up to the final data
analysis. Now think of something that would help put you and keep you into the ballet mood.
Practice is essential to ballet as it is to ALT. The results are based upon a real ALT test I ran for a
component supplier.

Background - Imagine you work as a reliability engineer for a manufacturer of a small

electromechanical component. A test lab with sensitive equipment and a reasonable test budget is
available to you. Periodic accelerated life tests are run to qualify new products and processes. The
information will usually go directly to management, marketing and perhaps to some select
customers as a detailed report. The following data was collected as part of the ALT run to
determine how long a product would last in the field. All conditions are accelerated unless
otherwise noted.

I - Planning the Test - Select a series of typical samples of the new product to be tested. We
assume that you wish to know the relationship, if any, between some of the product performance
variables and the device life as a function of applied stress. Now select the most important
stresses to be employed in the test and select the levels of the stress. Be sure to consider the worst
case customers, the tendency for customer abuse and any other field information.
Fifty conforming samples were selected randomly from pre-production. These samples
were only tested at room temperature and nominal conditions. Divide the 50 samples into three
groups of unequal size. One group was selected because it is at, or near, the low end of some
performance specification. Similarly, the second group was selected because it is in the middle of
the performance range and the last group was selected because it is at, or near, the high side of
some performance specification. Figure 7.1 shows these three selected groups. Note, we are
assuming that all the samples are in specification or at the edge of specification and that samples
actually cover the full range of the specification in a reasonable fashion. The statistical
distribution is shown by Figure 7.1. The bulk of the quantity is near the middle of the
specification with some parts marginally beyond the limits of specification. Normally, these out
of conformance parts would be removed at some quality conformance step. In our case, some of
the out-of-specification parts will be deliberately included in these accelerated tests for
information purposes. These out of specification units might add about 10 more samples to the
total in test.
The test levels will be selected as three different temperatures, 70ºC, 105ºC and 145ºC.
This stress will be combined with three different operating voltages, +12 V DC, +16 VDC and a
+12 VDC, alternating voltage condition, that changes direction every hour. A total of nine test
conditions can be run with three different test chambers. A definition of failure is selected as a
component change from the initial measure by more than 10%. All units will remain on test until
the test ends, a catastrophic failure of the part occurs or some major event. The + 12VDC
condition was selected to address a specific failure mode.

90
 Lower Middle
Sample Upper
Sample
Sample

Lower Upper
Spec. Spec.
Limit Limit
XXXXXX
XXXXXXXXX
XXXXXXXXXXXXXXXX
X X XXXXXXXXXXXXXXXXX X

Figure 7.1 – Distribution of a Measured Parameter

II - Setting up the Test - This stage of the test will involve getting the three chambers and test
equipment ready. At this step, the measuring devices that will be used are checked and
recalibrated as necessary. The test chambers are checked to be sure they actually operate as
desired and none of the equipment will go out of calibration before the test ends.
Next, a small repeatability study looks at the noise from measuring and re-measuring
some of the samples using five "golden units". These are also checked to verify the readings
remain consistent. These "golden units" are similar devices that have a long history of stable
measurements. One golden sample should be near the middle of the specification range. A second
sample should measure at the upper edge of specification. A third sample should measure at the
lower edge of specification. Of the remaining two golden samples, one should measure just "out
of specification" on the low end and one just "out of specification" at the high end. These will be
used for correlation during the periodic test stops that are used to make performance
measurements. All other equipment to be employed in the ALT will be given a brief check to
verify stability and sensitivity. All of the samples are labeled and then initial readings taken. A
noise figure is estimated for each sample from the three independent readings taken for each.

