Introduction To Quality and Reliability Engineering - R-Jiang

Springer Series in Reliability Engineering
RenyanJiang
Introduction
to Quality and
Reliability
Engineering
Series editor
Hoang Pham, Piscataway, USA
More information about this series at http://www.springer.com/series/6917
Renyan Jiang
Introduction to Quality and

Reliability Engineering
123
Renyan Jiang
School of Automotive and Mechanical
Engineering
Changsha University of Science
and Technology
Changsha
China
Additional material to this book can be downloaded from http://extras.springer.com.
ISSN 1614-7839 ISSN 2196-999X (electronic)

ISBN 978-3-662-47214-9 ISBN 978-3-662-47215-6 (eBook)
DOI 10.1007/978-3-662-47215-6
Jointly published with Science Press, Beijing

ISBN: 978-7-03-044257-4 Science Press, Beijing
Library of Congress Control Number: 2015939148
Springer Heidelberg New York Dordrecht London

Science Press, Beijing and Springer-Verlag Berlin Heidelberg 2015
This work is subject to copyright. All rights are reserved by the Publishers, whether the whole or part
of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microlms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specic statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publishers, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publishers nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
Springer-Verlag GmbH Berlin Heidelberg is part of Springer Science+Business Media

(www.springer.com)
Preface
Manufacturing businesses need to develop new products and improve current

products to better meet consumer needs in order to survive and grow in a erce
competitive environment. Customers have expectations regarding product perfor-
mance over time. Product quality and reliability are crucial competence factors and
hence the major concerns of manufacturing industries. To achieve world-class
quality, the manufacturer of a product must satisfy customer needs using various
models, tools, and techniques to help manage reliability and quality for new and
current products.
The life cycle of a product refers to several stages from its conception, through
design and manufacture, to service and disposal. Each stage can add value to the
product and the magnitude is characterized by a well-known smile curve shown in
Fig. 1. As can be seen from the gure, the efforts made in the pre-manufacturing
and post-manufacturing stages can result in greater value than the value resulting
from the efforts made in the manufacturing stage. This implies that manufacturing
businesses not only emphasize on the manufacturing stage of the product life cycle
but also need to get into the pre-manufacturing (design and development) and post-
manufacturing (post-sale support). In order to do this, engineers need to be educated
on product reliability and quality.
The education on quality and reliability engineering becomes essential to train
product engineers. This book is written as an introductory textbook for senior
undergraduate and postgraduate students in various engineering and management
programs and can be used as a reference book for researchers and engineers in
related elds. It provides readers with a primary training in quality and reliability
engineering in the real industrial context.
This book focuses on concepts, models, tools, and techniques of quality and
reliability in the context of product life cycle. These can be used for deciding the
reliability for new product, ensuring certain level of quality of the product,
assessing the quality and reliability of current products being manufactured, and
improving the reliability and quality of the product.
The book comprises 17 chapters organized into four parts and some extra
materials. The rst part consists of six chapters and aims to provide basic concepts
v
vi Preface
Fig. 1 Smile curve
ADDED VALUE
RESEARCH AND BRAND AND
MANUFACTURING
DEVELOPMENT SERVICES
LIFE CYCLE STAGES
and background materials such as product life cycle, basic concepts of quality and
reliability, common distribution models in quality and reliability, basic statistical
methods for data analysis and modeling and, models and methods for modeling
failure point processes.
The second part consists of ve chapters and deals with major quality and
reliability problems in product design and development phase. The covered topics
include design for X, design for quality, design for reliability, and reliability tests
and data analysis.
The third part consists of four chapters and deals with quality and reliability
problems in product manufacturing phase. The covered topics include product
quality variations, quality control at input, statistical process control, and quality
control at output.
The fourth part consists of two chapters and deals with product warranty and
maintenance.
The extra materials consist of three appendices and deal with some important
theories and tools, including multi-criteria decision making analysis techniques,
principal component analysis, and Microsoft Excel, with which a number of real-
world examples in this book can be computed and solved. Exercises for each
chapter are also included in extra materials.
This book is the main outcome of the Bilingual teaching program of the course
Quality and Reliability Engineering supported by the education ministry of
China (No. 109, 2010).
The publication of this book was nancially supported by the China National
Natural Science Foundation (No. 71071026 and No. 71371035), the Science
Publication Foundation of Chinese Academy of Sciences (No. 025, 2012), and the
Academic Work Publication Foundation of Changsha University of Science and
Technology, China.
The author would like to thank Prof. D.N. Prabhakar Murthy for his invaluable
support and constructive comments on the earlier outlines and manuscripts of this
book, and thank Profs. Dong Ho Park and Toshio Nakagawa for their comments
and suggestions on the manuscripts of this book.
Contents
Part I Background Materials
1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Product and Product Life Cycle . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Product Life Cycle . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 Technology Life Cycle of a Product . . . . . . . . . . . . . 4
1.3 Notions of Reliability and Quality. . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Product Reliability . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Product Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.3 Link Between Quality and Reliability . . . . . . . . . . . . 7
1.4 Objective, Scope, and Focus of this Book . . . . . . . . . . . . . . . 8
1.5 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Engineering Activities in Product Life Cycle . . . . . . . . . . . . . . . . 11

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Engineering Activities in Pre-manufacturing Phase . . . . . . . . . 11
2.2.1 Main Activities in Front-End Stage . . . . . . . . . . . . . . 11
2.2.2 Main Activities in Design and Development Stage . . . 12
2.3 Engineering Activities in Production Phase . . . . . . . . . . . . . . 15
2.3.1 Types of Production Systems . . . . . . . . . . . . . . . . . . 15
2.3.2 Production System Design . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Quality Control System Design. . . . . . . . . . . . . . . . . 19
2.3.4 Production Management . . . . . . . . . . . . . . . . . . . . . 22
vii
viii Contents
2.4 Engineering Activities in Post-manufacturing Phase. . . . . . . . . 22

2.4.1 Main Activities in Marketing Stage . . . . . . . . . . . . . . 22
2.4.2 Main Activities in Post-sale Support Stage . . . . . . . . . 22
2.4.3 Recycle, Refurbishing, and Remanufacturing . . . . . . . 23
2.5 Approach for Solving Quality and Reliability Problems . . . . . . 24
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Fundamental of Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Concepts of Reliability and Failure . . . . . . . . . . . . . . . . . . . . 27
3.2.1 Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 Failure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.3 Failure Mode and Cause . . . . . . . . . . . . . . . . . . . . . 29
3.2.4 Failure Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.5 Failure Severity and Consequences . . . . . . . . . . . . . . 30
3.2.6 Modeling Failures . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Reliability Basic Functions. . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.1 Probability Density Function . . . . . . . . . . . . . . . . . . 31
3.3.2 Cumulative Distribution and Reliability Functions. . . . 32
3.3.3 Conditional Distribution and Residual Life. . . . . . . . . 33
3.3.4 Failure Rate and Cumulative Hazard Functions. . . . . . 34
3.3.5 Relations Between Reliability Basic Functions . . . . . . 35
3.4 Component Bathtub Curve and Hockey-Stick Line . . . . . . . . . 36
3.5 Life Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.1 Measures of Lifetime . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.2 Dispersion of Lifetime. . . . . . . . . . . . . . . . . . . . . . . 40
3.5.3 Skewness and Kurtosis of Life Distribution . . . . . . . . 41
3.6 Reliability of Repairable Systems . . . . . . . . . . . . . . . . . . . . . 41
3.6.1 Failure-Repair Process . . . . . . . . . . . . . . . . . . . . . . . 41
3.6.2 Reliability Measures . . . . . . . . . . . . . . . . . . . . . . . . 43
3.6.3 Failure Point Process. . . . . . . . . . . . . . . . . . . . . . . . 44
3.7 Evolution of Reliability Over Product Life Cycle . . . . . . . . . . 46
3.7.1 Design Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.7.2 Inherent Reliability . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7.3 Reliability at Sale . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7.4 Field Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7.5 Values of Weibull Shape Parameter Associated
with Different Reliability Notions . . . . . . . . . . . . ... 48
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 48
Contents ix
4 Distribution Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Discrete Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.1 Basic Functions of a Discrete Distribution . . . . . . . . . 51
4.2.2 Single-Parameter Models . . . . . . . . . . . . . . . . . . . . . 52
4.2.3 Two-Parameter Models . . . . . . . . . . . . . . . . . . . . . . 53
4.2.4 Hypergeometric Distribution. . . . . . . . . . . . . . . . . . . 56
4.3 Simple Continuous Distributions . . . . . . . . . . . . . . . . . . . . . . 57
4.3.1 Weibull Distribution . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.2 Gamma Distribution . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3.3 Lognormal Distribution . . . . . . . . . . . . . . . . . . . . . . 59
4.4 Complex Distribution Models Involving Multiple Simple
Distributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.1 Mixture Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.2 Competing Risk Model . . . . . . . . . . . . . . . . . . . . . . 60
4.4.3 Multiplicative Model . . . . . . . . . . . . . . . . . . . . . . . . 61
4.4.4 Sectional Models . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5 Delay Time Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Statistical Methods for Lifetime Data Analysis . . . . . . . . . . . . . . . 67

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Reliability Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2.1 Sources and Types of Data . . . . . . . . . . . . . . . . . . . 67
5.2.2 Life Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2.3 Performance Degradation Data . . . . . . . . . . . . . . . . . 72
5.2.4 Data on Use Condition and Environment . . . . . . . . . . 72
5.3 Nonparametric Estimation Methods for Cdf . . . . . . . . . . . . . . 73
5.3.1 Complete Data Case . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.2 Grouped Data Case . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.3 Alternately Censored Data Case . . . . . . . . . . . . . . . . 74
5.4 Parameter Estimation Methods . . . . . . . . . . . . . . . . . . . . . . . 79
5.4.1 Graphical Method . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4.2 Method of Moments . . . . . . . . . . . . . . . . . . . . . . . . 80
5.4.3 Maximum Likelihood Method . . . . . . . . . . . . . . . . . 81
5.4.4 Least Square Method. . . . . . . . . . . . . . . . . . . . . . . . 83
5.4.5 Expectation-Maximum Method . . . . . . . . . . . . . . . . . 83
5.5 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.5.1 Chi Square Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.5.2 KolmogorovSmirnov Test . . . . . . . . . . . . . . . . . . . 85
5.6 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6.1 Likelihood Ratio Test . . . . . . . . . . . . . . . . . . . . . . . 86
5.6.2 Information Criterion. . . . . . . . . . . . . . . . . . . . . . . . 87
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
x Contents
6 Reliability Modeling of Repairable Systems . . . . . . . . . . . . . . . . . 89

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.2 Failure Counting Process Models . . . . . . . . . . . . . . . . . . . . . 90
6.2.1 Renewal Process. . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.2.2 Homogeneous Poisson Process . . . . . . . . . . . . . . . . . 91
6.2.3 Nonhomogeneous Poisson Process . . . . . . . . . . . . . . 91
6.2.4 Empirical Mean Cumulative Function . . . . . . . . . . . . 92
6.3 Distribution Models for Modeling Failure Processes . . . . . . . . 93
6.3.1 Ordinary Life Distribution Models . . . . . . . . . . . . . . 93
6.3.2 Imperfect Maintenance Models . . . . . . . . . . . . . . . . . 94
6.3.3 Variable-Parameter Distribution Models . . . . . . . . . . . 94
6.4 A Procedure for Modeling Failure Processes . . . . . . . . . . . . . 94
6.4.1 An Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.4.2 Modeling Procedure . . . . . . . . . . . . . . . . . . . . . . . . 95
6.5 Tests for Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.5.1 Graphical Methods . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.5.2 Tests with HPP Null Hypothesis . . . . . . . . . . . . . . . . 97
6.5.3 Tests with RP Null Hypothesis . . . . . . . . . . . . . . . . . 100
6.5.4 Performances of Trend Tests . . . . . . . . . . . . . . . . . . 101
6.6 Tests for Randomness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.6.1 Runs Above and Below Median Test . . . . . . . . . . . . 102
6.6.2 Sign Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.6.3 Runs Up and Down . . . . . . . . . . . . . . . . . . . . . . . . 104
6.6.4 MannKendall Test. . . . . . . . . . . . . . . . . . . . . . . . . 105
6.6.5 Spearman Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.6.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.7 Tests for Normality and Constant Variance . . . . . . . . . . . . . . 106
6.7.1 Tests for Normality . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.7.2 Tests for Constant Variance . . . . . . . . . . . . . . . . . . . 108
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Part II Product Quality and Reliability in Pre-manufacturing Phase
7 Product Design and Design for X . . . . . . . . . . . . . . . . . . . . . . . . 113

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2 Product Design and Relevant Issues . . . . . . . . . . . . . . . . . . . 113
7.2.1 Product Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
7.2.2 Key Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.2.3 Time-Based Product Design . . . . . . . . . . . . . . . . . . . 114
7.2.4 Design for Life Cycle . . . . . . . . . . . . . . . . . . . . . . . 115
7.2.5 Design for X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Contents xi
7.3 Design for Several Overall Performances . . . . . . . . . . . . . . . . 116

7.3.1 Design for Safety . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.3.2 Design for Environment. . . . . . . . . . . . . . . . . . . . . . 119
7.3.3 Design for Quality . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3.4 Design for Reliability . . . . . . . . . . . . . . . . . . . . . . . 120
7.3.5 Design for Testability . . . . . . . . . . . . . . . . . . . . . . . 120
7.4 Design for Production-Related Performances . . . . . . . . . . . . . 121
7.4.1 Design for Manufacturability . . . . . . . . . . . . . . . . . . 121
7.4.2 Design for Assembliability . . . . . . . . . . . . . . . . . . . . 121
7.4.3 Design for Logistics . . . . . . . . . . . . . . . . . . . . . . . . 122
7.5 Design for Use-Related Performances . . . . . . . . . . . . . . . . . . 123
7.5.1 Design for Serviceability . . . . . . . . . . . . . . . . . . . . . 123
7.5.2 Design for Maintainability . . . . . . . . . . . . . . . . . . . . 124
7.5.3 Design for Supportability . . . . . . . . . . . . . . . . . . . . . 125
7.6 Design for Retirement-Related Performances . . . . . . . . . . . . . 125
7.6.1 Design for Recyclability . . . . . . . . . . . . . . . . . . . . . 125
7.6.2 Design for Disassembliability . . . . . . . . . . . . . . . . . . 126
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
8 Design Techniques for Quality. . . . . . . . . . . . . . . . . . . . . . . . . . . 129

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.2 House of Quality and Quality Function Deployment . . . . . . . . 129
8.2.1 House of Quality . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8.2.2 Priorities of Engineering Characteristics . . . . . . . . . . . 131
8.2.3 Satisfaction Degrees of Customer Attributes . . . . . . . . 132
8.2.4 Quality Function Deployment. . . . . . . . . . . . . . . . . . 133
8.3 Cost of Quality and Loss Function . . . . . . . . . . . . . . . . . . . . 134
8.3.1 Quality Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.3.2 Loss Function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.3.3 Applications of Quality Loss Function. . . . . . . . . . . . 136
8.4 Experimental Optimum Method . . . . . . . . . . . . . . . . . . . . . . 137
8.4.1 Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.4.2 Specific Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.4.3 Design of Experiments . . . . . . . . . . . . . . . . . . . . . . 138
8.4.4 Data Analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.5 Model-Based Optimum Method . . . . . . . . . . . . . . . . . . . . . . 143
8.5.1 Constraint Conditions . . . . . . . . . . . . . . . . . . . . . . . 144
8.5.2 Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . 144
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
9 Design Techniques for Reliability . . . . . . . . . . . . . . . . . . . . . . . . 147

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.2 Process of Design for Reliability. . . . . . . . . . . . . . . . . . . . . . 147
9.3 Reliability Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
xii Contents
9.4 Reliability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

9.4.1 Change Point Analysis . . . . . . . . . . . . . . . . . . . . . . 149
9.4.2 FMEA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
9.4.3 System Reliability Analysis . . . . . . . . . . . . . . . . . . . 150
9.5 Reliability Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.5.1 Empirical Methods . . . . . . . . . . . . . . . . . . . . . . . . . 154
9.5.2 Physics of Failure Analysis Method . . . . . . . . . . . . . 155
9.5.3 Life Testing Method . . . . . . . . . . . . . . . . . . . . . . . . 157
9.5.4 Simulation Method . . . . . . . . . . . . . . . . . . . . . . . . . 157
9.6 Reliability Allocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
9.6.1 Reliability Allocation Methods for Nonrepairable
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 158
9.6.2 Reliability Allocation Methods for Repairable
Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9.7 Techniques to Achieve Desired Reliability . . . . . . . . . . . . . . . 162
9.7.1 Component Deration and Selection . . . . . . . . . . . . . . 162
9.7.2 Redundancy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.7.3 Preventive Maintenance . . . . . . . . . . . . . . . . . . . . . . 165
9.7.4 Reliability Growth Through Development . . . . . . . . . 166
9.8 Reliability Control and Monitoring . . . . . . . . . . . . . . . . . . . . 166
9.8.1 Reliability Control in Manufacturing Process . . . . . . . 166
9.8.2 Reliability Monitoring in Usage Phase. . . . . . . . . . . . 167
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
10 Reliability Testing and Data Analysis . . . . . . . . . . . . . . ....... 169

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 169
10.2 Product Reliability Tests in Product Life Cycle. . . . . ....... 169
10.2.1 Reliability Tests Carried Out During Product
Development Stage . . . . . . . . . . . . . . . . . . ....... 169
Manufacturing Phase . . . . . . . . . . . . . . . . . ....... 170
Usage Phase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
10.3 Accelerated Testing and Loading Schemes . . . . . . . . . . . . . . . 171
10.3.1 Accelerated Life Testing . . . . . . . . . . . . . . . . . . . . . 171
10.3.2 Accelerated Degradation Testing. . . . . . . . . . . . . . . . 172
10.3.3 Loading Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . 172
10.4 Accelerated Life Testing Data Analysis Models . . . . . . . . . . . 174
10.4.1 Life Distribution Models . . . . . . . . . . . . . . . . . . . . . 174
10.4.2 Stress-Life Relationship Models . . . . . . . . . . . . . . . . 176
10.4.3 Inverse Power-Law Model . . . . . . . . . . . . . . . . . . . . 177
10.4.4 Proportional Hazard Model . . . . . . . . . . . . . . . . . . . 177
10.4.5 Generalized Proportional Model . . . . . . . . . . . . . . . . 179
10.4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Contents xiii
10.5 Accelerated Degradation Testing Models . . . . . . . . . . . . . . . . 181

10.5.1 Physical-Principle-Based Models. . . . . . . . . . . . . . . . 182
10.5.2 Data-Driven Models . . . . . . . . . . . . . . . . . . . . . . . . 182
10.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.5.4 A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.6 Design of Accelerated Stress Testing. . . . . . . . . . . . . . . . . . . 187
10.6.1 Design Variables and Relevant Performances . . . . . . . 187
10.6.2 Empirical Approach for ALT Design. . . . . . . . . . . . . 189
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
11 Reliability Growth Process and Data Analysis . . . . . . . . . . . . . . . 193

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
11.2 TAF Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
11.3 Reliability Growth Plan Model . . . . . . . . . . . . . . . . . . . . . . . 195
11.3.1 Reliability Growth Plan Curve . . . . . . . . . . . . . . . . . 195
11.3.2 Duane Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
11.4 Modeling Effectiveness of a Corrective Action . . . . . . . . . . . . 197
11.4.1 Type of Failure Modes . . . . . . . . . . . . . . . . . . . . . . 197
11.4.2 Effectiveness of a Corrective Action . . . . . . . . . . . . . 197
11.5 Reliability Growth Evaluation Models . . . . . . . . . . . . . . . . . . 198
11.5.1 Software Reliability Growth Models
and Parameter Estimation. . . . . . . . . . . . . . . . ..... 199
11.5.2 Discrete Reliability Growth Models
for Complex Systems . . . . . . . . . . . . . . . . . . ..... 202
11.5.3 Continuous Reliability Growth Models
for Complex Systems . . . . . . . . . . . . . . . . . . . . . . . 204
11.6 Design Validation Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
11.7 A Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.7.1 Data and Preliminary Analysis . . . . . . . . . . . . . . . . . 209
11.7.2 Assessment and Prediction of Failure Intensity
of Each Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.7.3 Prediction of Unobserved Failure Modes . . . . . . . . . . 213
11.7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
11.7.5 Reliability Growth Plan Curve . . . . . . . . . . . . . . . . . 216
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Part III Product Quality and Reliability in Manufacturing Phase
12 Product Quality Variations and Control Strategies . . . . . . . . . . . . 221

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
12.2 Variations of Quality Characteristics and Their Effect
on Product Quality and Reliability . . . . . . . . . . . . . . . . . . . . 221
xiv Contents
12.2.1 Variations of Quality Characteristics

and Variation Sources . . . . . . . . . . . . . . . . . . ..... 221
12.2.2 Effect of Unit-to-Unit Variability on Product
Quality and Reliability. . . . . . . . . . . . . . . . . . ..... 223
12.2.3 Effect of Operating and Environmental Factors
on Product Reliability . . . . . . . . . . . . . . . . . . . . . . . 225
12.3 Reliability and Design of Production Systems. . . . . . . . . . . . . 226
12.3.1 Reliability of Production Systems . . . . . . . . . . . . . . . 226
12.3.2 Design of Production Systems . . . . . . . . . . . . . . . . . 228
12.4 Quality Control and Improvement Strategies. . . . . . . . . . . . . . 228
12.4.1 Inspection and Testing. . . . . . . . . . . . . . . . . . . . . . . 229
12.4.2 Statistical Process Control . . . . . . . . . . . . . . . . . . . . 229
12.4.3 Quality Control by Optimization . . . . . . . . . . . . . . . . 230
12.5 Quality Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
12.5.1 Principles of Quality Management . . . . . . . . . . . . . . 232
12.5.2 Quality Management Strategies. . . . . . . . . . . . . . . . . 232
12.5.3 ISO Quality Management System . . . . . . . . . . . . . . . 233
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
13 Quality Control at Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
13.2 Acceptance Sampling for Attribute . . . . . . . . . . . . . . . . . . . . 235
13.2.1 Concepts of Acceptance Sampling . . . . . . . . . . . . . . 235
13.2.2 Acceptance Sampling Plan . . . . . . . . . . . . . . . . . . . . 236
13.2.3 Operating-Characteristic Curve . . . . . . . . . . . . . . . . . 236
13.2.4 Average Outgoing Quality . . . . . . . . . . . . . . . . . . . . 237
13.2.5 Acceptance Sampling Based on Binomial
Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 237
13.2.6 Acceptance Sampling Based on Hypergeometric
Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 240
13.3 Acceptance Sampling for a Normally Distributed Variable . . .. 241
13.4 Acceptance Sampling for Lifetime . . . . . . . . . . . . . . . . . . .. 242
13.5 Acceptance Sampling for Variable Based on the Binomial
Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 245
13.6 Supplier Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 247
13.6.1 A Mathematical Model for Component Purchasing
Decision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 247
13.6.2 Supplier Selection Problem Involving Strategic
Partnership with Suppliers . . . . . . . . . . . . . . . . . . .. 248
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 249
Contents xv
14 Statistical Process Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
14.2 Control Charts for Variable . . . . . . . . . . . . . . . . . . . . . . . . . 251
14.2.1 Concepts of Control Charts . . . . . . . . . . . . . . . . . . . 251
14.2.2 Shewhart Mean Control Charts . . . . . . . . . . . . . . . . . 252
14.2.3 Range Chart. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
14.2.4 Errors of a Control Chart . . . . . . . . . . . . . . . . . . . . . 253
14.2.5 Average Run Length and Average Time to Signal . . . 254
14.3 Construction and Implementation of the Shewhart
Control Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
14.3.1 Construction of Trial Control Chart. . . . . . . . . . . . . . 256
14.3.2 Sampling Strategy. . . . . . . . . . . . . . . . . . . . . . . . . . 257
14.3.3 Nonrandom Patterns on Control Charts . . . . . . . . . . . 258
14.3.4 Warning Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
14.3.5 Out-of-Control Action Plan . . . . . . . . . . . . . . . . . . . 259
14.4 Process Capability Indices and Fraction Nonconforming . . . . . 260
14.4.1 Process Capability Indices . . . . . . . . . . . . . . . . . . . . 260
14.4.2 Fraction Nonconforming . . . . . . . . . . . . . . . . . . . . . 262
14.5 Multivariate Statistical Process Control Methods . . . . . . . . . . . 263
14.5.1 Multivariate Control Charts . . . . . . . . . . . . . . . . . . . 263
14.5.2 Multivariate Statistical Projection Methods . . . . . . . . . 263
14.6 Control Charts for Attribute . . . . . . . . . . . . . . . . . . . . . . . . . 264
14.6.1 Control Chart for Fraction Nonconforming. . . . . . . . . 264
14.6.2 Control Chart for the Number of Defects Per
Inspected Item . . . . . . . . . . . . . . . . . . . . . . . . . ... 265
14.6.3 Control Chart for the Average Number
of Defects Per Item . . . . . . . . . . . . . . . . . . . . . . ... 265
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... 266
15 Quality Control at Output. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
15.2 Optimal Screening Limit Problem . . . . . . . . . . . . . . . . . . . . . 267
15.2.1 Screening Limit Problem . . . . . . . . . . . . . . . . . . . . . 267
15.2.2 An Optimization Model . . . . . . . . . . . . . . . . . . . . . . 268
15.3 Screening Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
15.3.1 Types of Manufacturing Defects . . . . . . . . . . . . . . . . 270
15.3.2 Burn-in . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
15.3.3 Environmental Stress Screening . . . . . . . . . . . . . . . . 273
15.3.4 Comparison of ESS and Burn-in. . . . . . . . . . . . . . . . 273
15.4 Optimal Component-Level Burn-in Duration . . . . . . . . . . . . . 274
15.5 Optimal System-Level Burn-in Duration . . . . . . . . . . . . . . . . 277
15.5.1 Reliability Model . . . . . . . . . . . . . . . . . . . . . . . . . . 278
15.5.2 Cost Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
xvi Contents
Part IV Product Quality and Reliability in Post-manufacturing Phase
16 Product Warranty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
16.2 Product Warranties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
16.2.1 Concepts and Roles of Warranty. . . . . . . . . . . . . . . . 283
16.2.2 Maintenance-Related Concepts . . . . . . . . . . . . . . . . . 284
16.3 Warranty Policies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
16.3.1 Classification of Warranty Policies . . . . . . . . . . . . . . 284
16.3.2 Typical Warranty Policies . . . . . . . . . . . . . . . . . . . . 285
16.3.3 Special Policies for Commercial and Industrial
Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
16.3.4 Reliability Improvement Warranties. . . . . . . . . . . . . . 288
16.4 Reliability Models in Warranty Analysis . . . . . . . . . . . . . . . . 288
16.4.1 Reliability Characteristics of Renewal Process . . . . . . 289
16.4.2 Reliability Characteristics of Minimal Repair
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 290
16.4.3 Imperfect Repair Models for Modeling Effect
of Preventive Maintenance . . . . . . . . . . . . . . . . . . . . 290
16.4.4 Bivariate Reliability Models . . . . . . . . . . . . . . . . . . . 293
16.4.5 Bi-failure-Mode Models. . . . . . . . . . . . . . . . . . . . . . 294
16.5 Warranty Cost Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
16.5.1 Cost Analysis for Non-repairable Product
Under One-Dimensional FRW . . . . . . . . . . . . . .... 295
16.5.2 Cost Analysis for Repairable Product Under
One-Dimensional FRW . . . . . . . . . . . . . . . . . . . . . . 296
16.5.3 Cost Analysis for One-Dimensional PRW Policy . . . . 296
16.5.4 Cost Analysis for Two-Dimensional FRW Policy . . . . 297
16.6 Product Warranty Servicing . . . . . . . . . . . . . . . . . . . . . . . . . 299
16.6.1 Spare Part Demand Prediction . . . . . . . . . . . . . . . . . 299
16.6.2 Optimal RepairReplacement Decision . . . . . . . . . . . 300
16.6.3 Field Information Collection and Analysis . . . . . . . . . 300
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
17 Maintenance Decision Optimization . . . . . . . . . . . . . . . . . . . . . . . 303

17.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
17.2 Maintenance Policy Optimization . . . . . . . . . . . . . . . . . . . . . 303
17.2.1 Maintenance Tasks . . . . . . . . . . . . . . . . . . . . . . . . . 304
17.2.2 Timing of Maintenance Tasks . . . . . . . . . . . . . . . . . 307
17.2.3 Optimization of Maintenance Policies . . . . . . . . . . . . 308
Contents xvii
17.3 Repair-Replacement Policies. . . . . . . . . . . . . . . . . . . . ..... 308

17.3.1 Repair Cost Limit Policy and Its Optimization
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 309
17.3.2 Repair Time Limit Policy and Its Optimization
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 310
17.3.3 Failure Counting Policy with a Reference
Age and Its Optimization Model . . . . . . . . . . . ..... 311
17.4 Time-Based Preventive Replacement Policies . . . . . . . . ..... 313
17.4.1 Age Replacement Policy and Its Optimization
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 313
17.4.2 Periodic Replacement Policy with Minimal
Repair and Its Optimization Model . . . . . . . . . ..... 315
17.4.3 Block Replacement Policy and Its Optimization
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 315
17.4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 316
17.5 Inspection Policies . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 316
17.5.1 Inspection Policy with Perfect Maintenance
and Its Optimization Model . . . . . . . . . . . . . . ..... 317
17.5.2 Inspection Policy with Minimal Repair
and Its Optimization Model . . . . . . . . . . . . . . . . . . . 318
17.6 Condition-Based Maintenance . . . . . . . . . . . . . . . . . . . . . . . 319
17.7 System-Level Preventive Maintenance Policies . . . . . . . . . . . . 320
17.7.1 Group Preventive Maintenance Policy . . . . . . . . . . . . 320
17.7.2 Multi-level Preventive Maintenance Program . . . . . . . 322
17.7.3 Opportunistic Maintenance Policy . . . . . . . . . . . . . . . 322
17.8 A Simple Maintenance Float System . . . . . . . . . . . . . . . . . . . 324
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
Abbreviations
5M1E Materials, Manufacture, Man, Machine, Measurements,

and Environment
ADT Accelerated Degradation Testing
AGREE Advisory Group on Reliability of Electronic Equipment
AHP Analytic Hierarchy Process
AIC Akaike Information Criterion
ALT Accelerated Life Testing
AMSAA Army Materiel Systems Analysis Activity
AOQ Average Outgoing Quality
AQL Acceptable Quality Level
ARINC Aeronautical Radio, Inc.
ARL Average Run Length
ATS Average Time to Signal
CA Customers Attributes
CAD Computer-aided Design
CAE Computer-aided Engineering
CAQ Computer-aided Quality
CBM Condition-Based Maintenance
Cdf Cumulative Distribution Function
Chf Cumulative Hazard Function
CM Corrective Maintenance
CMMS Computerized Maintenance Management System
CNC Computer Numerical Control
CV Coefcient of Variation
DEA Data Envelopment Analysis
DFA Design for Assembliability
DFD Design for Disassembliability
DFL Design for Logistics
xix
xx Abbreviations
DFM Design for Manufacturability

DFMAIN Design for Maintainability
DFR Design for Reliability
DFRc Design for Recyclability
DFSp Design for Supportability
DFSv Design for Serviceability
DFT Design for Testability
DFX Design for X
DOE Design of Experiments
EC Engineering Characteristics
EDF Empirical Distribution Function
ELECTRE Elimination et Choice Translating Reality
EMM Expectation-Maximum Method
ESS Environmental Stress Screening
ETA Event Tree Analysis
FEF Fix Effectiveness Factor
FMEA Failure Mode and Effects Analysis
FMECA Failure Mode, Effect, and Criticality Analysis
FMS Flexible Manufacturing System
FRACAS Failure Reporting, Analysis, and Corrective Action Systems
FRW Free Replacement Warranty
FTA Fault Tree Analysis
HOQ House of Quality
HPP Homogeneous Poisson Process
IID (i.i.d.) Independent and Identically Distributed
ISO International Organization for Standardization
KMM KaplanMeier Method
LCC Life Cycle Cost
LCL Lower Control Limit
LHS Left-Hand Side
LSL Lower Specication Limit
LSM Least Square Method
MCDM Multi-Criteria Decision Making
MCF Mean Cumulative Function
MDT Mean Downtime
MLE Maximum Likelihood Estimation (Estimate)
MLM Maximum Likelihood Method
MRL Mean Residual Life
MROM Mean Rank Order Method
MSE Mean Squared Error
MTBF Mean Time between Failures
MTTF Mean Time to Failure
Abbreviations xxi
MTTR Mean Time to Repair

MVF Mean Value Function
NAM NelsonAalen Method
NHPP Nonhomogeneous Poisson Process
OC Operating Characteristic
OCAP Out-of-Control-Action-Plan
OEE Overall Equipment Effectiveness
PAA Part Average Analysis
PCA Principal Component Analysis
PCI Process Capability Index
PDCA Plan-Do-Check-Action
PDF (pdf) Probability Density Function
PEM Piecewise Exponential Method
PHM Prognostics and Health Management
PHM Proportional Hazard Model
PLC Product Life Cycle
PLM Product Lifecycle Management
PLS Partial Least Squares
PM Preventive Maintenance
Pmf Probability Mass Function
PRW Pro-rata Rebate Warranty
QFD Quality Function Deployment
RAMS Reliability, Availability, Maintainability, and Safety
(or Supportability)
RBD Reliability Block Diagram
RBM Risk-Based Maintenance
RCM Reliability-Centered Maintenance
RHS Right hand side
RMS Root mean square
RP Renewal Process
RPN Risk Priority Numbers
RUL Remaining Useful Life
SA Supportability Analysis
SPC Statistical Process Control
SSE Sum of Squared Errors
SSP Supplier Selection Problem
TAF Test-Analysis-and-Fix
TBF Time Between Failures
TBM Time-Based Maintenance
TFT Test-Find-Test
TOPSIS Technique for Order Preference by Similarity
to an Ideal Solution
TPM Total Productive Maintenance
TQM Total Quality Management
TTF Time to Failure
xxii Abbreviations
TTFF Time to the First Failure

TTT Total Time on Test
UCL Upper Control Limit
USL Upper Specication Limit
VOC Voice of Customer
WPM Weighted Product Model
WPP Weibull Probability Paper
WSM Weighted Sum Model
ZIP Zero-Inflated Poisson
Part I
Background Materials
Chapter 1
Overview
1.1 Introduction
Manufacturing businesses need to come with new products in order to survive in a

erce competitive environment. Product quality and reliability are important
competition factors. This book focuses on models, tools, and techniques to help
manage quality and reliability for manufactured and new products.
This chapter is organized as follows. Section 1.2 briefly discusses relevant
concepts of product. Section 1.3 presents notions of product reliability and product
quality. The objective, scope, and focuses of this book are presented in Sect. 1.4.
Finally, we outline the structure and contents of the book in Sect. 1.5.
1.2 Product and Product Life Cycle
1.2.1 Product
From a marketing perspective, a product is anything that can be offered to a

market that might satisfy a want or need; and from a manufacturing perspective,
products are purchased as raw materials and sold as nished goods. Based on the
type of consumer, products can be classied as consumer durables and industrial
products. The former is for consumers (e.g., cars) and the latter is for a business
and hence often termed as business-to-business product (e.g., engineering
machineries).
A product can be a single component or a complex system comprised of many
components. For the latter case, one can decompose a product into a hierarchy
Science Press, Beijing and Springer-Verlag Berlin Heidelberg 2015 3

R. Jiang, Introduction to Quality and Reliability Engineering,
Springer Series in Reliability Engineering, DOI 10.1007/978-3-662-47215-6_1
4 1 Overview
structure. A typical decomposition of a product includes six hierarchical levels,

i.e., material, part, component, assembly, subsystem, and system.
1.2.2 Product Life Cycle
The concept of product life cycle (PLC) is different for manufacturers and con-
sumers [4]. From the perspective of the manufacturer, the PLC refers to the phases
of a products life, from its conception, through design and manufacture to post-sale
service and disposal.
On the other hand, from the perspective of the consumer, the PLC is the time
from the purchase of a product to its discarding when it reaches the end of its useful
life or being replaced earlier due to either technological obsolescence or the product
being no longer of any use. As such, its life cycle involves only the following three
phases: acquisition, operations and maintenance, and retirement that leads to
replacement by a new one.
1.2.3 Technology Life Cycle of a Product
From the perspective of marketing, the PLC involves four phases (see Fig. 1.1):
Introduction phase with low sales
Growth phase with rapid increase in sales
Maturity phase with large and nearly constant sales, and
Decline phase with decreasing sales and eventually withdrawing from the
market.
It is desired to keep the maturity period going as long as possible. However, the
PLC gets shorter and shorter due to rapid technological change, global markets and
multiple vendor environments, erce competition and partnerships (or alliances)
environment and ever-increasing customer expectations.
Fig. 1.1 Technology life

cycle
Introduction Growth Maturity Decline
Market volume
Time
1.3 Notions of Reliability and Quality 5
1.3 Notions of Reliability and Quality
1.3.1 Product Reliability
Reliability of a product (or system) conveys information about the absence of

failures, and is usually dened as the probability that the system will perform its
intended function for a specied time period when operating under normal (or
stated) environmental conditions (e.g., see Ref. [4]). This denition deals with the
following four important points:
Intended function. It actually denes what the failure is. Failure can be a total
loss of function (termed as hard failure) or a part loss of performance when the
performance degrades to a specied level (termed as function failure or soft
failure).
Uncertainty or randomness of time to failure. It is reflected by the word
probability. The uncertainty is due to many factors, including the variability in
raw materials, manufacturing, operating, and environmental conditions.
Planning horizon or mission duration. It is reflected by the phrase specied
time period.
Use environment. It is reflected by the phrase normal or stated environmental
conditions. Product design and reliability assessment are based on a set of
nominal conditions such as usage intensity or load prole (a graph of load
versus time), operating environment, and maintenance activities. These condi-
tions determine stresses on the components and affect degradation rate. If these
conditions or activities are different from the nominal, the reliability perfor-
mance in eld will be different from those assessed in or derived from the
nominal conditions.
The reliability of a product is mainly determined by the decisions made during
the design and development and in the manufacturing of the product, and depends
on a number of factors, including manufacturing quality, operating environment
(e.g., heat, humidity, dust and chemical solvents), usage intensity (frequency and
severity), maintenance activities (e.g., frequency and depth of preventive mainte-
nance), and operators skills (e.g., see Ref. [3]).
The product reliability is important to both manufacturer and consumer. For the
manufacturer, the consequences due to product unreliability can be warranty cost,
customer satisfaction, product reputation, market share, and prot. Over the war-
ranty period, the cost to rectify failures is borne by the manufacturer but the owner
can have production and other losses. As such, effective management of product
reliability requires making proper reliability-related decisions using a PLC frame-
work. The main issues to address include the following [7]:
Why systems fail and this refers to reliability physics;
How to develop reliable systems and this refers to reliability-based design;
6 1 Overview
How to measure and test reliability in different stages and this refers to reliability
assessment and reliability testing;
How to maintain systems reliable and this refers to maintenance, fault diagnosis,
and prognosis.
1.3.2 Product Quality
Garvin [1] proposes the following ve criteria for dening the notion of quality:
(1) Judgmental criteria. Here, quality is associated with something universally
recognizable as a mark of high standard, achievement, or degree of excellence,
and hence is called the transcendent denition.
(2) Product-based criteria. Here, quality is dened in terms of some measurable
variable such as the acceleration of a car, efciency of an engine, or the like.
(3) User-based criteria. Here, quality is dened through tness for intended use.
For example, the user-based quality for a car may be smoothness of the ride,
ease of steering, etc.
(4) Value-based criteria. Here, quality is linked to the price of the product and its
usefulness or satisfaction.
(5) Manufacturing-based criteria: Here, quality is dened in terms of manufac-
tured items conforming to the design specication. Items that do not conform
either need some rectication action to make them conform or need to be
scrapped.
The product quality involves many dimensions. Garvin [1] suggests the fol-
lowing eight quality dimensions:
(1) Performance. This characterizes the primary operating characteristics or spe-
cic functions of the product. For a car, it can include acceleration, braking
distance, efciency of engine, emissive pollution generated, and so on.
(2) Features. These are the special or additional features of a product. For
example, for a car, the features include air conditioner, cruise control, or the
like.
(3) Aesthetics. This deals with issues such as appearance, feel, sound, and so on.
For a car, the body design and interior layout reflect the quality in this sense.
(4) Reliability. This is a measure of the product performing satisfactorily over a
specied time under stated conditions of use. Simply speaking, it reflects how
often the product fails.
(5) Durability. This is an indicator of the time interval after which the product has
deteriorated sufciently so that it is unacceptable for use. For a car, it may
correspond to corrosion affecting the frame and body to such a level that it is
no longer safe to drive.
1.3 Notions of Reliability and Quality 7
(6) Serviceability. This deals with all maintenance related issues, including fre-
quency and cost of maintenance, ease of repair, availability of spares, and so on.
(7) Conformance. This indicates the degree to which the physical and perfor-
mance characteristics meet some pre-established standards (i.e., design
requirements).
(8) Perceived quality. This refers to the perceptions of the buyers or potential
buyers. This impression is shaped by several factors such as advertising, the
reputation of the company or product, consumer report, etc.
A customer-driven concept of quality denes product quality as the collection of
features and characteristics of a product that contribute to its ability to meet or
exceed customers expectations or given requirements. Here, quality characteristics
are the parameters that describe the product quality. Excessive variability in critical
quality characteristics results in more nonconforming products or waste and hence
the reduction of variability in products and processes results in quality improve-
ment. In this sense, quality is sometimes dened as inversely proportional to
variability [2].
Product quality deals with quality of design and quality of conformance. The
quality of design means that the products can be produced in different levels of
quality, and this is achieved by the differences in design. The quality of confor-
mance means that the product conforms to the specications required by design,
and this is influenced by many factors such as manufacturing processes and quality-
assurance system.
Quality engineering is the set of operational, managerial and engineering
activities used to ensure that the quality characteristics of a product are at the
required levels.
1.3.3 Link Between Quality and Reliability
Due to variability in the manufacturing process, some items produced may not meet
the design specication and such items are called nonconforming. The performance
of nonconforming items is usually inferior to the performance of conforming items.
As a result, nonconforming items are less reliable than conforming items in terms of
reliability measures such as mean time to failure.
In a broad sense, reliability is one of quality dimensions and usually termed as
time-oriented quality or quality over time (e.g., see Ref. [6]). However, quality
is different from reliability in a narrow sense. This can be explained by looking at
quality and reliability defects [5]. Quality defects usually deal with decient
products (or components) or incorrectly assembled sets, which can be identied by
inspection against component drawings or assembly specications. In this sense,
quality is expressed in percentages. On the other hand, reliability defects generally
deal with the failures of a product in the future when it has been working well.
8 1 Overview
Therefore, reliability is expressed as the proportion of the survived items to the

population at a certain time. Simply speaking, quality is usually understood as
conformance-quality at present and reliability as non-failure in the future.
1.4 Objective, Scope, and Focus of this Book
Traditionally, quality, and reliability belong to two closely related disciplinary elds.
A main objective of this book is to provide a comprehensive presentation to these two
elds in a systematic way. We will discuss typical quality and reliability problems in
PLC, and present the models, tools, and techniques needed in modeling and ana-
lyzing these problems. The focus is on concepts, models, and techniques of quality
and reliability in the context of design, manufacturing, and operation of products.
The book serves as an introductory textbook for senior undergraduate or grad-
uate students in engineering and management elds such as mechanical engineer-
ing, manufacturing engineering, industrial engineering, and engineering
management. It can be also used as a reference book for product engineers and
researchers in quality and reliability elds.
1.5 Outline of the Book
The book comprises four parts and three appendices. Part I (comprising six chap-
ters) deals with the background materials with focus on relevant concepts, statistical
modeling, and data analysis. Part II (comprising ve chapters) deals with product
quality and reliability problems in pre-manufacturing phase; Part III (comprising
four chapters) and Part IV (comprising two chapters) deal with product quality and
reliability problems in manufacturing phase and post-manufacturing phase,
respectively. A brief description of each chapter or appendix is as follows.
This chapter provides an overview of the book. It deals with basic notions of
quality and reliability, their importance in the context of product manufacturing and
operation engineering, and the scope and focus of the book. Chapter 2 discusses
typical quality and reliability problems in each phase of PLC. Chapter 3 presents
the fundamentals of reliability, including basic concepts, reliability basic functions,
and various life characteristics and measures. Chapter 4 presents common distri-
bution models widely used in quality and reliability elds, and Chap. 5 discusses
statistical methods for lifetime data analysis with focus on parameter estimation and
model selection for lifetime distribution models. Chapter 6 presents models and
methods for modeling failure processes, including counting process models,
variable-parameter distribution models, and hypothesis tests for trend and
randomness.
The above six chapters are Part I of the book, which provides the background
materials. The following ve chapters are Part II of the book and focus on major
quality and reliability problems in the design and development phase of product.
1.5 Outline of the Book 9
Chapters 79 deal with product design relevant issues. Chapter 7 discusses

various design requirements in the context of PLC and outlines related techniques
to address those requirements. Chapter 8 focuses on design techniques for quality,
including house of quality, quality function deployment, and Taguchi method;
Chapter 9 focuses on design techniques for reliability, including specication of
reliability requirements, reliability prediction, and reliability allocation.
Chapters 1011 deal with product development relevant issues. Chapter 10 deals
with reliability tests and data analysis with focus on accelerated testing. Chapter 11
deals with reliability growth test with focus on reliability growth modeling and
prediction.
The following four chapters are Part III of the book and focus on statistical
quality control in the manufacturing phase.
Chapter 12 discusses sources of product quality variations and general methods
to improve product quality. Chapters 1315 deal with quality control at input,
statistical process control and quality control at output of product production pro-
cesses, respectively. The contents include acceptance sampling, supplier selection,
control chart techniques, process capability indices, burn-in, and environment stress
screening.
The last two chapters (Chaps. 16 and 17) are Part IV of the book and focus on
product warranty and maintenance decision optimization, respectively.
Online Appendix A presents typical multi-criteria decision making analysis
techniques, including Analytic Hierarchy Process and TOPSIS. Online Appendix B
presents a brief introduction to Microsoft Ofce Excel, with which a number of
real-world examples in this book can be computed and solved. Online Appendix C
presents Excel-based methods to nd eigenvalues and eigenvectors of a matrix and
to carry out a principal component analysis.
References
1. Garvin DA (1988) Managing quality: the strategic and competitive edge. The Free Press,
New York
2. Montgomery DC (2007) Introduction to statistical quality control, 4th edn. Wiley, New York
3. Murthy DNP (2010) New research in reliability, warranty and maintenance. In: Proceedings of
the 4th Asia-Pacic international symposium on advanced reliability and maintenance
modeling, pp 504515
4. Murthy DNP, Rausand M, sters T (2009) Product reliability: specication and performance.
Springer, London
5. Ryu D, Chang S (2005) Novel concepts for reliability technology. Microelectron Reliab
45(34):611622
6. Yang K, Kapur KC (1997) Customer driven reliability: integration of QFD and robust design.
In: Proceedings of 1997 annual reliability and maintainability symposium, pp 339345
7. Zio E (2009) Reliability engineering: old problems and new challenges. Reliab Eng Syst Saf
94(2):125141
Chapter 2
Engineering Activities in Product
Life Cycle
2.1 Introduction
In this chapter we discuss the main engineering activities in each phase of product
life cycle (PLC) with focus on quality and reliability activities. The purpose is to set
the background for the following chapters of this book.
From the manufacturers perspective, the life cycle of a product can be roughly
divided into three main phases: pre-manufacturing phase, manufacturing phase, and
post-manufacturing phase. We discuss the main engineering activities in each of
these phases in Sects. 2.22.4, respectively. An approach to solving quality and
reliability problems is presented in Sect. 2.5.
2.2 Engineering Activities in Pre-manufacturing Phase
The pre-manufacturing phase starts with identifying a need for a product, through a
sequence of design and development activities, and ends with a prototype of the
product. This phase can be further divided into two stages: front-end or feasibility
stage and, design and development stage. The main activities in these two stages are
outlined below [8].
2.2.1 Main Activities in Front-End Stage
The front-end stage mainly deals with product denition. Specically speaking, this
stage will dene the requirements of the product, its major technical parameters,
and main functional aspects and carries out the initial concept design. The main

12 2 Engineering Activities in Product Life Cycle
activities include generation and screening of ideas, product denition, project plan,
and project denition review.
Once the need for a product is identied, a number of ideas are generated and
some of them will be screened to further pursue. The screening deals with
answering questions such as whether the idea is consistent with the strategic focus
of the company, whether the market size, growth, and opportunities are attractive,
whether the product can be developed and produced, and whether there are issues
that may make the project fail.
Product denition states what characteristics the product should have in order to
meet the business objectives and customer needs. It rst translates feasible ideas
into technically feasible and economically competitive product concepts, and then
produces product concept through concept generation and selection. Two com-
monly used techniques to decide the best design candidate are design-to-cost and
life-cycle-cost analyses. The design-to-cost aims to minimize the unit manufac-
turing cost, whose cost elements include the costs of design and development,
testing, and manufacturing. The life-cycle-cost analysis considers the total cost of
acquisition, operation, maintenance, and discarding, and is used for expensive
products.
Project plan deals with planning the remainder of the new product development
project in detail, including time and resource allocation, scheduling of tasks, and so
on. A nal review and evaluation of the product denition and project plan is
conducted to decide whether to commit potentially extensive resources to a full-
scale development project.
2.2.2 Main Activities in Design and Development Stage
The design and development stage starts with the detail design of the products
form, then progresses to prototype testing and design renement through a test-
analysis-and-x (TAF) iterative process, and eventually ends with full product
launch.
2.2.2.1 Quality and Reliability Activities in Detail Design Stage
The initial efforts of the design stage aims to arrive at optimal product architecture.
The product architecture is the arrangement of the functional elements of a product
into several physical building blocks (e.g., modules), including mapping from
functional elements to physical components and specication of interfaces among
interacting physical components. Establishing the product architecture needs to
conduct functional decomposition and dene the functional relationships between
assemblies and components.
Once the product architecture is established, the design process enters the detail
design stage. In this stage, the forms, dimensions, tolerances, materials, and surface
2.2 Engineering Activities in Pre-manufacturing Phase 13
properties of all individual components and parts are specied; and all the drawings
and other production documents (including the transport and operating instructions)
are produced.
The detail design involves a detailed analysis for the initial design. Based on this
analysis, the design is improved and the process is repeated until the analysis
indicates that the performance requirements are met.
The detailed analysis involves simultaneously considering various product
characteristics such as reliability, maintainability, availability, safety, supportabil-
ity, manufacturability, quality, life cycle cost, and so on. Design for these char-
acteristics or performances are further discussed in Chap. 7.
Design for quality is an integrated design technique for ensuring product quality.
It starts with an attempt to understand the customers needs. Then, the House of
Quality is used to transform the customer needs into the technical requirements or
engineering specications of the product in the concept design stage; and the
quality function deployment (QFD) is used to determine more specic requirements
in the detail design stage. The Taguchi method can be used to determine important
design parameters. These techniques are discussed in detail in Chap. 8.
Design for reliability involves a number of reliability related issues, including
reliability allocation and prediction. Reliability allocation is the process to deter-
mine the reliability goals of subsystems and components based on the system
reliability goal, which includes the system-level reliability, maintainability, and
availability requirements (e.g., mean time between failures, mean time to repair,
mean down time, and so on). Reliability prediction is a process used for estimating
the reliability of a design prior to manufacturing and testing of produced items.
These are discussed in detail in Chap. 9.
2.2.2.2 Reliability Activities in Development Process
The development stage deals with component and product prototype testing. The
purpose is to rene the design. Using the TAF cycle, the initial design is revised and
improved to meet design requirements and specications. The reliability activities
involved in the development stage fall into the following three categories:
Reliability assessment,
Development tests, and
Reliability improvement.
Reliability assessment is basically concerned with evaluation of the current
reliability during the development process. It can be at any level from system down
to component. Reliability assessment requires test data from carefully designed
experiments and statistical analysis to estimate the reliability.
Development tests are carried out during the development stage to assess and
improve product reliability. Some of the tests carried out during product develop-
ment stage are as follows:
Testing to failure. This can be carried out at any level and each failure is
analyzed and xed.
Environmental and design limit testing. These tests are carried out at the extreme
conditions of its operating environment (including worst-case operating condi-
tions). All failures resulting from the test are analyzed through root-cause
analysis and xed through design changes.
Accelerated life testing. This involves putting items on test under conditions that
are far more severe than those normally encountered. It is used to reduce the
time required for testing.
Testing involves additional costs that depend on the type of tests, number of
items tested, and the test duration. On the other hand, more testing effort results in
better estimates of reliability and this in turn leads to better decision making. As a
result, the optimal testing effort must be based on a tradeoff between the testing
costs and benets derived through more accurate assessment of reliability. These
issues are further discussed in Chap. 10.
Reliability improvement can be achieved through stress-strength analysis,
redundancy design, reliability growth through a development program, and pre-
ventive maintenance (PM) regime design.
Stress-strength analysis assumes that both the strength of a component and the
stress applied to the component are random variables that are characterized by two
distribution functions, from which the probability of failure can be derived.
Different designs can have different distributions of stress and strength and hence
different reliabilities. As such, the stress-strength analysis can be used for the
purpose of reliability improvement.
Redundancy design involves using a multi-component module to replace a single
component. The reliability and cost of module increase with the number of com-
ponents and depend on the type of redundancy. Three different types of redundancy
are hot standby, cold standby, and warm standby. In hot standby, several identical
components are connected in parallel and work simultaneously. The module fails
when all the components fail. As a result, the module lifetime is the largest of all the
components. In cold standby, only one component is in use at any given time.
When it fails, it is replaced by a working component (if available) through a
switching mechanism. If the switch is perfect and the components do not degrade
when not in use, the module lifetime is the sum of lifetimes of all the components of
the module. In warm standby, one component works in a fully loading state and the
other components work in a partially loading state. The component in the partially
loading state has a longer expected life than the component in the fully loading
state. As a result, the warm standby module has a longer expected life than the hot
standby module when the other conditions are the same.
Reliability growth involves research and development effort where the product is
subjected to an iterative TAF process. During this process, an item is tested for a
certain period of time or until a failure occurs. Based on a root-cause failure
analysis for the failures observed during the test, design and/or engineering mod-
ications are made to improve the reliability. The process is repeated until the
2.2 Engineering Activities in Pre-manufacturing Phase 15
reliability reaches a certain required level. The reliability growth process also deals
with the reliability prediction of the system based on the test data and design
changes using various reliability growth analysis techniques, which are discussed in
detail in Chap. 11.
Finally, the eld failure probability can be considerably reduced by imple-
menting a well-designed PM regime, which species various PM activities in a
systematic and comprehensive way. The maintenance related concepts and issues
are discussed in Chaps. 16 and 17.
2.3 Engineering Activities in Production Phase
Three main activities involved in the production phase are production system
design, production system operation, and quality control for operations. In this
section, we rst introduce various types of production systems and then discuss
these three main activities.
2.3.1 Types of Production Systems
Based on production volume and production variety (i.e., number of different types
of products produced), the production system varies from factory to factory and
from product to product. Three common types of production systems are job shop
production, batch production, and mass production. We briefly discuss them below.
2.3.1.1 Job Shop Production
The job shop production system is characterized by the low production volume and
high production variety. Production equipments are mostly general purpose and
flexible to meet specic customer orders, and highly skilled labor is needed to
handle such equipments.
Flexible manufacturing systems (FMS) have been widely used in the job shop
production system. The main components of an FMS are computer numerical
controlled (CNC) machine tools, robots, automated material handling system,
automated storage and retrieval system, and computers or workstations. The FMS
can be quickly congured to produce a variety of products with changeable volume
and mix on the same system. However, it is complex as it is made up of various
different techniques, expensive as it requires a substantial investment of both time
and resources, and of low throughput. For more details about the FMS, see Ref. [1].
Fig. 2.1 Batch production 5

system
4
3 Set-up
Batch
2
Processing
1 Wait
0
Time
2.3.1.2 Batch Production
Batch production is suited for medium volume lot with moderate product variety. In
batch production, the production order is repeated at regular intervals as shown in
Fig. 2.1. Generally, production equipments are general purpose and suitable for
high production volume, and specially designed jigs and xtures are usually used to
reduce the setup time and increase the production rate. It requires that the skill level
of labor should be reasonably high but may be less compared to job shop
production.
2.3.1.3 Mass Production
Mass production is suited for large production volume and low production variety
with low cost per produced unit. The mass production process is characterized by
Mechanization to achieve high volume;
Elaborate organization of materials flow through various stages of
manufacturing;
Careful supervision of quality standards; and
Minute division of labor.
The mass production is usually in a continuous production or line production
way. The line production is a machining system designed for production of a
specic part type at high volume and low cost. Such production lines have been
widely used in the automotive industry.
2.3.2 Production System Design
A well-designed production system ensures to achieve low production cost, desired

productivity, and desired product quality. The main activities involved in the
production system design include (see Ref. [3] and the literature cited therein):
2.3 Engineering Activities in Production Phase 17
Supply chain design;

Production planning and process specications;
System layout on the plant floor; and
Equipment selection and tooling.
We briefly discuss each of these activities below.
2.3.2.1 Supply Chain Design
Two key elements in the production phase are obtaining raw materials and con-
verting raw materials into products (including manufacture and assembly of com-
ponents as well as assembly of assemblies). Raw materials and some components
and assemblies of a product are usually obtained from external suppliers, which
form a complex network termed as supplier chain. The supply chain design deals
with a variety of decisions, including supplier selection, transportation way,
inventory management policies, and so on. Various options form many combina-
tions, and each combination has different cost and performance. Given various
choices along the supply chain, the supply chain design aims to select the options so
as to minimize the total supply chain cost.
One key problem with supply chain design is to appropriately select suppliers.
Supplier selection is a multi-criteria decision making (MCDM) problem, which
involves many criteria such as quality, price, production time and direct cost added,
transportation, warehouse, and so on. Many methods have been developed for
solving the MCDM problems, and the main methods are presented in Online
Appendix A.
Once suppliers are selected, they will be managed through the activities of
several production functions (groups or departments), which include quality,
manufacturing, logistics, test, and so on. For details on supply chain design, see
Refs. [2, 6, 7].
2.3.2.2 Production Planning and Process Specications
There are many design parameters for a manufacturing system, such as number of
flow paths, number of stations, buffer size, overall process capability, and so on.
These depend on production planning, and further depend on process planning,
tolerance analysis, and process capability indicators.
Process planning determines the steps by which a product is manufactured. A
key element is setup planning, which arranges manufacturing features in a sequence
of setups that ensures quality and productivity.
In product design, the tolerance analysis deals with tolerance design and allo-
cation for each component of the product. In production system design, the toler-
ance analysis deals with the design and allocation of manufacturing tolerance,
which serves as the manufacturing process selection.
Different from the process capability indices that measure a specic processs
capability, process capability indicators attempt to predict a proposed production
systems performance. By identifying key drivers of quality in the production
system, these indicators can serve as guidelines for designing production systems
for quality.
2.3.2.3 System Layout
An important step in production system design is system layout. The system layout
impacts manufacturing flexibility, production complexity, and robustness.
Manufacturing flexibility is the capability of building several different products
in one system with no interruption in production due to product differences.
Manufacturing flexibility allows mass customization and high manufacturing uti-
lization. There exists a certain complex relation between flexibility and quality, and
use of robots can improve both flexibility and quality.
Production systems become more and more complex due to the demand for more
product functionality and variety. The manufacturing complexity is characterized
by the number of parts and products, the types of processes, and the schedule
stability. Generally, complexity negatively impacts manufacturing performance
measures, including quality.
Robustness deals with the capability against process drift and fluctuations in
operations. The process drift will lead to producing defective parts. Different
equipment and inspection allocation can have different machine processing time
and defective part arrival rate, and have different yields and drift rates. Sensitivity
analyses can be conducted to examine their interrelations for different design
candidates. The fluctuations in operations result from uncertain or inaccurate system
parameters and can damage product quality. Robust production system design aims
to minimize this damage.
2.3.2.4 Equipment Selection
Equipment determines machine operating characteristics (e.g., operating speed) and

reliability, and hence can impact the quality of produced products. As such, the
equipment selection aims to achieve a good tradeoff between productivity and
quality.
Both operational and quality failures exist in production processes. Operational
failures refer to machine breakdowns, and quality failures refer to production
of defective parts. The processing speed and buffer capacity affect these two types
of failures in a complex way. A quantitative model that considers these types of
failures is needed for equipment selection and operating speed optimization.
2.3.3 Quality Control System Design
The product production includes three elements: inputs (i.e., materials and labor of
operators), processes and outputs (i.e., nished products). Techniques to control
product quality evolve over time and can be divided into the following four
approaches:
Creating standards for producing acceptable products. It focuses on quality
testing at the output end of the manufacturing process.
Statistical quality control, including acceptance sampling with focus on the
input end of the manufacturing process as well as statistical process control with
focus on the manufacturing process.
Total production systems for achieving quality at minimum cost. It focuses on
the whole production system from raw materials to nished product, through
research and development.
Meeting concerns and preferences of consumers. It focuses on consumers needs
and involves the whole PLC.
As seen, the approach to product quality evolves from focusing on quality test
and control to focusing on the quality assurance and improvement. In other words,
the focus gradually moves from the downstream of the PLC toward the upstream of
the PLC. This is because xing a product quality problem in the upstream is much
more cost-effective than xing it in the downstream.
Quality control techniques can be divided into two categories: quality control for
product quality design and improvement, and quality control for production sys-
tems. The techniques in the second category include quality testing and statistical
quality control, and the techniques in the rst category include several basic
approaches. These are further discussed below.
2.3.3.1 Basic Approaches for Quality Design and Improvement
Basic design approaches for design and improvement of product and process
include QFD, design of experiments (DOE), and failure mode and effects analysis
(FMEA). We briefly discuss these issues here and further details are presented in
Chaps. 7 and 8.
QFD has been widely applied to both product design and production planning. It
rst translates customer requirements into product attributes for the purpose of
product design, and then further translates the product attributes into production
process requirements to provide guidelines for the design of the production process
and the design of the quality control process.
DOE approach is developed by Taguchi [9] for the parametric design of product.
The basic idea is to optimally select the combination of controllable (or design)
parameters so that the output performance is insensitive to uncontrollable factor
variation (or noise). The optimization is based on the data from a set of
well-designed experiments. As such, DOE has been widely applied to design or

quality improvement of product or/and process. For example, when DOE is used to
design a robust production system, physical experiments are rst carried out in a
production process, the experimental data are then analyzed to identify key process
parameters, and the key process parameters are optimized to achieve a desired
target. To avoid production disruptions, real experiments may not be conducted,
instead, one can use simulation and existing data.
FMEA is an important tool used to identify failure modes, analyze their effects,
and assess their risk. In a quality planning process, FMEA is often used to assess
the risks of candidate manufacturing processes so as to identify the best candidate.
FMEA has been widely applied to production planning and management to
improve quality and throughput.
2.3.3.2 Quality Control at Input
Statistical quality control is the application of statistical techniques to measuring

and evaluating the quality of a product or process. Two typical techniques are
acceptance sampling and statistical process control. We briefly discuss the accep-
tance sampling here and the statistical process control will be dealt with in the next
subsection.
The input materials are obtained from external suppliers in batches. Their quality
can vary from batch to batch and has a signicant impact on the conformance
quality of items produced. One way of ensuring high input quality is to test for the
quality and a batch is either accepted or rejected based on the outcome of the test.
The test is based on a small sample from the batch. The cost and relevant risks
associated with testing depend on the sample size as well as the type and duration of
tests. The key issue with acceptance sampling is sampling scheme design. More
details about the acceptance sampling are presented in Chap. 13.
2.3.3.3 Quality Control in Process
Quality control in process deals with quality inspection planning and statistical
process control. We rst look at inspection planning, which deals with quality
inspection in production systems. The principal issues with inspection planning
include quality failures, quality inspection, and the actions that may be taken in
response to inspection and measures of system performance.
The design variables of quality inspection system include the number and
locations of inspection stations, inspection plans (e.g., full inspection or sampling),
and corrective actions (e.g., rework, repair, or scrapping). The number and location
of inspection stations are dependent on both the production system and quality
control system; main influence factors include system layout and type of production
systems; and design constraints can be inspection time, average outgoing quality
limit, or budget limit.
When some controllable factors signicantly deviate from their nominal values,
the state of production process changes from in control to out of control. If the
change is immediately detected, then the state can be brought back to in control at
once so as to avoid the situation where many nonconforming items are produced.
The process control methods depend on the type of manufacturing system.
In the case of batch production, a process control technique is to optimize the
batch size. The expected fraction of nonconforming items increases and the setup
cost per item decreases as the batch size increases. As a result, an optimal batch size
exists to make the unit manufacturing cost minimal.
In the case of continuous production, a main process control technique is the use
of control charts to monitor product quality and detect process changes. It involves
taking small samples of the output periodically and plotting the sample statistics
(e.g., the mean, the spread, the number or fraction of defective items) on a chart.
A signicant deviation in the statistics is more likely to be the result of a change in
the state of the process. When this occurs, the process is stopped and the con-
trollable factors that have deviated are restored back to their nominal values before
the process is put into operation.
The cost of quality and accuracy of state prediction depend on the inspection
policy, nature, and duration of the testing involved as well as the control limits. The
design parameters of the inspection policy include the sampling frequency and
sample size. The inspection policy impacts not only quality but also productivity.
This is because normal production may be interrupted when a control chart gen-
erates an out-of-control signal, which can be either an indication of a real quality
problem or a false alarm. Generally, reducing the number of controls leads to better
productivity. Further discussion on control charts are presented in Chap. 14.
2.3.3.4 Quality Control at Output
Quality control at output deals with the quality inspection and testing of produced
items to detect nonconforming (nonfunctional or inferior) items and to weed them
out before the items are released for sale. For nonfunctional items, testing takes very
little time; but for inferior items, the testing can take a signicant length of time.
In either case, testing involves additional costs, and the cost of testing per unit is an
increasing function of the test period. As such, testing design needs to achieve a
tradeoff between the detection accuracy and test effort (i.e., time and cost).
For electronic products, the manufactured items may contain defects, and the
defects can be patent or latent. Environmental stress screening (ESS) can be
effective to force the latent defects to fail, and burn-in can be used to detect the
items with patent defects. Burn-in involves testing the item for a period of s. Those
items that fail during testing are scrapped or repaired. The probability that an item is
conforming after the burn-in increases with s. As such, the reliability of the item
population is improved through burn-in but this is achieved at the expense of the
burn-in cost and a useful life loss of s. A number of models have been developed to
nd the optimal testing scheme. These are discussed in Chap. 15.
2.3.4 Production Management
Production management focuses on the continuous improvement of product quality.

Quality improvements can be achieved by identifying and mitigating quality bot-
tlenecks, implementing an Andon system, batching products, and effectively
planning PM. A quality bottleneck is the factor that most impedes product quality.
Improving the bottleneck factor will lead to the largest improvement in product
quality. An Andon system is an alert system to indicate a quality or process problem
(e.g., part shortage, defect found, tool malfunction, etc.). The alert can be activated
manually by a worker or automatically by the production equipment itself. When an
alert is activated, the production line can be stopped if needed to correct the
problem. Production batching is usually used in multiple-product manufacturing
systems and can reduce changeover time and cost, and improve quality. Finally,
implementing an effective PM program for the production system can improve
productivity and quality.
2.4 Engineering Activities in Post-manufacturing Phase
The post-manufacturing phase can be further divided into three stages: marketing,
post-sale support, and retirement. We discuss the main activities in each of these
stages below.
2.4.1 Main Activities in Marketing Stage
Standard products involve a marketing stage and there is no such a stage for
custom-built products. Marketing deals with issues such as the logistics of getting
the product to markets, sale price, promotion, warranty, channels of distribution,
etc. To address these issues one needs to respond to external factors such as
competitors actions, economy, customer response, and so forth.
2.4.2 Main Activities in Post-sale Support Stage
The support service is necessary to ensure satisfactory operation of the product, and
can add value to the product from both manufacturers perspective (e.g., direct
value in initial sale of product) and customers perspective (e.g., extending the life
cycle, postponing product replacement, etc.). The support service includes one or
more of the following activities:
Providing spares parts, information, and training
Installation
2.4 Engineering Activities in Post-manufacturing Phase 23
Warranties
Maintenance and service contracts
Design modication and customization.
Among these activities, warranties and maintenance are two major issues. Here, we
briefly discuss these two issues and more details are presented in Chaps. 16 and 17.
Warranty is an assurance that the manufacturer offers to the buyer of its product,
and may be considered to be a contractual agreement between the buyer and
manufacturer (or seller). It species both the performance to be expected and the
redress available to the buyer if a failure occurs or the performance is unsatisfac-
tory. Usually, the manufacturer repairs or replaces the items that do not perform
satisfactorily or refunds a fraction or the whole of the sale price. Three important
issues associated with product warranty are warranty policy, warranty servicing cost
analysis, and warranty servicing strategy (e.g., repair or replace).
Maintenance is the actions to control the deterioration process leading to failure
of a system or to restore the system to its operational state through corrective
actions after a failure. As such, maintenance can be broadly divided into two
categories: PM and corrective maintenance (CM). Two important issues for the
manufacturer of a product are maintainability and serviceability design, and the
development of an appropriate PM program. The program will include various PM
actions with different intervals or implementation rules for the components and
assemblies of the product.
Carrying out maintenance involves additional costs to the buyer and is worth-
while only if the benets derived from such actions exceed the costs. This implies
that maintenance must be examined in terms of its impact on the system perfor-
mance. For more details about maintenance, see Ref. [5].
2.4.3 Recycle, Refurbishing, and Remanufacturing
Defective or retired products may return to the manufacturer, who can get prots
from the return through recycling, refurbishing, or remanufacturing. These have
signicant differences in the process and product performances.
Recycling is a process that involves disassembling the original product and
reusing components in other ways, and none of the original value is preserved.
Recycling often discards many of the parts, uses large amounts of energy and
creates much waste and burdens.
Refurbishing is servicing and/or renovation of older or damaged equipment to
bring it to a workable or better looking condition. A refurbished product is usually
worse than the new one in condition.
Remanufacturing is the process of disassembly and recovery. In remanufactur-
ing, the entire product is taken apart, all parts are cleaned and inspected, defective
parts are repaired or replaced, and the product is reassembled and tested. As such,
remanufactured products can be as good as the original ones if part conformity is

insured, and even exceeds the original factory standards if new repair technology is
applied or an original weakness/defect in design is identied and corrected in the
remanufacturing process. Remanufacturing not only reuses the raw materials but
also conserves the value added to the raw materials in the manufacturing process.
2.5 Approach for Solving Quality and Reliability Problems
Modern manufacturing deals with not only the technical aspects but also com-
mercial and managerial aspects. All these aspects need to be properly coordinated.
To effectively manage product quality and reliability requires solving a variety of
problems. These include:
deciding the reliability of a new product,
ensuring certain level of quality of the product,
assessing the quality and reliability of current products being manufactured, and
improving the reliability and quality of the current product.
Solving these problems generally involves the following four steps:
Step 1: Identify and clearly dene a real-world problem.
Step 2: Collect the data and information needed for developing a proper model
to assist the decision-making process.
Step 3: Develop the model for solving the problem.
Step 4: Develop necessary tools and techniques for analyzing the model and
solving the problem.
This approach can be jointly implemented with the plan-do-check-action
(PDCA) management cycle (e.g., see Ref. [4]). Here, Plan deals with establishing
the objectives and processes necessary to produce the expected output, Do means
to implement the plan, Check deals with studying the actual results and com-
paring them with the expected ones, and Action means corrective actions
(including adjustments) on signicant differences between actual and expected
results. The PDCA cycle is repeatedly implemented so that the ultimate goal is
gradually approached.
References
1. El-Tamimi AM, Abidi MH, Mian SH et al (2012) Analysis of performance measures of flexible
manufacturing system. J King Saud Univ Eng Sci 24(2):115129
2. Farahani RZ, Rezapour S, Drezner T et al (2014) Competitive supply chain network design: an
overview of classications, models, solution techniques and applications. Omega 45(C):92118
3. Inman RR, Blumenfeld DE, Huang N et al (2013) Survey of recent advances on the interface
between production system design and quality. IIE Trans 45(6):557574
References 25
4. International Organization for Standardization (2008) Quality management systems. ISO

9000:2000
5. Jiang R, Murthy DNP (2008) Maintenance: decision models for management. Science Press,
Beijing
6. Klibi W, Martel A, Guitouni A (2010) The design of robust value-creating supply chain
networks: a critical review. Eur J Oper Res 203(2):283293
7. Manzini R, Gamberi M, Gebennini E et al (2008) An integrated approach to the design and
management of a supply chain system. Int J Adv Manuf Technol 37(56):625640
8. Murthy DNP, Xie M, Jiang R (2003) Weibull models. Wiley, New York, pp 324347
9. Taguchi G (1986) Introduction to quality engineering. Asian Productivity Organization, Tokyo
Chapter 3
Fundamental of Reliability
3.1 Introduction
This chapter introduces reliability basic concepts, basic functions, and various life
characteristics and measures. We also discuss the evolution of product reliability in
different phases of the product life cycle.
The outline of the chapter is as follows. We start with a brief discussion of basic
concepts of reliability in Sect. 3.2. Reliability basic functions are presented in
Sect. 3.3, the bathtub failure rate curve is discussed in Sect. 3.4, and life charac-
teristics are presented in Sect. 3.5. Failure processes and characteristics of repair-
able systems are introduced in Sect. 3.6. Evolution of reliability over product life
cycle is discussed in Sect. 3.7.
3.2 Concepts of Reliability and Failure
3.2.1 Reliability
Reliability is the probability that an item performs specic functions under given
conditions for a specied period of time without failure. This denition contains
four elements. First, reliability is a probability of no failure, and hence it is a
number between zero and one. The probability element of reliability allows us to
calculate reliabilities in a quantitative way. The second element of reliability def-
inition deals with function and failure, which are two closely linked terms. A
failure means that a device cannot perform its function satisfactorily. There are
several concepts of failure and these are further discussed later. Third, reliability
depends on operating conditions. In other words, a device is reliable under given
conditions but can be unreliable under more severe conditions. Finally, reliability
usually varies with time so that the time to failure becomes a primary random

28 3 Fundamental of Reliability
variable. However, the time element of reliability is not applicable for one-shot
devices such as automobile air-bags and the like. In this case, reliability may be
dened as the proportion of the devices that will operate properly when used.
3.2.2 Failure
Failure can be any incident or condition that causes an item or system to be unable
to perform its intended function safely, reliably and cost-effectively. A fault is the
state of the product characterized by its inability to perform its required function.
Namely, a fault is a state resulting from a failure [1].
Some failures last only for a short time and they are termed as intermittent
failures, while other failures continue until some corrective action recties the
failures. Such failures are termed extended failures. Extended failures can be further
divided into complete and partial failures. A complete failure results in total loss of
function, while a partial failure results in partial loss of function. According to
whether a failure occurs with warning or not, the extended failures can be divided
into sudden and gradual failures. A complete and sudden failure is called a cata-
strophic failure and a gradual and partial failure is called a degraded failure.
Engineering systems degrade with time and usage. Figure 3.1 displays a plot of
the degradation amount (denoted as D(t)) versus time. Reliability-centered main-
tenance (RCM [2]) calls it the P-F curve. Here, the point P is called potential
failure where the item has an identiable defect or the degradation rate changes
quickly. If the defect or degradation continues, the potential failure will evolve into
a functional failure (i.e., the performance is lower than the required standard) at
time point F. The time interval between these two points is called the P-F interval.
A failure can be self-announced (e.g., the failure of light bulbs) or non-self-
announced (e.g., the failure of protective devices). In the case where the failure is
not self-announced, it can be identied only by an inspection. Such a failure is
called the hidden failure.
Fig. 3.1 P-F curve
P
D (t )
P-F interval
F
t
3.2 Concepts of Reliability and Failure 29
3.2.3 Failure Mode and Cause
A failure mode is a description of a fault whereby the failure is observed, or the way
in which the failure happens. It is possible to have several causes for the same
failure mode. Knowledge of the cause of failure is useful in the prevention of
failures. A classication of failure causes is as follows:
Design failure due to inadequate design.
Weakness failure due to weakness in the system so that it is unable to withstand
the stress encountered in the normal operating environment.
Manufacturing failure due to the item being not conforming to the design
specications.
Aging failure due to the effects of age and/or usage.
Misuse failure due to misuse of the system (e.g., operating in environments for
which it was not designed).
Mishandling failure due to incorrect handling and/or lack of care and
maintenance.
3.2.4 Failure Mechanism
Failure mechanism is a physical, chemical, or other process that leads to failure.

The failure occurs due to a complex set of interactions between the material and
other physical properties of the part and the stresses that act on the part. The process
through which these interact and lead to a failure is complex and different for
different types of parts. Dasgupta and Pecht [3] divide the mechanisms of failure
into two broad categories: overstress and wear-out mechanisms. In the overstress
case (see Fig. 3.2), an item fails only if the stress to which the item is subjected
exceeds the strength of the item. If the stress is below the strength, the stress has no
permanent effect on the item. A typical model associated with the overstress
mechanism is the stress-strength interference model.
Fig. 3.2 Overstress Strength Time to failure

mechanism
Stress, strength
Stress
t
Fig. 3.3 Wear-out

Endurance limit
mechanism
D (t )
Time to failure
In the wear-out case (see Fig. 3.3, where Dt indicates the cumulative damage
amount), the stress causes damage that accumulates irreversibly. The accumulated
damage does not disappear when the stress is removed, although sometimes
annealing is possible. The item fails when the cumulative damage reaches the
endurance limit. The deterioration process is a typical cumulative damage process.
3.2.5 Failure Severity and Consequences
The severity of a failure mode indicates the impact of the failure mode on the
system and the outside environment. A severity ranking classication scheme is as
follows [4]:
Catastrophic if failures result in death or total system loss.
Critical if failures result in severe injury or major system damage.
Marginal if failures result in minor injury or minor system damage.
Negligible if failures result in the injury or damage that is lower than marginal.
RCM [2] classies the failure consequence into four levels in descending order
of severity:
Failures with safety consequences,
Failures with environmental consequences,
Failures with operational consequences, and
Failures with non-operational consequences.
3.2.6 Modeling Failures
The basis of reliability analysis is modeling of failure. Modeling of failures can be

done at any level ranging from system level to component level and depends on the
goal (or purpose) that the model builder has in mind. For example, if the goal is to
determine the spare parts needed for components, then modeling of failure needs to
3.2 Concepts of Reliability and Failure 31
be done at the component level; and one might model failures at the system level if
the interest is in determining the expected warranty servicing cost.
Modeling of failures also depends on the information available. At the compo-
nent level, a thorough understanding of the failure mechanisms will allow building
a physics-based model. When no such understanding exists, one might need to
model the failures based on failure data. In this case the modeling is data-driven.
The data-driven approach is the most basic approach in reliability study.
3.3 Reliability Basic Functions
3.3.1 Probability Density Function
For an item, the time to failure, T, is a nonnegative random variable (i.e., T 0). If
there is a set of complete failure observations, we can incorporate the observations
into grouped data: (n1 ; n2 ; . . .; nm ), where ni is the number of failures in time
interval (ti1 ; ti ), 1 i m. Usually, t0 0; tm 1, and ti ti1 Dt with
Dt being the interval length. One can display the grouped data in a plot of ni versus
ti as shown inPFig. 3.4. This plot is termed as histogram of data.
m
Let n i 1 i denote the total number of failures. The relative failure fre-
n
quency in a unit time is given by
ni
fi ; Dt ti ti1 : 3:1
nDt
When n tends to innity and Dt tends to zero, the relative frequency histogram
will tend to a continuous curve. We denote it as f t and call it the probability
density function (pdf). A stricter denition of the pdf is given by
Prt \ T t Dt
f t lim 3:2
Dt!0 Dt
where PrA is the probability of event A.
Fig. 3.4 Histogram of 50

grouped data
40
30
n
20
10
0
0 50 100 150 200 250
t
Fig. 3.5 Plots of the Weibull 1.5

=3.8
pdf
1 =2.8
f (t )
=1.8
0.5
=0.8
0
0 1 2 3
t
The pdf has the following properties. First, it is nonnegative, i.e., f t 0. The
probability that the failure occurs in (t; t Dt) is f tDt, and hence the area under
the pdf curve is the total probability, which equals one, i.e.,
Z1
f tdt 1: 3:3
0
A typical pdf that has been widely used in reliability eld is the Weibull dis-
tribution given by

b t b1 t=gb
f t e ; b [ 0; g [ 0 3:4
g g
where b is the shape parameter and g is the scale parameter. Figure 3.5 shows the
plots of the Weibull pdf with g 1 and b 0:813:8, respectively. It is noted
that the distribution is less dispersive as the Weibull shape parameter b increases.
When b is about 3.44, the Weibull pdf is very close to the normal pdf with the same
mean and variance.
3.3.2 Cumulative Distribution and Reliability Functions
The cumulative distribution function (cdf) is the probability of event T t (i.e., the
item fails by time t). Letting Ft denote the cdf, we have
Zt
Ft PrT t f xdx: 3:5
0
Clearly, we have f t dFt=dt.

3.3 Reliability Basic Functions 33
Fig. 3.6 Relations between

Ft; Rt and f t
f (t )
F (t ) R (t )
The reliability (or survival) function is the probability of event T [ t (i.e., the
item survives to time t). Letting Rt denote the reliability function, we have
Z1
Rt PrT [ t f xdx: 3:6
t
Clearly, we have f t dRt=dt; Rt Ft 1, and
R0 F1 1; R1 F0 0: 3:7
The relations between the reliability function, cdf, and pdf are graphically dis-
played in Fig. 3.6.
For the Weibull distribution, the cdf and reliability function are given, respec-
tively, by
b b
Ft 1 et=g ; Rt et=g : 3:8
3.3.3 Conditional Distribution and Residual Life
If an item has survived to age x, the residual lifetime of the item is given by T x,
which is a random variable. We call the distribution of T x the conditional (or
residual life) distribution. The cdf of residual lifetime is given by
Ft Fx
Ftjx PrT tjT [ x ; t [ x: 3:9
Rx
Fig. 3.7 Pdf and conditional

density function
f (t|x )
f (t ), f (t|x)
f (t )
x t
The pdf and reliability function of residual lifetime are given, respectively, by
f t Rt
f tjx ; Rtjx : 3:10
Rx Rx
Figure 3.7 shows the relation between the underlying pdf and conditional pdf.
For the Weibull distribution, the reliability function of the residual life is given by
b
x=gb
Rtjx et=g : 3:11
3.3.4 Failure Rate and Cumulative Hazard Functions
If an item has survived to age t, the conditional probability of failure occurred at

interval (t; t Dt) is given by Prt \ T t DtjT [ t. The interval mean
failure rate is dened as the ratio of this probability and interval length Dt, and the
failure rate function is the limit of the interval mean failure rate when Dt tends to
zero. Let rt denote the failure rate function at age t. When Dt is innitesimal, we
have
Ft Dt Ft f tDt
Prt\T t DtjT [ t : 3:12
Rt Rt
From this, the failure rate function is given by
Prt\T t DtjT [ t f t
rt : 3:13
Dt Rt
It is noted that the failure rate is nonnegative and can take value in (0; 1).
Therefore, it is not a probability. The probability that the item will fail in t; t Dt,
given that it has not failed prior to t, is given by rtDt.
3.3 Reliability Basic Functions 35
The failure rate characterizes the effect of age on item failure explicitly. The plot
of rt versus t can be either monotonic or nonmonotonic. For the monotonic case,
an item is called positive aging if the failure rate is increasing, negative aging if the
failure rate is decreasing, or nonaging if the failure rate is constant.
For the Weibull distribution, the failure rate function is given by

b t b1
rt : 3:14
g g
Clearly, rt is increasing (or positive aging) when b [ 1, decreasing (or neg-

ative aging) when b\1 and constant (or nonaging) when b 1.
The cumulative hazard function is dened as
Zt
Ht rxdx: 3:15
0
For the Weibull distribution, the cumulative hazard function is given by

Ht t=gb .
3.3.5 Relations Between Reliability Basic Functions
The pdf, cdf, reliability function, and failure rate function are four basic reliability
functions. Given any one of them, the other three can be derived from it. This is
shown in Table 3.1.
To illustrate, we look at a special case where the failure rate is a positive constant
k. From Eq. (3.15), we have Ht kt. Using this in the last column of Table 3.1
we have
f t kekt ; Ft 1 ekt ; Rt ekt : 3:16
Equation (3.16) is the well-known exponential distribution.
Table 3.1 Relations between Derived f t Ft Rt rt

four basic functions \known
f t dFt=dt dRt=dt rteHt
Ft Rt 1 Rt 1 eHt
f xdx
0
Rt 1 Ft 1 Ft 1 Ft
rt f t=Rt f t=Rt f t=Rt
3.4 Component Bathtub Curve and Hockey-Stick Line
There is a close link between the shape of failure rate function (e.g., increasing or
decreasing) and aging property (or failure modes) of an item. The shape type of the
failure rate function is sometimes termed as failure pattern [2]. A well-known
nonmonotonic failure pattern is bathtub-shaped failure rate (or bathtub curve) as
shown in Fig. 3.8.
The bathtub curve can be obtained from observations for many nominally
identical nonrepairable items. It is composed of three parts, which correspond to
early failure phase, normal use phase, and wear-out phase, respectively. In the early
phase of the product use, the failure rate is usually high due to manufacturing and
assembly defects, and it decreases with time since the defects are removed. In the
normal use phase, the failure rate is low and the failure is mainly due to occasional
and random accidents or events (e.g., over-stress) so that the failure rate roughly
keeps constant. In the wear-out phase, the failure rate is high again due to accu-
mulation of damages, gradual degradation, or aging, and hence it increases with
time.
The time point where the failure rate quickly changes is called the change point
of the failure rate. The bathtub curve has two change points. The rst change point
can be viewed as the partition point between the early failure phase and the normal
use phase; and the second change point as the partition point between the normal
use phase and the wear-out phase. A produced item is usually subjected to a burn-in
test to reduce the failure rate in the early phase. In this case, the burn-in period
should not exceed the rst change point. The item can be preventively replaced
after the second change point so as to prevent the wear-out failure.
The desired failure pattern for a product should have the following features [5]:
The initial failure resulting from manufacturing or assembly defects should be
reduced to zero so that there are only random failures in the early phase. This
can be achieved by quality control.
The random failures should be minimized and no wear-out failure occurs during
the normal use period. This can be achieved by design and development
process.
Fig. 3.8 Bathtub curve and

hockey-stick line
r (t )
t
3.4 Component Bathtub Curve and Hockey-Stick Line 37
The occurrence of wear-out failure should be delayed to lengthen the useful life
of the product. This can be achieved by preventive maintenance.
This leads to a change from the bathtub curve to a hockey-stick line (i.e., the
dotted line shown in Fig. 3.8).
3.5 Life Characteristics
3.5.1 Measures of Lifetime
3.5.1.1 Mean Time to Failure
The mean time to failure (MTTF) describes the average of lifetime and is given by
Z1 Z1
l tf tdt Rtdt: 3:17
0 0
It is the rst-order moment of life T.

To derive the expression of MTTF for the Weibull distribution, we consider the
following integral:
Zt
mt; k tk f tdt; k 1; 2; . . .: 3:18
0
Letting z Ht gt b or t gz1=b , Eq. (3.18) can be written as below:

Zz
mt; k g k
zk=b ez dz: 3:19
0
Equation (3.19) can be expressed in terms of the gamma distribution function,

whose pdf is given by
1
gt; u; v tu1 et=v 3:20
vu Cu
where u is the shape parameter, v is the scale parameter, and Cu is the gamma
function evaluated at u. Compared with Eq. (3.20), Eq. (3.19) can be rewritten as
below:
mt; k gk C1 k=bGz; 1 k=b; 1 3:21

where Gz; u; v is the cdf of the gamma distribution with shape parameter u and
scale parameter v. Noting that G1 1, we have the MTTF of the Weibull life
given by

1
l m1; 1 gC 1 : 3:22
b
Microsoft Excel has standard functions to evaluate the gamma function and, the
pdfs and cdfs of the Weibull and gamma distributions. Specic details can be found
from Online Appendix B.
3.5.1.2 BX Life
The BX life is dened as

FBX X % x; x 2 0; 1: 3:23
The BX life with X 10 is called B10 life, which has been widely used in
industries. The BX life associated with x 1 e1 0:6321 is called the char-
acteristic life (denoted as tc ), which meets Htc 1.
Compared with the mean life, the BX life is more meaningful when an item is
preventively replaced at age BX to avoid its failure. In this case, the probability that
the item fails before t BX is x. This implies that this measure links the life with
reliability (i.e., 1 x). For the Weibull distribution, we have
BX g ln1 x1=b ; tc g: 3:24
Reference [6] denes a tradeoff BX life (denoted as BX ), which corresponds to

the maximum of tRt. It is noted that Rt is the probability for an item to survive
to age t and t is the useful lifetime of the item when it survives to and is preven-
tively replaced at age t. Therefore, the tradeoff BX life achieves a good tradeoff
between the useful life and reliability. The tradeoff BX life associated with the
Weibull distribution is given by
BX gb1=b ; x 1 e1=b : 3:25
3.5.1.3 Mean Residual Life
The mean residual life (MRL) is the expectation of the residual life and given by
Z1 Z1
1
lx t xf tjxdt Rtdt: 3:26
Rx
x x
3.5 Life Characteristics 39
Fig. 3.9 Plots of lx 160

=0.8
140
120
=1.0
100
(x )
80 =1.5
60
40 =2.5
20
0
0 50 100 150
x
For the Weibull distribution, from Eqs. (3.11), (3.18) and after some simpli-
cations, we have
" b !#
l x
lx 1G ; 1 1=b; 1 x: 3:27
Rx g
It is noted that the MRL is measured from age x, which is the lifetime already
achieved without failure. Combining the lifetime already achieved with the
expected remaining life, we have the expectation of the entire life given by
MLx lx x. It is called the mean life with censoring.
For g 100 and a set of values of b, Fig. 3.9 shows the plots of lx. As seen,
lx is increasing for b\1, constant for b 1 , and decreasing for b [ 1.
3.5.1.4 Useful Life Associated with Mission Reliability
Suppose the life of a nonrepairable item follows the Weibull distribution. After
operating for T time units (i.e., mission interval), the item is inspected. If the
inspection at ti i 1T indicates that the item is at the normal state, the mission
reliability that the item will survive the next mission interval is evaluated using the
conditional reliability given by
" #
RiT T b b b
b b
Rm iT exp i i 1 RTi i1 : 3:28
Ri 1T g
Assume that the mission reliability is required not to be smaller than a. For
b [ 1, Rm iT decreases with i so that the item has to be replaced after surviving to
a certain mission interval, say I, to ensure the required mission reliability. Clearly,
the I must meet the following relations:
b
I1b b
I b
RTI a; RTI 1 \ a: 3:29
Let x 2 I; I 1 be a real number, which meets:

b
x1b
RTx a: 3:30
As such, we have I intx, where intx is the nearest integer that is not larger
than x. The largest useful life of the item can be achieved when each inspection
indicates the normal state, and is equal to I T.
Example 3.1 Assume that the life of an item follows the Weibull distribution with
parameters b 2:5 and g 100. The duration of each mission is T 16, and
the required mission reliability is a 0:9. The problem is to calculate the largest
useful life of the item.
Solving Eq. (3.30) yields x 3:67, i.e., I 3. As such, the largest useful life
of the item equals 48. It is noted that the mean life is 88.73 and the tradeoff BX life
is 69.31 with X 32:97. This implies that the selection of life measure is appli-
cation-specic.
3.5.2 Dispersion of Lifetime
The variance of life is given by
Z1 Z1
r2 t l2 f tdt 2 tRtdt l2 : 3:31
0 0
It describes the dispersion of life and has a dimension of t2 . Its square root r is
termed as standard deviation and has the dimension of t. The ratio of standard
deviation and mean life is called coefcient of variation (CV) and is given by
q r=l: 3:32
It is independent of the timescale (i.e., dimensionless) and describes the relative

dispersion of the life.
For the Weibull distribution, from Eqs. (3.21) and (3.22) we have

2 1
r2 m1; 2 l2 g2 C 1 C2 1 : 3:33
b b
3.5 Life Characteristics 41
3.5.3 Skewness and Kurtosis of Life Distribution
The skewness of a life distribution is dened as
Z1
1
c1 3 t l3 f tdt: 3:34
r
0
It describes the symmetry of a distribution. For a symmetrical distribution (e.g., the

normal distribution), we have c1 0. However, a distribution is not necessary to
be symmetrical when c1 0. When the left [right] tail of the distribution is longer
than the right [left] tail, c1 \ 0 c1 [ 0 and the distribution is said to be left-skewed
[right-skewed]. For example, the right tail of the exponential distribution is longer
than the left tail, and hence it is right-skewed with c1 2.
The kurtosis of a life distribution is dened as
Z1
1
c2 4 t l4 f tdt 3: 3:35
r
0
It describes the relative peakedness or flatness of a probability distribution com-

pared with the normal distribution. Positive [negative] kurtosis indicates a relatively
peaked [flat] distribution. For normal distribution, c2 0; and for exponential
distribution, c2 6.
3.6 Reliability of Repairable Systems
When a repairable system fails, it can be restored to its work state by a repair and
then continues to work. This forms a failure-repair process. Depending on the depth
of repairs, the time between two successive failures usually decreases in a statistical
sense. In other words, the time between failures is not an independent and identi-
cally distributed random variable. As a result, the life distribution model and
associated life measures are generally no longer applicable for representing the
failure behavior of a repairable system. In this section, we briefly introduce basic
concepts and reliability measures of a repairable system.
3.6.1 Failure-Repair Process
Table 3.2 shows a failure-repair process of a repairable system, where sign +

indicates a running time. The process is graphically displayed in Fig. 3.10, where
Table 3.2 A failure-repair i t2i2 t2i1 t2i ui di mi di

process
1 0 3.47 3.96 3.47 0.49 3.47 0.49
2 3.96 13.34 14.28 9.38 0.94 6.43 0.71
3 14.28 19.04 19.56 4.76 0.51 5.87 0.64
4 19.56 22.61 23.43 3.05 0.82 5.17 0.69
5 23.43 30 [6:57 [5:45
Fig. 3.10 Failure-repair

process of a system
1
State
t1 t2 t3 t4 t5 t6 t7 t8
0
0 5 10 15 20 25 30
t
state 1 means working or up state and state 0 means failure or down state.
The start point of the ith up-down cycle is at t2i2 ; i 1; 2; 3; . . ., and the end point
is at t2i ; and the failure occurs at t2i1 . The uptime and the downtime are given,
respectively, by
ui t2i1 t2i2 ; di t2i t2i1 : 3:36
The downtime can be broadly divided into two parts: direct repair time and other
time. The direct repair time is related to the maintainability of the system, which is a
design related attribute; and other time depends on many factors such as the sup-
portability of the system.
Let si denote the direct repair time of the ith up-down cycle. The mean time to
repair (MTTR) can be dened as below:
1X i
si si : 3:37
i k1
Let Dt denote the total downtime by time t. The total uptime is given by
Ut t Dt: 3:38
Figure 3.11 shows the plots of Dt and Ut for the data of Table 3.2.
3.6 Reliability of Repairable Systems 43
Fig. 3.11 Total uptime and 30

downtime by time t 25
U (t ), D (t )
20
15 U (t )
10
5
D (t )
0
0 5 10 15 20 25 30
t
3.6.2 Reliability Measures
3.6.2.1 Mean Time Between Failures, Mean Time to Repair,

and Mean Downtime
Mean time between failures (MTBF) and mean downtime (MDT) can be evaluated
at the end of each up-down cycle (i.e., t2i ) and given, respectively, by
Ut2i Dt2i
mi ; di : 3:39
i i
For the data shown in Table 3.2, the MTBF and MDT are shown in the last two
columns of the table.
3.6.2.2 Availability
The pointwise (or instantaneous) availability can be computed as
Ut
At : 3:40
t
For the data shown in Table 3.2, Fig. 3.12 shows the plot of instantaneous
availability.
Depending on the depth of maintenance activities performed previously, the
MTBF, MDT, and availability may or may not be asymptotically constant when i or
t is large. Usually, At decreases at rst and then asymptotically tends to a con-
stant. When t ! 1, Eq. (3.40) can be written as
MTBF
A1 : 3:41
MTBF MDT
We call A1 the eld availability.