III - Administering the Test - This step starts with dividing up the approximately 60 samples
into nine groups. Six samples will be placed in each of the nine groups. The distribution of the six
samples will be as follows: Three samples will be from the middle of the distribution
specification range, one sample will come from the high sample range and one sample from the
low sample range. In addition, one "out-of-specification" sample (if available) will be placed in
with each of the nine different test conditions. Thus, these six samples at each condition cover the
range of devices. Three of the five remaining good samples and one bad sample will go into the
lowest temperature chamber bringing the total of to eight each at these test conditions. The
remaining five golden samples will be kept in reserve as needed for correlation. These will be
periodically rechecked during the ALT.
The test is planned for 1000 hours or about six weeks. Periodically, the test will be
stopped and the samples removed from the test chambers and allowed to cool to room
temperature and then the appropriate specifications will be measured. This measurement of all the
test samples and the golden (correlation) units will occur at zero hours, at 24 hours, 96 hours, 192
hours, 384 hours, 768 hours and 1000 operating hours or at the most convenient test times close

91
to these proposed test points. There is nothing magic about these times for check, but they are
spread apart enough to quickly identify the possibility of changing performance measures.
A check of the required measuring equipment ensures it will be available throughout the
test. A backup piece of measuring equipment may also be identified, and if necessary, initial
measurements could be made with this equipment as well as the original equipment. Thus, a
qualified alternate piece of measuring equipment would be available.

IV - Bring the Accelerated Life Test to a Successful Conclusion. - At this stage, the ALT
should have been running for at least 400 hours. During that time some slight changes of the good
samples and the out-of-specification samples have already been observed. Table 7.1 provides the
details of the results at this point in time. We will use the analysis of these results to make
decisions as we move ahead.

Table 7.1 - Preliminary Test Results

Sample Operating Test Time

Conditions
24 hrs. 96 hrs. 200 hrs. 384 hrs.

70ºC +0.08% -0.19 % -0.28% -0.37%

12 VDC change change change change

105ºC +0.14% -0.23% -0.36% -0.48%

12 VDC change change change change

145ºC +0.12% -0.30% -0.49% -0.61%

12 VDC change change change change

70ºC +0.07% -0.24% -0.44% -0.57%

16 VDC change change change change

105ºC +0.04% -0.29% -0.38% -0.46%

16 VDC change change change change

145ºC -0.01% -0.32% -0.47% -0.64%

16 VDC change change change change

70ºC 0.00% +0.02% -0.03% -0.06%

+12 VDC change change change change

105ºC +0.02% +0.06% -0.05% -0.10%

+12 VDC change change change change

145ºC +0.05% +0.01% -0.12% -0.19%

+12 VDC change change change change

The repeatability of the good samples is well within expectations so far. The golden
samples show about 20% smaller repeatability measures at this time. This sets a lower limit on
the noise of the measurements. None of the samples have met the +10% change for failure. All

92
 samples are below 1% change so far. Table 7.1 presents the average of all the originally good test
samples in each test condition.
Some conclusions can be drawn at this early point of the test. It is not too early to begin
such trend analysis. These conclusions might include:

A) The higher the temperature, the sooner the initial increase begins and the larger the
subsequent decrease becomes.

B) Voltage may have a slight effect. A good statistical test should be performed to look at
this possibility. A simple analysis with DOE may be sufficient.

C) The samples that alternate the polarity of the voltage show smaller changes.

V - Analyzing the Data - At the end of the 1000 hours, the data was summarized into Table 7.2
and then analyzed.

Table 7.2 – Life Test Results of Components

Sample Operating Test Time

Conditions
24 hrs. 96 hrs. 200 hrs. 384 hrs. 750 hrs. 1000 hrs.
70ºC +0.08% -0.19 % -0.28% -0.37% -1.16 % -1.78%
12 VDC change change change change change change

105ºC +0.14% -0.23% -0.36% -0.48% -1.51% -2.34%

12 VDC change change change change change change

145ºC +0.12% -0.30% -0.49% -0.61% -1.98% -3.05%#

12 VDC change change change change change change

70ºC +0.07% -0.24% -0.44% -0.57% -1.76% -2.88%

16 VDC change change change change change change

105ºC +0.04% -0.29% -0.38% -0.46% -2.38% -4.57%#

16 VDC change change change change change change

145ºC -0.01% -0.32% -0.47% -0.64% -3.42%# -5.89%#

16 VDC change change change change change change@

70ºC 0.00% +0.02% -0.03% -0.06% -0.32% -0.67%

+12 VDC change change change change change change

105ºC +0.02% +0.06% -0.05% -0.10% -0.45% -0.81%

+12 VDC change change change change change change

145ºC +0.05% +0.01% -0.12% -0.19% -0.87% -1.22%

+12 VDC change change change change change change

# - One initially out of specification sample changed more than 10%, the average of all samples is shown
@ An initially in specification sample changed more than 10%