Fig. 3.12 Instantaneous 1

availability
0.8
0.6
A (t )
0.4
0.2
0
0 5 10 15 20 25 30
t
For a new product, MTBF and MTTR can be estimated based on test data. In this
case, one can use the following to assess the inherent availability
MTBF
A : 3:42
MTBF MTTR
Since MDT [ MTTR, the inherent availability is always larger than the eld
availability.
3.6.3 Failure Point Process
Relative to uptime, downtime is small and can be ignored. In this case, the failure-
repair process reduces into a failure point process. A failure point process can be
represented in two ways. In the rst way, the process is represented by the time to
the ith failure, Ti , which is a continuous random variable and can be represented by
a distribution function. In the second way, the process is represented by the total
number of failures by time t, Nt, which is a discrete random variable. We briefly
discuss these two representations as follows.
3.6.3.1 Continuous Representation
If a repairable system is repaired to an as-good-as-new condition following each

failure, then the failure process is a renewal process and the times between failures
are independent and identically distributed (i.i.d.). In this case, the distribution of
Xi Ti Ti1 is the same as the distribution of X1 T1 .
If the repair only restores the system to an as-bad-as-old state, then the failure
process is called a minimal repair process, and the times between failures are no
longer i.i.d. In other words, Xi s distribution is generally different from the distri-
bution of T1 .
Usually, the repair restores the system to a state that is somewhere between
as-good-as-new and as-bad-as-old states. Such a repair is called general repair or
3.6 Reliability of Repairable Systems 45
Fig. 3.13 Cumulative 5

number of failures E [N (t )]
4
N (t )
2
0
0 5 10 15 20 25 30
t
imperfect repair. In this case, Xi s distribution is also different from the distribution
of T1 . The models and methods for modeling Ti or Xi are discussed in Chap. 6.
3.6.3.2 Discrete Representation
For a given item, the number of cumulative failures is given by Nt i;

t 2 ti ; ti 1 ; i 0; 1; 2; . . .. For the data in Table 3.2, Fig. 3.13 shows the plot of Nt
versus t. Since the downtime is neglected, the t is actually Ut (i.e., uptime).
For a set of nominally identical items, we can obtain several failure point pro-
cesses, from which we can estimate the expectation of Nt, ENt, which is
usually termed as cumulative intensity function, mean value function (MVF) or
mean cumulative function (MCF).
A typical model for MCF is the power-law model, given by
b
t
ENt : 3:43
g
It has the same expression as the Weibull cumulative hazard function but they have
completely different meanings.
Using a curve tting method such as least squared method, we can obtain the
estimates of b and g for the data in Table 3.2, which are b 1:1783 and
g 7:9095. The plot of the tted power-law model is also shown in Fig. 3.13 (the
smooth curve).
The rate of occurrence of failures (or failure intensity function) is dened as
mt dENt=dt: 3:44
For the power-law model given by Eq. (3.43), we have

b t b1
mt : 3:45
g g
It has the same expression as the Weibull failure rate function but with completely
different meanings.
3.6.3.3 System Bathtub Curve
The plot of mt versus t is often termed as the failure pattern of a repairable system
(see Ref. [7]). The plot can be bathtub-shaped. In this case, it is called the system
bathtub curve. The system bathtub curve is different from the component bathtub
curve. The former refers to the rate of occurrence of failures for a repairable system
and the latter refers to the failure rate function for a nonrepairable component. More
specically, the failure rate function represents the effect of the age of an item on
failure and is independent of maintenance actions. On the other hand, the rate of
occurrence of failure represents the intensity for a repairable system to occur the
next (or subsequent) failure, and strongly depends on the maintenance actions
completed previously.
3.7 Evolution of Reliability Over Product Life Cycle
The reliability of a product depends on technical decisions made during the design
and manufacturing phases of the product and is affected by many factors such as
use environment, operating conditions, maintenance activities, and so on. This
implies that product reliability evolves over time. In other words, the reliabilities
evaluated at different time points in the life cycle can be different.
According to the time points when the reliabilities are evaluated, there are four
different reliability notions [8]. They are design reliability, inherent reliability,
reliability at sale, and eld reliability. These notions are important for completely
understanding product reliability, and also useful for one to select an appropriate
model to model the reliability at a certain specic time point. We briefly discuss
these notions as follows.
3.7.1 Design Reliability
Design reliability is the reliability predicted at the end of design and development
phase. The design reliability is inferred based on the test data of product prototypes
and their components, and corrective actions taken during the development process.
The test data is obtained from strictly controlled conditions without being impacted
by actual operating conditions and maintenance activities. As such, the precision of
the prediction will depend on the prediction method and the agreement between the
test conditions and actual use conditions. If the prediction method is appropriate
3.7 Evolution of Reliability Over Product Life Cycle 47
and the test conditions are close to the actual operating conditions, the design
reliability can be viewed as the average eld reliability of product population.
Precise prediction of the reliability of new products in the design phase is
desirable since it can provide an adequate basis for comparing design options.
Specic methods of reliability prediction will be discussed in Chaps. 9 and 11.
3.7.2 Inherent Reliability
Inherent reliability is the reliability realized in manufacture phase. It is usually

evaluated using the life test data of the product after the product is manufactured.
Inherent reliability is different from design reliability due to the influence of
manufacturing process and deviation between the hypothesized and actual reli-
abilities of components.
Since manufacturing processes are inherently variable, the lifetimes of nominally
identical items (components or products) can be different. The life variation results
from unit-to-unit variability due to material properties, components quality varia-
tion, assembly errors, and others. Jiang and Murthy [9] develop the models to
explicitly model the effects of component nonconformance and assembly errors on
reliability. The specic details will be presented in Sect. 12.4.
3.7.3 Reliability at Sale
For a given product, there is a time interval from its assembly to customer delivery.
Usually, the customer delivery time is used as the origin of the product life. Before
this time point, the product is subjected to storage and transportation, which can
result in performance deterioration. The deterioration is equivalent to the product
having been used for a period of time. As a result, the reliability at sale is different
from the inherent reliability and depends on the packaging and packing, trans-
portation process, storage duration, and storage environment.
3.7.4 Field Reliability
The eld reliability is evaluated based on eld failure data. It is different from the
reliability at sale due to the influence of various extrinsic factors on the reliability.
These factors include
Usage mode (continuous or intermittent),
Usage intensity (high, medium, or low),
Usage load (e.g., large, medium, or small),

Operating environment (i.e., temperature, humidity, vibration, etc.),
Functional requirement (i.e., denition of functional failure threshold),
Maintenance activities (PM and CM), and
Operators skill and human reliability.
There can be two approaches to represent the joint effect of these factors on the
eld reliability. The rst approach is to t the data from the items that work under
similar conditions to an appropriate model, and the tted model is only applicable
for the items running under those working conditions. The other approach is to
build a multivariate reliability model such as the proportional hazard (or intensity)
model [10].
3.7.5 Values of Weibull Shape Parameter Associated

with Different Reliability Notions
Assume that inherent reliability, reliability at sale, and eld reliability of a non-
repairable component can be represented by the Weibull distribution with the shape
parameter being bI , bS and bF , respectively. The variation sources that impact the
inherent reliability is less than the variation sources that impact the reliability at
sale. Similarly, the variation sources that impact the reliability at sale is less than the
variation sources that impact the eld reliability. Larger variability results in larger
life spread, and smaller Weibull shape parameter. As a result, it is expected that
bF \ bS \ bI , which have been empirically validated (see Ref. [11]).
References
1. Blischke WR, Murthy DNP (2000) Reliability: modeling, prediction, and optimization. Wiley,
New York, pp 1314
2. Moubray J (1997) Reliability-centered maintenance. Industrial Press Inc, New York
3. Dasgupta A, Pecht M (1991) Material failure mechanisms and damage models. IEEE Trans
Reliab 40(5):531536
4. US Department of Defense (1984) System safety program requirement. MIL-STD-882
5. Ryu D, Chang S (2005) Novel concepts for reliability technology. Microelectron Reliab 45(3
4):611622
6. Jiang R (2013) A tradeoff BX life and its applications. Reliab Eng Syst Saf 113:16
7. Sherwin D (2000) A review of overall models for maintenance management. J Qual Maint Eng
6(3):138164
8. Murthy DNP (2010) New research in reliability, warranty and maintenance. In: Proceedings of
the 4th Asia-pacic international symposium on advanced reliability and maintenance
modeling, pp 504515
9. Jiang R, Murthy DNP (2009) Impact of quality variations on product reliability. Reliab Eng
Syst Saf 94(2):490496
References 49
10. Cox DR (1972) Regression models and life tables (with discussion). JR Stat Soc B 34
(2):187220
11. Jiang R, Tang Y (2011) Influence factors and range of the Weibull shape parameter. Paper
presented at the 7th international conference on mathematical methods in reliability,
pp 238243
Chapter 4
Distribution Models
4.1 Introduction
In this chapter we introduce typical distributional models that have been widely
used in quality and reliability engineering.
The outline of the chapter is as follows. We start with discrete distributions in
Sect. 4.2, and then present simple continuous distributions in Sect. 4.3. The con-
tinuous distributions involving multiple simple distributions are presented in
Sect. 4.4. Finally, the delay time model involving two random variables is presented
in Sect. 4.5.
4.2 Discrete Distributions
4.2.1 Basic Functions of a Discrete Distribution
Consider a nonnegative discrete random variable X that assumes integer values

from the set {0, 1, 2, }. Suppose that there is a set of observations of X given by
fnx ; x 0; 1; 2; . . .:g 4:1
where nx is a nonnegative integer. We call the data given by Eq. (4.1) the count
data. There are many situations where the count data arise, e.g., grouped failure
data, number of defects in product quality analysis, accident data in trafc safety
study, and so on.
The probability mass function (pmf) f x is the probability of event X x, i.e.,
f x PrX x: 4:2

52 4 Distribution Models
P given by Eq. (4.1), we have the empirical pmf given by fx nx =n,

For the data
where n 1 x0 nx .
The cumulative distribution function is dened as Fx PrX x, and the
reliability function as Rx PrX [ x PrX x 1. As a result, we have
f 0 F0 1 R0
4:3
Rx 1 Fx; f x Fx Fx 1; x [ 1:
For the data given by Eq. (4.1), we have the empirical cdf given by
P
Fx 1n xi0 ni .
The discrete failure rate function, rx, is dened as
f x
r0 f 0; rx ; x 1: 4:4
Rx 1
Many discrete distribution models have been developed in the literature (e.g., see
Refs. [1, 10]). Based on the number of distribution parameters, the discrete dis-
tributions can be classied into the following three categories:
Single-parameter models (e.g., geometric and Poisson distributions),
Two-parameter models (e.g., binomial, negative binominal, and zero-inflated
Poisson distributions), and
Models with more than two parameters (e.g., hypergeometric distribution).
4.2.2 Single-Parameter Models
4.2.2.1 Geometric Distribution
Suppose that there is a sequence of independent Bernoulli trials and each trial has
two potential outcomes: failure (or no) and success (or yes). Let p 2 0; 1
[q 1 p] denote the success [failure] probability in each trial. The geometric
distribution is the probability distribution of the event X failures before one suc-
cess trial. As such, the pmf of the geometric distribution is given by
f x qx p: 4:5
The cdf, reliability function, and failure rate function are given, respectively, by
Fx 1 qx1 ; Rx qx1 ; rx p: 4:6

4.2 Discrete Distributions 53
The mean and variance of X is given, respectively, by
l 1=p; r2 q=p2 : 4:7
There is a close link between the exponential distribution and the geometric
distribution. Suppose that a continuous random variable T 0 follows the
exponential distribution and the observed times are at tx x 1Dt. In this case,
we have
PrX x ektx ektx1 1 ekDt ekxDt : 4:8
Letting p 1 ekDt , Eq. (4.8) becomes Eq. (4.5). This implies that
X 0; 1; 2; . . . follows the geometric distribution.
4.2.2.2 Poisson Distribution
The Poisson distribution expresses the probability of a given number of events x

occurring in a xed interval of time if these events occur mutually independently
with a constant arrival rate k. As such, the Poisson distribution is often used to
predict the number of events over a specic time interval such as the number of
failures of a given fleet of vehicles in 1 week.
The pmf of the Poisson distribution is given by
kx ek
f x ; k [ 0: 4:9
x!
The mean and variance of X is given by
l r2 k: 4:10
This relation is called the Poisson equal-dispersion in the literature.
4.2.3 Two-Parameter Models
4.2.3.1 Binomial Distribution
Consider a sequence of n independent Bernoulli trials with a success probability

p 2 0; 1. Let X denote the number of successes. The pmf of the binomial dis-
tribution is given by
f x Cn; xpx 1 pnx 4:11

where Cn; x is the number of combinations choosing x items from n items. The
mean and variance of X is given, respectively, by
l np; r2 l1 p: 4:12
Example 4.1 Suppose that n = 10 items are tested and the success probability is
p = 95 %. Calculate the probability that the number of conforming items equals x.
The probability that the number of conforming items equals x is evaluated by
Eq. (4.11) and the results are shown in Table 4.1.
4.2.3.2 Negative Binomial Distribution
Consider a sequence of independent Bernoulli trials with the success probability

p. The experiment is stopped when the rth failure occurs, where r is a predened
number. Let X denote the number of successes so that the total number of trials is
N X r, which is a discrete random variable with the support r; r 1; . . .. The
event N x r is equivalent to the following two events:
(a) The r xth trial is a failure event with probability 1 p, and
(b) There are r 1 failures in the rst r x 1 trials. The probability of this
event is Cr x 1; xpx 1 pr1 .
As a result, the pmf of X is the product of the probabilities of these two events,
given by
f x Cx r 1; xpx 1 pr : 4:13
The mean and variance are given, respectively, by
l pr=1 p; r2 l=1 p: 4:14
The negative binomial distribution can be extended to the case where r is a positive
real number rather than an integer. In this case, Cx r 1; x is evaluated as
Cx r
Cx r 1; x : 4:15
Cx 1Cr
Table 4.1 Results for x f(x) x f(x) x f(x)

Example 4.1
0 9.8 1014 4 2.7 106 8 0.0746
1 1.9 1011 5 6.1 105 9 0.3151
2 1.6 109 6 0.0010 10 0.5987
3 8.0 108 7 0.0105
In the above denition, the number of failures is xed and the number of
successes is a random variable. The negative binomial distribution can be dened
differently. Let x denote the number of failure and r denote the number of successes.
The experiment is stopped at the rth success. In this case, the pmf of X is still given
by Eq. (4.13).
Example 4.2 Suppose we need to have 100 conforming items and the probability
that an item is conforming is 0.95. The problem is to determine how many items we
need to buy so that we can obtain the required number of conforming items with a
probability of 90 %.
For this example, the number of successes is xed and the number of failures is a
random variable so that the second denition is more appropriate. In this case, the
problem is to nd the value of x so that Fx 1 < 0.9 and Fx > 0.9 for r = 100
and p = 0.95. The computational process is shown in Table 4.2. As seen from the
table, we need to buy x r 108 items to ensure the probability of 90 % to have
100 conforming items.
The problem can be solved using the binomial distribution. Suppose we want to
buy n items with n [ 100. If the failure number is smaller than n 100, the
requirement can be met. The probability of this event is given by
Fn 100; n; 1 p, which must be larger than or equal to 90 %. The computa-
tional process based on this idea is shown in Table 4.3, and the solution is the same
as the one obtained from the negative binomial distribution.
4.2.3.3 Zero-Inflated Poisson Distribution
A count distribution is said to be Poisson-zero-inflated if its proportion of zeros f 0

exceeds the proportion of zeros of a Poisson distribution having the same mean k.
From Eq. (4.9), the proportion of zeros of the Poisson distribution is f 0 ek .
Real count data (e.g., the number of claims of sold cars in the warranty period) is
often Poisson-zero-inflated, and the zero-inflated Poisson (ZIP) distribution pro-
vides a way of modeling the excess zeros (e.g., see Ref. [11]).
Table 4.2 Results for x f(x) F(x) x f(x) F(x)

Example 4.2 (r = 100)
0 0.0059 0.0059 5 0.1701 0.5711
1 0.0296 0.0355 6 0.1489 0.7200
2 0.0747 0.1103 7 0.1127 0.8327
3 0.1271 0.2373 8 0.0754 0.9081
4 0.1636 0.4009
Table 4.3 Results from the n 105 106 107 108 109 110
binomial distribution
F(n) 0.5711 0.7200 0.8327 0.9081 0.9533 0.9779
Dene a special count distribution: f 0 1 and f x 0 for x [ 0. It repre-

sents perfect items. Imperfect items are represented by the Poisson distribution. Let
a 2 0; 1 denote the proportion of the imperfect items. The pmf of the ZIP model is
obtained by mixing these two distributions:
8 k
< 1 a ae ; for x 0
f x x k 4:16
: a k e ; for x [ 0:
x!
The mean and variance of X are given by
l ak; r2 ak2 k a2 k2 l1 k ak [ l: 4:17
Since f 0 ek 1 a1 ek [ 0, we have f 0 [ ek . This implies

that the proportion of zeros of the ZIP distribution is really larger than that of the
corresponding Poisson distribution. Especially, the ZIP distribution reduces to the
Poisson distribution when a 1.
4.2.4 Hypergeometric Distribution
The binomial distribution describes the probability of x successes in n drawn from

an innite population so that 0 x n. The hypergeometric distribution describes
the probability of x successes (or n x failures) in n drawn from a nite population
of size N, which contains m successes. Clearly, 0 x minn; m and
n x N m. This implies:
max0; n m N x minn; m: 4:18
The probability mass function is given by:
f x Cmx CNm
nx
=CNn 4:19
where CAB CA; B is the number of combinations choosing B items from A items.
The mean and variance are given, respectively, by
N mN n
l nm=N; r2 l : 4:20
NN 1
A typical application of the hypergeometric distribution is acceptance sampling.

Here, n items are drawn from N items. Among all the N items, there are m con-
forming items and N m defective items. The random variable X is the number of
conforming items in the drawn n items.
Table 4.4 Pmf of X x 5 6 7 8 9 10

f(x) 0.0001 0.0040 0.0442 0.2098 0.4313 0.3106
f(x; 0.9) 0.0015 0.0112 0.0574 0.1937 0.3874 0.3487
Example 4.3 Assume N; m; n = (50, 45, 10). In this case, we have x 2 5; 10.
The pmf of X is shown in the second row of Table 4.4. For purpose of comparison,
the last row shows the pmf from the binomial distribution with n 10 and p 0:9.
4.3 Simple Continuous Distributions
The univariate continuous distributions can be broadly divided into two categories:
simple distributions and complex distributions. The complex distributions can be
further divided into several sub-categories (e.g., see Ref. [12]). We present some
simple distributions in this section and several complex distributions that involve
two or more simple distributions in the following section.
4.3.1 Weibull Distribution
The Weibull pdf is given by Eq. (3.4), the mean is given by Eq. (3.22), and the
variance is given by Eq. (3.33). The Weibull distribution is mathematically tractable
with closed-form expressions for all the reliability basic functions. It is also flexible
since the failure rate function can be decreasing, constant, or increasing. The shape
parameter b represents the aging characteristics and the scale parameter g is the
characteristic life and proportional to various life measures. Jiang and Murthy [7]
present a detailed study for the properties and signicance of the Weibull shape
parameter.
The three-parameter Weibull distribution is an extension of the two-parameter
Weibull model with the cdf given by a piecewise function:
8
< 0; t 2 0; c
Ft tc b 4:21
: 1 exp ; t [ c
g
where c (> 0) is called the location parameter.

When lnT follows the three-parameter Weibull distribution, T follows the log-
Weibull distribution. The log-Weibull distribution has some excellent properties
and can be used as a life distribution [8].
The well-known Weibull transformations are given by:
x lnt; y ln lnRt lnHt: 4:22
Under these transformations, the two-parameter Weibull distribution can be written

as below:
y bx b lng: 4:23
It is a straight line in x-y plane. The plot of y versus x is called the Weibull
probability paper (WPP) plot. The Weibull transformations can be applied to any
other distribution with the nonnegative support but the resulting WPP plot is no
longer a straight line. For example, the WPP plot of the three-parameter Weibull
distribution is concave.
Due to y lnHt, we have jyj Ht for small t and y Ht for large
t. Similarly, due to x lnt, we have jxj t for small t t 2 0; 1 and x t for
large t t 1. As a result, the Weibull transformations produce an amplication
effect for the lower-left part of the WPP plot and a compression effect for the upper-
right part [3].
4.3.2 Gamma Distribution
The pdf of the gamma distribution is given by Eq. (3.20). Generally, there are no
closed-form expressions for the other three basic functions but Microsoft Excel has
standard functions to evaluate them (see Online Appendix B).
The k-order origin moment of the gamma distribution is given by:
Z1
mk xk gxdx vk Ck u=Cu: 4:24
0
As a result, the mean and variance are given b
l uv; r2 uv2 4:25
Somewhat similar to the Weibull distribution, its failure rate function is

decreasing, constant, or increasing when u\1, u 1 or u [ 1, respectively.
However, different from the Weibull distribution, when t ! 1, we have
f t f 0 t u1 1 1
rt ! lnf t0 ! :
1 Ft f t t v v
This implies that the failure rate of the gamma distribution tends to a constant rather
than zero or innity.
4.3 Simple Continuous Distributions 59
The gamma distribution has a long right tail. It reduces to the exponential
distribution when u 1; to the Erlang distribution when u is a positive integer; and
to the chi-square distribution with n degrees of freedom when u n=2 and v 2.
The chi-squared distribution is the distribution of the random variable
P
Qn ni1 Xi2 , where Xi 1 i n are independent and standard normal random
variables. The chi-squared distribution is widely used in hypothesis testing and
design of acceptance sampling plans.
4.3.3 Lognormal Distribution
The lifetime of a component (e.g., bearing) or structure subjected to corrosion or

fatigue failure usually follows the lognormal distribution given by
lnt ll t
Ft U Ufln l 1=rl g 4:26
rl el
where U: is the standard normal cdf. It is noted that ell is similar to the Weibull
scale parameter and r1 l is similar to the Weibull shape parameter. Therefore, we
call ll and rl the scale and shape parameters, respectively. The mean and variance
are given, respectively, by
l expll r2l =2; r2 l2 expr2l 1: 4:27
The lognormal distribution has a longer right tail than the gamma distribution.
The failure rate function is unimodal, and can be effectively viewed as increasing
when rl < 0.8, constant when rl 2 0:8; 1:0 and decreasing when rl [ 1 [9].
4.4 Complex Distribution Models Involving

Multiple Simple Distributions
In this section we look at complex models involving multiple simple distributions

(e.g., the Weibull distribution). More details about these models can be found in
Refs. [4, 5, 12].
4.4.1 Mixture Model
In a batch of products, some are normal while others are defective. The lifetime of
the normal product is longer than that of the defective product, and hence the
former is sometimes called the strong sub-population and the latter is sometimes
called the weak sub-population.
In general, several different product groups are mixed together and this forms a
mixture population. Two main causes for the mixture are:
(a) product parts can come from different manufacturers, and
(b) products are manufactured in different production lines or by different oper-
ators or by different production technologies.
Let Fj t denote the life distribution for sub-population j, and pj denote its
proportion. The life distribution of population is given by
X
n X
n
Ft pj Fj t; 0 \ pj \ 1; pj 1: 4:28
j1 j1
When n 2 and Fj t is the Weibull distribution, we call Eq. (4.28) the twofold
Weibull mixture. The main characteristics for this special model are as follows [6]:
The WPP plot is S-shaped.
The pdf has four different shapes as shown in Fig. 4.1.
The failure rate function has eight different shapes.
The mixture model has many applications, e.g., burn-in time optimization and
warranty data analysis. We will further discuss these issues in Chaps. 15 and 16.
4.4.2 Competing Risk Model
An item can fail due to several failure modes, and each can be viewed as a risk. All
the risks compete and the failure occurs due to the failure mode that rst reaches.
Such a model is termed as competing risk model. An example is the system
composed of n independent components without any redundant component. The
system fails when any component fails; or the system can survive to t only if each
component of the system survives to t.
Let Ti denote the time to failure of component i, and Ri t denote the probability
that component i survives to t. Similarly, let T denote the time to failure of the
Fig. 4.1 Shapes of pdf of the 0.012

twofold Weibull mixture 0.01
0.008
f (t )
0.006
0.004
0.002
0
0 50 100 150 200
t
4.4 Complex Distribution Models Involving Multiple Simple Distributions 61
system, and Rt denote the probability that the system survives to t. Clearly,
T minTi ; 1 i n. As a result, under the independent assumption we have
Y
n
Rt Ri t: 4:29
i1
If the ith item has an initial age ai at the time origin, Ri t should be replaced by
Ri t ai =Rai :
From Eq. (4.29), the system failure rate function is given by
X
n
rt ri t: 4:30
i1
This implies that the system failure rate is the sum of component failure rates.
If n items are simultaneously tested and the test stops when the rst failure
occurs, this test is called the sudden death testing. The test duration T is a random
variable and follows the n-fold competing risk model with Fi t F1 t; 2 i n.
In this case, the cdf of T is given by Ft 1 1 F1 tn .
Another special case of Eq. (4.29) is n 2 and is termed as the twofold com-
peting risk model. In this case, the item failure can occur due to one of two
competing causes. The time to failure T1 due to Cause 1 is distributed according
to F1 t, and the time to failure T2 due to Cause 2 is distributed according to
F2 t. The item failure is given by the minimum of T1 and T2 , and Ft is given by:
Ft 1 1 F1 t1 F2 t 4:31
When Fi t; i 1; 2, is the Weibull distribution, we obtain the twofold Weibull

competing risk model, which has the following characteristics [5]:
The WPP plot is convex.
The pdf has the four different shapes as shown in Fig. 4.1.
The failure rate function has three different shapes: decreasing, bathtub-shaped,
and increasing.
4.4.3 Multiplicative Model
Consider a system made up of n independent components. The system works as

long as any of the components work. In other words, the system fails only if all the
components fail. We call this model the multiplicative model. Using the same
notations as those in the competing risk model, the system life is given by
T maxTi ; 1 i n. Under the independent assumption, we have
Y
n
Ft Fi t: 4:32
i1
If the ith item has an initial age ai at the time origin, Fi t should be replaced by
1 Ri t ai =Rai .
The multiplicative model has two typical applications. The rst application is the
hot standby system, where n components with the same function simultaneously
operate to achieve high reliability. The second application is in reliability test,
where n items are simultaneously tested and the test stops when all the components
fail. In this case, the test duration T is a random variable and follows the n-fold
multiplicative model with Fi t F1 t; 2 i n; the cdf of T is given by
Ft F1n t.
Another special case of the model given by (4.32) is n 2. In this case, Ft is
given by
Ft F1 tF2 t: 4:33
If Fi t; i 1; 2, are the Weibull distribution, we obtain a twofold Weibull

multiplicative model. This model has the following characteristics [5]:
The WPP plot is concave.
The pdf has three different shapes: decreasing, unimodal, and bimodal.
The failure rate function has four different shapes: decreasing, increasing, uni-
modal, and unimodal-followed-by-increasing.
4.4.4 Sectional Models
A general sectional (or piecewise) model is dened as
Ft Gi t; t 2 ti1 ; ti ; 1 i n; t0 0; tn 1 4:34
where Gi t is an increasing function of t and meets the following:
G1 0 0; Gi ti Gi1 ti ; Gtn 1: 4:35
It is noted that a step-stress testing model has the form of Eq. (4.34).
4.4 Complex Distribution Models Involving Multiple Simple Distributions 63
Murthy et al. [12] dene an n-fold sectional model given by

(
1 k1 k1 R1 t; t 2 0; t1
Rt 4:36
ki Ri t; t 2 ti1 ; ti ; i 2
where ki [ 0 and Ri 1 Fi t are reliability functions. In terms of cdf, Eq. (4.36)

can be written as
(
ki Fi t; t 2 ti1 ; ti ; 1 i n 1
Ft 4:37
1 kn kn Fn t; t 2 tn1 ; 1:
For the distribution to be continuous at the break points, the model parameters
need to be constrained. We consider two special cases as follows.
4.4.4.1 Sectional Model Involving Two-Parameter

Weibull Distributions
Consider the model given by Eq. (4.36). Assume that Ri t 1 i n are the two-
parameter Weibull distribution; k1 1 and ki [ 0 for 2 i n. As such, the model
has 3n 1 parameters (assume that tis are known). To be continuous, the
parameters meet the following n 1 relations:
R1 t1 k2 R2 t1 ; ki Ri ti ki1 Ri1 ti ; 2 i n 1: 4:38
As a result, the model has 2n independent parameters.

Especially, when n 2 and b1 b2 b, Eq. (4.38) reduces to

expt=g1 b ; t 2 0; t1
Rt b ; k2 expgb b b
2 g1 t1 : 4:39
k2 expt=g2 ; t 2 t1 ; 1
This twofold Weibull sectional model has only three independent parameters.
4.4.4.2 Sectional Model Involving Three-Parameter

Weibull Distributions
Consider the model given by Eq. (4.37). Assume that F1 t is the two-parameter
Weibull distribution, Fi t (2 i n) are the three-parameter Weibull distribution
with the location parameter ci , and ki 1 for 1 i n. To be continuous, the
parameters meet the following n 1 relations:
Fig. 4.2 Distribution 1

functions of Eqs. (4.39) and
0.8
(4.41)
0.6
F (t )
0.4
0.2
0
0 5 10 15
t
Fi ti Fi1 ti ; 1 i n 1: 4:40
As such, the model has 2n independent parameters (if tis are known).
Especially, when n 2 and b1 b2 b, Eq. (4.40) reduces to
8
>
< 1 expt=g1 b ; t 2 0; t1
Ft g 4:41
>
: 1 exptc 2 b ; c2 1 2 t1 :
g2 ; t 2 t1 ; 1 g1
Example 4.4 The models given by Eqs. (4.39) and (4.41) can be used to model
simple step-stress testing data. Assume t1 8 and g1 ; b = (10, 2.3). When
g2 6:88, we have k2 2:2639 for Model (4.39); when g2 5, we have c2 4
for Model (4.41). Figure 4.2 shows the plots of the distribution functions obtained
from Models (4.39) and (4.41). As seen, they are almost overlapped, implying that
the two models can provide almost the same t to a given dataset.
4.5 Delay Time Model
The distribution models presented above involve only a single random variable. In
this section, we introduce a distribution model, which involves two random
variables.
Referring to Fig. 4.3, the item lifetime T is divided into two parts: normal and
defective parts. The normal part is the time interval (denoted as U) from the
beginning to the time when a defect initiates; and the defective part is the time
period from the defect initiation to failure, which is termed as delay time and
denoted as H. Both U and H are random variables.
The delay time concept and model are usually applied to optimize an inspection
scheme, which is used to check whether the item is defective or not. Suppose an
item is periodically inspected. If a defect is identied at an inspection (as Case 1 in
4.5 Delay Time Model 65
Fig. 4.3 Delay time concept

and periodic inspection 2
U H
Case
1
0
t
Fig. 4.3), the item is preventively replaced by a new one; if the item fails before the
next inspection (as Case 2 in Fig. 4.3), it is correctively replaced. As such, the
maintenance action can be arranged in a timely way and the operational reliability is
improved. For more details about the concept and applications of the delay time, see
Ref. [13] and the literature cited therein.
Suppose a single item is subjected to a major failure mode (e.g., fatigue) and the
failure process of the item can be represented by the delay time concept. Let Fu t
and Fh t denote the distributions of U and H, respectively. The time to failure is
given by T U H. Assuming that U and H are mutually independent and, Fu t
and Fh t are known, the distribution function of T is given by
Zt
Ft Fh t xdFu x: 4:42
0
Generally, Ft given by Eq. (4.42) is analytically intractable. A Monte Carlo

simulation approach can be used to nd Ft (e.g., see Ref. [2]). In this case, a set of
N random times of both U and H is rst generated (see Sect. B.3 of Online
Appendix B), and then a random sample of T is obtained. An approximation of Ft
can be obtained by tting the sample to an appropriate distribution. We illustrate the
approach with the following example.
Example 4.5 Assume that U follows the Weibull distribution with b 2:5 and
g 10, and H follows the gamma distribution with u 1:5 and v 1:4. From the
known conditions we have ET EU EH 10:9726. Assume that
T approximately follows the Weibull distribution with shape parameter b. Then, the
scale parameter is a function of b and is given by g ET=C1 1=b. Take a
random sample of size 500 for T and t the sample to the Weibull distribution. The
estimated shape parameter is b 2:8481, from which we have g 12:3144.
The methods to t a given dataset to a specic distribution will be discussed in

the next chapter.
References
1. Bracquemond C, Gaudoin O (2003) A survey on discrete lifetime distributions. Int J Reliab

Qual Saf Eng 10(01):6998
2. Jiang R (2013) Relationship between delay time and Gamma process models. Chem Eng
Trans 33:1924
3. Jiang R (2014) A drawback and an improvement of the classical Weibull probability plot.
Reliab Eng Syst Saf 126:135142
4. Jiang R, Murthy DNP (1995) Modeling failure-data by mixture of 2 Weibull distributions: a
graphical approach. IEEE Trans Reliab 44(3):477488
5. Jiang R, Murthy DNP (1995) Reliability modeling involving two Weibull distributions. Reliab
Eng Syst Saf 47(3):187198
6. Jiang R, Murthy DNP (1998) Mixture of Weibull distributionsparametric characterization of
failure rate function. Appl Stoch Models Data Anal 14(1):4765
7. Jiang R, Murthy DNP (2011) A study of Weibull shape parameter: properties and signicance.
Reliab Eng Syst Saf 96(12):16191626
8. Jiang R, Wang T (2013) Study of the log-Weibull lifetime distribution. In: Proceedings of
international conference on quality, reliability, risk, maintenance, and safety engineering,
pp 851854
9. Jiang R, Ji P, Xiao X (2003) Aging property of unimodal failure rate models. Reliab Eng Syst
Saf 79(1):113116
10. Johnson NL, Kemp AW, Kotz S (1992) Univariate discrete distributions, 2nd edn. John Wiley
and Sons, New York
11. Majeske KD, Herrin GD (1995) Assessing mixture-model goodness-of-t with an application
to automobile warranty data. In: Proceedings of annual reliability and maintainability
symposium, pp 378383
12. Murthy DNP, Xie M, Jiang R (2003) Weibull models. John Wiley and Sons, New York
13. Wang W (2008) Delay time modeling. In: Murthy DNP, Kobbacy (eds) Complex system
maintenance handbook. Springer, London
Chapter 5
Statistical Methods for Lifetime Data
Analysis
5.1 Introduction
In this chapter, we consider the problem to t a given dataset to a distribution

model. This problem deals with parameter estimation, hypothesis testing for
goodness of t, and model selection. The parameter estimation deals with deter-
mining the parameters of a distribution model based on given data; hypothesis
testing for goodness of t with assessing the appropriateness of the tted model; and
model selection with choosing the best model from a set of candidate models. We
present typical statistical methods to address these issues.
The outline of the chapter is as follows. We rst discuss various types of reli-
ability data in Sect. 5.2. Nonparametric estimation methods of cdf are presented in
Sect. 5.3 and parameter estimation methods are presented in Sect. 5.4. Section 5.5
deals with hypothesis testing and Sect. 5.6 with model selection.
5.2 Reliability Data
5.2.1 Sources and Types of Data
Data for reliability modeling and analysis are mainly from testing and use eld, and
sometimes from the published literature and experts judgments. The test data are
obtained under controlled conditions and the eld data are usually recorded and
stored in a management information system.
The data used for reliability modeling and analysis can be classied into the
following three types:
Data of time to failure (TTF) or time between failures (TBF);
Data of performance degradation (or internal covariate resulting from degra-
dation); and

68 5 Statistical Methods for Lifetime Data Analysis
Data of environments and use conditions (or external covariate resulting in

degradation).
We discuss these types of data in the following three subsections.
5.2.2 Life Data
5.2.2.1 Time to Failure and Time Between Failures
A non-repairable item fails (or is used) only once. In this case, a failure datum is an
observation of TTF, and the observations from nominally identical items are
independent and identically distributed (i.i.d.). A life distribution can be used to t
TTF data.
On the other hand, a repairable item can fail or be used several times. In this
case, a failure datum is an observation of time between successive failures
(including the time to the rst failure (TTFF) for short). Depending on the main-
tenance activities to restore the item to its working state, the TBF data are generally
not i.i.d. except the TTFF data. The model for modeling TBF data is usually a
stochastic process model such as the power-law model. In this chapter, we focus on
the i.i.d. data, and we will look at the modeling problem for the data that are not i.i.
d. in the next chapter.
5.2.2.2 Complete and Incomplete Data
According to whether or not the failure time is exactly known, life data can be
classied into complete data and incomplete data. For complete data, the exact
failure times are known. In other words, each of the data is a failure observation. It
is often hard to get a complete dataset since the test usually stops before all the
tested items fail or the items under observation are often in a normal operating state
when collecting data from eld.
The incomplete data arise from censoring and truncation. In the censoring case,
the value of an observation is only partially known. For example, one (or both) of
the start point and endpoint of a life observation is (are) unknown so that we just
know that the life is larger or smaller than a certain value, or falls in some closed
interval. A censoring datum contains partial life information and should not be
ignored in data analysis.
In the truncation case, the items are observed in a time window and the failure
information is completely unknown out of the observation window. For example,
the failure information reported before the automotive warranty period is known but
completely unknown after the warranty period. There are situations where the
failure information before a certain time is unknown. For example, if the time for an
item to start operating is earlier than the time for a management information system
5.2 Reliability Data 69
Censored observations
Observation window
Fig. 5.1 Truncated failure point process and censored observations
to begin running, the failure information of the item before this information system
began running is unknown. However, truncation usually produces two censored
observations as shown in Fig. 5.1, where sign indicates a failure. The censored
observation on the left is actually a residual life observation and is usually termed
left-truncated data and the censored observation on the right is usually termed right-
truncated data.
5.2.2.3 Types of Censoring Data
There are three types of censoring: left censoring, right censoring, and interval
censoring. In left censoring case, we do not know the exact value of a TTF
observation but know that it is below a certain value. We let tf denote the actual
failure time and t denote its known upper bound. A left censoring observation
meets the following:
0\ tf \ t : 5:1
Left censoring can occur when a failure is not self-announced and can be identied
only by an inspection.
In the right censoring case, we only know that the TTF is above a certain value.
Let t denote the known lower bound. A right censoring observation meets the
following:
t \ tf \1: 5:2
Right censoring can occur when an item is preventively replaced or the life test is
stopped before failure of an item.
Table 5.1 Grouped data Interval (0, t1 ) (t1 , t2 ) (tk1 , tk )

Failure number n1 n2 nk
In interval censoring case, we only know that the TTF is somewhere between
two known values. An interval censoring observation meets the following:
t \tf \t : 5:3
Interval censoring can occur when observational times are scheduled at some dis-
crete time points.
It is noted that both left censoring and right censoring can be viewed as special
cases of interval censoring with the left endpoint of the interval at zero or the right
endpoint at innity, respectively.
Grouped data arise from interval censoring. Suppose n items are under test, and
the state of each item (working or failure) is observed at time ti iDt, i 1; 2; . . ..
If an item is at the failure state at ti , then the exact failure time is unknown but we
know ti1 \tf ti . As a result, a grouped (or count) dataset can be represented by
Table 5.1, and the interval with tk 1 is called half-open interval.
When the sample size n is large, a complete dataset can be simplied into the
form of the grouped data. Such an example is the well-known bus-motor major
failure data from Ref. [4]. The data deal with the bus-motor major failure times. The
major failure is dened as serious accidents (usually involved worn cylinders,
pistons, piston rings, valves, camshafts, or connecting rod or crankshaft bearings) or
performance deterioration (e.g., the maximum power produced fell below a spec-
ied proportion of the normal value).
Table 5.2 shows the times to the rst through fth major failures of the bus-
motor fleet. The time unit is 1000 miles and tu is the upper bound of the nal
interval. It is noted that minor failures and preventive maintenance actions are not
shown and the total number of bus-motors varies, implying that a large amount of
information is missed.
5.2.2.4 Typical Life Test Schemes
There are three typical life test schemes. The rst scheme is test-to-failure. Suppose
n items are tested and the test ends when all the items fail. Times to failure are given
by
t1 ; t2 ; . . .; tn : 5:4
This dataset is a complete dataset with all failure times being known exactly. With
this scheme, the test duration is a random variable and equals the maximum of
lifetimes of the tested items.
Table 5.2 Grouped bus- ti1 ti 1st 2nd 3rd 4th 5th
motor failure data
0 20 6 19 27 34 29
20 40 11 13 16 20 27
40 60 16 13 18 15 14
60 80 25 15 13 15 8
80 100 34 15 11 12 7
100 120 46 18 16
120 140 33 7
140 160 16 4
160 180 2
180 220 2
tu 220 210 190 120 170
Total 191 104 101 96 85
number
To reduce the test time, the second test scheme (termed as Type I censoring or
xed time test) stops the test at a predetermined time t s. In this scheme, the
number of failure observations, k (0 k n), is a random variable, and the data are
given by

t1 ; t2 ; . . .; tk ; tj s ; k 1 j n : 5:5
where s means >.

The third test scheme (termed as Type II censoring or xed number test) stops
the test when the kth failure is observed, where k is a predetermined number. In this
scheme, the test duration (i.e., tk ) is a random variable and the data are given by

t1 ; t2 ; . . .; tk ; tj tk ; k 1 j n : 5:6
5.2.2.5 Field Data
For the manufacturer of a product, the eld information can be used as feedback to
learn about the reliability problems of a product and to improve future generations
of the same or similar product. For the user, the eld information can be used to
optimize the maintenance activities and spare part inventory control policy. Many
enterprises use a management information system to store the maintenance-related
information. Most such systems are designed for the purpose of management rather
than for the purpose of reliability analysis. As a result, the records are often
ambitious and some important information useful for reliability analysis is missed.
In extracting eld data from a management information system, there is a need to
differentiate the item age from the calendar time and inter-failure time. Figure 5.1
shows the failure point process of an item, with the repair time being ignored. The
Table 5.3 An alternately ti 110 151 255 343 404 438 644
censored dataset
di 1 1 0 0 1 1 0
ti 658 784 796 803 958 995 1000
di 0 1 0 1 0 0 1
ti 1005 1146 1204 1224 1342 1356 1657
di 1 1 1 1 1 1 1
horizon axis is calendar time, in which the failure times can be denoted as
T fTi ; i 1; 2; . . .g. The time between two adjacent failures Ti Ti1 can be
either item age (denoted as Xi ) if the failure is corrected by a replacement, or
inter-failure time (denoted as Yi ) if the failure is corrected by a repair so that the
item is repeatedly used. Though Xi and Yi look similar, they are totally different
characteristics. The Xi and Xi1 come from two different items and can be i.i.d.; Yi
and Yi1 come from the same item and are usually not i.i.d. As a result, models for
modeling T, X and Y can be considerably different.
When extracting life data of multiple nominally identical and non-repairable
items from a management information system, we will have many right-censored
data. If the data are reordered in ascending order, we will obtain an alternately
censored dataset. The alternately censored dataset is different from the Type-I and
Type-II datasets where the censored observations are always larger than or equal to
the failure observations. Table 5.3 shows a set of alternately censored data, where ti
is failure or censored time and di is the number of failures at ti . In practice we often
observe several failures at the same time. Such data are called tied data and result
from grouping of data or from coarse measurement (e.g., rounding-errors). In this
case, we can have di [ 1.
5.2.3 Performance Degradation Data
The performance of an item deteriorates with time and usage and leads to failure if
no preventive maintenance action is carried out. The degradation information is
usually obtained through condition monitoring or inspection, and is useful for life
prediction. The degradation can be measured by the variables or parameters (e.g.,
wear amount, vibration level, debris concentration in oil, noise level, etc.) that can
directly or indirectly reflect performance. Such variables or parameters are often
called covariates or condition parameters.
5.2.4 Data on Use Condition and Environment
Another type of information useful for life prediction is the use condition and
environment data (e.g., load, stress level, use intensity, temperature, humidity, etc.).
As an example, consider accelerated life testing. It deals with testing items under
more severe conditions than normal use conditions (e.g., more intensive usage,
higher stress level, etc.) so as to make the item fail faster. Under a constant-stress
accelerated testing scheme, the failure data are given by paired data
(si ; tij ; 1 i m; 1 j ni ), where si is the stress level and can be viewed as data on
use condition and environment.
5.3 Nonparametric Estimation Methods for Cdf
Probability plots are commonly used to identify an appropriate model for tting a
given dataset. To present the data on a plotting paper, the empirical cdf for each
failure time must be rst estimated using a nonparametric method. Some parameter
estimation methods such as the graphical and least square methods also require
estimating the empirical cdf.
Nonparameter estimation method of cdf depends on the type of dataset available.
We rst look at the case of complete data and then look at the case of incomplete
data.
5.3.1 Complete Data Case
Consider an ordered complete dataset given by:
t1 t2 . . . tn : 5:7
Let Fi denote the nonparametric estimate of Fti . Fi is a random variable and

follows the standard beta distribution with shape parameters i and n i 1. The
i
mean of Fi is given by n1 , and the median can be evaluated by
betainv0:5; i; n i 1, which is an Excel standard function to accurately evaluate
the median of the above-mentioned beta cdf (see Ref. [7]). The median of Fi can be
approximately evaluated by
i 0:3
Fi : 5:8
n 0:4
5.3.2 Grouped Data Case
Assume that there are di failures in the interval ti1 ; ti ; 1 i m. The total number
P
of failures by ti equals ni ij1 dj and the sample size is n nm . Fti can be
estimated by betainv0:5; ni ; n ni 1 or Eq. (5.8) with i being replaced by ni .
Table 5.4 Empirical cdf of ti di Fti Fti , (5.8) xi yi

grouped data for Example 5.1
20 6 0.0296 0.0298 2.9957 3.5038
40 11 0.0871 0.0873 3.6889 2.3953
60 16 0.1707 0.1708 4.0943 1.6755
80 25 0.3014 0.3015 4.3820 1.0254
100 34 0.4791 0.4791 4.6052 0.4274
120 46 0.7195 0.7194 4.7875 0.2400
140 33 0.8920 0.8918 4.9416 0.8000
160 16 0.9756 0.9754 5.0752 1.3118
180 2 0.9860 0.9859 5.1930 1.4517
220 2 0.9964 0.9963 5.3936 1.7264
Example 5.1 Consider the rst set of bus-motor failure data given in Table 5.2.
Based on the median estimate of Fi , the estimates of empirical cdf are shown in the
third column of Table 5.4. The estimates obtained from Eq. (5.8) are shown in the
fourth column. As seen, the estimates from the two methods are very close to each
other. The last two columns of Table 5.4 are the corresponding Weibull transfor-
mations obtained from Eq. (4.22). The WPP plot (along with the regression straight
line of the data points) is shown in Fig. 5.2.
5.3.3 Alternately Censored Data Case
Typical nonparametric estimators for alternately censored data include:

KaplanMeier method (KMM [8]),
NelsonAalen method (NAM [12]),
Mean rank order method (MROM [12]), and
Piecewise exponential method (PEM [9]).
We present and illustrate these methods as follows.
5.3.3.1 KaplanMeier Method
Consider an alternately censored dataset, which is arranged in ascending order. Let i

(1 i n) be the order number of the ith observation in all observations and j
(1 j m) be the order number of the jth failure observation and ij be corre-
sponding value of i. If a censored observation has the same value as a failure
observation, we always arrange it after the failure observation; if there are several
tied failure observations, each observation should itself have values of i and j.
For a complete dataset, we have ij j; for an incomplete dataset, we have
ij j 0. Conditional on the kth failure at tik , the total number of survival items just
5.3 Nonparametric Estimation Methods for Cdf 75
3
2
1
0
-1 0 1 2 3 4 5 6
y
-2
-3
-4
-5
x
Fig. 5.2 WPP plot of Example 5.1
prior to time tik is n ik 1 and the number of survival items just after tik is n ik .
n ik
The conditional reliability is given by Rk . As such, the empirical cdf at
n ik 1
tik is estimated as
Y
j
Fj 1 Rk ; t 2 tij ; tij 1 : 5:9
k1
It is noted that Fj is a staircase function of t with Fj tij Fj tij1 . In other words,

the empirical cdf has a jump at tij .
Example 5.2 Consider the alternately censoring dataset shown in Table 5.3. The
empirical cdf evaluated by KaplanMeier method is shown in the fth column of
Table 5.5, and the corresponding WPP plot is displayed in Fig. 5.3.
5.3.3.2 NelsonAalen Method
Consider the following dataset:

tj dj sj;1 sj;2 . . . sj;kj \tj1 dj1 ; j 0; 1; . . .; m 5:10
with t0 d0 dm1 0 and tm1 1. There are dj failure observations at tj ,

there are kj censored observations over [tj ; tj1 ), and the last failure observations are
P
at tm . The sample size is n m j0 dj kj .
The NelsonAalen method estimates empirical cumulative hazard function (chf).
Consider the interval t 2 tk ; tk1 . The increment of the chf is given by
Ztk1
nf tDt dk
DHk rtdt r ttk1 tk : 5:11
nRt Nk
tk
Table 5.5 Empirical cdf for i ti di j KMM NAM MROM PEM

the data in Table 5.3
1 110 1 1 0.0476 0.0465 0.0325 0.0465
2 151 1 2 0.0952 0.0930 0.0786 0.0930
3 255 0
4 343 0
5 404 1 3 0.1485 0.1448 0.1305 0.1416
6 438 1 4 0.2017 0.1966 0.1825 0.1936
7 644 0
8 658 0
9 784 1 5 0.2631 0.2561 0.2419 0.2483
10 796 0
11 803 1 6 0.3301 0.3208 0.3064 0.3102
12 958 0
13 995 0
14 1000 1 7 0.4138 0.4006 0.3852 0.3774
15 1005 1 8 0.4976 0.4804 0.4639 0.4603
16 1146 1 9 0.5813 0.5601 0.5427 0.5431
17 1204 1 10 0.665 0.6399 0.6215 0.6259
18 1224 1 11 0.7488 0.7195 0.7003 0.7087
19 1342 1 12 0.8325 0.7990 0.7790 0.7913
20 1356 1 13 0.9163 0.8781 0.8577 0.8734
21 1657 1 14 1 0.9552 0.9362 0.9534
0
0 2 4 6 8
-1
y
-2
-3
-4
x
Fig. 5.3 WPP plots obtained from different methods for Example 5.2
Here, r t is interval average failure rate, f t is interval average density function,

is interval average reliability function, Nk is the number of items just prior to
Rt
time tk , and dk is the number of items failed at tk . The empirical chf is given by
X
j X
j
dk
Hj DHk ; t 2 tj ; tj1 : 5:12
k1 k1
Nk
The empirical chf is a staircase function of t with Htj Hj1 and Htj Hj .
As such, the empirical cdf is evaluated by

Ftij 1 eHj : 5:13
Example 5.2 (continued): For the data in Table 5.3, the empirical cdf evaluated by
NelsonAalen method is shown in the sixth column of Table 5.5 and the corre-
sponding WPP plot is also displayed in Fig. 5.3.
5.3.3.3 Mean Rank Order Method
Consider the data structure that KaplanMeier method adopts. Let rj denote the
rank order number of the jth failure observation. For a complete dataset, we have
rj ij ij1 1 rj1 1: 5:14
For an alternately censored dataset, the above relation is revised as below:
rj rj1 1 dj : 5:15
Here, dj ( 0) is the additional rank increment caused by right-censored observa-

tions before tij . Let kj denote the total number of equivalently right-censored data in
[tij1 ; tij ). Here, the word equivalently means that the total number is not necessary
to be the actual number of censored observations in that interval because it may
include the part from the censored data before tij1 . Let Nj be the total number of
items just prior to time tij , with Nj n ij 1. Namely, there are Nj 1 obser-
vations that are in interval (tij ; 1). The failure times of these Nj 1 items (both
observed and unobserved) divide the interval (tij ; 1) into Nj subintervals. Each of
the kj censored observations could fall into the interval (tij1 ; tij ) or one of the other
Nj subintervals if running those items to failure. Assume that the probability of a
censored observation falling into each of these Nj 1 intervals is the same. Then,
the average failure number in (tij1 ; tij ), resulted from these kj censored observations,
is given by
ki
dj : 5:16
Nj 1
Using Eq. (5.16) to Eq. (5.15) and noting that Nj kj n rj1 , we have
N j kj 1 n rj1 1
rj rj1 rj1 : 5:17
Nj 1 n ij 2
This is the mean rank order estimator. The empirical cdf at tij can be evaluated by
Eq. (5.8) with i being replaced by rj , or by
Fj betainv0:5; rj ; n rj 1: 5:18
Due to use of average and median, this estimator is more robust than Kaplan
Meier and NelsonAalen methods, where the cdf tends to be overestimated due to a
jump in failure number. This can be clearly seen from Table 5.5. As such, we
recommend using this estimator.
the mean rank order method is shown in the seventh column of Table 5.5 and the
corresponding WPP plot is also displayed in Fig. 5.3.
5.3.3.4 Piecewise Exponential Method
All the above three methods ignore information in the exact position of censored
observations. The piecewise exponential method considers this information and can
be viewed as an improvement to the NelsonAalen method.
Consider the dataset given by Eq. (5.10) and the time interval (tj1 ; tj ]. Let kj
denote the average failure rate in this interval. The chf at tj is given by
X
j
Hj kk tk tk1 ; t0 0: 5:19
k1
Similar to the case in the NelsonAalen method, the empirical chf is a staircase
function of t. The empirical cdf is given by Eq. (5.13).
The remaining problem is to specify the value of kj . We rst present the fol-
lowing relation here and its proof will be presented after we discuss the maximum
likelihood method of parameter estimation:
kj dj =TTTj 5:20
where
X
kj1
TTTj sj1;l tj1 Nj tj tj1 : 5:21
l1
From Eq. (5.21), it is clear that the information in the exact position of a censored
observation is included in the estimator of the empirical cdf.
the piecewise exponential method is shown in the last column of Table 5.5 and the
corresponding WPP plot is also displayed in Fig. 5.3. As seen, the estimated cdf
values are smaller than those from the NelsonAalen method and the KaplanMeier
method.
5.3.3.5 Discussion
As seen from Table 5.5, the estimates obtained from all the four methods are fairly
close to each other for this example. It is worthwhile noting that the increments of
the empirical cdf from t 1000 to 1005 are 0.0838, 0.0798, 0.0787, and 0.0829 for
the KMM, NAM, MROM, and PEM, respectively. For such a small interval, a
small increment is more reasonable. In this sense, the MROM really provides better
estimates.
In addition, according to the WPP plots shown in Fig. 5.3, the Weibull distri-
bution is not an appropriate model for tting the dataset in this example.
5.4 Parameter Estimation Methods
For a given set of data and a given parametric model, the parameter estimation deals
with determining the model parameters. There are several methods to estimate the
parameters and different methods produce different estimates. Typical parameter
estimation methods are
Graphical method,
Method of moments,
Maximum likelihood method,
Least square method, and
Expectation-maximum method.
5.4.1 Graphical Method
The graphical parameter estimation method is useful for model selection and can be
used to get the initial estimates of the model parameters. Generally, the graphical
method is associated with a probability plot, and different distribution can have
different probability plot(s). That is, the graphical method is distribution-specic.
Table 5.6 Estimated Weibull Method b g l

parameters for Example 5.1
Graphical 2.3523 106.10 94.02
Moment 2.7264 108.94 96.91
MLM 2.8226 108.49 96.63
LSM 3.0564 110.95 99.16
Average 96.68
In this subsection, we focus on the WPP plot. This is because the characteristics of
the WPP plots of many Weibull-related models have been studied (see Ref. [11]).
The graphical method starts with the nonparametric estimate of cdf. Once this is
done, the data pair (tj ; Ftj ) are transformed by Eq. (4.22). The WPP plot of the
data can be obtained by drawing yj versus xj .
For the two-parameter Weibull distribution, we can t the WPP plot of the data
into a straight line y a bx by regression. Comparing it with Eq. (4.23), we have
the graphical estimates of the Weibull parameters given by
b b; g ea=b : 5:22
To illustrate we consider Example 5.1. The coefcients of the regression straight

line are given by (a; b) = (10.9718, 2.3523). Using these in Eq. (5.22) yields the
graphical estimates of the Weibull parameters shown in the second row of
Table 5.6.
The graphical estimation methods for other Weibull-related distributions can be
found in Ref. [11].
5.4.2 Method of Moments
For a complete dataset, the rst two sample moments can be estimated as
1X n
1 X n
m1 ti ; s2 ti m1 2 : 5:23
n i1 n 1 i1
For a grouped data with the interval length Dt, under the assumption that data points
are uniformly distributed over each interval, the rst two sample moments can be
estimated as
X
n
ni 1X n
ni
m1 ti Dt=2 ; s2 ti m1 3 ti1 m1 3 : 5:24
i1
n 3 i1 nDt
5.4 Parameter Estimation Methods 81
On the other hand, the theoretic moments (e.g., l and r2 ) of a distribution are the
functions of the distributional parameters. The parameters can be estimated through
letting the theoretic moments equal the corresponding sample moments. This
method is termed the method of moments. It needs to solve an equation system
using an analytical or numerical method.
For a single-parameter model, we use the rst order moment (i.e., mean); for a
two-parameter model, we can use both the rst- and second-order moments (i.e.,
mean and variance). Clearly, the method of moments is applicable only for situa-
tions where the sample moments can be obtained. For example, this method is not
applicable for Example 5.2.
To illustrate, we look at Example 5.1. From Eq. (5.24), the rst two sample
moments are estimated as (m1 ; s 96:91; 38:3729). Assume that the Weibull
distribution is appropriate for tting the data. We need to solve the following
equation system:
gC1 1=b m1 ; g2 C1 2=b s2 m21 : 5:25
Using Solver of Microsoft Excel we obtained the solution of Eq. (5.25) shown in the
third row of Table 5.6.
5.4.3 Maximum Likelihood Method
Let h denote the parameter set of a distribution function Ft. The likelihood
function of an observation is dened as follows:
Lt f t; h for a failure observation t,
Lt Ft; h for a left-censoring observation t ,
Lt Rt; h for a right-censoring observation t ,
Lt Fb; h Fa; h for an interval observation t 2 a; b.
For a given dataset, the overall likelihood function is given by
Y
n
Lh Li h 5:26
i1
where Li h is the likelihood function of the ith observation, and depends on the
distributional parameters. The maximum likelihood method (MLM) is based on the
idea that if an event occurs in a single sampling it should have the greatest prob-
ability. As such, the parameter set is determined by maximizing the overall like-
lihood function given by Eq. (5.26) or its logarithm given by
X
n
lnLh lnLi h: 5:27
i1
Compared with Eq. (5.26), Eq. (5.27) is preferred since Lh is usually very small.
The maximum likelihood estimates (MLE) of the parameters can be obtained using
Solver to directly maximize lnLh.
The MLM has a sound theoretical basis and is suitable for various data types. Its
major limitation is that the MLEs of the parameters may be nonexistent for dis-
tribution with the location parameter as the lower or upper limit of the lifetime. In
this case, the maximum spacing method can be used to estimate the parameters for a
complete dataset without ties (see Ref. [5]). For an incomplete dataset or a complete
dataset with ties, one needs to use its variants (see Ref. [7] and the literature cited
therein).
Using the MLM to t the Weibull distribution to the dataset of Example 5.1, we
obtained the estimated parameters shown in the fourth row of Table 5.6. The last
row of Table 5.6 shows the average of the mean life estimates obtained from
different estimation methods. As seen, the mean life estimate obtained from the
MLM is closest to this average.
We now prove Eq. (5.20) using the MLM. Consider the time interval (tj1 ; tj ].
There are totally Mj1 survival items just after tj1 ; there are kj1 censored
observations in this interval; their values are (sj1;l ; 1 l kj1 ), which meet
tj1 sj1;l \tj ; there are dj failure observations at tj , and the other Mj1 kj1 dj
observations are larger than tj . Assume that Ft can be approximated by the
exponential distribution with failure rate kj . To derive the overall likelihood
function, we divide the observations that have survived to tj1 into three parts:
censored observation in interval (tj1 ; tj ), failure observations at tj , and the other
observations that are larger than tj . Their likelihood functions are given, respec-
tively, by
X
kj1 X
kj1
lnL1 lnekj sj1;l tj1 kj sj1;l tj1 ;
l1 l1
lnL2 dj lnkj ekj tj tj1 dj lnkj dj kj tj tj1 ; and

kj tj tj1
lnL3 Mj1 kj dj lne kj Mj1 kj dj tj tj1 :
The total likelihood is given by
X
3
lnL lnLi dj lnkj kj TTTj : 5:28
i1
Letting d lnL=dkj 0, we have kj dj =TTTj , which is Eq. (5.20).

5.4 Parameter Estimation Methods 83
5.4.4 Least Square Method
The least square method (LSM) is a curve tting technique. Similar to the graphical
method, it needs the nonparametric estimate of cdf. Let Fj be the nonparametric
estimate at a failure time tj ; 1 j m, and Ft; h denote the cdf to be tted. The
parameters are estimated by minimizing SSE given by
X
m
SSE Ftj ; h Fj 2 : 5:29
j1
The least square estimates of the parameters for Example 5.1 are shown in the
fth row of Table 5.6.
5.4.5 Expectation-Maximum Method
The expectation-maximum method (EMM) is applicable for the incomplete data

case. It uses an iterative process to estimate the parameters. The method includes
two steps. The rst step is called the Expectation step. Given initial values of the
parameters, the expected value of a censoring observation can be computed. For
example, for a right-censoring observation t , the expected life (i.e., the mean life
with censoring) is given by
t t MRLt 5:30
where MRLt is the mean residual life function evaluated at t . Using t to replace
t , the incomplete dataset is transformed into an equivalently complete dataset.
The second step is the Maximum-step. It applies the MLM to the equivalently
complete dataset to estimate the parameters. After that, the expected life of a
censoring observation is updated using the new estimates of the parameters. The
process is repeated until convergence. Using an Excel spreadsheet program, the
iterative process can be completed conveniently.
Since the expectation step reduces the randomness of the censored data, the
model tted by this method tends to have smaller dispersion (i.e., overestimating b
for the Weibull distribution).
5.5 Hypothesis Testing
A statistical hypothesis test is a method using observed samples to draw a statistical

conclusion. Generally, it involves a null hypotheses and an alternative hypothesis
about the distributions of the observations or about some statistical property (e.g.,
trend or independence). A test statistic is dened based on the null hypothesis, and
then the distribution of the test statistic is derived. A probability threshold (or
signicance level) is selected, which is commonly 5 or 1 %. The null hypothesis is
either rejected or not rejected by comparing the observed value of the test statistic
with the selected threshold value, or by comparing the p-value of the test statistic
with the selected signicance level. More details about hypothesis test can be found
from statistics books (e.g., see Ref. [11]).
In this section we focus on the hypothesis testing for the goodness of t of a
tted distribution. We introduce two simple and popular tests: the chi square test
and the KolmogorovSmirnov test. The former requires that the sample size is
large, while the latter does not have such a limitation.
5.5.1 Chi Square Test
When the sample size n is not small and the dataset is given in the form of interval
data or can be transformed into interval data, the goodness of t of a tted model
(with m parameters) can be evaluated using the chi square statistic given by
X
k
ni Ei 2
v2 ; Ei nFti Fti1 : 5:31
i1
Ei
The smaller the v2 is, the better the tted model is. The goodness of t can be
measured by the p-value given by pv PrfQ [ v2 g, where Q is a chi-squared
random variable with degree of freedom k 1 m. The larger the p-value is (i.e.,
v2 is small), the better the goodness of t is. To accept a tted model, we usually
require pv 0:10:3 (e.g., see Ref. [2]). It is noted that this range is much larger
than the commonly used signicance level (0.01 or 0.05).
To illustrate, we look at Example 5.1. We merge the last two intervals to make
nk larger than or close to 5. In this case, k 9; m 2; v2 24:42 and
pv 0:0004. Since pv 0:1, we conclude that the Weibull distribution is not an
appropriate model for tting the rst bus-motor failure dataset.
Example 5.3 Consider the rst bus-motor failure dataset. It was mentioned earlier
that the bus-motor major failure is due to two failure modes: serious accidents and
performance deterioration. This implies that the twofold Weibull competing risk
model given by Eq. (4.31) can be appropriate for tting this dataset. The MLEs of
the model parameters are shown in the second row of Table 5.7. By merging the last
5.5 Hypothesis Testing 85
Table 5.7 Parameters of the twofold competing risk models for Example 5.3
b1 g1 b2 g2 ln L m AIC
Model 0 1.2939 279.84 4.2104 122.27 384.962 4 1166.887
Model 1 1 530.11 3.9341 118.66 385.182 3 1164.545
two intervals we have k 9; m 4; v2 1:2419 and pv 0:8716. As a result, the

twofold Weibull competing risk model is an appropriate model for tting this
dataset.
5.5.2 KolmogorovSmirnov Test
Consider a complete dataset. The KolmogorovSmirnov statistic is the maximum

difference between the empirical cdf and theoretical cdf given by
i i1
Dn max fmaxjFt j; jFt jg: 5:32
1in n n
If the sample comes from distribution Ft, then Dn will be sufciently small. The
null hypothesis is rejected at level a if Dn [ ka , where ka is the critical value at
signicance level of a. The critical value of the test statistic can be approximated by

a b
kc p 1 c 5:33
n n
The coefcient set (a; b; c) is given in Table 5.8; and the relative error (e; %) is
shown in Fig. 5.4. As seen, the relative error is smaller than 0.7 % for n 5.
5.6 Model Selection
Often one considers more than one candidate model and chooses the best one from
the tted models based on some criterion. Determination of candidate models can
be based on failure mechanism, experience, or graphical approach.
The numbers of parameters of candidate models can be the same or different. If
the numbers of parameters of candidate models are the same, we can directly
compare the performance measures of the tted models. The performance measure
is the logarithm maximum likelihood value if the parameters are estimated by
MLM; it is the sum of squared errors if the parameters are estimated by LSM. The
selection will give the model with largest log-likelihood value or the smallest sum
of squared errors.
Table 5.8 Coefcients of a 0.1 0.05 0.01

(5.33)
a 1.224 1.358 1.628
b 0.2057 0.2593 0.3753
c 0.6387 0.7479 0.8858
0.7
0.6
0.5
0.4
, %
0.3 =0.05 =0.1

=0.01
0.2
0.1
0
0 10 20 30 40
n
Fig. 5.4 Relative errors of Eq. (5.33)
If the numbers of parameters of candidate models are different, the performance-

based criterion is no longer appropriate since it favors the model with more model
parameters and hence results in possible over-tting. In this case, we need look at
other criteria. We introduce such two criteria as follows.
5.6.1 Likelihood Ratio Test
Suppose that there are two candidate models (denoted as Model 0 and Model 1,
respectively). Model 1 is a special case of Model 0. Namely, Model 1 is nested
within Model 0. For example, the exponential distribution is a special case of the
Weibull distribution or gamma distribution. The MLM is used to t the two can-
didate models to a given dataset. Let lnL0 and lnL1 denote their log-likelihood
values, respectively. The test statistic is given by:
D 2lnL0 lnL1 : 5:34
Model 0 is preferred if D is sufciently large. In many cases, the probability

distribution of D can be approximated by a chi-square distribution with degree of
freedom m0 m1 , where m0 and m1 are the numbers of parameters of Models 0 and
1, respectively. Model 0 is accepted if D is sufciently large or pv is sufciently
small (e.g., pv \0:1).
To illustrate, we look at Example 5.3. Model 0 is the twofold Weibull competing
risk model. Noting that in Model 0 b1 (=1.2939) is close to 1, it may be appropriate
5.6 Model Selection 87
to approximate F1 t by an exponential distribution. As such, Model 1 can be an

exponential-Weibull competing risk model.
The MLEs of the parameters of Model 1 are shown in the last row of Table 5.7.
Using the results in the sixth column of Table 5.7 to Eq. (5.34), we have
D 0:4384, which corresponds to pv 0:5079, which is much larger than 0.1. As
a result, the exponential-Weibull competing risk model is accepted.
The tted exponential-Weibull competing risk model has an appropriate physical
interpretation: the exponential distribution represents the serious accidents and the
Weibull distribution represents the performance deterioration.
5.6.2 Information Criterion
The information criterion is appropriate for model selection when the candidate
models are either nested or non-nested. A statistical model should have an appro-
priate tradeoff between the model simplicity and goodness of t. The Akaike
information criterion [1] incorporates these two concerns through giving a penalty
for extra model parameters to avoid possible over-tting. In terms of log-likelihood,
the Akaike information criterion (AIC) is dened as below:
AIC 2 lnL 2m: 5:35
Smaller AIC implies a better model. As such, the best model is the one with the
smallest AIC.
The AIC given by Eq. (5.35) is applicable for the cases where the sample size is
large and m is comparatively small. If the sample size is small relative to m, the
penalty given by AIC is not enough and several modications have been proposed
in the literature (e.g., see Refs. [3, 6]).
To illustrate, we look at Example 5.3. The values of AIC for the two candidate
models are shown in the last column of Table 5.7. The exponential-Weibull
competing risk model has a smaller AIC and hence is preferred. This is consistent
with the conclusion obtained from the likelihood ratio test.
References
1. Akaike H (1974) A new look at the statistical model identication. IEEE Trans Autom Control
19(6):716723
New York
3. Burnham KP, Anderson DR (2004) Multimodel inference understanding AIC and BIC in
model selection. Sociol Methods Res 33(2):261304
4. Davis DJ (1952) An analysis of some failure data. J Am Stat Assoc 47(258):113150
5. Ekstrm M (2008) Alternatives to maximum likelihood estimation based on spacings and the
KullbackLeibler divergence. J Stat Plan Infer 138(6):17781791
6. Hurvich CM, Tsai CL (1989) Regression and time series model selection in small samples.
Biometrika 76(2):297307
7. Jiang R (2013) A new bathtub curve model with a nite support. Reliab Eng Syst Saf 119:44
51
8. Kaplan EL, Meier P (1958) Nonparametric estimation from incomplete observations. J Am
Stat Assoc 53(282):457481
9. Kim JS, Proschan F (1991) Piecewise exponential estimator of the survivor function. IEEE
Trans Reliab 40(2):134139
10. Lehmann EL, Romano JP (2005) Testing statistical hypotheses, 3E edn. Springer, New York
11. Murthy DNP, Xie M, Jiang R (2003) Weibull models. Wiley, New York
12. Nelson W (1982) Applied life data analysis. Wiley, New York
Chapter 6
Reliability Modeling of Repairable
Systems
6.1 Introduction
Most of the models presented in Chaps. 3 and 4 are univariate life distributions.
Such models are suitable for modeling an i.i.d. random variable (e.g., time to the
rst failure), and represent the average behavior of the populations reliability
characteristics.
A repairable system can fail several times since the failed system can be restored
to its operating condition through corrective maintenance actions. If the repair time
is neglected, the times to failure form a failure point process. The time between the
i 1th failure and the ith failure, Xi , is a continuous random variable. Depending
on the effect of the maintenance actions, the inter-failure times Xi s are generally not
i.i.d. As such, we need new models and methods for modeling the failure process.
This chapter focuses on such models and methods.
There are two categories of models for modeling a failure process. In the rst
category of models, the underlying random variable is Nt, which is the number of
failures by t; and in the second category of models, the underlying random variable
P
is Xi or Ti ij1 Xj , which is the time to the ith failure. We call the rst category
of models the discrete models (which are actually counting process models) and the
second category of models the continuous models (which are actually variable-
parameter distribution models).
The model and method for modeling a given failure process depend on whether
or not the inter-failure times have a trend. As such, the trend analysis for a failure
process plays a fundamental role in reliability analysis of repairable systems. When
the trend analysis indicates that there is no trend for a set of inter-failure times, a
further test for their randomness is needed.
This chapter is organized as follows. We rst look at the failure counting
process models in Sect. 6.2, and then look at the distribution models in Sect. 6.3.

90 6 Reliability Modeling of Repairable Systems
A multi-step procedure for modeling failure processes is presented in Sect. 6.4.

Tests for trend are discussed in Sect. 6.5, and tests for randomness are discussed in
Sect. 6.6. Finally, we briefly introduce tests for normality and constant variance in
Sect. 6.7.
6.2 Failure Counting Process Models
A point process is a continuous time stochastic process characterized by events

(e.g., failures) that occur randomly. Let Nt denote the cumulative failure number
in time interval (0, t), which is a discrete random variable. A failure point process
satises the following:
N0 0;
Nt is nonnegative integer;
Nt is nondecreasing; and
for s \ t, Nt Ns is the number of failures in s; t.
Three typical counting processes are renewal process (RP), homogeneous
Poisson process (HPP), and nonhomogeneous Poisson process (NHPP). We briefly
outline them below.
6.2.1 Renewal Process
A failure counting process is a renewal process if the inter-failure times Xi s are a

sequence of i.i.d. random variables with distribution function Fx. The expected
number of renewals in 0; t is called renewal function and is given by
Zt
Mt Ft Mt xf xdx: 6:1
0
When t l, a well-known asymptotic relation for the renewal function is

" 2 #
t r
Mt 0:5 1 6:2
l l
where l and r are the mean and standard deviation of the inter-failure time. The
variance of Nt is given by
X
1
Vt 2n 1F n t Mt2 6:3
n1
where F n t is the n-fold convolution of Ft with itself.

6.2 Failure Counting Process Models 91
For a repairable system, a renewal process assumes that the system is returned to
an as new condition every time it is repaired. As such, the distribution of Xi is the
same as the distribution of X1 . For a multi-components series system, if each
component is replaced by a new one when it fails, then the system failure process is
a superposed renewal process. In general, a superposed renewal process is not a
renewal process. In fact, it is close to a minimal repair process when the number of
components is large.
6.2.2 Homogeneous Poisson Process
If the times between failures are independent and identically exponentially dis-
tributed, the renewal process reduces into a homogeneous Poisson process (also
termed as stationary Poisson process). In this case, Nt follows a Poisson distri-
bution with the Poisson parameter kt, where k is failure intensity.
6.2.3 Nonhomogeneous Poisson Process
The NHPP is also termed as nonstationary Poisson process. Its increment,

DN Nt s Nt; s [ 0, follows the Poisson distribution with mean
Mt; s Mt s Mt. When s is small, the NHPP meets the following
relation:
PrNt s Nt 1 mts 6:4
where mt is called the failure intensity function.

The NHPP arises when a complex system is subjected to a minimal repair
process. Let Ft be the distribution of time to the rst failure. Then Nt follows a
Poisson distribution with the Poisson parameter given by the chf
Ht ln1 Ft. The NHPP model is generally suitable for the purpose of
modeling data with a trend.
Especially, when Ft is the Weibull distribution, we obtain the well-known
power-law model given by
b
t
Mt Ht 6:5
g
where Mt ENt is the mean cumulative function. In this model, b provides

the following information:
if b 1, the failure arrivals follow a homogeneous Poisson process;
if b [ 1, the system deteriorates with time; and
if b \ 1, the system improves with time.
Depending on the time origin and observation window, the power-law model
can have two variants. If we begin the failure counting process at t d (either
known or unknown) and set this time as the time origin, then Eq. (6.5) can be
revised as
b
td b d
Mt : 6:6
g g
If d is unknown, Eq. (6.6) has three parameters.

Consider the case where the time is recorded with date and the time when an
item begins working is unknown. We set a certain date as the time origin and revise
Eq. (6.5) as

t t0 b
Mt 6:7
g
where t0 is an unknown parameter to be estimated.

The power-law model can be extended to a multivariate case so as to reflect the
influences of various factors (e.g., operational environment and maintenance his-
tory) on reliability. Such an extension is the proportional intensity model (see Refs.
[5, 7] and the literature cited therein). A proportional intensity model consists of
two parts: baseline part and covariate part. The power-law model is usually used as
the baseline intensity function.
6.2.4 Empirical Mean Cumulative Function
Suppose we have several failure point processes that come from nominally identical
systems with different observation windows 0; Ti ; 1 i n. Arrange all the
failure data in ascending order. The ordered data are denoted as
t1 s1 t2 s2 tJ sJ T maxTi ; 1 i n 6:8
where tj s are failure times (i.e., not including censored times) and sj is the number
of the systems under observation at tj . The nonparametric estimate of the MCF is
given by
M t0 0 0; M tj M tj1 1=sj ; 1 j J: 6:9
M t at tj has a jump, i.e., M tj M tj1 \ M tj . To smooth, we dene

the representative value of the MCF at tj as:
Mtj M tj1 M tj =2: 6:10
We call the MCF given by Eq. (6.10) the empirical MCF.

6.2 Failure Counting Process Models 93
For a given theoretical model Mh t such as the power-law model, the parameter
set h can be estimated by the MLM or LSM. The LSM is simple and estimates the
parameters by minimizing the sum of squared errors given by:
X
m
SSE Mh tj Mtj 2 : 6:11
j1
6.3 Distribution Models for Modeling Failure Processes
We consider three categories of models that can be used for modeling failure
processes in different situations. They are:
Ordinary life distribution models;
Imperfect maintenance models; and
Distribution models with the parameters varying with the numbers of failures or
system age.
We briefly discuss them below.
6.3.1 Ordinary Life Distribution Models
Ordinary life distribution models can be used to model the renewal process and
minimal repair process. When each failure is corrected by a replacement or perfect
repair, times to failure form a renewal process, whose inter-failure times are i.i.d.
random variables and hence can be modeled by an ordinary life distribution.
When each failure is corrected by a minimal repair, times to failure form a
minimal repair process. After a minimal repair completed at age t, the time to the
next failure follows the conditional distribution of the underlying distribution (i.e.,
the distribution of X1 T1 ). This implies that the distribution of inter-failure times
can be expressed in terms of the underlying distribution though they are not i.i.d.
random variables.
When each failure is corrected by either a replacement or a minimal repair, the
inter-failure times can be modeled by a statistical distribution. Brown and Proschan
[2] develop such a model. Here, the item is returned to the good-as-new state with
probability p and to the bad-as-old state with probability q 1 p. The parameter
p can be constant or time-varying. The process reduces into the renewal process
when p 1 and into the minimal repair process when p 0.
6.3.2 Imperfect Maintenance Models
When each failure is corrected by an imperfect maintenance, the time to the next
failure depends on the effects of prior maintenance actions. As such, the ordinary
life distribution is no longer applicable, and a category of imperfect maintenance
models can be used for modeling subsequent failures.
Preventive maintenance (PM) aims to maintain a working item in a satisfactory
condition. The PM is often imperfect, whose effect is in between the perfect
maintenance and minimal maintenance. As such, the effect of PM can be repre-
sented by an imperfect maintenance model.
There are a large number of imperfect maintenance models in the literature, and
Pham and Wang [11] present a review on imperfect maintenance models, and Wu
[16] provides a comprehensive review on the PM models (which are actually
imperfect maintenance models). Several typical imperfect maintenance models will
be presented in Chap. 16.
6.3.3 Variable-Parameter Distribution Models
This category of models assumes that Xi s can be represented by the same life
distribution family Fx; hi with the parameter set hi being functions of i or ti .
Clearly, when hi is independent of i or ti , the model reduces into an ordinary
distribution model.
For the bus-motor data shown in Table 5.2, Jiang [4] presents a normal variable-
parameter model, whose parameters vary with i; and Jiang [6] presents a Weibull
variable-parameter model, whose parameters are also functions of i. The main
advantage of these models is that they can be used to infer the life distribution after
a future failure.
6.4 A Procedure for Modeling Failure Processes
In this section, we present a multi-step procedure for modeling a failure point

process. Before presenting the procedure, we rst look at a numerical example to
illustrate the necessity of such a procedure.
6.4.1 An Illustration
Example 6.1 The data shown in Table 6.1 come from Ref. [12] and deal with
failure times (in 1000 h) of a repairable component in a manufacturing system.
6.4 A Procedure for Modeling Failure Processes 95
Table 6.1 A failure point process

i 1 2 3 4 5 6 7 8 9 10 11 12
xi 0.673 0.983 1.567 2.349 3.314 1.786 1.745 2.234 0.987 1.756 2.567 2.163
ti 0.673 1.656 3.223 5.572 8.886 10.672 12.417 14.651 15.638 17.394 19.961 22.124
Table 6.2 Maximum Assumption RP NHPP Nonstationary

likelihood estimates the
Weibull parameters b 2.8182 0.9440 3.4396
g or g1 2.0732 1.5911 2.4169
d 2.3609
lnL 12.7819 19.3208 9.6293
AIC 29.5638 25.2587
Under the assumption that the times to failure form an RP with the underlying
distribution being the Weibull distribution, we obtained the MLEs of the parameters
shown in the second column of Table 6.2.
Under the NHPP assumption with the MCF given by Eq. (6.5) (i.e., the power-
law model), we obtained the MLEs of the parameters shown in the third column of
Table 6.2 (for the MLE of the power-law model, see Sect. 11.5.3.1). The empirical
and tted MCFs are shown in Fig. 6.1.
From Table 6.2 and Fig. 6.1, we have the following observations:
the parameters of the tted models are signicantly different, but
the plots of Mt are close to each other.
A question is which model we should use. The answer to this question depends
on the appropriateness of the assumption for the failure process. This deals with
testing whether the failure process is stationary and whether the inter-failure times
are i.i.d. Such tests are called test for trend and test for randomness, respectively. As
a result, a procedure is needed to combine these tests to the modeling process.
6.4.2 Modeling Procedure
Modeling a failure point process involves a multi-step procedure. Specic steps are
outlined as follows.
Step 1: Draw the plot of the MCF of data and other plots (e.g., running arith-
metic average plot, which will be presented later). If the plots indicate that the trend
is obvious, implement Step 3; otherwise, implement Step 2.
14
12
10
8
M(t )
6
NHPP
4
Asymptotic RF
2
0
0 5 10 15 20 25
t
Fig. 6.1 Empirical and tted MCFs
Step 2: If the trend is not very obvious, carry out one or more tests for
stationarity to further check for trend. If no trend is conrmed, a further test for i.i.d.
assumption needs to carry out. If the i.i.d. assumption is conrmed, the data can be
modeled by an appropriate life distribution model.
Step 3: This step is implemented when the inter-failure times have a trend or
they are not i.i.d. In this case, the data should be modeled using nonstationary
models such as the power-law model, variable-parameter models, or the like.
6.5 Tests for Stationarity
Stationarity is time invariance of the data. For example, inter-failure times in

repairable systems undergoing reliability growth testing usually increases with time
statistically and inter-failure times in repairable systems in service can be
decreasing with time statistically. If such trends do not exist, the failure process of
the system is stationary. The objective of stationarity tests is to determine whether
the pattern of failures is signicantly changing with time so as to select appropriate
models for modeling the data.
When a repairable system is repaired to good-as-new condition following each
failure, then the failure process can be viewed as an RP. For an RP, the times
between failures are i.i.d. As mentioned earlier, the HPP is a special RP where inter-
failure times are i.i.d. exponential random variables.
In reliability trend tests, the null hypothesis H0 is that the underlying process
of the interarrival times is stationary. Since both RP and HPP are stationary pro-
cesses, trend tests can be divided into two categories: HPP null hypothesis and RP
null hypothesis. When the null hypothesis is HPP, rejecting H0 just implies that the
process does not follow an HPP, and does not necessarily imply that there exists a
trend in the process. However, if the null hypothesis is RP, rejecting H0 does imply
6.5 Tests for Stationarity 97
that there exists a trend in the process. However, when we cannot reject the null
hypothesis at the given level of signicance, it does not necessarily imply that we
accept the null hypothesis unless the test has a particularly high power (which is the
probability of correctly rejecting the null hypothesis given that it is false [10]). This
is because the conclusion is made based on the assumption that the null hypothesis
is true and depends on the signicance level (which is the probability the null
hypothesis to be rejected assumed that it is true [10]), whose value is commonly
small (0.05 or 0.01).
In this section, we present several tests for stationarity. We will use the data
shown in Table 6.1 to illustrate each test.
6.5.1 Graphical Methods
A plot of data helps get a rough impression for trend before conducting a quanti-
tative trend test. Such a plot is the empirical MCF. If the process is stationary, the
plot of empirical MCF is approximately a straight line through the origin.
Another useful plot is the plot of the running arithmetic average. Consider a set
P
of inter-failure times xi ; 1 i n. Let ti ij1 xj . The running arithmetic
average is dened as below:
ri ti =i; i 1; 2; . . .: 6:12
If the running arithmetic average increases as the failure number increases, the
time between failures is increasing, implying that the systems reliability gets
improved with time. Conversely, if the running arithmetic average decreases with
the failure number, the average time between failures is decreasing, implying that
the systems reliability deteriorates with time. In other words, if the process is
stationary, the plot of running arithmetic average is approximately a horizon line.
Figure 6.2 shows the plot of running arithmetic average for the data in Table 6.1.
As seen, the reliability gets improved at the beginning and then becomes stationary.
For this case, one could implement the second step or directly go to the third step.
6.5.2 Tests with HPP Null Hypothesis
Tests with HPP null hypothesis include Crow test, Laplace test, and Anderson-
Darling test.
1.5
r (i )
1
0.5
0
0 2 4 6 8 10 12 14
i
Fig. 6.2 Plot of running arithmetic average
6.5.2.1 Crow Test
This test is developed by Crow [3] and is based on the power-law model given by
Eq. (6.5). When b 1, the failure process follows an HPP. As such, the test
involves whether an estimate of b is signicantly different from 1. The null
hypothesis is b 1 and the alternative hypothesis is b 6 1.
For one system on test, the maximum likelihood estimate of b is
X
n
^ n=
b lnT=ti 6:13
i1
where n is the number of observed failures and T is the censored time, which can be
^ follows a chi-squared distribution
larger than or equal to tn . The test statistic 2n=b
with the degree of freedom of 2n. The rejection criterion for null hypothesis H0 is
given by
^ \ v2
2n=b ^
2n;1a=2 or 2n=b [ v2n;a=2 6:14
2
where v2k;p is the inverse of the one-tailed probability of the chi-squared distribution
associated with probability p and degree of freedom k.
Example 6.2 Test the stationarity of the data in Table 6.1 using the Crow test.
From Eq. (6.13), we have b ^ 0:9440 and 2n=b ^ 25:423. For signicant level
a 0:05, v22n;a=2 39:364 and v22n;1a=2 12:401. As a result, we cannot reject H0 .
6.5.2.2 Laplace Test
The alternative hypothesis of this test is the NHPP. Conditioning on tn , ti s

1 i n 1 are uniformly distributed on (0, tn ). Let
X
n1
U ti : 6:15
i1
The mean and variance of U are given by
lU tn n 1=2; r2U tn2 n 1=12: 6:16
The test statistic is the standard normal score given by Z U lU =rU . For large
n, Z approximately follows a standard normal distribution. The rejection criterion
for H0 is given by
Z \ za=2 or Z [ z1a=2 : 6:17
Example 6.2 (continued) Test the stationarity of the data in Table 6.1 using the
Laplace test.
From Eqs. (6.15) and (6.16), we have U 132.687, lU 121.682, rU
21.182, and Z 0.5280. For a 0:05, za=2 z1a=2 1:96. As a result, we
cannot reject H0 .
6.5.2.3 AndersonDarling Test
The AndersonDarling test for trend is based on the AndersonDarling test statistic
given by (see Ref. [8]):
1X n0 h t tn 1i i
i
AD n0 2i 1 ln ln 1 0 6:18
n0 i1 T T
where T is the censored time of observation process; n0 n if the process is

censored at time T [ tn , and n0 n 1 if the process is censored at time T tn .
The null hypothesis is the HPP and the alternative hypothesis is the NHPP. It is
one-sided and the null hypothesis is rejected if AD is greater than the critical value.
The asymptotic critical value is shown in the second column of Table 6.3.
AndersonDarling test.
Table 6.3 Asymptotical Signicant level (%) AD LR GAD

critical value of test statistic
5 2.49 1.65 2.49
1 3.86 2.33 3.86
From Eq. (6.18), we have AD 0:2981 and hence the null hypothesis is not
rejected for a 0:05.
6.5.3 Tests with RP Null Hypothesis
Tests with RP null hypothesis include Mann test, LewisRobinson test, and gen-
eralized AndersonDarling test.
6.5.3.1 Mann Test
This test is presented in Ref. [1] and is sometimes called reverse arrangement test or
pairwise comparison nonparametric test (see Refs. [14, 15]). The null hypothesis is
renewal process and the alternative hypothesis is nonrenewal process. The test
needs to compare all the interarrival times xj and xi for j [ i. Let uij 1 if xj [ xi ;
otherwise uij 0. The number of reversals of the data is given by
X
U uij : 6:19
i\j
Too many reversals indicate an increasing trend, too few reversals imply a
decreasing trend, and there is no trend if the number of reversals is neither large nor
small.
Under H0 , the mean and variance of U are given, respectively, by
lU nn 1=4; r2U n 2:5nn 1=36: 6:20
The test statistic is the standard normal score given by Z U lU =rU . For large
n (e.g., n 10), Z approximately follows a standard normal distribution. The
rejection criterion for H0 is given by Eq. (6.17).
Mann test.
From Eqs. (6.19) and (6.20), we have U 44, lU 33, rU 7:2915 and
Z 1:5086. As a result, we cannot reject H0 for a 0:05.
6.5.3.2 LewisRobinson Test
LewisRobinson test is a modication of Laplace test. The null hypothesis is

renewal process and the alternative hypothesis is nonrenewal process. The Lewis
Robinson test statistics is LR Z=CV, where Z is the standard normal score of the
Laplace test statistic and CV is the coefcient of variation for the observed inter-
arrival times. The critical value for rejecting H0 is shown in the third column of
Table 6.3.
LewisRobinson test.
Using the approach outlined above, we have Z 0.5280, CV 0.4051 and
LR 1.3034. As a result, we still cannot reject H0 for a 0:05.
6.5.3.3 Generalized AndersonDarling Test
The test statistic of the generalized AndersonDarling test is given by (see Ref. [8])
n
n 4x2 X i 2 1 1 2
GAD 2
qi ln qi ri ln 1 ri 6:21
r2 i1
i1 ni n
where
nxi 1 Xn1
qi ti ixi =tn ; ri 1; and r2 xi1 xi 2
tn 2n 1 i1
with

i 2 1
q2i ln j 0; qi ri ln 1 j 0:
i 1 i1 n i in
It is one-sided and the null hypothesis is rejected if GAD is greater than the critical
value, which is shown in the last column of Table 6.3.
AndersonDarling test.
From Eq. (6.21), we have GAD 1:3826 and hence cannot reject the null
hypothesis for a 0:05.
6.5.4 Performances of Trend Tests
The performances of the tests discussed above have been studied (see Refs. [8, 9,
15]), and the results are summarized in Table 6.4. It is noted that no test provides
very good performance for the decreasing case.
Table 6.4 Summary of performance of trend tests

Test Null Decreasing trend Increasing trend Bathtub trend
hypothesis case case case
Crow HPP Very good
Laplace HPP Good
Anderson HPP Good
Darling
Mann RP Good
Lewis RP Good
Robinson
Generalized RP Very good Very good
AD
6.6 Tests for Randomness
Randomness means that the data are not deterministic and/or periodic. Tests for
randomness fall into two categories: nonparametric methods and parametric
methods. In this section, we focus on nonparametric methods.
6.6.1 Runs Above and Below Median Test
Consider a sequence of n observations of a random variable X. Each observation is

classied into one of two categories: plus (or 1) and minus (or 0). A run is dened
as a sequence of identical observations that are different from the observation before
or/and after this run. Both the number of runs and their lengths can be used as
measures of the randomness of the sequence. Too few runs mean that some runs are
too long and too many runs result in short runs. As such, we only need to consider
the total number of runs.
In the run test, all the observations xi ; 1 i n are compared with the median
x0:5 . To specify the median, we arrange the data in ascending order, i.e.,
xj xj1 ; 1 j n 1: 6:22
If n is an odd number, the median of the data is given by x0:5 xn1=2 ; if n is an

even number, the median of the data is given by x0:5 xn=2 xn=21 =2.
To calculate the number of runs above and below the median, we compare each
observation with the median. Let ri Ixi [ x0:5 , where Ixi [ x0:5 1 if
xi [ x0:5 and Ixi [ x0:5 0 if xi \ x0:5 . Let M0 denote the number of times of
sign change from 1 to 0 or from 0 to 1. The total number of runs above and
below the median is equal to M M0 1.
6.6 Tests for Randomness 103
Under the null hypothesis that the data are random, the number of runs r is a
discrete random variable with mean and variance given by:
2n1 n n1 2n1 n n1 n
lM 1; r2M lM 1 : 6:23
n nn 1
P
where n1 ni1 ri . For n 10, r approximately follows the normal distribution.
The test statistic is the standard normal score given by
Z M lM =rM : 6:24
The critical values with signicance level a is given by
za=2 U1 a=2; z1a=2 U1 1 a=2: 6:25
The null hypothesis H0 is not rejected if Z 2 za=2 ; z1a=2 ; otherwise, rejected.

Example 6.3 Test the randomness of the data in Table 6.1 using the runs above and
below median test.
For this example, the median x0:5 1:771. The values of ri are shown in the
third column of Table 6.5. As seen, M0 5 and hence M 6. From Eqs. (6.23)
and (6.24), we have lM 7, rM 1:6514 and Z 0:6055. When a 0:05,
za=2 1:96 and z1a=2 1:96. As a result, the null hypothesis is not rejected.
Table 6.5 Tests for i xi ri Si Runs of Si

randomness
1 0.673 1 1 1
2 0.983 1 1 1
3 1.567 1 1 1
4 2.349 0 1 1
5 3.314 0 1 2
6 1.786 0 1 2
7 1.745 1 1 3
8 2.234 0 1 4
9 0.987 1 1 5
10 1.756 1 1 5
11 2.567 0 1 6
12 2.163 0
Median 1.771 M=6 P=7 R=6
6.6.2 Sign Test
Let Si signxi1 xi ; 1 i n 1, and m denote the number of nonzero Si . As

such, Si forms m Bernoulli tests. Let P denote the number of times of Si 1. Under
the null hypothesis that the data come from a random process, we expect that there
are roughly equal numbers for both positive and negative signs. The number of
positive signs P converges weakly to the normal distribution with mean and vari-
ance given by:
lP m=2; r2P m=12: 6:26
The test statistic is Z P lP =rP , and the critical value is given by Eq. (6.25).
Example 6.3 (continued) Test the randomness of the data in Table 6.1 using the
sign test.
The values of Si are shown in the fourth column of Table 6.5. From these values,
we have m 11, P 7 and Z 1:5667. As a result, the null hypothesis is not
rejected at the signicance level of 5 %.
6.6.3 Runs Up and Down
Let R denote the number of runs of Si (which is dened in Sect. 6.6.2). Under the
null hypothesis, R is approximately a normal random variable with mean and
variance given by:
lR 2m 1=3; r2R 16m 13=90: 6:27
The test statistic is the normal score Z R lR =rR and the critical values are
given by Eq. (6.25).
runs up and down test.
The numbers of runs of Si are shown in the last column of Table 6.5. We have
m 11, R 6 and Z 1:2384. As a result, the null hypothesis is not rejected at
the signicance level of 5 %.
6.6 Tests for Randomness 105
6.6.4 MannKendall Test
The null hypothesis of the MannKendall test is that the data are i.i.d. and the
alternative hypothesis is that the data have a monotonic trend. The test statistic is
n1 X
X n
S signxj xi : 6:28
i1 ji1
Under the null hypothesis and for a large n, S approximately follows the normal
distribution with zero mean and the variance given by
r2S nn 12n 5=18: 6:29
The standardized test statistic is given by
Z S signS=rS : 6:30
The critical values with signicance level of a are given by Eq. (6.25).
MannKendall test.
Using the approach outlined above, we have S 22 and rS 14:5831. From

Eq. (6.30) yields Z 1:44, which is smaller than 1.96, the critical value associated
with the signicance level of 5 %. Therefore, the null hypothesis is not rejected.
6.6.5 Spearman Test
Similar to the MannKendall test, the null hypothesis of the Spearman test is that
the data are i.i.d. and the alternative hypothesis is that the data have a monotonic
trend. The test statistic is
Pn
6 Rxi i2
D1 i1
6:31
nn2 1
where Rxi is the rank of xi in the sample with the rank of the smallest observation
being 1. Under the null hypothesis and for a large n, D approximately follows the
normal distribution with zero mean and the variance given by
r2D 1=n 1: 6:32

The standardized test statistic is given by Z D=rD , and the critical values with
signicance level a are given by Eq. (6.25).
Spearman test.
From Eqs. (6.31) and (6.32), we have D 0:4406, rD 0:3015 and

Z 1:4612. Once more, the null hypothesis is not rejected.
In terms of power, Spearman test and MannKendall test are better than the
other three tests. It is noted that Spearman test is simpler than MannKendall test.
As such, Spearman test achieves a good tradeoff between the power and simplicity.
6.6.6 Discussion
According to the results of Examples 6.2 and 6.3, the data in Table 6.1 can be
modeled by an appropriate distribution model. After tting the data to the normal,
lognormal and Weibull distributions using the MLM, it is found that the Weibull
distribution is the best in terms of log maximum likelihood value. The estimated
parameters are shown in the second column of Table 6.2.
However, one may directly implement the third step after carrying out the rst
step since Fig. 6.2 indicates that there is a trend at the early stage of the use. In this
case, the variable-parameter model can be used.
According to Fig. 6.2, the Weibull scale parameter can increase with i and tend
to a constant. Therefore, we assume that the shape parameter keeps unvarying and
the scale parameter is given by
gi g1 1 ei=d ; i 1; 2; . . .: 6:33
Using the MLM, we obtained the parameters shown in the last column of
Table 6.2. In terms of the AIC (see the last row of Table 6.2), the variable-
parameter Weibull model is much better than the two-parameter Weibull distri-
bution. This illustrates the importance of the graphical methods in modeling a
failure process.
6.7 Tests for Normality and Constant Variance
Some statistical analyses (e.g., regression analysis) sometimes need to test nor-
mality and constant-variance. We briefly discuss these two issues in this section.
6.7 Tests for Normality and Constant Variance 107
6.7.1 Tests for Normality
The chi square test and the KolmogorovSmirnov test discussed in Sect. 5.5 are
general methods for testing the goodness of t of a distribution, including the
normal distribution. The normal QQ plot and the skewnesskurtosis-based method
are two simple and specic methods for testing the normality.
The normal QQ plot is applicable for both complete and incomplete data and
can be easily generated using Excel. To be simple, we consider a complete ordered
sample: x1 x2 xn . The empirical cdf at xi can be evaluated by Fi
i=n 1 or Fi betainv0:5; i; n i 1. Let zi U1 Fi ; 0; 1. The normal
QQ plot is the plot of xi versus zi . If the data come from a normal distribution, then
the normal QQ plot of the data should be roughly linear.
Example 6.4 Test the normality of the data in Table 6.1 using the normal QQ plot.
Using the approach outlined above, we obtained the normal QQ plot of the data
shown in Fig. 6.3. As seen, the data points scatter roughly along a straight line,
implying that the normality hypothesis cannot be rejected.
The skewness and kurtosis (i.e., c1 and c2 , see Sect. 3.5.3) of a normal distri-
bution are zero. If the sample skewness and kurtosis are signicantly different from
zero, the data may not be normally distributed. The JarqueBera statistic (see Ref.
[13]) combines these two measures as
n 2
J c c22 =4 6:34
6 1
For large n, the normality hypothesis cannot be rejected if J\6.

Example 6.4 (continued) Test the normality of the data in Table 6.1 using the
skewnesskurtosis-based method.
2
x
0
-2 -1 0 1 2
z
Fig. 6.3 Normal QQ plot

The skewness and kurtosis of the data are c1 0:2316 and c2 0:0254,
respectively. From Eq. (6.34), we have J 0:1076, which is much smaller than 6,
implying that the normality hypothesis cannot be rejected.
6.7.2 Tests for Constant Variance
Suppose that the data xi ; yi ; i 1; 2; . . . are tted to a regression model y gx.