93
The following is a list of conclusions drawn from the data of Table 7.2:

A) It appears there is an initial short term increase in the performance measure by the 24 or 96
hour test point. This is followed by a consistent decrease in the measurement out to the end of the
life test. The initial increase may reflect a "wear-in" situation of the product. All of the samples
and test conditions showed the initial wear-in change. Failure analysis could not identify a
physical cause for such behavior. This behavior had been seen before in other ALTs and was
thought to be a characteristic of this family of components. Physics of Failure should be studied
for the initial wear-in. One can conclude there are two microscopic processes occurring for these
samples. The first process causes the increase, while the second process leads to a long term
decrease.

B) The higher the operating temperature, the sooner the initial increase in percentage. It also
appears that the initial increase may also be related to the operating voltage. Results from +12V
show a bigger initial increase than +16V. The alternating voltage shows the smallest initial
increase.

C) The higher the temperature and voltage, the larger the subsequent decrease in the readings as
measured by percentage. There also appears to be a slight correlation between the size of the
increase at 24 hrs. and the size of the decrease at 750 hours.

D) Voltage has an initial effect as well as temperature. A simple DOE analysis should be run to
confirm this interaction. I leave it to the reader to conduct this analysis. The assumptions of the
DOE may not all be accurate, so be a little cautious about conclusions.

E) The samples in the alternate voltage polarity show smaller changes under all test and stress
conditions. That is, these samples have a smaller increase and a smaller decrease than the
constant polarity conditions. The reversing voltage condition appears to be important. The
decreased change is typically 5 to 10 times below that of the normal 12 VDC test condition. The
low degradation under this special condition cannot all be explained by having survived only half
the time in the forward voltage direction. The reverse voltage must have some mollifying effect
on the failure mode or failure mechanism. Failure analysis did confirm there was something
different going on with this alternating voltage sample, but could not pinpoint a mechanism.

F) There was only one failure (more than a 10% change) among the 50 initially good samples at
1000 hours. There were three failures among the 10 samples that were initially out of
specification. The long term drift seemed larger for the units at the lower edge of specification.
This suggests what ever leads to low initial readings is related to the long term drift.

7.3 – Models of Degradation Testing

Figure 7.2 shows a typical degradation curve for this type of data. This will be typical of
all of the possible curves and so a family of curves are not shown. Degradation can be linear,
exponential or logarithmic. These models are shown in Equation 7.1. Each describes the basic
time dependence of a curve. We need to select the most appropriate model for the data.

The Change,  = A - B(t) Linear decrease (7.1)

 = A e  Bt Exponential decrease
 = A – B[ln(t)] Log decrease
 = A - B(t) N Non-Linear decrease

94
Other forms of degradation exist and the reader is referred to Nelson [7] for additional models.
Fitting the data shown in Table 7.2 to one of the equations shown in Equation 7.1 would lead to
possible parameters. These degradation models can also be combined with stress terms to show
the impact of these terms.
Now Figure 7.2 shows a short term increase and a long term decrease. Therefore, we
select a model that can cover both events. This form of the degradation will be:

 = A + B(t) - e C (t ) (7.2)

This model yields the following for the unknowns A, B and C with this data set.

A = 0.99698, B = - 0.00159 and C = - 0.006202

Pick another model such as a simple linear model:

 = A - B(t)

and A = -0.002334 with B = + 0.0000139

This simple model is a poor fit at low time since it doesn’t address the increase. A more complex
model would be needed. Use the power model below:

 = A - B(t) N

Yields A = -0.003226, B = 4.41 x10 8 and N = 1.861

This model is also a poor fit at low time.