The residuals are calculated by di yi gxi . A good regression model requires
that the residuals should have equal variance.
The equal variance can be veried by the TukeyAnscombe plot (see Ref. [13]),
which is a plot of di versus gxi . If the points in the plot are randomly distributed
without trend, the constant-variance hypothesis cannot be rejected; otherwise,
rejected. As such, the problem becomes the one to test the trend and randomness of
the residuals.
References
1. Ascher H, Feingold H (1984) Repairable systems reliability: modeling, inference,

misconceptions and their causes. Marcel Dekker, New York
2. Brown M, Proschan F (1983) Imperfect maintenance. J Appl Prob 20(4):851859
3. Crow LH (1974) Reliability analysis for complex, repairable systems. In: Proschan F, Serfling
RJ (eds) Reliability and biometry. SIAM, Philadelphia, pp 379410
4. Jiang R (2011) New approach to the modeling of motor failure data with application to the
engine overhaul decision process. J Risk Reliab 225(3):355363
5. Jiang R (2012) A general proportional model and modelling procedure. Qual Reliab Eng Int
28(6):634647
6. Jiang R (2012) Weibull process model with application for modeling bus-motor failures. Inf J
15(12 B):55415548
7. Jiang ST, Landers TL, Rhoads TR (2006) Assessment of repairable-system reliability using
proportional intensity models: a review. IEEE Trans Reliab 55(2):328336
8. Kvaly JT, Lindqvist BH (1998) TTT-based tests for trend in repairable systems data. Reliab
Eng Syst Saf 60(1):1328
9. Kvaly JT, Lindqvist BH, Malmedal H (2001) A statistical test for monotonic and non-
monotonic trend in repairable systems. Paper presented at European conference on safety and
reliabilityESREL, pp 15631570
10. Lehmann EL, Romano JP (2005) Testing statistical hypotheses, 3E edn. Springer, New York
11. Pham H, Wang HZ (1996) Imperfect maintenance. Eur J Oper Res 94:425438
12. Regattieri A (2012) Reliability evaluation of manufacturing systems: methods and
applications. Manufacturing System. http://www.intechopen.com/books/manufacturing-
system/reliability-evaluation-of-manufacturing-systemsmethods-and-applications. Accessed
16 May 2012
13. Thode HC (2002) Testing for normality. Marcel Dekker, New York
References 109
14. Tobias PA, Trindade D (2011) Applied reliability, 2nd edn. Van Nostrand Reinhold,
New York
15. Wang P, Coit DW (2005) Repairable systems reliability trend tests and evaluation. In:
Proceedings of 51st annual reliability and maintainability symposium, pp 416421
16. Wu S (2011) Preventive maintenance models: A review. In: Ouali MS, Tadj L, Yacout S et al
(eds) Replacement models with minimal repair. Springer-Verlag, London, pp 129140
Part II
Product Quality and Reliability
in Pre-manufacturing Phase
Chapter 7
Product Design and Design for X
7.1 Introduction
The life cycle of the product starts from identication of a need. Product design
transforms the need into the idea that produces the desired product. Traditionally, the
product design focused mainly on the acquisition phase of the products life cycle
and was completed purely based on the consideration of product functionality (see
Ref. [5]). To produce a competitive product, product design needs to consider a wide
range of requirements, including product features, cost, quality, reliability, manu-
facturability, and supportability. These requirements are often conflicting. Design
for X (DFX for short) is a set of design methodologies to address these requirements.
In this chapter we briefly discuss the DFX in the context of product life cycle.
The outline of the chapter is as follows. We start with a brief discussion of
product design and relevant issues in Sect. 7.2. Section 7.3 deals with designs for
safety, environment, quality, and reliability. Designs for production-related per-
formances are discussed in Sect. 7.4; designs for use-related performances are
discussed in Sect. 7.5, and designs for retirement-related performances are dis-
cussed in Sect. 7.6.
7.2 Product Design and Relevant Issues
7.2.1 Product Design
Product design is the process of creating a new product. This process is generally
divided into the following ve distinct stages:
Product planning stage. It starts with setting up a project team and denes major
technical parameters and product requirements in terms of performance, costs,
safety, etc.

114 7 Product Design and Design for X
Concept design stage. During this stage, several design concepts are generated
and evaluated to determine whether the product requirements can be met and to
assess their levels of technology and risks. The basic outcome of this stage is
one or more product concepts or options for further consideration. A life cycle
cost (LCC) analysis can be carried out for each design option.
System-level design stage. In this stage, more details are specied, detailed
analysis is carried out, and subsystems begin to take shape for the selected
concept(s).
Detail design stage. During this stage, all components and parts are dened in all
details and most of the manufacturing documentation is produced.
Design renement stage. In this stage, one or more product prototypes are made
and tested so as to nd possible design defects and accordingly modify the design.
The modern product design needs to use numerous working methods and
software packages such as computer-aided design (CAD), computer-aided engi-
neering (CAE), and computer-aided quality (CAQ). These packages are usually
integrated to a product lifecycle management (PLM) system (see Ref. [6] and the
literature cited therein).
7.2.2 Key Issues
Product design aims to create a product with excellent functional utility and sales
appeal at an acceptable cost and within a reasonable time. This deals with the
following three aspects:
Excellent functional utility and sales appeal. This actually deals with product
quality, including reliability and other performance characteristics. Design for X
can be used to address this issue.
Acceptable cost. The cost is evaluated through considering all cost elements
involved in product life cycle. Design for life cycle addresses this issue.
Reasonable time. Product design has become a regular and routine action and
time-to-market has to become shorter and shorter. A time-based product design
approach is used to address this issue.
These approaches are further discussed as follows.
7.2.3 Time-Based Product Design
The main purpose of time-based product design is to reduce the time to market. The
basic approach is to make key participants (e.g., marketing, research and devel-
opment, engineering, operations, and suppliers) be involved as early as possible.
This implies (a) use of a team-based concurrent design process and (b) early
7.2 Product Design and Relevant Issues 115
involvement of key participants. The cross-functional teamwork is due to the fact

that product design requires various expertise and decision-making skills; and early
involvement will considerably facilitate the early identication and prevention of
design problems.
Other technologies to reduce the time to market include computer-aided design,
rapid prototyping, virtual reality, and so forth.
7.2.4 Design for Life Cycle
The objective of design for life cycle is to maximize the life cycle value of prod-
ucts users and minimize the LCC of the product. To maximize the life cycle value,
the design needs to take into account various performance characteristics by using a
methodology of Design for X, where X stands for key performance charac-
teristics of the product.
To minimize LCC, the design needs to take into account various activities and
costs that involve in various phases of the life cycle. Life-cycle assessment is a key
activity of design for life cycle. It assesses materials, services, products, processes,
and technologies over the entire life of a product and, identies and quanties
energy and materials used as well as wastes released to the environment. The main
outcome of the life-cycle assessment is LCC of the product, which is often used as
the decision objective to choose the best design alternative. Therefore, the LCC
analysis is usually carried out in the early stage of product design and has become a
common practice in many organizations.
Life cycle cost is composed of numerous cost elements. The main cost elements
for the manufacturer include research and development cost, manufacturing cost,
marketing cost, operation and maintenance cost, environmental preservation cost,
disposal, and recycle cost. For the user they include purchase cost, operation and
maintenance cost, environmental preservation cost, and residual value of the
product at retirement, which is an income. The LCC model of a product represents
cost elements and their interrelationships.
7.2.5 Design for X
In addition to the LCC and time to market, the manufacturer is also concerned about
other design factors such as manufacturability, assembliability, testability, and so
on. On the other hand, several factors (e.g., price, quality, safety, serviceability,
maintainability, etc.) impact the purchase decisions of customers. These imply that
product design needs to consider many performance requirements or characteristics.
Design for X addresses the issue of how to achieve the desired performances
through design. There is a vast literature on DFX, (see Refs. [15]), and the lit-
erature cited therein.
Fig. 7.1 Classication of performances (Xs)
The performance requirements for a product can be roughly divided into two
categories: overall and phase-specic. The overall performances are those that are
related to more than one phase of the product life cycle (e.g., safety, environment,
quality and reliability). The phase-specic performances are those that are related to
a certain specic phase of the product life cycle. For example, manufacturability is a
production-related performance and maintainability is a use-related performance.
As such, the phase-specic performances can be further divided into three sub-
categories: production-related, use-related, and retirement-related. Figure 7.1 dis-
plays this classication and main performances in each category.
7.3 Design for Several Overall Performances
In addition to LCC, basic overall performances of a product include safety, envi-

ronment, quality, reliability, and testability. We briefly discuss design for these
performances as follows.
7.3.1 Design for Safety
Safety is referred to the relative protection from exposure to various hazards (e.g.,
death, injury, occupational illness, damage to the environment, loss of equipment,
and so on). Hazards are unsafe acts or conditions that could lead to harm or damage
7.3 Design for Several Overall Performances 117
to humans or the environment. Human errors are typical unsafe acts that can occur
at any time throughout the product life cycle; and unsafe conditions can be faults,
failures, malfunctions, and anomalies.
Risk is usually dened as the product of the likelihood or probability of a hazard
event and its negative consequence (e.g., level of loss of damage). As such, a risk
can be characterized by answering the following three questions:
What can happen?
How likely will it happen?
If it does happen, what are the consequences?
Clearly, risk results from a hazard (which is related to the rst question) but a
hazard is not necessary to produce risk if there is no exposure to that hazard (which
is related to the second question).
The levels of risk can be classied as:
Acceptable risk without immediate attention;
Tolerable risk that needs immediate attention; and
Unacceptable risk.
A product is considered safe if the risks associated with the product are assessed to
be acceptable.
Products must be produced safely and be safe for the user. System safety aims to
optimize safety by the identication of safety related risks, eliminating or con-
trolling them by design and/or procedures, based on acceptable system safety level.
A system safety study is usually carried out in both the concept design and
system-level design phases. The main issues of the system safety study are hazard
analysis, specication of safety requirement, and mitigation of safety risk through
design. We discuss these three issues as follows.
7.3.1.1 Hazard Analysis
Hazard analysis includes hazard identication, risk estimation, and risk evaluation.
Preliminary hazard analysis is a procedure to identify potential hazardous condi-
tions inherent within the system by engineering experience. It also determines the
criticality of potential accidents. Functional hazard assessment is a technique to
identify hazardous function failure conditions of part of a system and to mitigate
their effects. Risk estimation deals with quantifying the probability of an identied
hazard and its consequence value. If the risk is unacceptable, risk mitigation
measures must be developed so that the risk is reduced to an acceptable level. Risk
evaluation aims to validate and verify risk mitigation measures.
Three typical hazard analysis techniques are failure mode, effect and criticality
analysis (FMECA), fault tree analysis (FTA), and event tree analysis (ETA).
FMECA is an extension of the failure mode and effects analysis (FMEA). FMEA
aims to identify failure modes and, their causes and effects. The criticality analysis
uses a risk priority number (which is the product of risk and the probability that the
hazard event would not be detected) to quantify each failure mode. It facilitates the
identication of the design areas that need improvements.
FTA is a commonly used method to derive and analyze potential failures and
their potential influences on system reliability and safety. FTA builds a fault tree
with the top event being an undesired system state or failure condition. The analysis
helps to understand how systems can fail and to identify possible ways to reduce
risk. According to this analysis, the safety requirements of the system can be further
broken down.
ETA builds an event tree based on detailed product knowledge. The event tree
starts from a basic initiating event and provides a systematic coverage of the time
sequence of event propagation to its potential outcomes. The initiating event is
usually identied by hazard analysis or FMECA.
7.3.1.2 Specication of Safety Requirements
According to hazard analysis, a set of the safety requirements of the product must
be established in early stages of the product development. The product design must
achieve these safety requirements.
Hierarchical design divides the product into a number of subsystems. In this
case, the product safety requirement will be further allocated to appropriate safety-
related subsystems.
7.3.1.3 Design Techniques for Safety
Safety must be built into a product by considering safety at all phases. Typical
design techniques for a product to satisfy its safety requirements are redundancy
design, fail-safe design, and maintainability design. Maintainability design will be
discussed in Sect. 7.5 and hence we briefly discuss the rst two techniques below.
The redundancy design is a kind of fault tolerance design. Fault tolerance means
that a product can operate in the presence of faults. In other words, the failure
of some part of the product does not result in the failure of the product. All methods
of fault tolerance are based on some form of redundancy. The redundancy design
uses additional component to provide protection against random component fail-
ures. It is less useful for dependent failures. Design diversity is an appropriate way
to deal with dependent failures. The redundancy design will be further discussed in
Chap. 9.
Products can be designed to be fail-safe. A fail-safe device will cause no harm to
other devices or danger to personnel when a failure occurs. In other words, the fail-
safe design focuses on mitigating the unsafe consequences of failures rather than on
avoiding the occurrence of failures. Use of a protective device is a typical approach
of fail-safe design. For example, the devices that operate with fluids usually use
safety valves as a fail-safe mechanism. In this case, the inspection of protective
devices will be an important maintenance activity.
7.3.2 Design for Environment
Design for environment (DFE) is a systematic consideration of design issues

associated with environmental safety and health over the full product life cycle.
DFE involves many disciplines such as environmental risk management, occupa-
tional health and safety, pollution prevention, resource conservation, waste man-
agement, and so on. The goals of DFE include minimization of the use of
nonrenewable resources, effective management of renewable resources, and mini-
mization of toxic release to the environment.
DFE ensures that the designed product is environmentally friendly. Through
appropriately designing products and processes, a manufacturer can reduce costs
and increase prots by recapturing pollutants and reducing solid waste. As such,
environmental concerns require considering environmental criterion, environmental
impact metrics, and other issues such as disassembly and recyclability during the
design stages. The environmental criterion is the environmental attribute of the
product, and can be translated into environmental impact metrics, which can be
further used to assist design decision-making.
DFE tools include DFE guidelines, product assessments, and product steward-
ship metrics. The guidelines cover product usage, product consumable supplies,
shipment packaging, manufacturing processes, and end-of-life product strategies.
The product assessments help to measure results and to identify target improvement
opportunities. The product stewardship metrics include material conservation and
waste reduction, energy efciency, and design for environmental and manufacturing
process emissions.
7.3.3 Design for Quality
Quality must be designed in the product; and poor design cannot be compensated
through inspection and statistical quality control. Design for quality is a set of
methodologies to proactively assure high quality by design. It aims to offer
excellent performances to meet or exceed customer expectations.
There are many design guidelines for quality. These include:
using quality function deployment (QFD) to capture the voice of the customer
for product denition,
using Taguchi method to optimize key parameters (e.g., tolerances),
reusing proven designs, parts, and modules to minimize risk,
simplifying the design with fewer parts, and
using high-quality parts.
The QFD and Taguchi method will be further discussed in the next chapter.
7.3.4 Design for Reliability
Reliability must be designed into products and processes using appropriate meth-
ods. Design for reliability (DFR) is a set of tools or methodologies to support
product and process design so that customer expectations for reliability can be met.
DFR begins early in the concept stage, and involves the following four key
activities:
Determining the usage and environmental conditions of the product and dening
its reliability requirements. The requirements will be further allocated to
assemblies, components and failure modes, and translated into specic design
and manufacturing requirements using the QFD approach.
Identifying key reliability risks and corresponding mitigation strategies.
Predicting the products reliability so that different design concepts can be evaluated.
Performing a reliability growth process. The process involves repeatedly testing
for prototypes, failure analysis, design changes, and life data analysis. The
process continues until the design is considered to be acceptable. The accept-
ability can be further conrmed by a reliability demonstration test.
The rst three activities will be further discussed in Chap. 9 and the reliability
growth by development will be discussed in detail in Chap. 11.
The product design obtained after these activities may be modied based on the
feedbacks from manufacturing process and eld usage.
7.3.5 Design for Testability
Testability is the ability of a product to accurately determine its functionality by

test. High testability can considerably reduce the time of performing test. In
development stage, testability of a product facilitates the development of test pro-
grams and can help reduce the test cost and time. In production phase, testability of
the product provides an interface for production test. In usage phase, testability can
help nd and indicate the presence of faults, and record diagnostic information
about the nature of the faults. This diagnostic information can be used to locate the
source of the failure. In such a way, testability helps reduce time to repair.
Design for testability (DFT) refers to a class of design methods to make test
generation and diagnosis simple and easy. DFT has influences on the product deign.
For example, testability will affect key decisions such as product structure, design,
and selection of components and assemblies, and the manufacturing technologies.
The nature of the tests required to perform determines the type of the test equipment
and may impact equipment investment decision and test development. As such,
DFT should be considered from the concept stage of the product.
Two important aspects with DFT are (a) to make the product testable and (b) to
make the test effective. Product being testable refers to accessibility; and test being
effective refers to identication of defective products for production test or isolation

of fault (including a high level of fault coverage) for fault diagnosis.
7.4 Design for Production-Related Performances
In the production phase, an important consideration is to reduce manufacturing cost.

This is achieved through product and manufacturing system designs. The main
product performance requirements related to this phase are manufacturability, as-
sembliability, and logistics.
7.4.1 Design for Manufacturability
Manufacturability is a design attribute for the designed product to be easy and cost-
effective to build. Design for manufacturability (DFM) uses a set of design
guidelines to ensure the manufacturability of the product. It is initiated at the
conceptual design.
DFM involves various selection problems on structure, raw material, manufac-
ture method and equipment, assembly process, and so on. The main guidelines of
DFM include:
Reducing the total number of parts. For this purpose, one-piece structures or
multi-functional parts should be used. Typical manufacturing processes asso-
ciated with the one-piece structures include injection molding and precision
castings.
Usage of modules and standard components. The usage of modules can simplify
manufacturing activities and add versatility; and the usage of standard compo-
nents can minimize product variations, reduce manufacture cost and lead times.
Usage of multi-use parts. Multi-use parts can be used in different products with
the same or different functions. To develop multi-use parts, the parts that are used
commonly in all products are identied and grouped into part families based on
similarity. Multi-use parts are then created for the grouped part families.
There are many other considerations such as ease of fabrication, avoidance of
separate fasteners, assembly direction minimization, and so on.
7.4.2 Design for Assembliability
Assembliability is a design attribute for a product to be easy to assemble. By design

for assembliability (DFA), a product is designed in such a way that it can be eco-
nomically assembled using appropriate assembly methods. Clearly, DFA overlaps
with DFM. In other words, some design guidelines for DFM (e.g., modularity design
and minimization of the total number of parts) are also applicable for DFA.
The basic guidelines of DFA are
to ensure the ease of assembly, e.g., minimizing assembly movements and
assembly directions; providing suitable lead-in chamfers and automatic align-
ment for locating surfaces and symmetrical parts;
to avoid or simplify certain assembly operations, e.g., avoiding visual obstruc-
tions, simultaneous tting operations, and the possibility of assembly errors.
7.4.3 Design for Logistics
For a manufacturing company, logistics deals with the management of the flow of
resources (e.g., materials, equipment, product and information, etc.) from procure-
ment of the raw materials to the distribution of nished products to the customer.
The product architecture has an important influence on the logistic performance
of the product. For example, the make-or-buy decision for a specic part will result
in considerably different logistic activities. Design for logistics (DFL) is a design
method that aims at optimizing the product structure to minimize the use of
resources.
A considerable part of the product cost stems from purchased materials and
parts. DFL designs a product to minimize total logistics cost through integrating the
manufacturing and logistic activities. As such, DFL overlaps with DFM and DFA,
and hence some guidelines of DFM and DFA (e.g., modular design and usage of
multi-use parts) are also applicable for DFL.
The logistic system usually consists of three interlinked subsystems: supply
system, production system, and distribution system. A systematic approach is
needed to scientically organize the activities of purchase, transport, storage, dis-
tribution, and warehousing of materials and nished products.
The supply system depends on nature of the product and make-or-buy decision
of its parts, and needs to be flexible in order to match different products.
In a production system, two key approaches to achieve desired logistics perfor-
mance are postponement and concurrent processing. The postponement means to
delay differentiation of products in the same family as late as possible (e.g., painting
cars with different colors); and the concurrent processing means to produce multiple
different products concurrently. The main benet of delaying product differentiation
is more precise demand forecasts due to aggregation of forecasts for each product
variant into one forecast for the common parts. The precise demand forecasts can
result in lower stock levels and better customer service. Product designs that allow for
delaying product differentiation usually involve a modular structure of the product,
and hence modularity is an important design strategy for achieving desired logistic
performance. However, the postponement may result in higher manufacturing costs,
adjustment of manufacturing processes, and purchase of new equipment.
7.4 Design for Production-Related Performances 123
The concurrent processing aims to minimize lead times. This is achieved by

redesigning products so that several manufacturing steps can take place in parallel.
This may refer to product line restructuring so that the many models and versions of
end products are assembled from relatively independent assemblies and auxiliary
systems, which can be manufactured concurrently.
Key considerations for the distribution system are economic packaging and
transportation. To achieve this purpose, products should be designed in such ways
so that they can be efciently packed, stored, and transported. Distribution centers
localization can reduce total inventory and the relevant costs of storing and moving
materials and products through the supply chain. The localization may require
product redesign and distribution center modications, and hence should be con-
sidered when designing the product.
7.5 Design for Use-Related Performances
The performance requirements in the usage phase can be divided into two cate-
gories: user-focused and post-sale-focused. The user-focused performances include
user-friendliness, ergonomics, and aesthetics. These can have high priorities for
consumer durables. The post-sale-focused performances include reliability, avail-
ability, maintainability, safety, serviceability, supportability, and testability and
have high priorities for capital goods. Among these performances, reliability,
availability, maintainability and safety or supportability (RAMS) are particularly
important. Design for RAMS involves the development of a service system
(including a preventive maintenance program) for the product.
Some of the post-sale-focused performances have been discussed earlier, and the
others are briefly discussed as follows.
7.5.1 Design for Serviceability
Serviceability is the ability to diagnose, remove, replace, adjust, or repair any

component or subsystem with relative ease. Design for serviceability (DFSv) is a
methodology to make designed products be easy to service and maintain. DFSv
starts with determination of serviceability requirements. The serviceability of a
design is reviewed using a checklist. The main outcome of the review is a list of
possible opportunities for improvement. A set of design guidelines can be used to
address these opportunities and the results are integrated into the new design.
The main considerations for serviceability include:
Location. A good practice to design the components that are likely to fail or
need servicing is to make them close to the assembly surface. This can reduce
the cost of the most frequent service operations.
Simplication. For example, minimization of the number of layers of compo-

nents can reduce the number of components removed to gain access to a specic
part; and minimization of the number of connections between subassemblies can
reduce the time and complexity to remove and install subassemblies.
Standardization. This deals with use of standard components. The benets of
standardization include component cost reduction, availability of parts, and
reduction in the use of specialty tools.
Ease of repair. Good practices include ergonomics design and development of a
modular product structure. The latter can considerably simplify repair operation
by removing the whole module instead of individual and embedded
components.
7.5.2 Design for Maintainability
Maintainability is the relative ease and economy to restore a failed item to a

specied working condition using prescribed procedures and resources. It has
overlap with serviceability. Main difference is that the serviceability focuses on
preventive maintenance and the maintainability on corrective maintenance. Design
for Maintainability is a design methodology to assure that the product can be
maintained throughout its life cycle at reasonable expense without any difculty.
Maintainability characteristics of a product are dened by maintainability
requirements, which can be quantitative and qualitative. A key quantitative indi-
cator is mean time to repair (MTTR). For a specied fault, time to repair includes
the times required for fault localization, fault removal, adjustment, calibration, and
verication. Different faults have different times to repair. As such, MTTR is
estimated through considering all possible faults.
Maintainability design guidelines address qualitative requirements of maintain-
ability, including:
Safety requirements. These deal with the avoidance of injury to personnel and
damage to the equipment during maintenance and servicing. For example, sharp
edges, corners, or protrusions should be avoided.
Accessibility requirements. These deal with having sufcient space or clearance
for adequate viewing and hand access during maintenance and servicing.
Assembliability and dis-assembliability requirements. These shall facilitate the
operations of disassembly and reassembly during maintenance and servicing.
Testability requirements. These shall help to detect and isolate failure.
Other requirements to support repair operations. For example, the weight and
dimension shall be within reasonable ranges; the needs for special tools are
minimized; reference designations, handles and guide pins for alignment are
provided if necessary.
7.5 Design for Use-Related Performances 125
7.5.3 Design for Supportability
Product support is an essential factor for achieving customer satisfaction in many

industries. It includes maintenance support and logistics support. Maintenance sup-
port includes four aspects and deal with personnel (e.g., training), tool (e.g., main-
tenance and other facilities), material (e.g., various spares and consumables), and
information (e.g., installation, operating and maintenance instructions, modication
instructions and checkout procedure). Logistics support focuses on the aspect of
material of maintenance support. The required spares inventory depends on product
maintenance concept (which is the set of rules prescribing what maintenance is
required and how demand for it is activated) and specic quantities can be optimized.
Supportability is ability that a product is easy and economical to service and
support. Design for supportability (DFSp) is a set of techniques to assure the
supportability of the product through design.
To implement DFSp, support requirements are reviewed and evaluated to ensure
that basic mission-related elements are designed to be supportable in an effective
and efcient manner, and all system requirements are adequately addressed through
design of maintenance and logistics support infrastructure. If requirements are met,
the design is approved and the program enters into the next stage; otherwise, the
appropriate changes are initiated. Design review and evaluation are done through
supportability analysis (SA). SA evaluates:
alternative repair policies that are subject to the constraints specied by the
maintenance concept, and
equipment design characteristics in terms of logistic support requirements.
Through SA, the logistics and maintenance support resources for a given design
conguration are identied, supportability requirements and design criteria are
established, and various design alternatives can be evaluated.
7.6 Design for Retirement-Related Performances
The end-of-life products can be discarded, recycled, refurbished, or remanufac-

tured. In this section, we focus on the case of recycling. In this case, main per-
formance requirements for a product are its recyclability and disassembliability. We
briefly discuss them as follows.
7.6.1 Design for Recyclability
Generally, it is not possible to recycle a product completely in an economical way.

Therefore, the recycling aims to maximize the recycling resources (e.g., number of
reusable parts) and to minimize the potential pollution (e.g., amount of waste) of the
remainder. Design for recyclability (DFRc) is a set of design techniques to achieve

this objective. It is usually implemented during design evaluation stage.
Two basic considerations associated with DFRc are dismantling techniques and
recycling costs. Dismantling requires the knowledge of the destination or recycling
possibility of the disassembled parts. This has to consider possible advances in
recycling and re-engineering techniques from the time when a product is designed
to the time when it reaches the end of its life. The design for dismantling aims to
remove the most valuable parts and maximize the yield of each dismantling
operation.
Material compatibility is a major issue for product retirement and deals with the
concept of clumping. A clump is a collection of components and/or subassemblies
that share a common characteristic based on user intent. The designer may need to
clump components that are not compatible due to certain constraints. If the post-life
intent of the product is to be recycled, the mechanical connections among the
components should be easily broken (e.g., snap ts, screws, etc.) when the materials
in the clump are not compatible.
Another issue for DFRc is material recognition. It requires technology capable of
identifying materials, including the proportion and type of materials. Fourier
Transform Infra-Red-based equipment has been developed for identifying plastics
and some ller materials.
7.6.2 Design for Disassembliability
Disassembly is a process of systematic removal of desirable parts from an assembly

without impairment of the parts due to the process. Design for disassembliability
(DFD) is a set of design methods and techniques to make a product be easily
disassembled. For example, the exotic materials that are difculty to recycle should
be avoided; and parts that have plastic and metal fused together should not be used
since they are difcult to separate.
Design for disassembliability also considers disassembly method and sequence.
Two basic methods of disassembly are reverse assembly and using brute force. In
the case of reverse-assembly, if a fastener is screwed in, then it is screwed out; if
two parts are snap t together, then they are snapped apart. While in the case of
brute force, parts are just pulled or cut.
References
1. Dombrowski U, Schmidt S, Schmidtchen K (2014) Analysis and integration of design for X

approaches in lean design as basis for a lifecycle optimized product design. Procedia CIRP
15:385390
2. Gatenby DA, Foo G (1990) Design for X (DFX): key to competitive, protable products. AT &
T Tech J 69(3):213
References 127
3. Huang GQ, Shi J, Mak KL (2000) Synchronized system for Design for X guidelines over the
WWW. J Mater Process Tech 107(13):7178
4. Keys LK (1990) System life cycle engineering and DFX. IEEE Trans CHMT 13(1):8393
5. Kuo TC, Huang SH, Zhang HC (2001) Design for manufacture and design for X: concepts,
applications, and perspectives. Comput Ind Eng 41(3):241260
6. Saaksvuori A, Immonen A (2008) Product lifecycle management, 3rd edn. Springer, Berlin
Chapter 8
Design Techniques for Quality
8.1 Introduction
We mentioned two different concepts of quality in Chap. 1. One is customer-driven

quality concept, which denes quality as the ability of a product to meet or exceed
customer expectations; and the other emphasizes the reduction of variability in
important quality characteristics, which denes quality as inversely proportional to
variability [6]. Two design techniques for quality that are closely related to the
above two quality concepts are quality function deployment (QFD) and Taguchi
method [8]. QFD is a product development process which is based on the notion of
house of quality (HOQ [4]). HOQ is a design approach to translate customer
expectations into engineering characteristics so that the customers expectations can
be met. Taguchi method focuses on variability of important quality characteristics.
Two key concepts or techniques with the Taguchi method are quality loss and
robust design by experimental optimization. These techniques are not only appli-
cable for product design but also for process design as well as quality improvement
of product and process. In this chapter we present these techniques in the context of
product design.
The outline of the chapter is as follows. Section 8.2 deals with HOQ and QFD, and
Sect. 8.3 deals with quality loss function. Experimental optimum method is discussed
in Sect. 8.4 and the model-based optimum method is discussed in Sect. 8.5.
8.2 House of Quality and Quality Function Deployment
8.2.1 House of Quality
As mentioned earlier, the design and development process of a product starts with
requirement denition and ends with a prototype version of the product that meets
customer needs. The HOQ is a design technique developed to identify and

130 8 Design Techniques for Quality
Correlations
ECs
Benchmarking
Importance
Relationships
Effects
CAs
between ECs
and CAs
Evaluation for competing products
Target levels
Fig. 8.1 The house of quality
transform customer needs into technical specications. It is based on the belief that
a product should be designed to reflect customer needs.
Figure 8.1 shows an HOQ in the product planning stage. The customers needs
or attributes (CA) are represented on the left-hand side (LHS) of the HOQ, and are
usually qualitative and vague. The relative importance or weight of a CA helps to
identify critical CAs and to prioritize design efforts. The weight of a CA can be
determined using various approaches such as the AHP (see Online Appendix A).
For example, for CA i, a score of si 1; 2; . . .; 9 can be assigned to it based on
the customers preference. The weight of CA i can be calculated by
si
xi P
m 8:1
sk
k1
where m is the number of CAs.

The technical specications or the engineering characteristics (EC) to meet the
CAs are listed on the ceiling of the HOQ. The ECs are the design requirements that
affect one or more of the CAs.
The roof is a diagonal matrix, which indicates the correlations among the ECs.
The correlation can be assessed by the design team in a subjective manner. For
example, for ECs i and j, a score of sij 0; 1; ; 2; . . . or 9 can be assigned to
8.2 House of Quality and Quality Function Deployment 131
represent their correlation degree based on the experts judgment. The correlation
coefcient between the ECs i and j can be calculated by
qij sij =9; i \ j; i 1; 2; :::; n 1 8:2
where n is the number of ECs.

The main body of the HOQ is a relationship matrix that indicates how strongly a
certain EC covers a certain CA. The strengths assigned to the relationships between
CAs and ECs are also assessed in a subjective manner. For CA i and EC j, a score
of sij (0, 1, , or 9) can be assigned to represent the strength of their relationship
based on the experts judgment. For example, if there is a very strong relationship
between CA i and EC j, we take sij 9; if the relationship is relatively strong, we
take sij 3; and if there is no relationship, we take sij 0. The strength can be
further normalized as below:
rij sij =9 2 0; 1; i 1; 2; . . .; m; j 1; 2; . . .; n: 8:3
The right-hand side (RHS) of the HOQ is the comprehensive effects from all ECs
for all CAs, and also may include a competitive benchmarking value for each CA.
The bottom part of the HOQ may give the competing products performance,
comprehensive evaluation, and conclusions about how the designing product is
superior to the competing products. The target levels of ECs are determined using
all the information in the HOQ.
There can be different versions for the LHS, RHS, and bottom part of the HOQ,
depending on specic applications. For example, there can be a correlation matrix
for the CAs, which is usually placed on the LHS of the HOQ.
The HOQ helps transform customer needs into engineering characteristics, pri-
oritize each product characteristic, and set development targets. To achieve these
purposes, an evaluation model is used to evaluate the importance rating of each EC,
and another evaluation model is used to evaluate the comprehensive effect of all the
ECs on each CA. As such, the future performance of the designing product can be
predicted by aggregating these effects. We discuss these models as follows.
8.2.2 Priorities of Engineering Characteristics
The importance ratings or priorities for the ECs can be evaluated using the CAs
relative importance and the relationship matrix. If the correlations among the CAs
can be ignored, the priorities of the ECs can be calculated by (see Ref. [3])
X
m
pj xi rij ; 1 j n: 8:4
i1
The normalized weights are given by
X
n
wj pj = pk ; 1 j n: 8:5
k1
Example 8.1 Consider four CAs and ve ECs. Their relationship matrix is shown
in the top part of Table 8.1. Using Eq. (8.4) yields the priorities of the ECs shown in
the third row from the bottom; and the normalized weights are shown in the second
row from the bottom. The last row shows the ranking number of each EC. As seen,
EC 5 is the most important and EC 4 is the least important.
8.2.3 Satisfaction Degrees of Customer Attributes
The satisfaction degree of a CA can be evaluated based on the relationship matrix of

the HOQ. We present a simple approach as follows.
The starting point is the relationship matrix with rij representing the satisfaction
degree of CA i resulting from EC j. This implies that dij 1 rij is a measure of
dissatisfaction. Let Si denote the total satisfaction degree of CA i and Di 1 Si
denote the total dissatisfaction degree. Assume that the effects of ECs are mutually
independent, we have
Y
n Y
n
Di 1 Si dij 2 0; 1; Si 1 1 rij 2 0; 1: 8:6
i1 i1
The overall performance of design is given by
X
m
S x i Si : 8:7
i1
Generally, the effects of the ECs on CA i are not independent. Therefore,

Eq. (8.6) gives the upper bound of Si , and Eq. (8.7) gives the upper bound of S. As
such, the results obtained from Eq. (8.6) can be modied based on experts
judgments.
The model given by Eq. (8.6) can be viewed as a special case of the multipli-
cative utility model of Keeney [5], which is given by
Y
n
1 KUi 1 Kaj uij ; aj 2 0; 1; K [ 1 8:8
j1
8.2 House of Quality and Quality Function Deployment 133
Table 8.1 Relationship xi EC 1 EC 2 EC 3 EC 4 EC 5

matrix and priorities of ECs
for Example 8.1 CA 1 0.22 0.56 0 0.78 0 0
CA 2 0.13 0 0.67 0.56 0 0
CA 3 0.35 0 0.44 0 0 0.78
CA 4 0.30 0.56 0 0 0.67 0.56
pj 0.2912 0.2411 0.2444 0.2010 0.4410
wj 0.2053 0.1699 0.1723 0.1417 0.3108
Rj 2 4 3 5 1
Table 8.2 Satisfaction CA 1 CA 2 CA 3 CA 4 S

degrees of CAs for
Example 8.1 Si 0.9032 0.8548 0.8768 0.9361 0.8975
where uij is the utility of attribute j, Ui is the overall utility, aj is the attribute weight
and K is a constant. When uij 1, the overall utility should equal 1, i.e.,
Y
n
1K 1 Kaj : 8:9
j1
As such, K is the nonzero solution of Eq. (8.9). Clearly, Eq. (8.8) reduces into
Eq. (8.6) when aj 1, K 1 and uij rij .
Example 8.1 (continued) Consider the relationship matrix shown in Table 8.1. The
satisfaction degrees of the CAs calculated from Eq. (8.6) are shown in Table 8.2. As
seen, CA 4 can be met well by the design but CA 2 is not met well. The overall
performance of design equals 0.8975, which is equivalent to an 8-point in terms of
the 9-point scale used in AHP.
8.2.4 Quality Function Deployment
Quality function deployment is a series of HOQs, where the ECs of the current
HOQ become the CAs of the next HOQ. Each HOQ relates the variables of one
design stage to the variables of the subsequent design stage. The process stops at a
stage when the design team has specied all the engineering and manufacturing
details. In this way, the QFD ensures quality throughout each stage of the product
development and production process.
Typically, QFD is composed of four HOQs. The rst HOQ transforms the
customers needs to the engineering characteristics (or design requirements); the
second HOQ transforms the engineering characteristics to parts characteristics (or
part requirements); the third HOQ transforms the part characteristics to
technological requirements, and the fourth HOQ transforms the technological

requirements to production requirements.
QFD does not conne to product design. It can also be applicable for process or
system design. Reference [1] presents a literature review on QFD.
8.3 Cost of Quality and Loss Function
8.3.1 Quality Costs
The quality costs include prevention cost (e.g., process improvement and training
costs), appraisal or evaluation cost (e.g., inspection or test costs), external loss (e.g.,
warranty cost and sale loss) and internal loss (e.g., scrap and rework costs). These
cost elements are highly correlated. For example, as product quality level increases,
the prevention and appraisal costs increase but the internal and external losses
decrease. The traditional viewpoint with quality is to nd an optimal quality level
so that the total quality cost achieves its minimum. However, the modern viewpoint
with quality thinks that the continuous quality improvement is more cost-effective
based on the following reasons:
It will result in an improved competitive position so that the product will be sold
at a higher price and will have an increased market share.
It can result in decrease in failure costs and operational cost. These lead to an
increase in prots.
As a result, the quality should be continuously improved and should not adhere to
an optimal quality level.
8.3.2 Loss Function
Let Y denote the quality characteristic and, LSL and USL denote the lower and
upper specication limits, respectively. Items that conform to the design speci-
cations are called conforming and those that do not are called nonconforming or
defective. The quality loss depends on the value of Y (denote the value as y).
Traditionally, the loss is thought to be zero when y falls inside the specication
limits; otherwise, the loss is a positive constant. As such, the conventional quality
loss function is a piecewise step function as shown in Fig. 8.2. Such a function
implies that any value of Y within the specication limits is equally desirable.
Taguchi [8] considers that any deviation from a predetermined target value T
represents an economic loss to the society. The loss can be incurred by the man-
ufacturer as warranty or scrap costs; by the customer as maintenance or repair costs;
or by the society as pollution or environmental costs [2]. As such, there can be
8.3 Cost of Quality and Loss Function 135
Target value
L (y)
Conventional loss function Taguchi loss function
LSL USL
Fig. 8.2 Taguchi quality loss function
a quality cost for any conforming product as long as its quality characteristic is not
at the target value.
Using the Taylor expansion series, Taguchi proposes a quadratic loss function to
model the loss of the deviation of the quality characteristic from its target value:
Ly Ky T2 8:10
where K is a coefcient to be specied. Clearly, reducing the variability leads to a

small cost of quality. This is why Montgomery [6] denes quality as inversely
proportional to variability though the phrase is not mathematically strict.
The above function is only suitable for the situations where the deviation in
over- and under-directions can cause the same loss. A product can have an
asymmetric loss function. In this case, a piecewise loss function can be dened as
Ly K1 y T2 for y\T; Ly K2 y T2 for y [ T: 8:11
For a batch of products, let Fy denote the distribution of Y with mean l T

and standard deviation r. For the conforming items, Y follows the doubly-truncated
normal distribution with the support y 2 LSL; USL and the density function is
given by
f y /y; l; r=1 2UD; 0; r 8:12
where / denotes the normal density function and D l LSL USL l.

Assume that the loss function is given by Eq. (8.10). The average quality loss per
conforming item is given by
ZUSL
lL Ky l2 f ydy KV 8:13
LSL
where
V r2 f1 2D=r/D=r; 0; 1=2UD=r; 0; 1 1g: 8:14
Equations (8.13) and (8.14) clearly show the benet of reducing the variability.
When the target T is a nite value, it is called nominal-the-best case where the
quality characteristic Y should be densely distributed around the target value. Two
other cases are smaller-the-better and larger-the-better. If Y is non-negative, a
smaller-the-better quality characteristic has the target value T 0 so that we have
Ly Ky2 . A larger-the-better quality characteristic can be transformed into a
smaller-the-better quality characteristic using the transformation Y 1 so that we
have Ly K=y2 .
8.3.3 Applications of Quality Loss Function
In a part batch production environment, Yacout and Boudreau [10] assess the
quality costs of the following quality policies:
Policy 1: Nothing is done to control or prevent variations
Policy 2: 100 % inspection
Policy 3: Prevention by statistical process control (SPC) techniques
Policy 4: A combination of Policies 2 and 3.
Policy 1 can have large external loss due to delivering nonconforming units to
customers. The quality loss function with a relatively large value of K can be used
to evaluate the expected costs in the in-control state and out-of-control states,
respectively. The total cost per cycle is the sum of the costs in the two states.
The quality costs of Policy 2 mainly include inspection cost and internal loss due
to reworking or scraping nonconforming units. The internal loss can be evaluated
using the quality loss function with a relatively small value of K. Relative to Policy
1, this policy costs less to discover the nonconforming units and to prevent them
from reaching the customer.
The quality costs of Policy 3 mainly include external loss and prevention cost.
The prevention involves the use of control charts, which are used to detect whether
the process is in-control or not. If an out-of-control state is detected, the assignable
causes can be corrected at the end of the cycle. This reduces the fraction non-
conformance and hence also reduces the external loss. The prevention costs are
sampling costs, costs of investigating false and true alarms, and correction costs. If
the detection and correction of assignable causes can effectively reduce the
occurrence of out-of-control state, this policy will result in quality improvement.
The quality costs of Policy 4 mainly include internal loss, inspection cost and the
prevention cost. Compared with Policy 2, it has a smaller internal loss; and com-
pared with Policy 3, it does not have the external loss.
The optimal quality policy can be determined through evaluating the effects of
different quality policies on the quality costs and the outgoing quality.
8.4 Experimental Optimum Method 137
8.4 Experimental Optimum Method
8.4.1 Basic Idea
Taguchi [8] divides the design phase into three stages: systematic design, para-
metric design, and tolerance design; and develops an experimental optimum method
for the parametric design. The basic idea is to divide the parameters that impact the
performances of a product (or process) into two categories: controllable and
uncontrollable. The controllable parameters are design variables and the uncon-
trollable parameters are called the noises (e.g., manufacturing variation, environ-
mental and use conditions, and degradation or wear in components and materials).
The problem is to nd the optimal levels of the controllable parameters so that the
performance is not sensitive to the uncontrollable parameters. The method needs to
carry out a set of experiments. To reduce the experimental efforts and obtain
sufcient information, the experiment design is stressed. This approach is called
Taguchi method, which is based on orthogonal array experiments.
Taguchi method is an experiment-based optimization design method, and
applicable for the situation where no good mathematical model is available for
representing the product performance. Taguchi method does not conne to the
parametric design of the product. In fact, it is applicable for optimizing any engi-
neering process. The word parameter or variable can be a design option or a
type of part.
8.4.2 Specic Procedure
Taguchi method involves a multi-step procedure. Depending on the complexity of

the problem, the steps can be grouped into four phases: plan, implementation,
analysis, and validation.
The plan phase involves the following actions or issues:
Identifying the quality characteristics, and dening the objective function to be
optimized;
Identifying the controllable factors (i.e., design parameters) and their levels;
identifying the noise factors and their levels if applicable;
Designing the experiments and identifying the testing conditions for each
experiment; and
Dening the data analysis procedure.
The implementation phase deals with conducting the designed experiments to
collect data on the effect of the design parameters on the performance measure.
The analysis phase deals with analyzing the data obtained from the experiments
and predicting the performance of the product or process under the optional con-
ditions. The main outcomes of the analysis include:
Determining the effect of the different parameters on the performance;

Identifying possible factor interactions;
Determining the optimum levels for the controllable parameters; and
Predicting the performance at the optimal levels of the controllable parameters.
Analysis on the data collected from the experiments can be used to select new
parameter values to optimize the performance characteristic. As such, the validation
phase deals with performing the verication experiment for the predicted optimal
levels and performance and planning the future actions.
Here, two key issues are experiment design and data analysis. We separately
discuss them in the next two subsections.
8.4.3 Design of Experiments
The experimentation requires time and resources. The experiment design is to nd

the best settings of parameters so that the necessary data can be obtained with a
minimum amount of experimentation. The method for experiment design depends
on the number of controllable parameters. Generally, a factorial design can be
appropriate if the number of parameters is small and a random design can be
appropriate if the number of parameters is large. When there are an intermediate
number of variables, few interactions between variables and only a few variables
contribute signicantly, Taguchis orthogonal array experiments are the most
appropriate (see [7]).
8.4.3.1 Factorial Design
Let Pi ; 1 i n denote the controllable parameters and ki ; 1 i n denote the

number of levels of Pi . A factorial design considers all the combinations ofQlevels
n
for all the parameters. As such, the factorial design requires totally i1 ki
experiments, and hence is only applicable for the case where n is small. An
advantage of the factorial design is that it can be used to determine the interactions
between variables.
8.4.3.2 Random Design
When the number of controllable parameters is large, the total number of experi-
ments to be completed can be specied as a constraint condition. For a given
variable Pi and number of levels ki , let pj 1 j ki denote the probability for level
Pj
j to happen. Further, let q0 0; qj l1 pl and qki 1. The level of Pi can be
randomly generated by li j if
qj1 \r\qj 8:15
where r is a uniform random number within 0 and 1. The experiments with the
required number can be obtained by repeatedly using Eq. (8.15).
8.4.3.3 Design of Experiments of Taguchi Method
Taguchi method is based on orthogonal array experiments, with which each vari-
able and each level will be tested equally. The array is selected by the number of
parameters and the number of levels. Here, the parameters can be controllable and
uncontrollable. The experiment design for controllable and uncontrollable param-
eters is separately conducted.
Design for controllable parameters. Once the controllable parameters have been
determined, the levels of these parameters must be determined. Determining the
levels of a variable needs rst to specify its minimum and maximum, and then to
specify the number of levels taking into account the change range and the cost of
conducting experiments. The number of levels for all parameters in the experi-
mental design is usually chosen to be the same so as to facilitate the selection of the
proper orthogonal array. Once the orthogonal array is determined, necessary
adjustment is allowed for different parameters to have different numbers of levels.
The proper orthogonal array can be selected based on the number of parameters
(n) and the number of levels (k) as indicated in Table 8.3, where the subscript of the
array indicates the number of experiments to be completed. Once the name of the
array has been determined, the predened array can be looked up (see Refs. [7],
[8]). For example, when n 4 and k 3, the array is L9 . The corresponding
combinations of levels are shown in Table 8.4.
Design for uncontrollable parameters. Uncontrollable parameters or external
factors affect the performance of a product or process, and the experiments may
reflect their effects on the performance in two different ways. The rst way is to
conduct several trials (i.e., repeated tests) for a given combination of controllable
parameters (i.e., an experiment). The second way is to explicitly consider a set of
uncontrollable parameters, and to generate several combinations of levels of these
Table 8.3 Selection of k\n 2 3 4 5 6 7 8 9 10

orthogonal array
2 L4 L4 L8 L8 L8 L8 L12 L12 L12
3 L9 L9 L9 L18 L18 L18 L18 L27 L27
4 L16 L16 L16 L16 L32 L32 L32 L32 L32
5 L25 L25 L25 L25 L25 L50 L50 L50 L50
Table 8.4 Orthogonal i P1 P2 P3 P4

array L9
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1
parameters. The outcome can be an orthogonal array called the noise matrix. For a
given experiment for controllable parameters, all the combinations specied in the
noise matrix will be tested. As a result, each experiment corresponds to several
trials.
8.4.3.4 Mixed Experiment Design
A mixed experiment design is a mixture of the factorial design, random design, and
Taguchi design. For example, if one considers the number of experiments specied
by the Taguchi design is somehow small, a given number of additional experiments
can be carried out based on random design.
Finally, it is worth indicating that an experiment is not necessary to be
physical. In other words, the experiment can be computational, including simula-
tion. In this case, the experiment design is still needed to change the values of
parameters in an appropriate way.
8.4.4 Data Analysis
The data analysis deals with three issues: calculation of signal-to-noise ratios,
evaluation of the effects of the different parameters, and optimization of levels of
the controllable parameters. We separately discuss these issues in the next three
subsections.
8.4.4.1 Calculating Signal-to-Noise Ratio
In Taguchi method, signal-to-noise ratio is used as an objective to determine the

best control factor levels. The denition of signal-to-noise ratio is dependent on the
property of performance characteristic Y, which can be smaller-the-better, nominal-
the-best and larger-the-better.
We rst look at the nominal-the-best case. Let yij ; 1 i n; 1 j ki denote the

measured performance characteristic of the jth trial of the ith experiment. The mean
and variance of ki trials are given, respectively, by
1X ki
1 X ki
li yij ; r2i yij li 2 : 8:16
ki j1 ki 1 j1
The signal-to-noise ratio is dened as

2 2
l l
SNi 10 log10 i2 4:343 ln i2 : 8:17
ri ri
A large signal-noise ratio implies good robustness and hence is desired.

For the smaller-the-better case, the ideal value is zero, and the signal-to-noise
ratio is dened as
SNi 10 log10 r2i 4:343 lnr2i 8:18
where
1X ki
r2i y2 : 8:19
ki j1 ij
The larger-the-better case can be transformed into the smaller-the-better case by

letting zij y1
ij . As such, the signal-to-noise ratio is given by Eqs. (8.18) and (8.19)
with yij being replaced by zij .
8.4.4.2 Evaluating Effects of Parameters
After obtaining the signal-to-noise ratio for each experiment, the average signal-to-
noise ratio can be calculated for each level of a given parameter. Let SNij denote the
average of the signal-to-noise ratios for the jth level of parameter Pi . For example,
for the experiments shown in Table 8.4, we have
SN1 SN2 SN3 SN3 SN6 SN9 SN2 SN6 SN7

SN11 ; SN23 ; SN42 :
3 3 3
Once the averages of the signal-to-noise ratios are obtained, the range of the
averages for parameter Pi can be calculated as
Di max SNij min SNij : 8:20

j j
Table 8.5 Experiment results i Trial 1 Trial 2 Trial 3 li ri SNi

and signal-to-noise ratios
1 29.62 27.55 29.01 28.73 1.06 28.63
2 23.56 22.93 23.8 23.43 0.45 34.34
3 18.82 17.97 19.02 18.60 0.56 30.47
4 25.12 25.72 23.05 24.63 1.40 24.90
5 24.79 24.64 24.07 24.50 0.38 36.19
6 30.07 28.43 30.42 29.64 1.06 28.91
7 18.92 21.1 21.49 20.50 1.39 23.41
8 32.21 32.59 33.94 32.91 0.91 31.17
9 25.14 22.89 23.58 23.87 1.15 26.32
Table 8.6 Averages of Level P1 P2 P3 P4

signal-to-noise ratios
1 31.15 25.65 29.57 30.38
2 30.00 33.90 28.52 28.89
3 26.97 28.57 30.02 28.85
Di 4.18 8.26 1.50 1.53
Rank 2 1 4 3
A large value of Di implies that Pi has a large effect on the output. As such, the
effects of the parameters can be ranked based on the values of Di .
The correlation coefcient between (SNij ; 1 j k) and (SNlj ; 1 j k) can
represent the interaction between Pi and Pl .
Example 8.2 Assume that the problem involves four variables and each variable
has three levels, and the performance characteristic is nominal-the-best. The
orthogonal array is L9 shown in Table 8.4. Each experiment is repeated for three
times and the results are shown in the second to fourth columns of Table 8.5. The
values of mean, standard deviation and signal-to-noise ratio for each experiment are
shown in the last three columns of Table 8.5.
From the signal-to-noise ratios of the experiments, we can obtain the average
signal-to-noise ratios and the results are shown in Table 8.6. The range of the
averages for each parameter is shown in the second row from the bottom, and the
rank numbers of the parameters are shown in the last row. According to the ranking,
we can conclude that P2 has the highest effect and, P3 and P4 have the lowest effect
on the output.
From the average signal-to-noise ratios shown in Table 8.6, we can calculate
their correlation coefcients and the results are shown in Table 8.7. As seen, there
can be weak interactions between P1 and P4 , P2 and P3 , and P2 and P4 .
Table 8.7 Interactions P2 P3 P4

between parameters
P1 0.10 0.52 0.73
P2 0.79 0.76
P3 0.20
8.4.4.3 Optimal Combination of Levels
For parameter Pi , the best level l meets
SNil max SNij : 8:21

j
For Example 8.2, Table 8.6 shows that the best level combination is (1, 2, 3, 1)
for (Pi , 1 i 4). It is noted that there is not such a combination in Table 8.4, and
the combination that is closest to this combination is Experiment 5. A supple-
mentary test may be conducted to verify this combination.
An approximate method can be used to predict the performance under the
optimal combination. It is noted that the optimal combination is obtained through
changing the P1 s level of Experiment 5 from Level 2 to Level 1. Referring to the
second column of Table 8.4, the performance increment resulting from changing
P P
Level 2 of P1 to Level 1 of P1 can be estimated as Dy 13 3i1 yi 13 6i4 yi ,
where y can be l or r. As such, the performance under the optimal combination can
be estimated as y y5 Dy. The computational process is shown in Table 8.8, and
the last row gives the predicted performances.
8.5 Model-Based Optimum Method
As mentioned earlier, the Taguchi method is applicable for the situation where no
good mathematical model is available. There are many problems where mathe-
matical models can be developed to optimize the design of products or processes. In
this section, we use the tolerance design problem as an example to illustrate the
model-based optimum method.
Tolerance design is an important issue in mass production. The design variables
are tolerances for individual components, which are subject to the constraints from
machine tools capabilities and functional requirements, and the objective function
is the total cost per produced unit [9]. As such, two key issues are (a) specifying the
constraint conditions and (b) developing the cost model.
Table 8.8 Estimated P1 l r SN

performances of the optimal
combination Experiment 5 24.50 0.38 36.19
Level 1 23.59 0.6900
Level 2 26.26 0.9467
Predicted 21.83 0.1233 44.96
8.5.1 Constraint Conditions
In a tolerance chain, there exists a resultant dimension that is derived from the
primary dimensions. The tolerance of the resultant dimension is a function of the
tolerances of the primary dimensions. The statistical method is commonly
employed for tolerance analysis. It is based on the assumption that the primary
dimensions are normally distributed.
Let xr and tr denote the resultant dimension and its tolerance, respectively; xi
denote the ith dimension and ti denote its tolerance; and xr Gxi ; 1 i n
denote the dimension relationship. From the dimension relationship and statistical
method, the tolerance relationship is given by
" #1=2
Xn
@G 2
tr ti : 8:22
i1
@xi
Due to the capability of machine tools, ti has a lower bound timin . The tolerance
of the resultant dimension determines the performance of assembly, and hence an
upper bound trmax has to be specied for tr . As a result, the constraints are given by
tr trmax ; ti timin ; 1 i n: 8:23
8.5.2 Objective Function
In the tolerance design problem, two main cost elements are manufacturing cost and
quality loss. The manufacturing cost decreases and the quality loss increases with
tolerances in a complex way.
We rst examine the manufacturing cost, which consists of manufacturing cost
and assembly cost. Generally, small tolerances result in an increase in the manu-
facturing cost due to using precision machine and measuring devices. The toler-
ance-cost relationship can be obtained by tting empirical data into a proper
function. Two typical functions are:
ct a b=tc ; ct a b=ect : 8:24

8.5 Model-Based Optimum Method 145
The assembly cost is usually not sensitive to the tolerances of components, and
hence can be excluded for the optimal tolerance design problem. As such, the total
manufacturing cost of an assembly is given by
X
n
CM ci ti : 8:25
i1
We now look at the quality loss. For an assembly, the functional requirement is
the resultant dimension. Let X denote the actual resultant dimension, and f x
denote the distribution of X for a batch of products. Assume that X follows the
normal distribution with mean xr and the standard deviation r. We further assume
that the process capability index Cp 3r
tr
is a constant, which is larger than or equal
to 1 (for process capability indices, see Sect. 14.4). Letting A 3C1 2 , we have
p
r2 Atr2 : 8:26
For a given value of x, the loss function is given by Eq. (8.11). For a batch of
products, the average loss is given by
Z1
K1 K2 2
Ltr Lxf xdx r KAtr2 8:27
2
1
where K K1 K2 =2.
Under the assumption that CM and Ltr are mutually independent, the total cost
is given by
X
n
CT ci ti KAtr2 : 8:28
i1
The optimal tolerances can be obtained by minimizing the total cost subject to the
constraints given by (8.23).
Example 8.3 Consider an assembly consisted of a shaft (x1 ) and a hole (x2 ). Design
variables are the tolerances of the shaft and hole, i.e., t1 and t2 . The clearance is
p
xr x2 x1 and the tolerances of xr is given by tr t12 t22 . The lower bound of
t1 and t2 is 0.05 mm; and the upper bound of tr is 0.2 mm.
The empirical data for manufacturing costs are shown in Table 8.9. It is found
that the negative exponential model in Eq. (8.24) is suitable for tting the data and
the parameters are shown in the last three rows of Table 8.9.
Table 8.9 Empirical data for t1 (mm) Cost t2 (mm) Cost

manufacturing costs
0.008 9.52 0.010 16.19
0.024 8.09 0.020 11.72
0.075 3.76 0.039 9.13
0.150 2.59 0.075 4.01
0.250 1.81 0.150 3.24
0.250 2.59
a1 1.7771 a2 2.7036
b1 9.1995 b2 17.7806
c1 18.3191 c2 29.8213
Assume that Cp 1, i.e., A 1=9. When tr \0:2, excess clearance has a larger
loss than insufcient clearance. Therefore, assume that K1 130 and K2 520 so
that K 325 and Ltr 36:11tr2 . As a result, the total cost is given by
CT KAt12 t22 a1 b1 ec1 t1 a2 b2 ec2 t2 :
The optimal solution is t1 0:13 and t2 0:15, and corresponds to tr 0:20
and CT 6:86. It is found that the solution is insensitive to the value of Cp in this
example.
References
1. Chan LK, Wu ML (2002) Quality function deployment: a literature review. Eur J Oper Res
143(3):463497
2. Ganeshan R, Kulkarni S, Boone T (2001) Production economics and process quality: a
Taguchi perspective. Int J Prod Econ 71(1):343350
3. Han CH, Kim JK, Choi SH (2004) Prioritizing engineering characteristics in quality function
deployment with incomplete information: a linear partial ordering approach. Int J Prod Econ
91(3):235249
4. Hauser JR, Clausing D (1988) The house of quality. Harv Bus Rev 66(3):6373
5. Keeney RL (1974) Multiplicative utility functions. Oper Res 22(1):2234
7. Roy RK (2001) Design of experiments using the Taguchi approach: 16 steps to product and
process improvement. Wiley, New York
8. Taguchi G (1986) Introduction to quality engineering. Asian Productivity Organization,
Tokyo
9. Wu CC, Chen Z, Tang GR (1998) Component tolerance design for minimum quality loss and
manufacturing cost. Comput Ind 35(3):223232
10. Yacout S, Boudreau J (1998) Assessment of quality activities using Taguchis loss function.
Comput Ind Eng 35(1):229232
Chapter 9
Design Techniques for Reliability
9.1 Introduction
Design for reliability (DFR) is a process to ensure that customer expectations for
reliability are fully met. It begins early in the concept stage and focuses on iden-
tifying and designing out or mitigating potential failure modes. Many reliability
activities are conducted to determine, calculate, and achieve the desired reliability.
Main reliability-related issues considered in the design stage include
Specication of reliability requirements
Reliability analysis
Reliability prediction
Reliability allocation, and
Reliability improvement.
In this chapter, we focus on these issues.
The outline of the chapter is as follows. In Sect. 9.2, we discuss the process to
implement DFR. Sections 9.39.7 deal with each of the above-mentioned issues,
respectively. Finally, we briefly discuss reliability control and monitoring in
manufacturing and usage phases in Sect. 9.8.
9.2 Process of Design for Reliability
DFR is a well-dened process to incorporate various reliability activities into the

design cycle so that reliability is designed in products and processes using appro-
priate methods or tools. The DFR process involves the following steps:
Step 1: Specify reliability requirements and goals of the product based on
customers needs, environment, and usage conditions.

148 9 Design Techniques for Reliability
Step 2: Carry out a qualitative reliability analysis to identify key reliability risks
and risk reduction strategies.
Step 3: Perform a reliability prediction so as to quantitatively evaluate design
options and identify the best.
Step 4: Allocate the product reliability requirements to its components (or failure
modes).
Step 5: Achieve the desired reliability using various reliability improvement
strategies such as deration, redundancy, preventive maintenance, and reliability
growth by development.
The DFR process continues to the manufacturing phase to control the reliability
of the produced product, and to the usage phase to monitor the eld reliability of the
product so as to obtain necessary information for further design or process changes.
9.3 Reliability Requirements
In this section, we look at the following two issues:

Determining system-level reliability requirements, and
Allocating the system-level reliability requirements to the components.
Required product reliability depends on the usage rate, operating environment,
the voice of the customer (VOC [1]), and many other variables. The environmental
and usage conditions can be determined based on surveys and measurement. The
VOC can be obtained from contracts, competitive analysis, and other consider-
ations. The methods to quantify the VOC include the Kano models [2], afnity
diagrams [3], and pair-wise comparisons (see Online Appendix A).
The reliability requirements of a product are stated for some nominal usage
condition and depend on whether the product is repairable or not. For nonrepairable
products, the reliability goal is usually a minimum value of required reliability in a
given time interval, or a minimum value of required MTTF, or a maximum value of
allowed failure rate; for repairable products, the reliability goal is usually a mini-
mum value of required availability, a maximum number of failures in a given time
interval, or a minimum value of required MTBF or MTTF.
Once the product (or system-level) reliability requirements are specied, these
requirements should be further allocated down to subsystem level, assembly level,
component level, and even to failure mode level to ensure that the product reli-
ability requirements are met. This deals with reliability allocation.
The allocation process starts from a few design options, each of which can be
decomposed into several hierarchical levels. A qualitative analysis (such as an
FMEA analysis) is rst carried out to evaluate each option so as to identify feasible
design options.
A quantitative analysis is then carried out for each feasible design option. The
quantitative analysis involves system reliability analysis and reliability prediction.
9.3 Reliability Requirements 149
The main purpose is to compare feasible design options and prepare necessary
information for reliability allocation.
Once the quantitative analysis is carried out, the system-level reliability
requirements are allocated to the elements of the other hierarchical levels. The
specied reliability requirements at each hierarchical level then are further trans-
lated into manufacturing requirements using the QFD discussed in the previous
chapter.
For more details about the reliability allocation process, see Refs. [46].
9.4 Reliability Analysis
Reliability analysis can be qualitative or quantitative. Two typical qualitative

methods used in DFR are change point analysis and FMEA; and a typical quan-
titative method is system reliability analysis using reliability block diagram or/and
fault tree diagram.
9.4.1 Change Point Analysis
A product can be
a completely new product,
an upgrade of an existing product,
an existing product introduced to a new market or application, or
a product existing in the market but being new to the company.
Different types of product result in different changes in design, manufacturing,
usage environment, performance requirements, and so on. Changes imply risks, and
hence a thorough change point analysis will help to identify and understand design
and/or application changes introduced with this new product and associated risks in
a qualitative way.
9.4.2 FMEA
FMEA connects given initiating causes to their end consequences. For a given
design option, the main objectives of an FMEA are
to identify the items or processes to be analyzed,
to identify their functions, failures modes, causes, effects, and currently used
control strategies,
to evaluate the risk associated with the issues identied by the analysis, and
to identify corrective actions.
In the FMEA, functional analysis plays a key role in identications of potential

failures. It helps to understand various functions and associated performance criteria
of the system and its each functional block, and to identify interrelationships
between the functions.
Risk assessment is an important issue in the FMEA. Risk is dened as P S,
where P is the likelihood or frequency of occurrence for a given cause of failure and
S is the consequence (or severity) of effect of failure. Considering the possibility of
detection of the cause of failure, risk priority number (RPN) is dened as
RPN P S D 9:1
where D is the probability that the current control scheme cannot detect or prevent
the cause of failure.
FMEA can be extended to FMECA to include a criticality analysis. MIL-STD-
1629A [7] presents the procedures for conducting a FMECA. For each potential
failure mode, a criticality matrix is established to evaluate risk and prioritize cor-
rective actions. The horizontal axis of the matrix is the severity of the potential
effects of failure and the vertical axis is the likelihood of occurrence. For each
potential failure and for each item, a quantitative criticality value is calculated based
on failure probability analysis at a given operating time under the constant failure
rate assumption.
SAE J1739 [8] divides FMEA into design FMEA, process FMEA, and
machinery FMEA. The design FMEA is used to improve designs for products and
processes, the process FMEA can be used in quality control of manufacturing
process, and the machinery FMEA can be applied to the plant machinery and
equipment used to build the product. RCM is actually a systematic application of
the machinery FMEA.
9.4.3 System Reliability Analysis
Many applications (e.g., risk assessment, reliability prediction, etc.) require carrying
out a system reliability analysis. In system reliability analysis, system failure is
modeled in terms of the failures of the components of the system. There are two
different approaches to link component failures to system failures. They are bottom-
up and top-down approaches, respectively.
In the bottom-up approach, one starts with failure events at the component level
and then proceeds to the system level to evaluate the consequences of such failures
on system performance. FMEA uses this approach.
In the top-down approach, one starts at the system level and then proceeds
downward to the component level to link system performance to failures at the
component level. Fault tree analysis (FTA) uses this approach. A similar graphical
tool is reliability block diagram (RBD). In FTA or RBD, the state of the system can
be expressed in terms of the component states through the structure function. The
9.4 Reliability Analysis 151
difference between FTA and RBD is that the former is failure-oriented and the latter
is success-oriented.
9.4.3.1 Fault Tree Analysis
A fault tree is composed of basic events, top event, and logic gates. The basic
events are the bottom events of the fault tree, the top event is some particular system
failure mode, and the gates serve to permit or inhibit the passage of fault logic up
the tree. The inputs of the gate are the lower events, and the output is a higher event.
As such, the gates show the relationships between the input events and the output
event, and the gate symbol denotes the type of relationship. A fault tree shows the
logical interrelationships of basic events that lead to the top event.
A cut set is a combination of basic events that can cause the top event. A
minimal cut set is the smallest combination of basic events that result in the top
event. All the minimal cut sets for the top event represent the ways that the basic
events can cause the top event. Through identifying all realistic ways in which the
undesired event can occur, the characteristics of the top event can be calculated.
The fault tree includes only those faults that contribute to this top event and are
assessed to be realistic. It is often used to analyze safety-related systems.
9.4.3.2 Reliability Block Diagram
In a RBD, the system is divided into blocks that represent distinct elements
(components or modules). These elemental blocks are then combined according to
system-success pathways. Each of the blocks is often comprised of units placed in
series, parallel, or a combination of both. Based on the RBD, all system-success
pathways are identied and the overall system reliability can be evaluated.
A RBD is developed for a given system function. If the system has more than
one function, each function must be considered individually. The RBD cannot
effectively deal with complex repair and maintenance strategies and hence the
analysis is generally limited to the study of time to the rst failure.
9.4.3.3 Structure Function
Structure function is a mathematical representation of the reliability structures of

system. It links component reliability to system reliability. Both system and its
components are characterized as being in one of two statesworking or failed.
Let XS t denote the state of the system at time t, and Xi t, 1 i n, denote the
state of component i at time t, and Xt X1 t; X2 t; . . .; Xn t denote the state
of the n components at time t. Xs t 1 [Xi t 1] when the system [component i]
is in the working state; otherwise Xs t 0 [Xi t 0] when it is in the failed state.
The state of the system is given by a function uXt, which is called the
structure function with
XS t uXt: 9:2
The form of uX depends on the RBD. The reliability structure of most systems
can be represented as a network involving series, parallel, and k-out-of-n connec-
tions. For the system with series structure, the system fails whenever a component
fails. In this case, the structure function is given by
Y
n
uX Xi : 9:3
i1
For the system with parallel structure, the system fails only when all the com-
ponents fail. In this case, the structure function is given by
Y
n
uX 1 1 Xi 9:4
i1
For the k-out-of-n system, the system is functioning

P if at least k of n (identical or
similar) components are functioning. Let y ni1 Xi , which represents the number
of components in working state. The structure function is given by

1; if y k
uX : 9:5
0; if y\k
Specially, when k 1, the k-out-of-n system reduces into the system with parallel
structure; and when k n, the k-out-of-n system reduces into the system with series
structure. In the latter case, the components are not necessary to be identical or similar.
A component is said to be irrelevant if the system state is not affected by the state
of the component; and a system is said to be coherent if it does not have irrelevant
components.
Example 9.1 Consider a two-out-of-three system. In this case, y X1 X2 X3 .
The event y 2 corresponds to the following four mutually independent events:
no component fails so that Xs X1 X2 X3 ,
the third component fails so that Xs X1 X2 1 X3 ,
the second component fails so that Xs X1 1 X2 X3 , and
the rst component fails so that Xs 1 X1 X2 X3 .
As a result, the structure function of the system is given by their sum, i.e.,
/X X1 X2 X1 X3 X2 X3 2X1 X2 X3 : 9:6
9.4 Reliability Analysis 153
9.4.3.4 Relation Between Structure Function and Reliability Function
The reliability function of a system can be derived from its structure function.
Assume that component failures are statistically independent, and components are
new and working at t 0. To be simple, we focus on the distribution of time to the
rst failure of the system.
The reliability functions of the components are given by
Ri t PrfXi t 1g 9:7
for 1 i n. The reliability function of the system is given by

RS t PrfXS t 1g: 9:8
Let FS t and Fi t denote the failure distributions for the system and component i,
respectively. We have RS t 1 FS t and Ri t 1 Fi t. Since the compo-
nent and system states are binary valued, we have
RS t PrfuXt 1g EuXt: 9:9
This can be written as
RS t EuXt uEXt upt 9:10
where pt is the vector R1 t; R2 t; . . .; Rn t. As a result, we have
FS t 1 upt: 9:11
For the system with series structure, the system reliability function is the com-
peting risk model by Eq. (4.29). For the system with parallel structure, the system
distribution function is the multiplicative model by Eq. (4.32). For the k-out-of-n
system with n identical components, the system reliability function is given by
X
n
RS t Cn; xpx 1 pnx ; p Ri t 9:12
xk
where Cn; x is the number of combinations choosing x items from n items.

Example 9.2 Consider a two-out-of-n system with p Ri t ekt and n 2; 3
and 4, respectively. The reliability function of the system is
RS;2 t e2kt for n 2, which is an exponential distribution. This two-out-of-
n system is actually a system with series structure.
RS;3 t RS;2 t3 2ekt for n 3, whose density function is unimodal and
failure rate is increasing.
RS;4 t RS;2 6 8ekt 3e2kt for n 4, whose density function is also
unimodal and failure rate is also increasing.
0.8
RS,n (t ) 0.6
n =4
0.4 n =3
n =2
0.2
0
0 1 2 3
t
Fig. 9.1 Plots of RS;n t for Example 9.2
This implies that the k-out-of-n system with exponential components can be aging.
For k 1, Fig. 9.1 shows the plots of RS;n t. As seen, the system reliability gets
considerably improved as n increases. In fact, the B10 life equals 0.0527, 0.2179,
and 0.3863 for n 2, 3 and 4, respectively. As a result, B10 n 3=B10 n 2
4:1 and B10 n 4=B10 n 2 7:3.
9.5 Reliability Prediction
The overall product reliability should be estimated early in the design phase.
The reliability prediction uses mathematical models and component reliability data to
estimate the eld reliability of a design before eld failure data of the product are
available. Though the estimates obtained from reliability prediction can be rough
since real-world failure data are not available, such estimates are useful for identifying
potential design weaknesses and comparing different designs and their life cycle costs.
The reliability prediction requires knowledge of the components reliabilities,
the design, the manufacturing process, and the expected operating conditions.
Typical prediction methods include empirical method, physics of failure method,
life testing method, and simulation method (e.g., see Refs. [9, 10]). Each method
has its advantages and disadvantages and can be used in different situations. We
briefly discuss them as follows.
9.5.1 Empirical Methods
The empirical methods (also termed as the part count approach) are based on the
statistical analysis of historical failure data collected in the eld, and can be used to
quickly obtain a rough estimation of product eld reliability. The empirical methods
assume:
9.5 Reliability Prediction 155
The product is comprised of several independent components in series,

Failures of the components are mutually independent, and
The prediction only considers the normal life period with a constant failure rate.
For the electronic products, the infant mortality failure rate associated with the
bathtub curve can be eliminated by improving the design and production processes,
and the wear-out period is never reached due to quick advances in technology. As
such, only the normal life period needs to be considered. In this case, the reliability
prediction is equivalent to predicting the failure rate at the steady-state stage or the
failure rate associated with random failure mode.
In almost all empirical methods, the predicted failure rate is composed of two
parts: basic part and corrected part. In the MIL-HDBK-217 predictive model [11,
12], the basic part is the failure rate under the reference conditions (i.e., typical or
average operational conditions), and the influence of actual operational conditions
is represented by a set of factors. As such, the failure rate under specic operating
conditions is predicted as:
X
n
k kb;i pS pT pE pQ pA 9:13
i1
where kb;i is the basic failure rate of the ith component, which usually come from the
reliability databases or handbooks for similar components, pS ; pT ; pE ; pQ and pA
are revising coefcients to reflect the effects of stresses, temperature, environment,
quality specications, and components complexity, respectively. The coefcient is
1 if the actual conditions are consistent with the reference conditions; otherwise,
larger or smaller than 1.
There are differences between different prediction methods. For example,
Telcordia predictive method [13, 14] allows combining the historical data with data
from laboratory tests and eld tracking data, and its corrected part only considers
the quality factor, electrical stress factor and temperature stress factor. The corrected
part of RDF 2000 method [15] is based on the mission proles, which compre-
hensively reflect the effect of mission operational cycling, ambient temperature
variation and so on.
The empirical methods are simple and can be used in early design phases when
information is limited. However, the data may be out-of-date and it is hard to
adequately specify the revising coefcients.
9.5.2 Physics of Failure Analysis Method
A physics of failure model relates the life characteristic of a component to the

stresses (e.g., humidity, voltage, temperature, etc.), and the reliability is predicted
based on such models. The model often contains the parameters to be specied and
the parameters can be determined from design specications or from test data.
When the component has multiple failure modes, the components failure rate is the
sum of the failure rates of all failure modes. Similarly, the systems failure rate is
the sum of the failure rates of the components involved. Several popularly models
are briefly outlined as follows (for more details, see Ref. [16] and the literature cited
therein).
Arrhenius model. This model describes the relation between the time-to-failure
and temperature. It is based on the phenomenon that chemical reactions can be
accelerated by increasing the operating temperature. The model is given by
LT AeEa =Tk 9:14
where LT is the life characteristic (e.g., MTBF or median life), T is the absolute
temperature, k is the Boltzmann constant (1=11605), A is a constant to be
specied, and Ea is the activation energy, which depends on the product or material
characteristics.
Eyring model. The Eyring model is given by
LT; S AT a eBS=T 9:15
where LT; S is the life characteristic, T is the absolute temperature, S is another

stress (e.g., mechanical stress, humidity, or voltage), A, B and a are constants.
The Eyring model can be viewed as an extension of the Arrhenius model and has
several variants. When the stress is the voltage (V), the life-stress relation is an
inverse power function of V, given by
LT; V LTV b 9:16
where LT is given by Eq. (9.14). The model given by Eq. (9.16) can be extended
to include the third stress (e.g., humidity, H) with the inverse power relation
given by
LT; V; H LT; VH c : 9:17
where LT; V is given by Eq. (9.16), and c is a constant. A variant of Eq. (9.17)
(also termed as the corrosion model) is given by
LT; V; H LTf VeaH 9:18
where LT is given by Eq. (9.14), f V is a function of V, and a is a constant.

Model for fatigue failure. Fatigue failures can occur due to temperature cycling
(represented by the cycling frequency f ) and thermal shock (represented by the
maximum temperature Tmax and temperature range during a cycle DT ). Each stress
cycle produces damage to the item, and the damage is accumulated. The item fails
9.5 Reliability Prediction 157
when the cumulative damage exceeds its critical value. The number of cycles to
failure is given by
LTmax
Nf 9:19
f a DbT
where LTmax has the form of Eq. (9.14), a and b are constants.
The physics of failure methods can provide accurate prediction, and needs
detailed component manufacturing information (e.g., material, process, and design
data) and operational condition information (e.g., life cycle load prole). Due to the
need for detailed information, complex systems are difcult to be modeled physi-
cally and hence it is only applicable for components.
The physics of failure models have important applications in accelerated life test
design and data analysis. This will be further discussed in the next chapter.
9.5.3 Life Testing Method
Life testing methods are used to determine reliability by testing a relatively large
sample of units operating under normal or higher stresses. The data can be tted to
an appropriate life distribution using statistical methods discussed in Chap. 5, and
reliability metrics can be estimated from the tted life distribution model.
The prediction results obtained from the life testing method are usually more
accurate than those from the empirical method since the prediction is based on
failure data from particular products.
The life testing method is product-specic. It is particularly suited to obtain
realistic predictions at the system level because the prediction results at the system
level obtained from the empirical and physics of failure methods may be inaccurate
due to the fact that the assumptions can be unrealistic. However, the life testing
method can be costly and time-consumed.
9.5.4 Simulation Method
In some situations (e.g., dealing with large systems or correlated failures, or

requiring testing a highly reliable item to failure), the above-mentioned prediction
methods are complex or too expensive. In this case, simulation is an effective tool to
simplify the prediction. For more details about reliability simulation, see Ref. [17].
9.6 Reliability Allocation
Reliability allocation aims to establish target reliability for each level in product
structure. It rst allocates the entire target reliability of a product to its subsystem,
and then allocates the sub-target reliability of each subsystem to its components.
Similar to reliability prediction, the underlying assumptions used in reliability
allocation are:
(a) all components in the system are in series,
(b) components failures are mutually independent, and
(c) failure rates of the components are constant.
The allocation methods depend on whether the system is nonrepairable or
repairable. We separately discuss these two cases as follows.
9.6.1 Reliability Allocation Methods

for Nonrepairable Systems
Several allocation methods available for nonrepairable systems include equal

apportionment method, ARINC method and AGREE method (e.g., see Ref. [18]).
The equal apportionment method takes the same reliability for each subsystem.
Let Rs denote the target reliability of the system, Ri denote the reliability allocated
to the ith subsystem, and n denote the number of subsystems. Letting Ri R0 ,
we have
Y
n
Rs Ri Rn0 ; Ri R1=n
s : 9:20
i1
The ARINC method is developed by Aeronautical Radio, Inc. It assumes that

current failure rates of the subsystems are known (obtained from existing failure
data or prediction standards) and the following inequality holds
X
n
c
ki [ ks 9:21
i1
c
where ki is the current failure rates of subsystem i and ks is the required system
failure rate. To reach the system failure rate goal, some improvement efforts must
9.6 Reliability Allocation 159
c
be made to reduce ki to ki . The ARINC method reduces the current failure rates by
equal percentages. The required failure rate reduction factor is given by
X
n
c
r ks = ki \1: 9:22
i1
As such, ki is calculated as
c
ki rki : 9:23
The AGREE method is developed by Advisory Group on Reliability of

Electronic Equipment. The method takes into consideration of the complexity (in
terms of number of components in each subsystem) and importance of each sub-
system (in terms of importance factor in 0 and 1). The method includes two steps.
In the rst step, only the complexity is considered; and in the second step, the
importance of each subsystem is considered.
We rst look at the rst step. All the subsystems are assumed to be equally
important and equal failure rates are allocated to all components in the system. Let
Rs s denote the required system mission reliability in the mission time (or oper-
ating time) s, and ks denote the required system failure rate, Since the required
system reliability at s is Rs s eks s , the required system failure rate is given by
ks lnRs s=s.
Letting ni denote the number of components P of the ith subsystem, then the total
number of components of the system is N ni1 ni . Let k0 denote the failure rate
allocated to each component, which is given by
ks lnRs s
k0 : 9:24
N sN
As such, the failure rate allocated to the ith subsystem is given by
ki ni k0 : 9:25
Equation (9.25) indicates that the failure rate allocated to each subsystem is pro-
portional to the number of components it contains (i.e., complexity).
We now look at the second step. Let wi (2 0; 1) denote the importance of the
w
ith subsystem and ki denote the failure rate allocated to the ith subsystem after
considering subsystem importance. The importance is subjectively determined
based on experts judgments. If subsystem i is important, a large value is assigned
to wi ; otherwise, a small value is assigned.
If wi 1, high subsystem failure rate cannot be tolerated; if wi 0, a subsystem
failure actually has no impact on the system. This implies that the allocated failure
rate should be inversely proportional to wi [19]. Based on this idea, the failure rate
adjusted after considering the importance is given by
w
X
n
ki
ki kki =wi ; k ks = : 9:26
i1
wi
Example 9.3 Consider a safety-related system with system reliability goal being
103 failure per year. The system comprises three subsystems, and the complexity
and importance information of the subsystems are shown in the second and third
columns of Table 9.1, respectively. The problem is to allocate the target failure rate
to each subsystem.
For this example, N 7 and k0 ks =N 0:1429 103 . The values of ki

(i 1; 2; 3) are shown in the fourth column. Combining the values of ki with the
importance factors, we have k 0:4774. The adjusted failure rates are shown in the
last column of Table 9.1.
9.6.2 Reliability Allocation Methods for Repairable Systems
The reliability goal of repairable systems can be described by availability, number

of failures in a given time interval, MTBF, or MTTF. The techniques to achieve the
reliability goal include reducing the failure rate and/or the downtime. Downtime
can be reduced by maintainability design, eliminating logistic delays, or/and
developing improved repair methods. The outputs of allocation can be failure rate,
repair rate (i.e., 1/MTTR) or availability of subsystems.
Though most systems are repairable, there are few allocation methods available
for repairable systems. The most popular method is the repairable systems allo-
cation method, which is somehow similar to the equal apportionment method used
for nonrepairable systems. It assigns equal availabilities for all subsystems so that
the system availability goal can be reached. In this subsection, we focus on this
method. For other methods, see Refs. [19, 20] and the literature cited therein.
Assume that the system availability goal is As . Let li and si denote the mean
time to failure and the mean time to repair for the ith subsystem, respectively.
Under the exponential distribution assumption, the failure rate (ki 1=li ) and
repair rate (1=si ) are constants. Assume that si is known. The problem is deter-
mine the values of ki s.
Table 9.1 Failure rates allocated to subsystems for Example 9.3

Subsystem Number of components Importance factor ki ; 103 w
ki ; 103
1 2 0.56 0.2857 0.2436
2 2 1.00 0.2857 0.1364
3 3 0.33 0.4286 0.6200
9.6 Reliability Allocation 161
The availability of the ith subsystem is given by
1
Ai : 9:27
1 ki s i
The equal-availability implies that the ki si is a constant. Let k denote this constant,
which will be derived later. As such, the failure rate allocated to the ith subsystem is
given by
k
ki : 9:28
si
We now derive the expression of k. The failure rate for the system is given by
X
n
ks ki : 9:29
i1
The mean time to failure for the system is given by ls 1=ks . The expected
downtime for each system failure is given by
X
n
ki nk
ss si : 9:30
i1
ks ks
Noting ks ss nk, we have the system availability given by
1 1
As : 9:31
1 ks ss 1 nk
As a result, from Eq. (9.31) we have
1=As 1
k : 9:32
n
Example 9.4 Consider the system discussed in Example 9.3 and assume that the
reliability goal is As 0:999. Mean times to repair information of subsystems are
shown in the second row of Table 9.2. The problem is to determine the failure rate
of each subsystem.
From Eq. (9.32), we have k 0:3337 103 ; and from Eq. (9.28), we have the
values of ki shown in the last row of Table 9.2.
Table 9.2 Failure rates Subsystem 1 2 3

allocated to subsystems for
Example 9.4 si , 103 1 0.5 2
ki , 103 0.3337 0.6673 0.1668
9.7 Techniques to Achieve Desired Reliability
When the predicted reliability is poorer than the required reliability target, the
reliability must be improved by design and development. The techniques to achieve
the desired reliability include component deration and selection, redundancy
design, preventive maintenance (including condition monitoring), and reliability
growth through development. We separately discuss these techniques as follows.
9.7.1 Component Deration and Selection
The probability of failure of the product can be decreased by limiting its maximum
operational and environmental stresses (e.g., temperature, pressure, etc.) to the level
below the capabilities of the components or by adopting the components with larger
capabilities. The former is called the component deration and the latter is called the
component selection. The criteria for the components (or materials) selection are
components reliability and its ability to withstand the expected environmental and
operational stresses. Components with better load bearing ability are preferred.
Selection of high performance components or materials increases service life and
reduces the maintenance cost but increases manufacturing cost. As such, the
selection decisions can be optimally made based on a life cycle costing analysis.
The component deration and selection require the information on component
reliability and operational and environmental stresses, and the failure probability
can be quantitatively evaluated using the stressstrength interference model if
component failure is due to overstress mechanism.
Because of manufacturing variability, the strength of a component, X, may vary
signicantly. For example, fracture and fatigue properties of engineering material
usually exhibit greater variability than the yield strength and the tensile strength. As
such, the strength is a random variable with distribution [density] function FX x
[fX x]. When the component is put into use, it is subjected to a stress, Y, which is
also a random variable. Let FY y [fY y] denote the distribution [density] function
of Y. If X is larger than Y, then the strength of the component is sufcient to
withstand the stress, and the component is functional. When a shock occurs, the
stress may be larger than the strength so that the component fails immediately
because its strength is not sufcient to withstand the stress to which it is subjected.
Assume that X and Y are independent. The reliability R that the component will not
fail when put into operation can be obtained using a conditional approach.
Conditional on Y y, we have
PfX [ YjY yg 1 FX y: 9:33

9.7 Techniques to Achieve Desired Reliability 163
On removing the conditioning, we have
Z1
R PfX [ Yg 1 FX yfY ydy: 9:34
0
Optionally, Eq. (9.34) can also be written as
Z1
R PfY\Xg FY xfX x dx: 9:35
0
Usually, one needs to use numerical methods to evaluate the integral in

Eq. (9.34) or (9.35). However, when both X and Y follows the normal distribution,
the reliability is given by R PrX Y 0. Let Ul1 ; r1 denote the stress dis-
tribution and Ul2 ; r2 denote the strength distribution. Then, the reliability can be
directly evaluated by
q
R 1 U 0; l2 l1 ; r21 r22 : 9:36
Clearly, the reliability increases as r2 decreases (i.e., the component has small
variability in strength) and as l2 l1 increases (i.e., the component has large
ability or margin of safety).
In the above discussion, the distributions of stress and strength are assumed to be
independent of time. However, the strength Xt usually degrades with time so that
it is nonincreasing and the stress Yt can change with time in an uncertain manner.
In this case, the time to failure (T) is the rst time instant that Xt falls below Yt,
i.e.,
T minftjXt Yt\0g 9:37
Different characterizations of Xt and Yt lead to different models. For more

details, see Sect. 17.6 of Ref. [21].
Example 9.5 Assume that the strength X is a constant x and the stresses occur at
random points in time due to external shocks. The shocks occur according to a
stationary Poisson process with k 2:5, and the stress Y resulting from a shock is a
random variable with distribution Gy, which follows the lognormal distribution
with ll 1:2 and rl 0:8, 1.0 and 1.5, respectively. The problem is to examine
the effect of the strength on the failure rate.
We rst derive the expression of the failure rate function and then study the
effect of the strength on the failure rate. Let T denote the time to failure and Ft
denote its distribution function. The probability for the item to survive to n shocks
is given by
PfT [ tjNt ng Gxn : 9:38
The reliability function is given by
X
1
Rt PfT [ tg pn Gxn 9:39
n0
where
ktn ekt
pn : 9:40
n!
Using Eq. (9.40) to Eq. (9.39) and after some simplication, we have
Rt expk1 Gxt: 9:41
This implies that the time to failure is given by an exponential distribution with the
failure rate function
rt k1 Gx: 9:42
Since Gx increases as x increases (i.e., the component becomes stronger), the

failure rate decreases as the strength increases and k decreases (shocks occur less
frequently).
Figure 9.2 shows the plots of the failure rate function (where ly ell rl =2 is the
2
mean of the stress). As seen, the failure rate quickly decreases as the strength
increases. For large x=ly (e.g., [ 2), the failure rate decreases as the dispersion of
2.5
l =0.8
2
l =1.0
1.5
(x )
0.5 l =1.5
0
0 1 2 3 4 5
x / y
Fig. 9.2 Effect of strength on failure rate function

9.7 Techniques to Achieve Desired Reliability 165
the stress (represented by rl ) decreases. This illustrates that the failure rate in the
normal usage phase can be controlled through design.
9.7.2 Redundancy
Redundancy uses several identical or similar components to perform the same

function within the product. It has been extensively used in electronic products and
safety systems to achieve high reliability when individual components have low
reliability.
As mentioned in Chap. 2, there are three basic types of redundancy: hot standby,
cold standby, and warm standby. In the hot standby, all the units simultaneously
work to share the load so that each unit is derated and its life is longer than the
average time obtained when they are separately used. In the cold standby, the
average time of the system is the sum of the mean lifetimes of all the units but a
sensor is needed to sense the failure and a switching mechanism is needed to
replace the failed unit by one standby unit (if available). Both the sensor and the
switching mechanism may fail. In the warm standby, the standby units work in
partial load so that its mean life is larger than the mean life of the unit when it works
in full load. When the unit operating in full load fails, the load of the standby units
are adjusted from partial load to full load. A k-out-of-n system is a type of more
general redundancy.
The redundancy design needs to determine the redundancy type and the values
of k and n, which influence the complexity, reliability, product and production cost.
Usually, constraints (e.g., weight, volume, etc.) are imposed to control the com-
plexity. In addition, the design of a redundancy system needs to consider the
possibility of common cause failures, and diversity of redundancy ways can reduce
common cause failures.
9.7.3 Preventive Maintenance
The component reliability target can be achieved through the use of preventive
maintenance. Typical preventive maintenance actions include inspection, replace-
ment, and condition monitoring.
The failure of safety-related components can be non-self-announced. In this
case, periodic inspection must be performed to check their state. The key parameter
to be determined is the inspection interval (RCM calls it failure-nding interval). A
shorter interval leads to a smaller failure downtime, which is the time interval
between occurrence and identication of a failure. On the other hand, too frequent
inspections will lead to more production disturbances. As such, two issues for
scheduling the inspection are to optimize the inspection interval and to group the
inspection and preventive maintenance actions (e.g., see Ref. [22]).
Another way to improve reliability involves replacing the component preven-

tively. Preventive replacement is especially suitable for the components with high
failure rates. The key parameter to be determined is the preventive replacement
time.
Condition monitoring techniques can be used to detect whether specic failure
conditions (degradation or latent failure conditions) are present or not. When such
conditions are revealed, the problems are rectied and the failures are prevented.
Maintenance decision optimization problem will be discussed in detail in
Chap. 17.
9.7.4 Reliability Growth Through Development
In the development stage of a product, a reliability growth program is implemented

to improve the reliability of the product or its components. The reliability growth is
achieved by a test-analysis-and-x process. The process starts with testing the
product prototypes or components to failure in order to collect the failure data
(including modes of failure, time to failure, and other relevant information). Once
failures occur, the failed items are carefully examined and failure analysis (or root
cause analysis) is performed so as to discover the causes of failure and identify
proper corrective actions (i.e., x or design changes).
Reliability growth analysis can be used to decide if a reliability goal has been
met as well as whether and how much additional testing is required. The analysis
involves life test data analysis and calculation of reliability metrics. The ultimate
reliability of the product is inferred based on the test observations with considering
the effectiveness of taken corrective actions. The process is repeated until the
product reliability targets are achieved. Specic models and methods for conducting
these analyses will be discussed in detail in Chap. 11.
9.8 Reliability Control and Monitoring
The DFR efforts will continue to the manufacturing and usage phases. In this
section, we briefly discuss the reliability activities in these two phases.
9.8.1 Reliability Control in Manufacturing Process
When the product goes into production, the DFR efforts will focus primarily on
reducing or eliminating problems introduced by the manufacturing process, and
include the activities such as quality inspections, supplier control, routine tests,
9.8 Reliability Control and Monitoring 167
measurement system analysis, and so on. Relevant techniques associated with these
activities will be discussed in Chaps. 1215.
9.8.2 Reliability Monitoring in Usage Phase
Continuous monitoring and eld data analysis are needed to observe and analyze
the behavior of the product in its actual use conditions. The experiences and lessons
obtained from this process are useful for further improvements or in future projects.
Failure reporting, analysis and corrective action systems (FRACAS) is a tool
used to capture such knowledge throughout the product development cycle. The
basic functions of FRACAS include data reporting, data storage, and data analysis.
When a failure is reported, failure analysis is carried out to identify the root cause of
failure. Once this is done, the corrective action is identied using an appropriate
approach such as identify-design-optimize-validate approach or dene-measure-
analyze-improve-control approach. In such a way, FRACAS accumulates a lot of
information useful for resolving reliability-related issues during the product life
cycle. For example, a model for eld reliability can be obtained through failure data
analysis and the tted model can be used to predict expected failures under war-
ranty and demand of key spare parts. In addition, eld data analysis helps to
identify reliability bottleneck problems, which is useful for improving future gen-
erations of the same or similar product.
References
1. Grifn A, Hauser JR (1993) The voice of the customer. Marketing Sci 12(1):127
2. Kano N, Seraku N, Takahashi F et al (1984) Attractive quality and must-be quality. J Jpn Soc
Qual Control 14(2):3948
3. Straker D (1995) A toolbook for quality improvement and problem solving. Prentice Hall,
New York
4. Murthy DNP, Rausand M, Virtanen S (2009) Investment in new product reliability. Reliab
Eng Syst Saf 94(10):15931600
5. Murthy DNP, sters T, Rausand M (2009) Component reliability specication. Reliab Eng
Syst Saf 94(10):16091617
6. Murthy DNP, Hagmark PE, Virtanen S (2009) Product variety and reliability. Reliab Eng Syst
Saf 94(10):16011608
7. US Department of Defense (1980) Procedures for performing a failure mode, effects and
criticality analysis. MILHDBK1629A
8. Society of Automotive Engineers (2000) Surface vehicle recommended practice. J1739
9. Denson W (1998) The history of reliability prediction. IEEE Trans Reliab 47(3):321328
10. OConnor PDT, Harris LN (1986) Reliability prediction: a state-of-the-art review. In: IEEE
Proceedings A: Physical Science, Measurement and Instrumentation, Management and
Education, Reviews 133(4):202216
11. US Military (1992) Reliability prediction of electronic equipment. MIL-HDBK-217F Notice 1
12. US Military (1995) Reliability prediction of electronic equipment. MIL-HDBK-217F Notice 2
13. Telcordia (2001) Reliability prediction procedure for electronic equipment. SR-332 Issue 1
14. Telcordia (2006) Reliability prediction procedure for electronic equipment. SR-332 Issue 2
15. IEC TR (2004) 62380. Reliability data handbookuniversal model for reliability prediction of
electronics components. PCBs and equipment. International Electrotechnical Commission,
Geneva-Switzerland
16. Escobar LA, Meeker WQ (2006) A review of accelerated test models. Stat Sci 21(4):552577
17. Minehane S, Duane R, OSullivan P et al (2000) Design for reliability. Microelectron Reliab
40(810):12851294
18. US Military (1998) Electronic reliability design handbook, Revision B. MIL-HDBK-338B,
pp 619
19. Amari SV, Hegde V (2006) New allocation methods for repairable systems. In: Proceedings of
2006 annual reliability and maintainability symposium, pp 290295
20. Kuo W, Wan R (2007) Recent advances in optimal reliability allocation. IEEE Trans Syst Man
Cybernet Part A 37(2):143156
21. Murthy DNP, Xie M, Jiang R (2003) Weibull models. Wiley, New York
Beijing
Chapter 10
Reliability Testing and Data Analysis
10.1 Introduction
Different types of reliability tests are conducted at different stages of product

development to obtain information about failure modes and evaluate whether the
reliability goal has been achieved. To reduce the test time, tests are often conducted
at higher stress levels than those normally encountered. Such tests are called
accelerated tests. In this chapter, we focus on accelerated testing-related issues,
including relevant concepts, data analysis and modeling, and test design.
This chapter is organized as follows. We rst introduce various types of reli-
ability tests in product life cycle in Sect. 10.2. Accelerated testing and loading
schemes are discussed in Sect. 10.3. Accelerated life testing (ALT) models and
accelerated degradation testing (ADT) models are discussed in Sects. 10.4 and 10.5,
respectively. Finally, we discuss accelerated testing design in Sect. 10.6.
10.2 Product Reliability Tests in Product Life Cycle
According to the time when testing is conducted, testing can be grouped into three
categories: developmental testing, manufacturing testing, and eld operational
testing.

Development Stage
The tests carried out during product development stage focus on discovering failure
modes and improving reliability, and provide information on degradation and

170 10 Reliability Testing and Data Analysis
reliability of failure modes. The tests include development tests, reliability growth
tests, and reliability demonstration tests.
Development tests are conducted at material, part, and component levels, and can
be divided into performance testing and life testing. The performance testing
includes critical item evaluation and part qualication testing as well as environ-
mental and design limit testing; and the life testing includes testing to failure, ALT,
and ADT.
The critical item evaluation and part qualication testing is conducted at part
level. It deals with testing a part under the most severe conditions (i.e., maximum
operating stress level, which is larger than the nominal operating stress level)
encountered under normal use in order to verify that the part is suitable under those
conditions. The tests to be performed depend on the product. For example, for
electronic components the temperature and humidity tests are the most commonly
conducted tests.
The environmental and design limit testing is conducted at part, subsystem, and
system levels and at the extreme stress level (i.e., the worst-case operating condi-
tions with the stress level larger than the maximum operating stress level under
normal use). It applies environmentally induced stresses (e.g., vibration loading due
to road input for automotive components) to the product. The test can be conducted
using accelerated testing with time-varying load. These tests aim to assure that the
product can properly perform at the extreme conditions of its operating prole. Any
failures resulting from the test is analyzed through root cause analysis and xed
through design changes.
Life testing deals with observing the times to failure for a group of similar items.
In some test situations (e.g., one-shot devices), one observes whether the test item is
success or failure rather than the time of failure.
Reliability growth testing is conducted at system or subsystem level by testing
their prototypes to failure under increasing levels of stress. Each failure is analyzed
and some of the observed failure modes are xed. The corrective actions lead to
reduction in failure intensities, and hence reliability is improved.
Reliability demonstration testing is conducted at system or subsystem level. The
purpose is to demonstrate that the designed product meets its requirements before it
is acceptable for large volume production or goes into service. It deals with testing a
sample of items under operational conditions.
This chapter focuses on life testing-related issues and the reliability growth
testing-related issues are discussed in the next chapter.