+0.5%
Change X 100 200 400 800 1000
from
Initial X
-0.5% X
X

-1.0%

-2.0% X

X
Figure 7.2 – The Change in Characteristics at 145ºC and 12V

95
If an Eyring Model is employed to relate the change of characteristics to the applied stresses, we
can express it as Equation 7.3. This is:

Ea

 = D(Voltage) N e KT (7.3)

The best fit unknowns for this stress-life equation and the whole data set is:

N = 2.10, E a = 0.102 eV and D = 0.000308.

This model assumes no interaction between the two variables of voltage and temperature. This is
not known in the case of the samples and there isn’t enough data to determine if interaction exists
between the two variables.
No simple model stress versus life model can be deduced for the alternating voltage
condition because only one test conditions was employed in test. The overall analysis of this
degradation data can go many different directions from here. It is now left to the reader to take
the analysis and support or deny any other theories or conclusions.

Example 7.2 - The Life of Insulating Oil - Table 7.3 contains the time to failure of an insulating
fluid at various levels of high voltage stress. The original data set is from Nelson [7, pages 89 and
129] and he treats this as an example of the Power Weibull degradation model. In this model, the
characteristic life is a power of the stress. This is:

 = e C (V)  B = A(V)  B (7.4)

Table 7.3 - Life of Insulation Fluid in Minutes

Stress Time to Insulation Failure

26 KV 5.8, 1579.5, 2323.7 minutes
28 KV 68.9, 108.3, 110.3, 426.1, 1067.6 minutes
30 KV 7.7, 17.1, 20.5, 21.0, 22.7, 43.4, 47.3, 139.1, 144.1, 175.9, 194.9 minutes
32 KV 0.27, 0.40, 0.69, 0.79, 2.8, 3.9, 9.9, 14.0, 15.9, 27.8, 53.2, 82.9, 89.3, 100.6, 215.1
34 KV 0.19, 0.78, 0.96, 1.3, 2.8, 3.2, 4.2, 4.7, 4.9, 6.5, 7.4, 8.0, 8.3, 12.1, 31.8, 32.5, 33.9,
36.7, 72.9 minutes
36 KV 0.35, 0.59, 0.96, 0.99, 1.7, 2.0, 2.1, 2.6, 2.7, 2.9, 3.7, 4.0, 5.4, 13.8, 25.5 minutes
38 KV 0.09, 0.39, 0.47, 0.73, 0.74, 1.1, 1.4, 2.4 minutes

A review of the data suggests one point appears in question. It is the test failure time at
5.8 minutes at 26 KV. This appears far too early as compared to the remaining two times at 26
KV. When the data at 28 KV is also considered, this time to failure still seems too short. A
statistical analysis of this point confirms this point does not appear to be part of the data set at 26
KV. The remaining data (less the 5.8 minute point) was analyzed by the MLE via Weibull. Table
7.4 is the summary of the times to failure from the raw data of Table 7.3. Based upon the results
shown in Table 7.4, a three parameter Weibull analysis was selected. The goodness-to-fit is also
shown in this table. The two-point version of 26 KV has a low goodness-to-fit, but the two-point
version was selected for the overall analysis because the 5.8 minute time to failure is so far out of
what might be expected at 26 KV.

96
 Probability Weibull

99.00
IPL/Weib
Data 1
90.00 28
F=5 | S=0
30
F=11 | S=0
32
50.00 F=15 | S=0
34
F=19 | S=0
36
Unreliability

F=15 | S=0

10.00

5.00

James McLinn
Rel-Tech
3/18/2010
3:46:33 PM
1.00
0.10 1.00 10.00 100.00 1000.00 10000.00
Time

Beta=0.7534, K=2.8487E-28, n=17.3040

Figure 7.3 – Basic Analysis of Oil Degradation

Table 7.4 - MLE Weibull Analysis of Insulating Fluid

Stress    L MLE Goodness-of-Fit Measure
26 KV 3 points 1.74 1912 -409 -24.8
--------------- -------------------------------------------------------------
26 KV 2 points 13.75 4447.5 -2323.9 -14.7
28 KV 3.82 1572.2 -1068 -36.9
30 KV 4.22 297.9 -195 -62.3
32 KV 4.12 279.8 -215 -83.4
34 KV 4.43 95.2 -72.9 -84.1
36 KV 4.17 32.8 -25.5 -50.9
38 KV 1.15 0.88 +0.07 -6.6