Manufacturing Phase
Tests carried out during manufacturing phase are called manufacturing tests.
Manufacturing tests are used to verify or demonstrate nal-product reliability or to
10.2 Product Reliability Tests in Product Life Cycle 171
remove defective product before shipping. Some such tests include environmental
stress screening and burn-in. These are further discussed in Chap. 15.
10.2.3 Reliability Tests Carried Out During

Product Usage Phase
Field operational testing can provide useful information relating to product reli-
ability and performance in the real world. The testing needs the joint effort of the
manufacturer and users. A useful tool to collect eld reliability information is
FRACAS, which was mentioned in the previous chapter.
10.3 Accelerated Testing and Loading Schemes
10.3.1 Accelerated Life Testing
Accelerated testing has been widely used to obtain reliability information about a
product and to evaluate the useful life of critical parts in a relatively short test time.
There are two ways to accelerate the failure of a product [4]:
The product works under severer conditions than the normal operating condi-
tions. This is called accelerated stress testing.
The product is used more intensively than in normal use without changing the
operating conditions. This is called accelerated failure time. This approach is
suitable for products or components that are not constantly used.
Accelerated stress testing (also termed as ALT) is used for situations where
products are constantly used, such as the components of a power-generating unit.
Such tests are often used to evaluate the useful life of critical parts or components of
a system. Accelerated testing with the evaluation purpose is sometimes termed as
quantitative accelerated testing. In this case, the results in accelerated stress testing
are related to the normal conditions by using a stress-life relationship model. The
underlying assumption for such models is that the components operating under
normal conditions experience the same failure mechanism as those occurring at
accelerated stress conditions. As such, the range of the stress level must be chosen
from operational conditions to the maximum design limits. Since the results are
obtained through extrapolation, the accuracy of the inference depends strongly on
the adequacy of the stress-life relationship model and on the degree of extrapolation
(i.e., difference between the test stress and the normal stress). Compared with
accelerated stress testing, accelerated failure time testing is preferred since it does
not need the stress-life relationship for purpose of extrapolation.
If the purpose is to identify failure modes rather than to evaluate the lifetime,
very high stress can be used and the testing is called highly accelerated stress testing
or qualitative accelerated testing. Usually, a single stress is increased step-by-step
from one level to another until the tested item fails. While the test time can be
considerably reduced, the testing introduces new failure modes, and the interactions
among different stresses may be ignored (see Ref. [8]).
10.3.2 Accelerated Degradation Testing
For highly reliable components, it is infeasible to test the components to failures. If

the performance of the component slowly degrades, a failure can be dened when
the amount of degradation reaches a certain level. As such, instead of observing
time to failure, ADT observes degradation amount as a function of time. The
information of time to failure can be obtained through extrapolation using a specic
degradation process model.
10.3.3 Loading Schemes
The models or/and methods for ALT data analysis depend on the loading scheme.
According to the number of stresses and whether the stresses change with time,
there are the following three typical loading schemes (see Ref. [9]):
Single factor constant stress scheme. It involves only a single stress, each test
item experiences a xed stress level but different items can experience different
stress levels.
Multiple factors constant stress scheme. It involves several stresses and the
levels of stresses remain unchanged during testing.
Time-varying stress scheme. It involves one or more stresses and the stresses
change with time. A typical example is the step-stress testing.
Accelerated testing usually involves a single stress factor. In this case, there are
three typical stress test plans: constant stress test plan, step-stress test plan, and tests
with progressive censoring [11].
Under a constant stress test (see Fig. 10.1), the test is conducted at several stress
levels. At the ith stress level, ni items are tested. The test terminates when a
prespecied criterion (e.g., a prespecied test time or failure number) is met. Time
to failure depends on the stress level.
In the constant stress test plan, many of the test items will not fail during the
available time if the stress level is not high enough. The step-stress test plan can
avoid this problem. Referring to Fig. 10.2, items are rst tested at a constant stress
level s1 for t1 period of time. The surviving items will be tested at the next higher
level of stress for another specied period of time (i.e., ti ti1 ). The process is
10.3 Accelerated Testing and Loading Schemes 173
n3
s, f (t ) s3
n2
s2
n1 s1
t3 t2 t t1
Fig. 10.1 Constant stress test plan
s3
s2
s, F (t )
s1
F (t )
t1 t t2 t3
Fig. 10.2 Step-stress test plan
continued until a prespecied criterion is met. A simple step stress test involves
only two stress levels.
In tests involving progressive censoring (see Fig. 10.3), a fraction of the sur-
viving items are removed at several prespecied time instants to carry out detailed
studies relating to the degradation mechanisms causing failure (e.g., to obtain the
degradation measurement of a certain performance). Two typical progressive cen-
soring schemes are progressive type-I and type-II censoring schemes.
In the type-I censoring scheme, n items are put on life test at time zero, and ni of
Ki surviving items (ni \Ki ) are randomly removed from the test at the prespecied
censoring time ti 1 i m 1. The test terminates at the prespecied time tm or
at an earlier time instant when the last item fails. This is a general xed-time
censoring scheme, as the case shown in Fig. 10.3.
In the type-II censoring scheme, the censoring time ti is the time when the ki th
failure occurs. As such, the test terminates at the occurrence of the km th failure so
that the test duration is a random variable. This is a general xed-number censoring
scheme.
n1 n2 n3
t1 t2 t3 tm
Fig. 10.3 Type-I censoring scheme with progressive censoring
10.4 Accelerated Life Testing Data Analysis Models
Referring to Fig. 10.4, ALT data analysis involves two models. One is the distri-
bution model of lifetime of a product at a given stress level and the other is the
stress-life relationship model, which relates a certain life characteristic (e.g., mean
life, median life or scale parameter) to stress level.
10.4.1 Life Distribution Models
Depending on the test plan, the life distribution models can be a simple model or a
piecewise (or sectional) model. We separately discuss them as follows.
10.4.1.1 Simple Models
Let s0 denote the nominal stress level of a component in the normal use conditions.
The components are tested at higher stress levels s. The time to failure is a random
variable and depends on s so that the distribution function can be written as
Ft; s; h. Here, t is called the underlying variable, s is sometimes called covariate,
and h is the distributional parameters set. Some distributional parameters are
functions of stress while others are independent of stress.
Under the constant stress plan, the lifetime data obtained at different stress levels
are tted to the same distribution family (e.g., the Weibull and lognormal distri-
butions). For the Weibull distribution with shape parameter b and scale parameter
g, it is usually assumed that the shape parameter is independent of stress and the
scale parameter depends on stress. For the lognormal distribution with parameters
ll and rl , the variable can be written as
1=r
lnt ll t l
x ln : 10:1
rl l
el
As seen from Eq. (10.1), ell is similar to the Weibull scale parameter and 1=rl is
similar to the Weibull shape parameter. Therefore, it is usually assumed that rl is
10.4 Accelerated Life Testing Data Analysis Models 175
8
7
6
5
L (s )= (s )
s
4
3
2 f (t;s 0)
1
0
0 10 20 30 40 50 60 70 80 90
t
Fig. 10.4 ALT models
independent of stress, and ll is a function of stress. The life data analysis techniques
discussed in Chap. 5 can be used to estimate the distribution parameters.
10.4.1.2 Sectional Models
Consider the step-stress test plan with k stress levels. Let si (ti ti1 ) denote the
test duration at stress level si and Fi t denote the life distribution associated with
the constant stress test at si . Assume that Fi ts come from the same distribution
family Ft. Fi t can be derived based on the concept of initial age.
When the test begins, the test item is new so that the item has an initial age of
zero. Therefore, we have F1 t Ft for t 2 0; t1 . Now we consider F2 t
dened in t 2 t1 ; t2 . At t t1 , the surviving item is no longer new since it has
operated for t1 time units at s1 . Operating for t1 time units at s1 can be equiva-
lently viewed as operating for c2 (c2 \ t1 ) time units at s2 . The value of c2 can be
determined by letting
F1 t1 ; h1 F2 c2 ; h2 10:2
where hi is parameter set. When Fi t is the Weibull distribution with the common
shape parameter b and different scale parameter gi , from Eq. (10.2) we have
t1
c2 g : 10:3
g1 2
Similarly, for the lognormal distribution we have
c2 t1 expl2 l1 10:4
As a result, we have
F2 t Ft t1 c2 ; h2 : 10:5
Generally, ci i 2 is determined by
Fi1 ti1 ; hi1 Fi ci ; hi 10:6
and
Fi t Ft ti1 ci ; hi : 10:7
This is a k-fold sectional model with the cdf being continuous.
10.4.2 Stress-Life Relationship Models
The stress-life relationship model is used to extrapolate the life distribution of the
item at stress level s0 . It relates a life characteristic L to the stress level s. Let
Ls ws denote this model. Generally, ws is a monotonically decreasing
function of s.
Relative to the life characteristic (e.g., MTTF or scale parameter) at the normal
stress level, an acceleration factor can be dened as
di Ls0 =Lsi : 10:8
Using the stress-life relationship or acceleration factor model, the life distribu-
tion at stress level s0 can be predicted. The accuracy of the life prediction strongly
depends on the adequacy of the model. As such, the key issue for ALT data analysis
is to appropriately determine the stress-life relationship.
Stress-life relationship models fall into three broad categories [4]: physics of
failure models, physics-experimental models, and statistical models. Physics of
failure models have been discussed in Sect. 9.5.2. We look at the other two cate-
gories of models as follows.
A physics-experimental model directly relates a life estimate to a physical
parameter (i.e., stress). For example, the relation between the median life and
electric current stress is given by
t0:5 aJ b 10:9
where J is the current density. The relation between the median life and humidity
stress is given by
t0:5 aH b or t0:5 aebH 10:10
where H is the relative humidity. For more details about this category of models,
see Ref. [4].
The statistical models are also termed as empirical models. Three typical
empirical models are inverse power-law model, proportional hazard model (PHM),
and generalized proportional model. We discuss these models in the following three
subsections.
10.4.3 Inverse Power-Law Model
Let T0 and Ts denote the time to failure of an item at stress levels s0 and s [s0 ,
respectively. Ts is related to T0 by the inverse power-law relationship
s c
0
Ts T0 10:11
s
where c is a positive constant to be estimated. Especially, when c 1, it reduces to

the traditional ALT model, where the lifetime is inversely proportional to stress.
For the Weibull distribution, under the assumption that the shape parameter is
independent of stress, Eq. (10.11) can be written as
s c
0
gs g0 or lngs clns0 lns lng0 : 10:12
s
It implies that the plot of lngs versus lns is a straight line.

For the lognormal distribution, under the assumption that rl is independent of
stress, Eq. (10.11) can be written as
s c
ell;s ell;0 or ll;s clns0 lns ll;0 :
0
10:13
s
It implies the plot of ll;s versus lns is linear.
10.4.4 Proportional Hazard Model
The PHM is developed by Cox [2] for modeling the failure rate involving covar-
iates. Let Z zi ; 1 i k denote a set of covariates that affects the failure rate of
an item, Z0 z0i ; 1 i k denote the reference values of the covariates, and k0 t

(termed as the baseline failure rate) denote the failure rate function when Z Z0 .
The PHM assumes that the failure rate function at arbitrary Z is proportional to the
baseline failure rate, i.e.,
rt; Z k0 tuZ 10:14
where uZ is a function of Z. From Eq. (10.14), we have
lnrt; Z lnk0 t lnuZ: 10:15
To facilitate the linearization, uZ has the following two forms:

" # !
Xk X
k
uZ exp bi zi z0i exp b0 bi z i 10:16
i1 i1
Pk
where b0 i1 bi z0i , and
k bi
Y Y
k
zi
uZ b0 zbi i 10:17
i1
z0i i1
Q
where b0 ki1 zb i
0i . As such, lnuZ is a linear function of Z (for Eq. (10.16)) or
lnZ (for Eq. (10.17)).
When Z does not change with time, Eq. (10.14) can be written as
Ht; Z H0 tuZ 10:18
where Ht; Z is the cumulative hazard function and H0 t; Z is the baseline

cumulative hazard function.
The PHM has two kinds of applications. In the rst kind of applications, one is
interested in the values of bi s, which quantify the effects of the covariates on the
failure rate. In this case, k0 t does not need to be specied. In the other kind of
applications, one is interested in quantifying the failure rate and hence k0 t must be
specied, which is usually assumed to be the Weibull failure rate given by

b t b1
k0 t : 10:19
g g
In the ALT context, the PHM is particularly useful for modeling the effects of
multiple stresses on the failure rate. For example, when a product is subjected to
two different types of stresses such as sa and sb , the covariates can be z1 sa ,

z2 sb and z3 sa sb , which describes the interaction between sa and sb .
10.4.5 Generalized Proportional Model
The PHM has been extended to more general cases, e.g.,

proportional degradation model [3]
proportional intensity model [12], and
proportional residual life model [13].
These models can be written as the following general form [6]:
Yt; Z y0 tuZ et; Z 10:20
where Yt; Z can be hazard rate, lifetime, residual life, failure intensity function,
cumulative failure number or wear amount, t is the items age or a similar variable,
y0 t is a deterministic function of t, uZ [0 is independent of t and meets
uZ0 1, and et; Z is a stochastic process with a zero mean and standard
deviation function rt; Z. As such, the model consists of baseline part y0 t,
covariate part uZ and stochastic part et; Z.
The proportional intensity model is particularly useful for representing the
failure process of a repairable component or system. More details about the gen-
eralized proportional model can be found in Ref. [6].
10.4.6 Discussion
In general, the reliability obtained from ALT data can be viewed as an approxi-
mation of the inherent reliability. This is because it is hard for the test conditions to
be fully consistent with the actual use conditions. As such, the accelerated testing is
used for the following purposes:
identifying problems,
comparing design options, and
obtaining rough estimate of the reliability at component-level.
Denition of stress is another issue that needs attention. A stress can be dened
in different ways, and hence the function form of a stress-life relationship depends
on the way in which the stress is dened. For example, according to the Arrhenius
Table 10.1 Failure time data 130 C, n = 100 250 C, n = 5

Test time Observation Test time Observation
900 1 failed 500 1 failed
1000 99 removed 700 1 failed
800 2 removed
950 1 failed
model, the temperature is measured using the Celsius scale T while the temperature
stress is usually written as
s 1000=T o C 273: 10:21
It is noted that a large stress level T has a small value of s in Eq. (10.21).
Finally, when involving multiple failure modes and multiple stresses the ALT
data analysis and modeling are much more complex than the cases discussed above.
Example 10.1 The data shown in Table 10.1 come from Example 6.8 of Ref. [4]
and deal with the times to failure or censoring. The experiment is carried out at two
temperature stress levels and the sample sizes n are different. Assume that the
design temperature is 70 C. The problem is to nd the life distribution of the
component at the design stress.
Assume that the time to failure follows the Weibull distribution and the shape
parameter is independent of the stress. Since the stress is temperate, the stress-life
relationship model can be represented by the Arrhenius model given by
gs aecs 10:22
where s is given by Eq. (10.21). For the purpose of comparison, we also consider
the Weibull PHM as an optional model. For the model associated with Eq. (10.16),
we have
gs g0 ebs=b ; b\0: 10:23
Letting g0 a and c b=b, Eq. (10.23) becomes Eq. (10.22). For the model
associated with Eq. (10.17), we have
gs g 0 s c : 10:24
Noting that a small s implies a large stress level, Eq. (10.24) is consistent with the
inverse power-law model given by Eq. (10.11). As a result, we consider the
optional models given by Eqs. (10.23) and (10.24).
Table 10.2 Estimated Models b c g0 lnL l0

parameters and predicted Equation (10.23) 5.4464 1.7039 33.94 32.7684 4499.0
lifetime
Equation (10.24) 5.4462 3.7220 79.04 32.7684 3913.0
Using the maximum likelihood method for all the observations obtained at all
the stress levels, we have the results shown in Table 10.2. As seen, though the two
models have almost the same values of b and lnL, the predicted values of MTTF
(l0 ) have a relative error of 13 %. This conrms the importance to use an appro-
priate stress-life relationship model. In this example, the appropriate model is the
Arrhenius model given by Eq. (10.23).
10.5 Accelerated Degradation Testing Models
For some products, there is a gradual loss of performance, which accompanies one
or more degradation processes. We conne our attention on the case where only a
single degradation process is involved, which is usually a continuous stochastic
process.
Let Yt; s denote the performance degradation quantity at time t and stress level
s. Failure is dened at a specied degradation level, say yf , so that the time to
failure is given by
Yt; s wt; s yf or ts w1 yf ; s: 10:25
If the life of an item is sufciently long so that the time of testing to failure is still
long even under accelerated stress conditions, one can stop the test before it fails
and extrapolate the time to failure based on the observed degradation measurement
using a tted degradation process model. This is the basic idea of ADT.
The ADT data analysis involves three models. One is the life distribution of
ts given by Eq. (10.25). Let Ft; s denote this distribution. The second model is
the stress-life relationship model that represents the relationship between ts and
s. The third model represents how Yt; s changes with t for a xed stress level s,
and it is called the degradation process model. The rst two models are similar to
the ALT models. As such, we focus on the degradation process model in this
section.
The models for degradation can be roughly divided into two categories: phys-
ical-principle-based models and data-driven models. We separately discuss them as
follows.
10.5.1 Physical-Principle-Based Models
A physical-principle-based model is a stochastic process model with a known

mean degradation specied by physical principles. Elsayed [4] presents a few
specic degradation process models. For example, a resistor degradation model is
given by
Yt=Y0 1 atb 10:26
where Yt is the resistance at t and Y0 is the initial resistance. A laser degradation

model is given by
Yt=Y0 expatb 10:27
where Yt is the value of a degradation parameter at t and Y0 is the original value

of the degradation parameter.
10.5.2 Data-Driven Models
The data-driven models are also called statistical or empirical models. Two general
degradation process models are additive and multiplicative models. The general
additive degradation model has the following form:
Yt lt et 10:28
where lt is a deterministic mean degradation path and et represents random

variation around a mean degradation level. A specic model of the additive model is
the well-known Wiener process model. Its mean degradation function is lt ht
and et is a zero-mean normal distribution. Another specic model developed by
Jiang [5] has the mean degradation function given by
lt atb ect : 10:29
The model can have an inverse-S-shaped mean degradation path and the Wiener
process model can be viewed as its special case (achieved when b 1 and c 0).
In the general additive model, the degradation path can be nonmonotonic.
We now look at the multiplicative models. Let c Yt Dt Yt
Yt denote the
degradation growth rate. The multiplicative degradation model assumes that the
degradation growth rate is a small random perturbation et, which can be described
10.5 Accelerated Degradation Testing Models 183
by a distribution. As such, the general multiplicative degradation model can be

written as
Yt Dt Yt1 et or Yt Dt Yt etYt: 10:30
The interpretation of this model is that the degradation increment is proportional

to the total amount of degradation already present with a random proportional
coefcient. This model can assure the monotonicity of the degradation path as the
degradation increment is nonnegative. The processes that may be expected to fol-
low the multiplicative degradation model include crack growth and propagation
processes and some chemical reaction processes.
A typical multiplicative degradation model is the lognormal degradation model,
which can be derived from Eq. (10.30). Suppose a degradation process is observed
at time instants ti iDt; 0 i n. The original value is Y0 y0 [ 0 and the
value of Yt at ti is Yi . According to the multiplicative degradation model, the total
degradation amount at tn is given by
Y
n
Yn y 0 1 ei : 10:31
i1
Since ei is small, we have ln1 ei ei . As such, Eq. (10.31) can be written as

X
n X
n
lnYn =y0 ln1 ei ei : 10:32
i1 i1
According to the central limit theorem, lnYn =y0 approximately follows the normal
distribution so that Yn =y0 approximately follows the lognormal distribution. As a
result, the amount of degradation Yt approximately follows a lognormal degra-
dation model at any time t. Assume that rl is independent of t and ll is dependent
on t. The mean degradation function is given by
lt expll t r2l =2 10:32
and the median log lifetime is given by
lny0:5 ll t: 10:33
There are other degradation process models. Two such models are the gamma
and Weibull process models. The stationary gamma process model assumes that the
degradation increment follows the gamma distribution with shape parameter uDt
and scale parameter v. The mean degradation function is given by
lt uvDt: 10:34
The Weibull process model assumes that Yt follows the Weibull distribution
with shape parameter bt and scale parameter gt for a given t [7]. The mean
degradation function is given by
lt gtC1 1=bt: 10:35
Finally, the proportional degradation model mentioned in Sect. 10.4.5 is also a

data-driven model.
10.5.3 Discussion
The underlying assumption for ADT is that the failure results from one or more
observable degradation processes. A crucial issue is to appropriately select the
degradation measurement based on engineering knowledge.
Different from ALT, ADT needs to extrapolate the time to failure. Predicted
failure times can be considerably underestimated or overestimated if an improper
degradation process model is tted. Therefore, the necessity to appropriately
specify the degradation process model must be emphasized.
Once the degradation process model is assumed, the lifetime distribution model
is implicitly specied. This implicitly specied lifetime model may not match the
explicitly assumed lifetime distribution in some characteristic such as the shape of
failure rate function (see Ref. [1]). In this case, the assumption for the degradation
model or assumption for the lifetime model should be adjusted to make them
consistent.
Finally, it is benecial to incorporate ALT with ADT. This is because insuf-
cient failure data can be supplemented by degradation data to increase product
reliability information. The progressive censoring test plans discussed in
Sect. 10.3.3 can be used for this purpose.
10.5.4 A Case Study
10.5.4.1 Background and Data
The data shown in Table 10.3 come from a type-I censoring ADT for electrical
insulation. The degradation measurement is the breakdown strength in kV. The
breakdown strength decreases with time in weeks, and depends on temperature (i.e.,
stress). The degradation tests are conducted at four stress levels. Failure threshold is
the breakdown strength of a 2 kV. In Ref. [10], the problem is to estimate the
median lifetime at the design temperature of 150 C. Here, we also consider the
lifetime distribution.
Table 10.3 Breakdown strength data in kV [10]

t 180 C 225 C 250 C 275 C t 180 C 225 C 250 C 275 C
1 15 14.5 11 11.5 16 15.3 11 11.5 5
15.5 15 12.5 13 16 12.5 12 5.5
16.5 15.5 14.5 14 17 13 12 6
17 16 15 14 18.5 14 12 6
2 13 12.5 11.5 11.5 32 12 9.5 10 2.4
13.5 12.5 12 12.5 12.5 11 10.5 2.5
14 13 12 13 13 11 10.5 2.7
16 13.5 12.5 13 16 11 11 2.7
4 13.5 12.5 12 9.5 48 13 10.5 6.9 1
13.5 12.5 12 10 13.5 11.5 7 1.2
17.5 13 13 11 13.6 12 7.9 1.5
17.5 15 13.5 11.5 16.5 13.5 8.8 1.5
8 15 10.5 11.5 5.5 64 12.5 10 6.7 1
15 13 11.5 6 13 10.5 7.3 1.2
15.5 13.5 12 6 16 11 7.5 1.2
16 14 12.5 6.5 16.5 11.5 7.6 1.5
10.5.4.2 Lognormal Degradation Model
The electrical insulation failure process is similar to the crack growth and propa-
gation process, and hence the lognormal degradation process model appears
appropriate. To specify this model, we need to specify the process parameters ll t
and rl .
Let yt denote the breakdown voltage observed at time t. Noting that expll t
is the median value of Yt, the function form of ll t can be obtained through
examining the shape of the data plot of lny versus t. It is found that lny can be
approximated by a linear function of t and hence ll t can be written as
ll t a t=b: 10:36
For a given stress level, the parameters (a; b; rl ) can be estimated using the max-
imum likelihood method. Once these parameters are specied, the median lifetime
at each stress level can be obtained from Eq. (10.36) by letting
ll t lny0:5 t ln2. As such, the median life can be estimated as
t0:5 ba ln2: 10:37

10.5.4.3 Median Lifetime at Design Stress
To estimate the median lifetime at design stress, we rst t the four median lifetime
data at different stress levels to the Arrhenius model given by
t0:5 ecds or lnt0:5 c ds 10:38
where s is given by Eq. (10.21). The estimated model parameters are c 11:1485
and d 8:4294. From the tted model, the median log breakdown voltage at the
design temperature 150 C equals 8.7792 and the corresponding median lifetime
equals 6497.5 weeks. The extrapolation process of median lifetime is graphically
shown in Fig. 10.5.
10.5.4.4 Lifetime Distribution at Design Stress
Let r0 and l0 t denote the parameters of the lognormal degradation model at the
design stress. The life distribution at the design stress is given by
Ft Uln2; l0 t; r0 10:39
where U: is the normal cdf. As such, the problem is to specify r0 and l0 t.

We rst look at r0 . It is noted that the values of rl in Table 10.4 varies with
stress level in a nonmonotonic way. Instead of assuming that it monotonically
12
10
8
ln(t 0.5 )
0
0 1 2 3
s
Fig. 10.5 Extrapolation of median lifetime
Table 10.4 Estimates of 180 C 225 C 250 C 275 C

parameters and lifetimes
a 2.7328 2.6029 2.5687 2.4134
b 649.30 254.73 106.54 25.13
rl 0.1105 0.0990 0.0841 0.2412
t0:5 1324.33 486.48 199.83 43.24
varies with stress, it is better to assume that it does not vary with stress. Under this
assumption, we need to re-estimate the model parameters. The results are rl
0:1477 and the values of a and b at different stress levels are almost the same as
those shown in Table 10.4. This implies that the lognormal degradation model is
insensitive to the value of rl in this example.
We now look at l0 t. Its function form is the same as the one given by
Eq. (10.36) with parameters a0 and b0 . In Eq. (10.36), a ll 0; in Eq. (10.38),
ell ecds . As such, a is a linear function of s. Based on the data of a and b in
Table 10.4, regression and extrapolation yield a0 2:8675. From Eq. (10.36), b has
a dimension of the lifetime, and hence the relation between b and s follows the
Arrhenius model. Through regression and extrapolation, we have b0 2962:23. As
such, from Eq. (10.39) the life distribution at the design stress is given by
Ft Uln2; a0 t=b0 ; r0 Ut; b0 a0 ln2; b0 r0 : 10:40
Clearly, it is the normal distribution with l 6440.93 and r 437.55. This yields
another estimate of median lifetime, i.e., 6440.93. The relative error between this
estimate and the estimate obtained earlier is 0.9 %.
The tted lifetime distribution provides more reliability information than a single
estimate of the median lifetime. For example, it is easy to infer B10 5880:19 from
the tted model.
10.6 Design of Accelerated Stress Testing
Design of accelerated stress testing is a complex optimization problem involving a

number of decision variables. Different test schemes require different test resources
(e.g., time, cost, etc.), produce different amounts of reliability information, and
result in different estimation accuracies. Due to the complexity of the problem, one
has to depend on the experiences with assistance of mathematical models. In this
section we focus on the design of the single factor constant stress test scheme. In a
similar way, one can deal with the test design problem involving multiple factors
through incorporating the empirical approach to be presented in this section with
the Taguchi experiment design method discussed in Chap. 8.
10.6.1 Design Variables and Relevant Performances
Assume that stress type, the design stress level s0 , and the extreme stress level su are
known. The design variables are the following:
Number of stress levels k ( 2),
Magnitude of each stress level si with s0 s1 \s2 \ \sk su ,
Test time ti with t1 t2 . . . tk , and

Number of items ni with n1 n2 nk .
Clearly, the total number of the design variables is 3k for a given k.
Assume that there is the prior knowledge about the stress-life relationship as well
as the type of life distribution at the normal use conditions. The stress-life rela-
tionship is given by T0 wsTs , where ws is actually acceleration factor with
ws0 1. Let Fi t denote the cdf of lifetime Ti at stress level si . As such, the prior
life distribution at the ith stress level is given by
Fi t F0 wsi t: 10:41
Main performances associated with a given test scheme are the required test
efforts and obtainable information amount. Main measures for the test efforts are the
required total test time and cost. At the ith stress level, the probability that an item
will fail by ti equals Fi ti , and the expected test time is given by
Zti
si 1Fi tdt: 10:42
0
The required total test time is given by
X
k
Ttt ni s i : 10:43
i1
Let c1 denote the cost per test item and c2 denote the test cost per unit test time.
The required total test cost can be computed as
X
k
C c1 c2 si ni : 10:44
i1
The obtainable information content can be represented by the total expected

number of failures given by
X
k
mf ni Fi ti : 10:45
i1
The larger it is the greater the reliability information content.

10.6 Design of Accelerated Stress Testing 189
It is noted that the information quality of a failure observation at a lower stress

level is higher than the information quality of a failure observation at a higher stress
level since the test conditions are closer to the normal use conditions. An equivalent
information weight can be used to represent the effect of this factor. The weight can
be dened as the reciprocal of the acceleration factor given by
wi 1=wsi : 10:46
To reflect the effect of the information weight on the reliability information quality,
we dene an equivalent total number of failures as
X
k
nf wi ni Fi ti 10:47
i1
It can be used as a performance measure to compare different design options

generated from an empirical approach.
10.6.2 Empirical Approach for ALT Design
Consider the single factor constant stress ALT scheme. The design variables can be
determined using an empirical approach [4]. Specic details are as follows.
10.6.2.1 Number of Stress Levels
Assume that the function form of the stress-life relationship model is known but
there are m unknown parameters. To specify these parameters, k m is required.
This implies that m is the lower bound of k.
Let n denote the total number of test items at all the stress levels and ta (a 0:5)
denote the a-fractile of time to failure, which is desired to estimate. At each stress
level, it is desired that the expected failure number is not smaller than 5. As such,
the upper bound of k is given by na 5k or k na=5. As a result, we have
m k na=5: 10:48
It is noted that a large value of k will make the design problem much more complex.
As such, many test schemes take k m 1 or m 2.
10.6.2.2 Magnitudes of Stress Levels
We rst look at the highest stress level sk . Clearly, sk should not exceed the extreme
stress level su and the validation range of the ALT model, which is determined
based on engineering analyses. A large value of sk results in a shorter test time but
poorer information quality.
The basic criterion to determine the other stress levels is that they should bias
toward s0 . A preliminary selection can be given by the following relation:
si s0 pi ; p sk =s0 1=k : 10:49
If we use Eq. (10.49) to determine the intermediate stress levels for the case
study in Sect. 10.5.4, they would be 175, 203, and 236, respectively. If using an
equal-space method, they would be 181, 213, and 244. It is noted that the stress
levels used in the case study are closer to the ones obtained from the equal-space
method.
If tests at different stress levels are conducted simultaneously, the total test
duration is determined by t1 , which depends on s1 . In this case, s1 can be deter-
mined based on the total test duration requirement, and Eq. (10.49) can be revised
as
si s1 pi1 ; p sk =s1 1=k1 : 10:50
If we use Eq. (10.50) to determine the intermediate stress levels for the case
study in Sect. 10.5.4, they would be 200 and 230, respectively.
10.6.2.3 Duration of Test at Each Stress Level
Usually, we have t1 t2 tk so that t1 can be determined based on the total

test duration requirement. Once t1 and s1 are specied, a1 F1 t1 is specied. If
it is small, we can take ai Fi ti [ a1 so that sufcient reliability information
can be obtained without signicantly increasing test time. As such, the test
durations at other stress levels can be empirically determined by the following
relation:
i1
Fi ti a1 2a a1 : 10:51
k1
It implies that a1 \F2 t2 \ \Fk tk 2a 1.

10.6 Design of Accelerated Stress Testing 191
10.6.2.4 Number of Test Items at Each Stress Level
It is preferred to allocate more units to low stress levels so as to obtain nearly equal
failure number at each stress level. It is desirable that the following relation can be
met:
ni Fi ti 5: 10:52
Example 10.2 In this example, we consider the test scheme in Example 10.1 but the
progressive censoring is not allowed. The following conditions maintain unchan-
ged: the stress-life relationship, the design stress level, number of stress levels, and
the upper stress level. To calculate the cost, we assume that c1 1 and c2 0:1.
One or more of the following variables allow changed: s1 ; t1 t2 , n1 and n2 . We

consider the following design options and evaluate their performances:
Option 1: The values of s1 , t1 , n1 and n2 are the same as those in Example 10.1.
The corresponding performances are shown in the third row of Table 10.5.
Option 2: Suppose the total test cost is required equal to 100,000, and only the
test time is adjusted to meet this requirement. The corresponding performances
are shown in the fourth row of Table 10.5. As seen, the information content
reduces 12 % relative to Option 1.
Option 3: The cost constraint is the same as that in Option 2, but only n1 is
adjusted. The results are shown in the fth row. In terms of the equivalent
failure number nf , this option is obviously outperforms Option 2 since the
information content only reduces 1.6 % relative to Option 1.
Option 4: According to Eq. (10.49), s1 132:3. This implies that s1 can be
slightly increases. This option is the same as Option 1 except s1 140. The
results are shown in the sixth row. In terms of the equivalent failure number, this
option is obviously outperforms Option 1 (with the information content
increasing 22 % relative to Option 1).
Option 5: This option is similar to Option 2 but there are two different points: (a)
s1 140 rather than 130, and (b) n1 and n2 are determined by ni Fi ti 5. The
results are shown in the last row. As seen, its performances signicantly out-
perform the other options (with the information content increasing 130 % rel-
ative to Option 1).
Table 10.5 Performance of Option Design variables Performances

design options
s1 t1 n1 n2 Ttt , 103 C, 103 nf
1 130 1000 100 5 103.8 103.9 1.3467
2 130 960 100 5 99.9 100 1.1850
3 130 1000 96 5 99.9 100 1.3257
4 140 1000 100 5 103.7 103.8 1.6496
5 140 1254 77 5 99.9 100 3.0931
This illustrates the usefulness of the performance measures presented in

Sect. 10.6.1 and the potential to improve the performances by appropriately
designing the test scheme.
References
1. Bae SJ, Kuo W, Kvam PH (2007) Degradation models and implied lifetime distributions.
2. Cox DR (1972) Regression models and life tables (with discussion). JR Stat Soc B 34(2):187220
3. Ebrahem MAH, Higgins JJ (2006) Non-parametric analysis of a proportional wearout model
for accelerated degradation data. Appl Math Comput 174(1):365373
4. Elsayed EA (1996) Reliability engineering. Addison Wesley Longman, New York
5. Jiang R (2010) Optimization of alarm threshold and sequential inspection scheme. Reliab Eng
Syst Saf 95(3):208215
6. Jiang R (2012) A general proportional model and modelling procedure. Qual Reliab Eng Int
28(6):634647
7. Jiang R, Jardine AKS (2008) Health state evaluation of an item: a general framework and
graphical representation. Reliab Eng Syst Saf 93(1):8999
8. Lu Y, Loh HT, Brombacher AC et al (2000) Accelerated stress testing in a time-driven product
development process. Int J Prod Econ 67(1):1726
9. Meeker WQ, Hamada M (1995) Statistical tools for the rapid development and evaluation of
high-reliability products. IEEE Trans Reliab 44(2):187198
10. Nelson W (1981) Analysis of performance-degradation data from accelerated tests. IEEE
Trans Reliab 30(2):149155
11. Nelson, Wayne B (2004) Accelerated testingstatistical models, Test Plans, and
DataAnalysis, John Wiley & Sons, New York
12. Percy DF, Alkali BM (2006) Generalized proportional intensities models for repairable
systems. IMA J Manag Math 17(2):171185
13. Wang W, Carr M (2010) A stochastic ltering based data driven approach for residual life
prediction and condition based maintenance decision making support. Paper presented at 2010
prognostics and system health management conference, pp 110
Chapter 11
Reliability Growth Process
and Data Analysis
11.1 Introduction
During the product development phase, the reliability of product can be improved
by a test-analysis-and-x (TAF) process. This process is called the reliability
growth process. A challenging issue in this process is to predict the ultimate reli-
ability of the nal product conguration based on all the test observations and taken
corrective actions. This needs to use appropriate reliability growth models. In this
chapter we focus on the reliability growth models and data analysis. Reliability
demonstration test to verify the design is also briefly discussed.
This chapter is organized as follows. We discuss the TAF process in Sect. 11.2.
Reliability growth plan model, corrective action effectiveness model, and reliability
growth evaluation models are presented in Sects. 11.311.5, respectively. We
discuss reliability demonstration test in Sect. 11.6. Finally, a case study is presented
in Sect. 11.7.
11.2 TAF Process
Referring to Fig. 11.1, the reliability growth process involves testing one or more
prototype systems under operating stress conditions to nd potential failure modes.
The testing is conducted in several stages and the test stress level can gradually
increase from the nominal stress to overstress. When a failure occurs before the
stage test ends, the failed part is replaced by one new (which is equivalent to a
minimal repair) and the test is continued. The stage test ends at a prexed test time
or a prexed number of failures.
The observed failure modes are then analyzed, and the outcomes of the analysis
are design changes, which lead to new gurations. The new gurations are tested in
the next test stage.

194 11 Reliability Growth Process and Data Analysis
8
7 1st stage 2nd stage
6
System 2
5
System 1
N (t )
4
System 3
3
2
1
0
0 50 100 150 200
t
Fig. 11.1 Reliability growth tests of multiple stages for several systems
For a given test stage, multiple failure point processes are observed. For a given
prototype system, the inter-arrival times in different test stages are independent but
nonidentically distributed. The reliability of the current conguration is assessed
based on the observed failure processes. If the assessed reliability level is unac-
ceptable for production, then the system design is modied, and the reliability of the
new conguration is predicted based on the observed failure processes and planned
corrective actions.
If the predicted reliability is still unacceptable, the growth testing is continued;
otherwise, the new conguration may need to experience a reliability demonstration
test to verify the design. This is because the effectiveness of the last corrective
actions and possibility of introducing new failure modes are not observed due to the
time and budget constraints.
According to the time when the corrective actions are implemented, there are
three different reliability growth testing strategies:
Test-nd-test strategy. This strategy focuses on discovering problems and the
corrective action is delayed to the end of the test.
Test-x-test strategy. In this strategy, the corrective action is implemented once
problems are discovered.
Test-x-nd-test strategy. This strategy is a combination of the above two
strategies. It xes some problems during the test and the other problems are
delayed to x until the end of the test.
The reliability growth process involves three classes of models:
Reliability growth plan models,
Corrective action effectiveness models, and
Reliability growth evaluation models.
We separately discuss these models in the following three sections.
11.3 Reliability Growth Plan Model 195
11.3 Reliability Growth Plan Model
11.3.1 Reliability Growth Plan Curve
Reliability growth planning deals with program schedules, amount of testing,

required resources, and so on. The planning is based on a reliability growth plan
curve, which describes the relation between the achieved reliability goal and test
duration. This curve is constructed early in the development program and is used to
evaluate the actual progress of the reliability program based upon the reliability data
generated during the reliability growth testing.
The planned growth curve is constructed based on reliability growth curves in
various stages. This necessitates setting stage reliability growth goals. The stage
reliability growth curves are estimated based on initial conditions, assumed growth
rate, and planned management strategy. Once they are specied, the overall growth
curve can be obtained through tting all the stage growth curves into the Duane
model, which will be discussed in the next subsection. Figure 11.2 shows the
relation between the stage growth curves and the overall growth curve. An illus-
tration for estimating the planned growth curve will be presented in Sect. 11.7.5.
It is not possible for a program to exactly follow the planned growth curve. As
such, the growth process will be monitored by comparing the planed growth curve
with the observed growth curve, and the plan will be accordingly adjusted (e.g.,
adjustment of the time frame, reassignment of resources, etc.).
11.3.2 Duane Model
The function form of the planned growth curve is the well-known Duane model [6].
Suppose that the nth failure of a system occurs at tn . The interval MTBF is given by
ln tn =n. Duane nds the following empirical relation between ln and tn :
Time for Growth Demonstration

corrective actions test period test period
Achieved
Overall MTBF
( )
growth curve Stage growth

curve
Stage 1 Stage 2 Stage 3
Fig. 11.2 Overall and stage reliability growth curves

lnln a b lntn : 11:1
Equation (11.1) can be written as
ln atnb ; a ea : 11:2
The value of a depends on the initial reliability level at the start of testing and
b 2 0; 1 represents the rate of growth. A large value of b implies a large rate of
growth.
If the reliability growth occurs continuously (e.g., the case in the test-x-test
strategy), the reliability achieved by t can be represented by instantaneous MTBF.
Let Mt denote the expected number of failures in (0, t). The instantaneous failure
intensity is given by mt dMt=dt. The instantaneous MTBF is given by
lt 1=mt.
Using ln tn =n to Eq. (11.2) and replacing n and tn by Mt and t, respectively,
we have Mt t1b =a and
mt 1 b=atb 1 bm0 t 11:3
where m0 t Mt=t is the interval average failure intensity over (0, t). As such,
the instantaneous MTBF is given by
a b
lt 1=mt t : 11:4
1b
Letting a lgb0 , Eq. (11.4) becomes:

0
b
l0 t
lt : 11:5
1 b g0
An interpretation for Eq. (11.5) is that the MTBF achieved at t g0 is 1=1 b

times of the initial MTBF l0 after the system is tested and continuously improved
for g0 time units. As a result, the required test time and associated test resources can
be planned based on the overall reliability growth model.
For the ith test stage with t 2 si1 ; si , if corrective actions are implemented at
the end of this test stage (i.e., the case in the test-nd-test strategy), the reliability of
the current conguration is assessed by the interval MTBF over si1 ; si rather
than the instantaneous MTBF at the end of this test stage.
11.4 Modeling Effectiveness of a Corrective Action 197
11.4 Modeling Effectiveness of a Corrective Action
11.4.1 Type of Failure Modes
There are two categories of failure modes:

assignable cause failure modes, which can be eliminated by corrective actions,
and
non-assignable cause failure modes, which cannot be eliminated due to insuf-
cient resources (e.g., time, budget, and knowledge) to nd and x the cause.
This implies that a failure mode can be corrected only if it is an assignable cause
failure mode.
11.4.2 Effectiveness of a Corrective Action
Different corrective actions can have different effects on reliability. The predicted
reliability can be inaccurate if the effectiveness of a corrective action is not
appropriately modeled (e.g., see Ref. [11]). There are two kinds of methods to
model the effectiveness of corrective actions: implicit (or indirect) and explicit (or
direct).
The implicit methods use extrapolation to predict the reliability after the cor-
rective actions are implemented. It is a kind of empirical methods. For example, the
Duane model uses the instantaneous MTBF to predict the MTBF of the next
conguration. Most of discrete reliability growth models fall into this category.
The explicit methods use a specic value called the x effectiveness factor (FEF)
to model the effectiveness of a corrective action. FEF is the fractional reduction in
the failure intensity of a failure mode after it is xed by a corrective action.
Therefore, it takes a value between 0 and 1. Specially, it equals 0 if nothing is done,
and equals 1 if the failure mode is fully removed.
Specically, let d denote FEF of a corrective action, k0 and k1 denote the failure
intensities before and after the corrective action is implemented, respectively. The
FEF is dened as
k0 k1 k1
d 1 : 11:6
k0 k0
This implies that given the values of d and k0 , one can calculate the value of k1
using Eq. (11.6), which is k1 1 dk0 . Comparing this with Eq. (11.3), d is
somehow similar to the growth rate b.
Suppose that t0 is a failure observation occurred before the mode is corrected.
This failure would occur at t1 if it were corrected at t 0. It is noted that the mean
life is inversely proportional to the failure rate for the exponential distribution.
Under the exponential distribution assumption, t1 can be calculated as
k0 t0
t1 t0 : 11:7
k1 1d
In this way, a failure time observed before the corrective action is implemented can
be transformed into a failure time that is equivalent to the failure time observed for
the new conguration. The benet to do so is that we can predict the life distri-
bution of the new conguration by tting the equivalent failure sample to a life
distribution.
For a given corrective action, FEF can be quantied based on subjective judg-
ment or/and historic data. It is often difculty for experts to specify such infor-
mation, and the historic data may not be suitable for the current situation. To
address this problem, one needs to properly consider the root cause of the failure
mode and the features of corrective action. A sensitivity analysis that considers
different FEF values can be carried out.
Example 11.1 Suppose that the strength of a component is considered to be not
strong enough. The corrective action is to use a stronger component to replace the
weak component used currently. Assume that the lifetime of the components fol-
lows the exponential distribution and the mean lifetimes of the original and new
components are 0 = 600 h and 1 = 1000 h, respectively. The problem is to
calculate the FEF of the corrective action.
The failure rates of the original and new components are k0 1=l0 and
k1 1=l1 , respectively. From Eq. (11.6), we have
k1 l
d 1 1 0 0:4:
k0 l1
11.5 Reliability Growth Evaluation Models
Reliability growth models are used to evaluate the improvement achieved in reli-
ability. According to the type of product, the models can be classied into two
categories: software reliability growth models and the reliability growth models for
complex repairable systems that are comprised of mechanical, hydraulic, electronic,
and electric units. These two categories of models are similar in the sense that the
growth takes place due to corrective actions. However, the corrective actions for
software can be unique and objective but they can be multidimensional and sub-
jective for complex systems.
11.5 Reliability Growth Evaluation Models 199
The reliability growth models for complex systems can be further classied into
two classes: discrete and continuous. Discrete models describe the reliability
improvement as a function of a discrete variable (e.g., the test stage number); and
continuous models describe the reliability improvement as a function of a contin-
uous variable (e.g., the total test time).
Reliability growth models can be parametric or nonparametric. The parametric
models are preferred since they can be used to extrapolate the future reliability if
corrective actions have been planned but have not been implemented.
11.5.1 Software Reliability Growth Models

and Parameter Estimation
11.5.1.1 Models
During the software testing phase, a software system is tested to detect software
faults remaining in the system and to x them. This leads to a growth in software
reliability. A software reliability growth model is usually used to predict the number
of faults remaining in the system so as to determine when the software testing
should be stopped.
Assume that a software failure occurs at random time and the fault caused the
failure is immediately removed without introducing new faults. Let Nt denote the
cumulative number of failures detected in the time interval 0; t, and
Mt ENt; mt dMt=dt: 11:8
We call Mt the mean value function of Nt, and mt the failure intensity
function, which represents the instantaneous fault detection or occurrence rate.
According to Ref. [9], a software reliability growth model can be written as the
following general form:
Mt M1 Gt 11:9
where M1 is the expected cumulative number of faults to be eventually detected

and Gt meets all characteristics of a cdf. Since software faults can be found after a
relatively long test time, Gt should have a long right tail. The distributions with
this characteristic include the exponential distribution, lognormal distribution,
inverse Weibull distribution, and Pareto Type II (or Lomax) distribution. The
inverse Weibull and Pareto Type II distributions are given respectively by
Gt expg=tb 11:10
and

t b
Ft 1 1 ; b; g [ 0: 11:11
g
The software reliability growth models can be used to estimate the number of
unobserved failure modes for complex systems. This will be illustrated in
Sect. 11.7.
11.5.1.2 Parameter Estimation Methods
The model parameters can be estimated using the maximum likelihood method and
least squares method. Consider the failure point process t1 t2 tn \T ,
where ti is the time to the ith failure, and T is the censored time. The distribution of
time to the rst failure is given by
f t mt expMt: 11:12
Conditional on t ti1 , the distribution of time to the ith failure is given by
f tjti1 mt expMti1 Mt: 11:13
The log-likelihood function is given by
X
n X
n
lnL lnmti MT n lnM1 M1 GT lngti 11:14
i1 i1
where gt dGt=dt. From Eq. (11.14), the maximum likelihood estimate of M1

is given by
M1 n=GT: 11:15
Since GT\1, we have M1 [ n. Using Eq. (11.15) to Eq. (11.14) and after some
simplications, we have
Xn
0 gti
lnL lnL n lnn n ln : 11:16
i1
GT
As such, the parameters of Gt can be estimated by maximizing lnL0 given by

Eq. (11.16) or minimizing the sum of squared errors given by
X
n
SSE Mti i 0:52 11:17
i1
Table 11.1 Failure times in 9 21 32 36 43 45 50 58 63

days
70 71 77 78 87 91 92 95 98
104 105 116 149 156 247 249 250 337
384 396 405 540 798 814 849
40
30
M(t )
20
10
0
0 200 400 600 800 1000
t
Fig. 11.3 Observed and tted inverse Weibull reliability growth curves
subject to the constraint given by Eq. (11.15). Here, we take the empirical estimate
of Mti as i 0:5 (since Mti i 1).
Example 11.2 The data set shown in Table 11.1 comes from Ref. [8]. The problem
is to t the data to an appropriate reliability growth model.
In this example, T tn . The observed data are displayed in Fig. 11.3 (the dotted
points). It indicates that a reliability growth model with an inverse-S-shaped growth
curve is desired, and hence the lognormal and inverse Weibull models can be
appropriate. For the purpose of illustration, we also consider the exponential and
Pareto models as candidate models.
The maximum likelihood estimates of the parameters of the candidate models
and associated values of lnL and SSE are shown in Table 11.2. As seen, the best
model is the inverse Weibull model. The reliability growth curve of the tted
inverse Weibull model is also shown in Fig. 11.3, which indicates a good agree-
ment between the empirical and tted growth curves.
According to the tted model, there are about 9 faults remaining in the system.
The time to the next failure can be estimated from Eq. (11.13) with Mt being the
tted inverse Weibull model. The expected time to the next failure is 1021 h.
Table 11.2 Estimated Model M1 b or l g or r lnL SSE

parameter and performances
Exponential 34.83 1 226.97 128.64 286.8
Inverse 43.09 0.7033 109.49 124.83 99.9
Weibull
Lognormal 36.62 1.2163 4.9627 125.48 134.7
Pareto 40.05 1.1865 216.77 127.18 173.5
11.5.2 Discrete Reliability Growth Models

for Complex Systems
Let Rj , Fj and kj are the reliability, unreliability, and failure intensity (or failure rate)
of the item at stage j j 1; 2; . . ., respectively. There are two general models
for modeling reliability growth process. One is dened in terms of Rj (or
Fj 1 Rj ) and applicable for attribute data (i.e., the data with the outcome of
success or failure), and the other is dened in terms of kj (or lj 1=kj ) and
applicable for the exponential life data. There are several specic models for each
general model. Different models or different parameter estimation methods can give
signicantly different prediction results. As such, one needs to look at several
models and select the best (e.g., see Ref. [7]).
11.5.2.1 Models for Attribute Data
In this class of models, the outcome of a test for an item (e.g., one-shot device) is
success or failure. Suppose that nj items are tested at the jth stage and the number of
successes is xj . As such, the stage reliability Rj is estimated as rj xj =nj . The
corrective actions are implemented at end of each stage so that Rj1 is not smaller
than Rj statistically. As such, Rj increases with j. A general reliability growth model
in this context can be dened as
Rj R1 hSj; j 1; 2; . . . 11:18
where R1 is the maximum obtainable reliability, h 2 0; R1 is a constant, which

represents the rate of growth, and Sj is a discrete survival function with the
support j 1. Specially, when R1 1, the model reduces into
Fj hSj: 11:19
Two specic models are the inverse power and exponential models given
respectively by
Rj R1 h=jk ; Fj hekj1 ; j 1; 2; . . .: 11:20
Example 11.3 The data of this example come from Ref. [1]. A TAF process
comprises of 12 stages with nj 20 at each stage. The numbers of successes xj are
14, 16, 15, 17, 16, 18, 17, 18, 19, 19, 20 and 19, respectively. The corrective action
is implemented at the end of each stage. The problem is to estimate the reliability of
the product after the last corrective action.
Table 11.3 Estimated Model h k R1 SSE R13

parameters and reliability for
Exponential 0.2942 0.1730 0.013733 0.9631
Example 11.3
Inverse power 0.3253 0.5836 1 0.024707 0.9272
Inverse power 4.0705 0.0280 4.7524 0.015427 0.9641
Fitting the data to the models in Eq. (11.20), we have the results shown in
Table 11.3. In terms of the sum of squared errors, the exponential model provides
better tting to the data. As a result, the reliability is estimated as R13 0:9631.
It is noted that the average reliability evaluated according to the data from the
last four stages is 0.9625. This implies that the inverse power model obviously
underestimates the reliability. However, if we take R1 as an arbitrary real number,
the inverse power model provides a reasonable estimate of reliability (see the last
row of Table 11.3).
11.5.2.2 Models for Lifetime Data
In this class of models, the outcome of a test is an observation of time to failure.