97
 Oil Degradation
99.00
Weibull
Data 1

90.00 W 2 MLE - SRM M

F=15 / S=0
Data 2

W 2 MLE - SRM M
F=19 / S=0
Data 3
50.00
W 2 MLE - SRM M
Cum Percent Failure

F=15 / S=0
Data 4

W 2 MLE - SRM M
F=11 / S=0
Data 5

W 2 MLE - SRM M
10.00 F=5 / S=0

5.00

James Mclinn
Rel-Tech
1.00 3/18/2010 16:04
0.10 1.00 10.00 100.00 1000.00 10000.00
Minutes





Figure 7.4 – Independent Analysis of Oil Degradation Data

Table 7.4 shows that the time offset for the two-point analysis at 26 KV is consistent with the rest
of the data in the table. The original data set (three-points) version at 26 KV is not consistent with
the rest of the data. The data at 38 KV must also be suspect because the times are so short and the
slope, , is very different from all of the other slopes. The data at 26 KV should not be used in
any analysis of stress versus life, since it also has a very different slope. Better yet, additional data
points should have been obtained at 26 KV when the test was originally performed. The use of
only three points at 26 KV seems to be too few for reliable results. Some additional testing is
typically required with questionable results. Further analysis will be done without this one stress
condition. Figures 7.3 and 7.4 shows the five stress levels ranging from 28 KV to 36 KV. These
seem to be similar to each other, but one is a common slope while the other is an independent
analysis of the data.
Employing the power Weibull model as shown in Equation 7.4, yields the following fit to
the unknowns for this equation.

A = 3.04 x 10 24 with B = 15.33

98
 Figure 7.5 shows the same data set as Figure 7.4, except that the data is presented as a 3
parameter Weibull analysis. This model was selected as the best fit approach to the degradation
data. Most of the Weibull slopes range from 0.47 to 0.87.

Oil Degradation
99.00
Weibull
Data 1

90.00 W 3 MLE - SRM M

F=15 / S=0
Data 2

W 3 MLE - SRM M
F=19 / S=0
Data 3
50.00
W 3 MLE - SRM M
Cum Percent Failure

F=15 / S=0
Data 4

W 3 MLE - SRM M
F=11 / S=0
Data 5

W 3 MLE - SRM M
10.00 F=5 / S=0

5.00

James Mclinn
Rel-Tech
1.00 3/18/2010 16:38
0.01 0.10 1.00 10.00 100.00 1000.00 10000.00
Minutes







Figure 7.5 – The Three Parameter Analysis of Oil Degradation

Figure 7.6 shows a graph of the log of Stress versus the Log of Life which was employed
to calculate the unknowns for Equation 7.3. This graph suggests that there may still be some non-
linear behavior present with these five stress levels. As a whole, the results from these five stress
levels are acceptable when the data from 26 KV and 38 KV are excluded. When the two extremes
are included in the stress versus life model, it is not possible to produce a simple model for stress
versus life. Otherwise, this example met the requirements outlined in this section. This data set is
not atypical of such ALT results. Most of the time, additional test samples would be run to
solidify the questions at highest and lowest stress conditions.

99
 Ln Stress Versus Ln Life
6

4
Ln Life

0
3.3 3.4 3.5 3.6
Ln Stress

Figure 7.6 Ln Stress versus Ln Life for the Oil Degradation

This graph shows the there is a reasonable relationship between the applied stress son the oil and
the expected life. At higher stresses, this relationship may break down quickly as new failure
modes may arise. At lower stress this relationship may extend for a few KV, but it is expected
that it would become non-linear at some point.

100
References
1. M. Sobel and J.A. Tischendorf, "Acceptance Sampling with New Life Test Objectives",
Proceedings of the 5th. National Symposium on Reliability and Quality Control in Electronics,
Philadelphia, Jan. 9-11, 1959, pp. 108-118
2. B. Epstein and M. Sobel, "Sequential Life Tests in the Exponential Case", Annals of
Mathematical Statistics, Vol. 26 (1955), pp. 82-93
3. Statistical Techniques for Life Testing, PB171580, U.S. Department of Commerce, Office of
Technical Services, Washington 25, DC, July 1961
4. Sobel, M., "Statistical Techniques for Reducing the Experiment Time in Reliability Studies",
Bell System Technical Journal, 35, pp. 179-202, Jan. 1956
5. Yurkowski, W., Schafer, R.E., and Finkelstein, J.M., "Accelerated Testing Technology" Rome
Air Development Center Technical Report, RADC TR-67-420, Griffiss AFB, 1967
6. AGREE, Reliability of Military Electronic Equipment, Report of the Advisory Group on
Reliability of Electronic Equipment, Office of the Assistant Secretary of Defense, Washington
DC, June 4, 1957
7. Nelson, W. , Accelerated Testing, Wiley, 1990, New York
8. Kececioglu, Dr. Dimitri and Sun, Feng-Bin, Environmental Stress Screening, Its
Quantification, Optimization and Management, Prentice Hall, 1995, Saddle River N.J.
9. Doty, Leonard, Reliability for the Technologies, ASQ Quality Press, 1989, Milwaukee
10. O'Connor, Patrick, Practical Reliability Engineering, John Wiley & Sons, 4 th edition, 2002,
New York
11. Nelson, Wayne, "Weibull Analysis of Reliability Data with few or No Failures", Journal of
Quality Technology, Vol. 17, No. 3, July 1985, pp. 140-146
12. Life Data Analysis Reference, published by the ReliaSoft Corp., 1450 S. Eastside Loop,
Tucson, Arizona 85710-6703
13. Abernethy, Dr. Robert, The New Weibull Handbook, published by ASME, 1999, New York
14. McLinn, James A., The Weibull Analysis Primer, published by the Reliability Division of
ASQ, email JMReL2@Aol.com for a copy
15. McLinn, James A., "Improving the Product Development Process", Proceedings of the 48th.
Annual Quality Congress, Las Vegas, pp. 507-516, published by ASQ, 1994
16. Raheja, Dev, Assurance Technologies, Wiley, 1991, New York
17. Jensen, Finn, Electronic Component Reliability, Wiley, 1995, New York
18. Carter, A.D.S., Mechanical Reliability and Design, Wiley, 1997, New York
19. Reliability Prediction Procedure for Electronic Equipment, published by Bellcore (now
called Telcordia Technologies) of Morristown, New Jersey 1 800- 521-2673
20. Kim, K.J. and Boardman, T.J., “Estimation of Mixture Weibull Reliability”, Reliability
Review, Volume 18, no. 2, June, 1998
21. Feinberg, A. and Widon, A, "Connecting parametric aging to catastrophic failure through
Thermodynamics", Transactions of the IEEE, Volume 45, Number 1, March 1996, pp. 28-33
22. Accelerated Life Data Analysis Software produced by the ReliaSoft Corp., 1450 S. Eastside
Loop, Tucson, Arizona 85710-6703
23. McLinn, James A., “A Short History of Reliability”, Reliability Review, Volume 30, no. 1,
March, 2010
24. McLinn, James A., Mechanical Design Reliability Handbook, published by the Reliability
Division of ASQ, email JMReL2@Aol.com for a copy
25. Nachlas, Joel, Binney, Blair, and Graber, Stephen, “Aging Acceleration under Multiple
Stresses”, Annual Proceedings of the RAMS, published by the IEEE, 1985, pp.438-440
26. Dodson, Bryan and Schwab, Harry, Accelerated Testing, SAE, 2006, Warrendale
27. Peck, D.S., “Comprehensive Models for Humidity Testing Correlation”, IRPS, #24, pp. 44-
50, 1986

101