Suppose that nj items are tested at the jth stage and the test duration is Tj . When a
failure occurs, the failed item is restored by minimal repair. Assume that the number
of failures is xj and the failure intensity during the interval between two successive
x
corrective actions is constant. The failure intensity is estimated as kj nj Tj j . Similar
to the models in Eqs. (11.18) and (11.19), two general models are given respec-
tively by
kj k1 hSj; kj k0 Sj; j 1: 11:21
Similar to the models in Eq. (11.20), two specic models are given respectively by
kj k1 h=jk ; kj k1 ekj1 ; j 1: 11:22
Optionally, the reliability growth model can be dened in terms of MTBF. Such
a model is the extended geometric process model [10]. Let Zj Tj Tj1 j
1; 2; . . . and lj EZj . In the context of reliability growth, the MTBF (i.e., lj ) is
increasing and asymptotically tends to a positive constant l1 (i.e., the maximum
obtainable MTBF); and the stochastic process X fl1 Zj g is stochastically
decreasing and tends to zero. If X follows a geometric process with parameter
l Z
a 2 0; 1, then Y f 1aj 1 j g is a renewal process with mean h and variance r2 .
As such, the mean function of the stochastic process Z fZj g is given by
lj EZj l1 haj1 l1 hej lnajj1 : 11:23

Table 11.4 Test data and j Tj xj kj , 102 lj , (11.22) lj , (11.24)

predicted failure intensity for
Example 11.4 1 100 27 1.0800 92.59 92.59
2 75 16 0.8533 117.19 117.19
3 148.32 123.72
This model can be viewed as a variant of Eq. (11.18) with the reliability being
replaced by MTBF. We dene the following two general models in terms of MTBF
lj l1 hSj; lj l1 Fj; j1 11:24
where Sj is a discrete survival function and Fj is a discrete cdf. They has a

common feature: lj ! l1 as j ! 1.
Noting lj 1=kj , the inverse power-law model in Eq. (11.22) can be written as
k
lj hkj 1 jk ! 1=k1 , which is somehow similar to the extended geometric process
model given by Eq. (11.23). The negative exponential model in Eq. (11.22) can be
written as lj k11 ekj1 ! 1. This implies that the negative exponential model
may overestimate the MTBF.
Example 11.4 Twenty-ve systems are tested in two stages. The test duration and
failure number of each stage are shown in the second and third columns of
Table 11.4, respectively. Under the constant-failure-intensity assumption, the esti-
mated failure intensities are shown in the third column. The problem is to predict
the MTBF to be observed in the third stage.
Fitting the estimated failure intensities to the exponential model in Eq. (11.22)
yields k 0:2356 and the predicted MTBF equals 148.32. Fitting the estimates of
1=kj to Eq. (11.24) with Fj being the discrete exponential model yields l1
126:08 and k 1:3257, and the predicted MTBF equals 123.72. We will show later
that the estimate obtained from Eq. (11.24) is superior to the estimate obtained from
Eq. (11.22).
11.5.3 Continuous Reliability Growth Models

for Complex Systems
The Duane model given by Eq. (11.2) may be the most important continuous
reliability growth model. Several variants and extensions of this model are pre-
sented as follows.
11.5.3.1 Crow Model and Parameter Estimation Methods
The Crow model [3] is given by
Mt t=gb 11:25
where t is the cumulative time on test and Mt is the cumulative number of

observed failures. When b\1, the system is improved; and when b [ 1, the system
is deteriorated.
The other names of this model include the NHPP model, Army materiel systems
analysis activity (AMSAA) model, and power-law model. It is actually the model
given by Eq. (11.3) with the following relations:
b 1 b; g a1=1b : 11:26
The maximum likelihood method and least spares method can be used to esti-
mate the parameters of the Crow model. The least squares method has been pre-
sented in Sect. 6.2.4, and the maximum likelihood method is outlined as follows:
Consider the failure point processes of n nominally identical systems:
ti1 ti2 tiJi Ti ; 1 i n 11:27
where Ti is a censored observation. Assume that the underlying life distribution is

the Weibull distribution with parameters b and g. Under the minimal repair
assumption, the conditional density and reliability functions in t [ ti;j 1 are given
by
fc t f t=Rti;j 1 ; Rc t Rt=Rti;j 1 11:28
where

b t b1
Mt t=gb ; mt ; Rt expMt; f t mtRt:
g g
11:29
P
n
The log-likelihood function is given by lnL lnLi , where
i1
X
Ji X
Ji b
Ti
lnLi lnRc Ti lnfc tij lnmtij : 11:30
j1 j1
g
The maximum likelihood estimates of the parameters can be obtained by directly

maximizing the log-likelihood function lnL.
Specially, when n 1, we have
X
J1
1=b
b J1 = lnT1 =t1j ; g T1 =J1 : 11:31
j1
A more special case is that n 1 and J1 1. In this case, we have
b 1= lnT1 =t1 ; g T1 ; mT1 1=T1 lnT1 =t1 : 11:32
Example 11.5 The data shown in Table 11.5 come from Ref. [5] and deal with the
reliability growth process of a system. Here, J 40 and T tJ . Crow ts the data
to the power-law model using the maximum likelihood method. The estimates
parameters are: b; g 0:4880; 1:6966. The failure intensity and MTBF at the
test end are estimated as
mtJ 5:994 103 ; MTBF 1=mtJ 166:83:
It is noted that MTBF can be estimated by tJ1 tJ , where tJ1 can be estimated
through letting MtJ1 J 1. Using this approach, we have tJ1 tJ 169:02,
which is slightly larger than the maximum likelihood estimate with a relative error
of 1.3 %.
Using the least square method, the estimated parameters are:
b; g 0:4796; 1:4143. The failure intensity and MTBF are estimated by
mtJ 6:034 103 ; MTBF 1=mtJ 165:73:
The relative error between the MTBF estimates obtained from the two methods is
0.66 %.
Table 11.5 Reliability growth test data (in hours) for Example 11.5
0.7 2.7 13.2 17.6 54.5 99.2 112.2
120.9 151.0 163.0 174.5 191.6 282.8 355.2
486.3 490.5 513.3 558.4 678.1 699.0 785.9
887.0 1010.7 1029.1 1034.4 1136.1 1178.9 1259.7
1297.9 1419.7 1571.7 1629.8 1702.3 1928.9 2072.3
2525.2 2928.5 3016.4 3181.0 3256.3
11.5.3.2 Piecewise Power-Law Model
The piecewise power-law model is developed by Calabria et al. [2]. It is assumed

that the failure process of each prototype in each test stage is a NHPP with failure
intensity given by:
bj
mj t t=gj bj 1 : 11:33
gj
The interval MTBF is given by lj bj C1 1=bj and the interval failure intensity
at the jth stage is given by kj 1=lj . If the test is no longer conducted after the last
corrective actions are implemented, the failure intensity can be predicted by tting the
estimated stage failure intensities [or interval MTBFs] to the models given by
Eq. (11.22) [or Eq. (11.24)]. The future failure intensity [or MTBF] is extrapolated
using the tted model in a way similar to the one in Example 11.4.
11.5.3.3 Power-Law Model for a System with Multiple Failure Modes
Consider a repairable system with K failure modes. The system failure intensity is
the sum of failure intensities from independent failure modes, i.e.,
X
K
ks ki 11:34
i1
where ki is the failure intensity of mode i evaluated at the end of a given stage, and
ks is the system failure intensity at the end of this stage.
There are two methods to evaluate ki . One is to assume that the failure intensity
for each mode is constant. In this case, the failure intensity of mode i is estimated as
ni
ki 11:35
nT
where ni is the number of failures of mode i during the current test stage, n is the
number of tested items, and T is the test duration of this stage.
The other method is to assume that the failure intensity for each mode can be
represented by the power-law model and the parameters are given by
X
ni
1=bi
b i ni = lnT=tij ; gi nT=ni : 11:36
j1
Equation (11.36) comes from Eq. (11.31). As a result, the failure intensity of mode i
in the current stage is given by
1
ki : 11:37
gi C1 1=bi
When bi 1, the intensity estimated from Eq. (11.37) is the same as the one
estimated from Eq. (11.35). However, when bi 6 1, the intensity estimates from
Eqs. (11.35) and (11.37) are different.
11.6 Design Validation Test
Though the reliability predicted in the development stage is more accurate than the
reliability predicted in the design stage, it is still inaccurate due to the following
reasons:
the judgment for FEFs is subjective,
the assumption for the failure process may be unrealistic,
the test conditions and environment can be different from the real operating
conditions,
test observations are insufcient,
the test prototypes may have manufacturing defects, and
the repairs may have quality problems.
As such, the predicted reliability should be viewed as an approximation of the
product inherent reliability. This necessitates a reliability demonstration test to
conrm or validate the prediction obtained from the reliability growth analysis.
Two key issues with the demonstration test are when and how this test is conducted.
The rst issue deals with the relationship between the demonstration test and the
reliability growth testing. To make the results reliable, the demonstration test should
have sufcient test time. On the other hand, the reliability growth testing and
demonstration testing usually require common test facilities and resources, and are
subject to the constraint on the total test time. This implies that more growth testing
can lead to a higher reliability level but will reduce demonstration test time and lead
to a lower demonstration condence, as shown in Fig. 11.2. As such, there is a need
to achieve an appropriate balance between growth testing and demonstration
testing.
The second issue deals with the design of demonstration test. The test plan
involves the determination of number of tested items, test duration, and accept-
ability criterion. Several factors that affect the test plan include the nature of the test
item, the type of demonstration, and the availability of test resources.
To test as more system interfaces as possible, the demonstration test should be
carried out on system or its critical subsystems. The test conditions must be as close
to the expected environmental and operating conditions as possible.
11.6 Design Validation Test 209
Standard test plans used in demonstration testing assume that the failure rate is
constant. The demonstration test is actually an acceptance sampling test, which will
be discussed in detail in Chap. 13.
11.7 A Case Study
11.7.1 Data and Preliminary Analysis
The data shown in Table 11.6 come from Ref. [12] and deal with failure occurrence
times of 13 failure modes from a developmental test, where 25 items are tested for
175 h.
A test-x-nd-test strategy is used. Specically, Mode 8 is partially corrected at
100 h, Modes 1, 5, 7, 8, 9, 10, and 11 are partially corrected at the test end, and the
other six modes are not corrected. As such, the test consists of two stages: t 2
0; 100 and t 2 100; 175. A simple prediction analysis for the future failure
intensity has been carried out in Example 11.4.
In this section, we carry out a detailed analysis for the data using the power-law
model for a system with multiple failure modes discussed in Sect. 11.5.3.3. We also
predict the number of unobserved failure modes using the software reliability
growth models presented in Sect. 11.5.1.
11.7.2 Assessment and Prediction of Failure

Intensity of Each Mode
The failure modes can be divided into three categories:

Failure modes without corrective actions,
Failure modes with the corrective actions, which are implemented at t 175,
and
Failure modes with the corrective actions, which are implemented at t 100.
We separately discuss each case as follows.
11.7.2.1 Failure Intensities of Failure Modes

Without Corrective Actions
For this category of failure modes, we only need to assess their interval failure
intensities based on the power-law model using the maximum likelihood method.
The results are shown in Table 11.7, where l1 gC1 1=b and k 1=l1 .
Table 11.6 Failure modes and occurrence times of failures and corrective actions
Mode FEF System Failure times Corrective action time
1 0.5 19 106.3 175
2 0 7 107.3
15 100.5
25 10.6
3 0 3 79.4
6 67.3, 67.1, 70.8, 162.3
14 102.7
17 100.8, 126.3
22 13.9, 64.1, 84.4
23 39.2, 48.5, 45.6, 53.3, 68.7
4 0 20 97.0
25 36.4
5 0.8 21 70.8 175
6 0 3 148.1
7 70.6
8 118.1
22 18.0
7 0.7 15 37.8 175
8 0.8 6 74.3 100
7 65.7, 93.1
9 0.5 13 90.8 175
10 0.5 13 99.2 175
11 0.5 17 130.8 175
12 0 5 169.1
15 114.7
16 6.6
17 154.8, 5.7, 21.4
23 125.4, 140.1
13 0 7 102.7
Table 11.7 Parameters of Mode b g l1 k, 103

power-law model and failure
intensities for failure modes 2 0.7797 2654.94 3065.53 0.3262
without corrective actions 3 1.0250 270.48 267.76 3.7347
4 0.9258 2678.32 2776.28 0.3602
6 1.0689 971.96 947.12 1.0558
12 0.8049 720.78 813.11 1.2298
13 1.8763 972.98 863.75 1.1577
Sum 7.8646
11.7 A Case Study 211
From the table we have the following observations:

The values of b of Modes 3, 4, and 6 are close to 1, and the values of b of the
other three modes are not close to 1. This implies that the constant-failure-
intensity assumption is not always true.
About 47.5 % of the total failure intensity comes from Mode 3 and hence it is a
reliability bottleneck.
An upper bound of MTBF can be obtained through neglecting the effect of the
other two categories of failure modes. Using the value from the last row of
Table 11.7 yields that the upper bound equals 127.15 h. This is consistent with
the result obtained in Example 11.4, where the model given by Eq. (11.24) gives
l1 126:08. This conrms the conclusion that the estimate obtained from
Eq. (11.24) is superior to the estimate obtained from Eq. (11.22) for
Example 11.4.
11.7.2.2 Intensities of Failure Modes with Corrective

Actions Implemented at t 175
For this category of failure modes, we rst assess their interval failure intensities in
t 2 0; 175, and then predict the intensities using Eq. (11.6). The assessed inten-
sities are shown in the fth column of Table 11.8 and the predicted intensities are
shown in the last column. The last row shows the predicted MTBF (i.e., 93.19)
without considering the contribution of Mode 8.
It is noted that the predicted MTBF is smaller than the observed MTBF (i.e.,
25 175=43 101:74). This is because the observed MTBF is estimated based on
a constant-intensity assumption. Actually, if no corrective action, the total failure
intensity obtained from the power-law model is 13:80 103 and the intensity after
taking account of the corrective actions is 10:73 103 . As a result, the reduction
in the intensity is 3:07 103 , and the average FEF equals 0.2863. In this sense,
the reliability gets improved.
Table 11.8 Parameters of power-law model and failure intensities for failure modes corrected
at t = 175
Mode b g l1 k, 103 1 dk, 103
1 2.0059 870.85 771.73 1.2958 0.6479
5 1.1051 3221.61 3103.87 0.3222 0.0644
7 0.6525 24286.12 33040.9 0.0303 0.0091
9 1.5241 1446.32 1303.22 0.7673 0.3837
10 1.7617 1087.87 968.51 1.0325 0.5163
11 3.4351 446.68 401.50 2.4906 1.2453
Sum 2.8667
MTBF 93.19
An interesting nding is that the corrected failure modes generally have larger
values of b than those failure modes without corrective actions. In fact, the average
of the bs values in Table 11.7 is 1.0801 and the average in Table 11.8 is 1.7474.
This implies that the value of b can provide useful clue in failure cause analysis. It
also conrms the observation that the constant-failure-intensity assumption is not
always true.
11.7.2.3 Intensity of Failure Mode with Corrective

Action Implemented at t 100
The corrective action for Mode 8 is implemented at t 100 and its effectiveness
has been observed in the second test stage when no failure occurs with this mode.
The FEF value (=0.8) indicates that the corrective action cannot fully remove this
mode. Clearly, we cannot directly assess the failure intensity of this mode in the
second test stage based on the observed data due to no failure observation available.
A method to solve this problem is outlined as follows.
Simultaneously consider the data from the two stages. In this case, the overall
likelihood function consists of two parts. The rst part is the likelihood function in
the rst test stage with parameters b1 and g1 , and the second part is the likelihood
function in the second test stage with parameters b2 and g2 . Assume that
b1 b2 b. The MTTF is given by
li gi C1 1=b; i 1; 2: 11:38
Letting ki 1=li , Eq. (11.6) is revised as
g2 g1 =1 d: 11:39
Maximizing the overall likelihood function yields the estimates of b and g1 , and the
failure intensity after considering the effect of the corrective action is given by [4]
k2 1 d=g1 C1 1=b: 11:40
Table 11.9 shows the estimated model parameters and predicted failure intensity
for both the constant-intensity and power-law models. In terms of the log maximum
Table 11.9 Predicted failure Model b g2 k2 , 103 lnL

intensity for mode 8
Constant intensity 1 4791.67 0.2087 28.4239
Power-law 3.8236 9380.08 0.1179 26.2099
MTBF 92.17
likelihood value, the power-law model is more appropriate than the constant-
intensity model. It is noted that the predicted failure intensity from the power-law
model is much smaller than the one obtained from the constant-intensity model.
This implies that the constant-intensity assumption can lead to unrealistic estimate
of the failure intensity when b is not close to 1.
The total failure intensity from all the failure modes is now 10:85 103 , and
hence the eventual MTBF is 92.17 h.
11.7.3 Prediction of Unobserved Failure Modes
If we continue the testing, more failure modes may be found. Each failure mode has
a contribution to the overall system failure intensity. As such, there is a need to
consider the influence of unobserved failure modes on reliability. To address this
issue, we need to look at the following two issues:
To estimate the cumulative number of failure modes expected in future testing,
and
To estimate the contribution of the unobserved failure modes to the total failure
intensity.
We look at these issues as follows.
11.7.3.1 Modeling Cumulative Number of Failure Modes
Let ti denote the earliest occurrence time of mode i, Mt denote the expected
number of failure modes observed by time t, and M1 denote the expected total
number of failure modes in the system. The new failure mode introduction process
in complex system is somehow similar to the software reliability growth process
and hence the software reliability growth models can be used to model this process.
For the case study under consideration, the earliest occurrence times of failure
modes are summarized in Table 11.10. Since the sample size is small, we consider
two simple models: (a) Gt is the exponential distribution, and (b) Gt is the
Pareto distribution with b 1. The maximum likelihood estimates of the param-
eters and performances are shown in Table 11.11. Figure 11.4 shows the plots of
the tted models. As seen, the growth curves are different for large t.
Table 11.10 Failure mode Mode 12 2 3 6 4 7 8

rst occurrence times
ti 5.7 10.6 13.9 18 36.4 37.8 65.7
Mode 5 9 10 13 1 11
ti 70.8 90.8 99.2 102.7 106.3 130.8
Table 11.11 Estimated Model M1 g lnL0 SSE Ic

model parameters
Exponential 15.15 89.57 65.2537 7.2332 122.8858
Pareto 22.35 125.82 65.3978 6.8025 122.3761
20
Pareto
15
Exponential
M(t )
10
0
0 100 200 300 400 500
t
Fig. 11.4 Estimate of unobserved failure modes
In terms of lnL0 , the exponential model provides better t to the data; in terms
of SSE, the Pareto model provides better t to the data. We combine these two
criteria into the following criterion:

SSE
Ic n ln 2 lnL0 : 11:41
n
The last column of Table 11.11 shows the values of Ic . As seen, the Pareto model
provides better t to the data in terms of Ic .
If the test is continued, the expected time to the occurrence of the jth failure
mode can be estimated from the tted Pareto model, which is given by
gj
sj ; j [ 13: 11:42
M1 j
For example, the 14th failure mode may appear at about t 211 test hours.
11.7.3.2 Contribution of Unobserved Failure Modes to Failure Intensity
Let kj denote the failure intensity of mode j and assume
kj k=sj 11:43
where k is a positive constant. The contribution of the rst j failure modes to the
total failure intensity is given by
1
Exponential
Pareto
0.8
0.6
C(j )
0.4
0.2
0
0 5 10 15 20 25
j
Fig. 11.5 Effect of unobserved failure modes on failure intensity
X
j X
M1 X
j X
M1
Cj kl = kl s1
l = s1
l : 11:44
l1 l1 l1 l1
As such, the contribution of the unobserved failure modes to the total failure
intensity is given by 1 Cj. Figure 11.5 shows the plots of Cj for the two tted
growth models.
Let kc denote the current estimate of the intensity (with J identied modes)
without considering the contribution of the unobserved failure modes; and ka
denote the additional intensity from the unobserved failure modes. We have
kc CJ
; kc ka kc =CJ; l0J CJlJ 11:45
ka 1 CJ
where lJ is the current estimate of MTBF without considering the contribution of

0
the unobserved failure modes; and lJ is the revised estimate of MTBF to consider
the effect of the unobserved failure modes.
For the current example, J 13, lJ 92:17, CJ 0:9601 (Pareto model) or
0.9855 (Exponential model). As such, l0J 88:49 (Pareto model) or 90.84
(Exponential model). Clearly, the intensity contribution of new modes is small so
that the additional test to correct the new modes will not lead to a signicant
improvement in the estimate of MTBF.
11.7.4 Discussion
For the data shown in Table 11.6, an intuitive estimate of MTBF without consid-
ering the effect of corrective actions is given by 25 175=43 101:74. It appears
140
Stage 2
120
100 Stage 1
80
m (t )
60
40
20
0
0 50 100 150 200 250 300
t
Fig. 11.6 Reliability growth plan curve for the case study
that the predicted reliability is an underestimate of MTBF. The two main causes to
have this impression are as follows.
The intuitive estimate comes from the constant-intensity model and has not
considered the effect of unobserved failure modes. However, the values of b
associated with nine failure modes are larger than 1, and six of them have
b 1. As a result, the prediction based on the constant-intensity assumption
may give an unrealistic estimate.
The observed reliability growth may partially come from the corrective actions
for the manufacturing and repair quality problems of the tested items. If this is
the case, the predicted reliability may be an underestimate since the manufac-
turing quality in mass production is expected to be better than the manufacturing
quality of prototypes.
Finally, it is important to differentiate the instantaneous intensity from interval
intensity. The instantaneous intensity is suitable for the case where improvement in
reliability continuously occurs and its value at the end of a stage should be viewed
as a prediction for the next stage. If the conguration is unchanged for a given test
stage, no reliability growth occurs in this stage. In this case, we should use the
interval intensity to evaluate the reliability.
11.7.5 Reliability Growth Plan Curve
We can use the data in Table 11.6 to derive a reliability growth plan curve. For the
purpose of illustration, we make the constant-intensity assumption though it can be
untrue. The estimates of MTBF in the two stages are shown in Table 11.4, and the
estimates are graphically displayed in Fig. 11.6 (the four dots). Fitting these points
to the Duane model given by Eq. (11.4) using the least squares method yields
a 42:12 and b 0:1981. Letting g0 100 yields l0 84:11. As such, the

reliability growth plan model can be written as
84:11 t b
lt : 11:46
1 b 100
The growth plan curve is shown in Fig. 11.6 (the continuous curve).
References
New York
2. Calabria R, Guida M, Pulcini G (1996) A reliability-growth model in a Bayes-decision
framework. IEEE Trans Reliab 45(3):505510
3. Crow LH (1974) Reliability analysis for complex, repairable systems. In: Proschan F, Serfling
RJ (eds) Reliability and biometry. SIAM, Philadelphia, pp 379410
4. Crow LH (2004) An extended reliability growth model for managing and assessing corrective
actions. In: Proceedings of annual reliability and maintainability symposium, pp 7380
5. Crow LH (2006) Useful metrics for managing failure mode corrective action. In: Proceedings
of annual reliability and maintainability symposium, pp 247252
6. Duane JT (1964) Learning curve approach to reliability monitoring. IEEE Trans Aero 2
(2):563566
7. Fries A, Sen A (1996) A survey of discrete reliability-growth models. IEEE Trans Reliab 45
(4):582604
8. Hossain SA, Dahiya RC (1993) Estimating the parameters of a non-homogeneous Poisson-
process model for software reliability. IEEE Trans Reliab 42(4):604612
9. Jiang R (2009) Required characteristics for software reliability growth models. In: Proceedings
of 2009 world congress on software engineering, vol 4, pp 228232
10. Jiang R (2011) Three extended geometric process models for modeling reliability deterioration
and improvement. Int J Reliab Appl 12(1):4960
11. Meth MA (1992) Reliability-growth myths and methodologies: a critical view. In: Proceedings
of annual reliability and maintainability symposium, pp 337342
12. OConnor PDT (2002) Practical reliability engineering, 4th edn. Wiley, New York
Part III
Product Quality and Reliability
in manufacturing Phase
Chapter 12
Product Quality Variations and Control
Strategies
12.1 Introduction
Manufacturing is the process of transforming inputs (raw materials, components,

etc.) into nished products [1]. The process used for manufacturing a product
depends on the demand for the product. If the demand is high, it is economical to
use a continuous production process; otherwise, it is more economical to use a
batch production process. The major challenge in this process is to retain the
designed-in performance. Two key issues are to control product quality and to
improve the production process. Product quality problem results from variations in
quality characteristics and the reliability of production systems signicantly impacts
the variations. Strategies to retain the desired product performance include testing,
statistical process control, and process optimization. In this chapter, we focus on
these issues.
The outline of the chapter is as follows. Section 12.2 deals with variations of
quality characteristics and their effect on product quality and reliability. The reli-
ability of production systems is analyzed in Sect. 12.3. Typical quality control and
improvement strategies are discussed in Sect. 12.4. Finally, we briefly discuss
quality management-related issues in Sect. 12.5.
12.2 Variations of Quality Characteristics and Their Effect

on Product Quality and Reliability
12.2.1 Variations of Quality Characteristics and Variation

Sources
Quality characteristics are the parameters that describe the product quality such as
length, weight, lifetime, number of defects, and so on. The data on quality

222 12 Product Quality Variations and Control Strategies
characteristics can be classied as attribute data (which take discrete integer values,
e.g., number of defects) and variable data (which take continuous values, e.g.,
lifetime).
Quality characteristics of a product are usually evaluated relative to design
specications. Specications are the desired values of quality characteristics on the
product or its components. The specications are usually specied by nominal
value, lower specication limit, and upper specication limit. Components or
products are nonconforming or defective if one or more of the specications are not
met.
Despite the efforts made during the design and development phases to ensure
optimal production and assembly characteristics, no production system is able to
produce two exactly identical outputs. Unit-to-unit difference in quality character-
istics is referred as variability.
The variability results from differences or variations in input materials, perfor-
mance of manufacturing equipment, operator skills, and other factors. These factors
are called sources of variation and can be divided into six aspects: Materials,
Manufacture, Man, Machine, Measurements and Environment, which are simply
written as 5M1E (e.g., see Ref. [2]). To discover the key variation sources of a
given quality problem, one can use a 5M1E approach. The approach rst generates
a check list for each aspect of the 5M1E based on empirical knowledge, and then
uses the check list to identify the potentially possible causes. The quality problem
can be solved by removing the impacts of those causes on the quality variability.
The causes that cause quality variation can be roughly classied into two types:
random cause (also termed as common cause or background noise), and
assignable cause (also termed as special causes).
The random causes are many small and unavoidable causes, and result in
inherent variability. A process that is only subjected to random causes is said to be
in statistical control. In practice, most of the variability is due to this type of causes.
Generally, nothing can be done with these causes except to modify the process.
Therefore, these causes are often called the uncontrollable causes.
The assignable causes include improperly adjusted machines, operator errors,
and defective raw material. The variability due to assignable causes is generally
large so that the level of process performance is usually unacceptable. A process
that is operating in the presence of assignable causes is said to be out of control. The
variability due to this type of causes can be controlled through effective quality
control schemes and process modication such as machine adjustment, mainte-
nance, and training for operators.
The probability that an item produced is nonconforming depends on the state of
manufacturing process. When the state is in control, the probability that an item
produced is nonconforming is very small although the nonconformance cannot be
avoided entirely. When the state changes from in-control to out-of-control due to
one or more of the controllable factors deviating signicantly from their target
12.2 Variations of Quality Characteristics 223
values, the probability of occurrence of nonconformance considerably increases. In

this case, some action has to be initiated to get the out-of-control state back to in
control.
Lifetimes observed in eld for nominally identical items (components or prod-
ucts) can be very different. This results from variability in various failure-related
factors. These factors roughly fall into two categories:
manufacturing variations, and
operating and environmental conditions.
The variability due to manufacturing variations (including raw materials variability)
is called unit-to-unit variability, and the other sources of variability are called as
external noise.
12.2.2 Effect of Unit-to-Unit Variability on Product Quality

and Reliability
12.2.2.1 Effect of Variability on Product Quality
As mentioned earlier, the lifetimes of nominally identical items can be different due
to unit-to-unit variability. The product reliability realized in manufacture phase is
called the inherent reliability, which is usually evaluated using the life test data of
the product after the product is manufactured. The test data are obtained from
strictly controlled conditions without being impacted by actual operating conditions
and maintenance.
Assume that the life follows the Weibull distribution with parameters b and g.
The life variability can be represented by the coefcient of variation r=l given by
s
C1 2=b
r=l 1: 12:1
C2 1 1=b
It is easy to show that a large value of b corresponds to a small variability.

Further, assume that s is the maximum acceptable life of a product, which can be
written as kl. For given s and l, k is a known constant. In this case, the non-
conformance fraction is given by
p 1 expfkC1 1=bb g: 12:2
Figure 12.1 shows the plots of r=l and p (with k 0:5) versus b. It clearly
shows that the smaller the variability is, the better the quality is. This is consistent
with the conclusion obtained by Jiang and Murthy [5].
Fig. 12.1 Plot of p and r=l 2.5

versus b
2
1.5
p,
1 p
0.5
0
0 1 2 3 4 5 6 7

12.2.2.2 Effect of Variability on Product Reliability
The items that do not conform to design specications are nonconforming. There
are two types of nonconforming items. In the rst case, the item is not functional
and this can be detected immediately after it is put in use. This type of noncon-
formance is usually due to defects in assembly (e.g., a dry solder joint). In the other
case, the item is functional after it is put in use but has more inferior characteristics
(e.g., a shorter mean life) than the conforming item. Such items usually contain
weak or nonconforming components, and cannot be detected easily. Jiang and
Murthy [4] develop the models to explicitly model the effects of these two types of
nonconformance on product reliability. We briefly outline them as follows:
We rst look at the case of component nonconformance. Let F1 t denote the life
distribution of a normal component and the proportion of the normal product is p.
Assume that the life distribution of the product with weak components is F2 t and
the proportion of the defective product is q 1 p. The life distribution of the
product population is given by
G1 t pF1 t qF2 t: 12:3
The relative life spread can be represented by the coefcient of variation of

lifetime, and a small value of the coefcient of variation is desired. We now
examine the relative life dispersion of the item population. To be simple, we
consider the case where the life of each subpopulation follows the Weibull distri-
bution with parameters bi and gi (i 1; 2). Let qi denote the coefcient of variation.
The mean and variance of the mixture are given, respectively, by
l pl1 ql2 ; r2 pr21 qr22 pql2 l1 2 : 12:4
From Eq. (12.4), we have
pr21 qr22 pql21 1 d2 pq21 qq22 d2 pq1 d2

q2 12:5
l21 p qd2 p qd2
12.2 Variations of Quality Characteristics 225
where d l2 =l1 . After some simplications, Eq. (12.5) can be written as:
pq1 d2 1 q21 qd2 q22 q21

q2 q21 : 12:6
p qd2
Since the normal item has longer life and smaller life dispersion, we have d\1
and q1 q2 . From Eqs. (12.4) to (12.6), we have l\l1 and q2 [ q21 , implying that
the life of the item population is smaller than the life of the normal item and has
larger life dispersion.
We now look at the case of assembly errors without component nonconfor-
mance. Assume that the life distribution of the product with assembly error is F3 t
and the proportion of such products is r. For this case, the life distribution of the
product population is given by
G2 t 1 1 F1 t1 rF3 t: 12:7
Considering the joint effect of both assembly errors and component noncon-
formance, the life distribution of the product population is given by
G3 t 1 1 G1 t1 rF3 t: 12:8
In general, the product inherent reliability is represented by G3 t. The mean life

derived from G3 t is smaller than the one associated with F1 t; the life dispersion
derived from G3 t is larger than the one associated with F1 t; and the failure rate
associated with G3 t can be nonmonotonic [4].
12.2.3 Effect of Operating and Environmental Factors

on Product Reliability
Field reliability of product depends on the inherent reliability of product and,

operating and environmental conditions. A large proportion of unanticipated reli-
ability problems result from unanticipated failure modes caused by environmental
effects.
Accurately modeling eld reliability needs to have knowledge of products
usage prole, sequence of operation, use environments, preventive maintenance
regime, and their joint effect on reliability. The generalized proportional intensity
model discussed in Chap. 9 can be an appropriate model to represent the failure
intensity.
12.3 Reliability and Design of Production Systems
12.3.1 Reliability of Production Systems
A complex production system is composed of a variety of components. Some of

components are subjected to catastrophic failure and wear. The excessive wear can
lead to poor quality of products and hence may also be regarded as a system failure
if it affects product quality severely. Therefore, the failures of production systems
include the catastrophic failure of components of the system, and product quality
deterioration due to the wear of components of the system. The catastrophic failure
(or hard failure) usually occurs at the early stage of production; and the degradation
failure (or quality failure) usually occurs after the production system has operated
for a relatively long period of time.
12.3.1.1 Modeling Hard Failure
A production system and its components are generally repairable. Therefore, we

need to use a failure process model to represent the stochastic behavior of failure
and repair of the system. For a given component, its failure data can be collected for
characterizing the failure process. If the failure process is stationary, the tted
power law model has b 1. In this case, the interfailure times can be tted to a
distribution. Suppose that there are n components, each of which can cause a
system hard failure, and the failures of components are mutually independent. In
this case, the failure of each component can be modeled by a life distribution, and
the reliability function associated with the system hard failures is given by
Y
n
Rf t Rj t 12:9
j1
where Rj t is the reliability function of the jth components.

Example 12.1 The data shown in Table 12.1 comes from Ref. [7] and deals with
interfailure times of an electric motor, which is a key component of the transfer
system of a valve assembly line. In the table, sign + indicates a censored
observation. A stationarity test indicates the null assumption that the failure process
is stationary is not rejected. Therefore, the data can be tted to a life distribution.
Fitting the data to the Weibull distribution yields the maximum likelihood esti-
mates: b 3:6708 and g 2495:8. The tted model can be used to schedule the
Table 12.1 Times to failure 1124 667 2128 2785 700+ 2500+ 1642 2756
(h)
3467 800+ 2489 2687 1974 1500+ 1000+ 2461
1945 1745 1300+ 1478 1000+ 2894 1500+ 1348
3097 1246 2497 2674 2056 2500+
12.3 Reliability and Design of Production Systems 227
preventive maintenance actions. For example, if the motor is preventively replaced

at the tradeoff BX life (which is 1751.3 h), 76.15 % of failures can be prevented.
12.3.1.2 Modeling Quality Failure
Let t 2 0; T denote the operating time of a continuous production line and pt

denote the probability or proportion that the item is conforming. The production
line is stopped at time T due to a corrective or preventive maintenance action. The
expected proportion of conforming items is given by
ZT
1
pT ptdt: 12:10
T
0
For batch production, the items are produced in lots of size Q. At the start of
each lot production, the state is in control. Let pi denote the probability that an
item is conforming. The expected proportion of conforming items is given by
1X
Q
pQ pi: 12:11
Q i1
Generally, pT [pQ ] decreases as T [Q] increases and depends on the components

wears and maintenance scheme.
The tool wear is accumulated with use, and signicantly affects the product
dimensional deviation. Let Wt denote the aggregated component wear after
operated for t time units. When t is large, the component aggregated wear
approximately follows a normal distribution with mean function lt and standard
deviation rt. Let wt dlt=dt denote the wear rate. For a mechanical com-
ponent under normal production conditions, the initial wear rate w0 is generally
high, then the wear rate will reduce and tend to a constant w1 . As such, the wear
rate can be approximated by:
wt w1 w0 w1 et=g ; g [ 0: 12:12
Equation (12.12) does not reflect the wear behavior in the wear-out stage and hence
is applicable only for the early and normal wear stages. The accumulated wear
amount is given by
Zt
Wt wxdx w1 t gw0 w1 1 et=g : 12:13
0
The quality failure can be modeled by relating WT or WQT to pT or pQ .

12.3.2 Design of Production Systems
The design of production system has a signicant impact on the fraction of con-
formance when the process is in control. Important issues for production system
design include supply chain design, production planning, system layout, equipment
selection, and production management.
Supply chain design involves supplier selection and contract specication. It also
deals with choosing the shipping mode and warehouse locations.
Production planning deals with the issues such as manufacturing tolerance
allocation, process planning, and process capability analysis (which will be dis-
cussed in Chap. 14) to predict the performance of a production system.
Main considerations for system layout are the systems flexibility and robust-
ness. The flexibility is the capability of producing several different products in one
system with no interruption in production due to product differences. Flexibility is
desired since it enables mass customization and high manufacturing utilization. A
robust production system is desired so as to minimize the negative influence of
fluctuations in operations on product quality. This can be achieved through using
the Taguchi method to optimally choose the nominal values of controllable factors.
Production equipment determines operating characteristics (e.g., production line
speed) and reliability. The speed impacts both quality and productivity. A high line
speed can increase productivity but harm quality. As such, the speed is a key factor
for equipment selection and needs to be optimized to achieve an appropriate
tradeoff between quality and productivity.
Production management focuses on the continuous improvement of product
quality. Quality improvements can be achieved by identifying and mitigating
quality bottlenecks. A quality bottleneck is the factor that can signicantly impact
product quality. Improving the bottleneck factor will lead to the largest improve-
ment in product quality.
Machine breakdowns affect product quality. Preventive maintenance improves
the reliability of production system and in turn improves quality. This necessitates
effectively planning preventive maintenance to mitigate machine deterioration.
12.4 Quality Control and Improvement Strategies
Various statistical techniques have been developed to control and improve quality.
Major strategies for quality control and improvement in a production system are
shown in Fig. 12.2. As seen, the quality control and improvement strategies fall into
the following three categories:
inspection and testing for raw materials and nal product,
statistical process control, and
quality control by optimization.
12.4 Quality Control and Improvement Strategies 229
Fig. 12.2 Techniques for

Test and Controllable Statistical
quality control and
optimization inputs process control
improvement in a production
system
Production Output: quality

process characteristics
Uncontrollable
Product test
inputs
Since quality is obtained by design and manufacturing activities, optimization can

be the most effective to quality improvement and reduction of variability among
these techniques. In this section, we briefly outline these techniques and illustrate
the optimization technique by examining the optimal lot size problem.
12.4.1 Inspection and Testing
Acceptance sampling is a technique to control the quality of the inputs of the

production process, and can also be applied to the nal product. Acceptance
sampling for raw materials, parts, components, or subassemblies that usually come
from other manufacturers or suppliers is called incoming inspection, and acceptance
sampling for the nal product is called outgoing inspection. We will discuss
incoming inspection in detail in the next chapter.
The nal products are often subjected to other tests such as environment stress
screening and burn-in to eliminate the defective products. These will be discussed
in detail in Chap. 15.
12.4.2 Statistical Process Control
The production process can be controlled using statistical process control tech-
niques, which include off-line and online quality-control techniques, depending on
the type of manufacturing process. In continuous production, the process often rst
operates in the in-control state and produces acceptable product for a relatively long
period of time, and then assignable causes occur so that the process shifts to an out-
of-control state and produces more nonconforming items. The change from in-
control to out-of-control can be detected through regularly inspecting the items
produced and using control charts. A control chart is a graphical tool used to detect
the process shifts. When a out-of-control state is identied, appropriate corrective
actions can be taken before many nonconforming units are manufactured. The
control chart technique will be discussed in detail in Chap. 14.
In batch production, the production system is set up and may be subjected to a
preventive maintenance before going to production, and hence the process starts in
control and can go to out of control during the production of a lot. As the lot size
increases, the expected fraction of nonconforming items in a lot increases and the
set-up cost per manufactured unit decreases. Therefore, the optimal lot size can be
determined by a proper tradeoff between the manufacturing cost and the benets
derived through better outgoing quality. This approach deals with quality control by
optimization. We look at the optimal lot size problem as follows.
12.4.3 Quality Control by Optimization
Let Q denote the lot size. At the start of each lot production, the process is in
control. The state can change from in-control to out-of-control. If the state is in out-
of-control state, the process will remain there until completion of the lot. Since the
expected fraction of nonconforming items increases and the setup cost per item
decreases as Q increases, an optimal batch size exists to minimize the expected
manufacturing cost per conforming item.
Let p0 [p1 ] denote the probability of occurrence of nonconforming items when
the manufacturing process is in control [out of control]. Clearly, we have p0 p1 .
Let N 2 0; Q denote the state change point, after which the process is out of
control. Since the probability of N 0 is zero, N is a random positive integer.
When 1 N\Q, the process ends with the out-of-control state; otherwise, the
process ends with the in-control state. Assume that the probability that the in-
control state changes to out-of-control state is q. For 1 N Q 1, we have
pi PrfN ig pi1 q 12:14
where p 1 q. The probability of N Q is given by
pC PrfN Qg pQ1 : 12:15
P
Q1
It is noted that pi pC 1.
i1
Conditional on N i 2 1; Q 1, the expected number of nonconforming
items equals
ni p0 i p1 Q i p1 Q p1 p0 i: 12:16
12.4 Quality Control and Improvement Strategies 231
Removing on the condition, we have the expected number of nonconforming items

given by
X
Q1
N1 pini Qp0 pC : 12:17
i1
The expected fraction of conforming items in a lot is given by

N1
pQ 1 : 12:18
Q
We now look at the cost elements. The setup cost depends on the state in the end
of previous run. It is cs if the state in the end of previous run is in control, and an
additional cost d is needed if the state in the end of previous run is out of control.
The probability that needs the additional cost is given by
X
Q1
pA pi 1 pQ1 : 12:19
i1
As such, the expected setup cost is given by

c0 cs dpA : 12:20
Let c1 denote the cost of producing an item (including material cost and labor
cost) and c2 denote the penalty cost of producing a nonconforming item. The
penalty cost depends on whether the nonconforming item has been identied before
being delivered to the customer. If yes, it includes disposal cost; if not, it includes
warranty cost. These costs are independent of Q.
The expected total cost is given by
CQ c0 c1 Q c2 1 pQ Q: 12:21
The total cost per conforming item is given by
c0 c1 Q c2 1 pQ Q
JQ : 12:22
pQ Q
The optimal lot size Q is achieved so that JQ achieves its minimum.

Example 12.2 Let q 0:005, p0 0:01 and p1 0:3. Assume that c1 1,
c2 0:1, cs 20 and d 5. The problem is to nd the optimal lot size.
Using the approach outlined above yields Q 201. Other relevant parameters
are shown in Table 12.2.
Table 12.2 Results for pC pA pQ N1 c0 CQ JQ

Example 12.2
0.3670 0.6330 0.8832 23.5 23.17 226.51 1.2760
12.5 Quality Management
Quality management aims to ensure an organization to achieve consistent product

quality. In this section, we briefly introduce principles of quality management, total
quality management, ISO quality management system, six sigma quality, and its
implementation process.
12.5.1 Principles of Quality Management
Well-known pioneers in quality include W. Edwards Deming, Joseph M. Juran, and

Armand Feigenbaum (e.g., see Ref. [6]). Deming emphasizes statistics and the role
of management. His recommendations for quality management are known as
Demings 14 points. Juran emphasizes organization for change and implementation
of improvement through managerial breakthrough. His approach to quality focuses
on planning, control, and improvement. Feigenbaums approach to quality focuses
on quality leadership, quality technology, and organizational commitment.
The basic principles of quality management that are widely recognized can be
summarized into one focus, three human or organization related factors, and
four approaches.
The one focus is customer focus. It means that an organization should strive to
understand customer needs and try to meet and exceed the expectations of customers.
The three human or organization related factors are leadership, involvement of
people, and mutually benecial supplier relationship. Specically, the leaders of an
organization should create and maintain such an internal environment, in which
people can become fully involved in achieving the organizations quality objective;
the abilities of people at all levels of an organization are completely used for the
benet of the organization; and the relationship between an organization and its
suppliers should be mutually benecial.
The four approaches are process approach, system approach, fact-based
approach, and continual improvement. Specically, activities and related resources
in an organization should be managed as a process; all interrelated processes in
achieving the quality objectives of an organization should be identied, understood,
and managed as a system; decisions should be made based on data analysis and
information; and the efforts to improve the overall performance of an organization
should never end.
12.5.2 Quality Management Strategies
12.5.2.1 Total Quality Management
Total quality management (TQM) is a strategy for implementing and managing

quality improvement activities in an organization. The core principles of TQM are
12.5 Quality Management 233
customer focus, involvement of all employees, and continuous improvement. It

emphasizes on widespread training and quality awareness.
TQM typically involves three kinds of teams with different focuses. A high-level
team deals with strategic quality initiatives, workforce-level teams focus on routine
production activities and cross-functional teams address specic quality improve-
ment issues.
Effectiveness of TQM is limited since only the continual improvement approach
in the four approaches of the quality management principles is emphasized.
12.5.2.2 Six Sigma Quality
In statistics, sigma is usually used to denote the standard deviation; which repre-
sents the variation about the mean. Assume that a quality characteristic follows a
normal distribution with mean l and the standard deviation r, and the specication
limits are l D. If D 6r, the probability that a product is within the specica-
tions is nearly equal to 1. As such, the six sigma concept can be read as nearly
perfect, defect-free performance or world-class performance.
Six sigma quality is a systematic and fact-based process for continued improve-
ment. It focuses on reducing variability in key product quality characteristics.
The six sigma implementation process involves ve phases. The rst phase is
design or dene phase. It involves identifying one or more project-driven
problems for improvement. The second phase is measure phase. It involves col-
lecting data on measures of quality so as to evaluate and understand the current state
of the process. The third phase is analyze phase. It analyzes the data collected in
the second phase to determine root causes of the problems and to understand the
different sources of process variability. The fourth phase is improve phase. Based
on the results obtained from the previous two phases, this step aims to determine
specic changes to achieve the desired improvement. Finally, the fth phase is
control phase. It involves the control of the improvement plan.
12.5.3 ISO Quality Management System
ISO 9000 series are the quality standards developed by the International
Organization for Standardization [3]. These standards focus on the quality system
with components such as management responsibility for quality; design control;
purchasing and contract management; product identication and traceability;
inspection and testing; process control; handling of nonconforming, corrective and
preventive actions; and so on. Many organizations require their partners or suppliers
to have ISO 9000 certication.
According to a number of comparative studies for the actual performance of the
enterprises with and without ISO 9000 Certication, its effectiveness strongly
depends on the motivation for the certication, i.e., just getting a pass or really
wanting to get an improvement in quality. This is because much of the focus of ISO
9000 is on formal documentation of the quality system rather than on variability
reduction and improvement of processes and products. As such, the certication
only certies the processes and the system of an organization rather than its product
or service.
References
New York, pp 492493
2. Han C, Kim M, Yoon ES (2008) A hierarchical decision procedure for productivity innovation
in large-scale petrochemical processes. Comput Chem Eng 32(45):10291041
3. International Organization for Standardization (2008) Quality management systems. ISO
9000:2000
4. Jiang R, Murthy DNP (2009) Impact of quality variations on product reliability. Reliab Eng
Syst Saf 94(2):490496
7. Regattieri A (2012) Reliability evaluation of manufacturing systems: methods and applications.
Manufacturing system. http://www.intechopen.com/books/manufacturing-system/reliability-
evaluation-of-manufacturing-systemsmethods-and-applications. Accessed 16 May 2012
Chapter 13
Quality Control at Input
13.1 Introduction
The input material (raw material, components, etc.) is obtained from external
suppliers in batches, and the quality can vary from batch to batch. Acceptance
sampling is a way to ensure high input quality. It carries out tests with a small
sample from a batch. The batch is either accepted or rejected based on the test
outcome. According to the nature of quality characteristics, acceptance sampling
plans can be roughly divided into two types: acceptance sampling for attribute
(where the outcome of test is either normal or defective) and acceptance sampling
for variable (where the outcome of test is a numerical value).
As a quality assurance tool, acceptance sampling cannot be used to improve the
quality of the product. A way for manufacturers to improve the quality of their
products is to reduce the number of suppliers and to establish a strategic partnership
with their suppliers [5]. This deals with the supplier selection problem.
In this chapter, we focus on acceptance sampling and supplier selection. The chapter
is organized as follows. Section 13.2 deals with acceptance sampling for attribute.
Acceptance sampling for a normally distributed variable and acceptance sampling
for lifetime are discussed in Sects. 13.3 and 13.4, respectively. Acceptance sampling
for variable can be transferred to acceptance sampling for attribute and this is discussed
in Sect. 13.5. Finally, we discuss the supplier selection problem in Sect. 13.6.
13.2 Acceptance Sampling for Attribute
13.2.1 Concepts of Acceptance Sampling
Suppose a supplier supplies a lot of items to a manufacturer. The decision of whether

the manufacturer accepts or rejects the lot is made based on the number of defective

236 13 Quality Control at Input
items in a sample taken randomly from the lot. The lot is accepted if the number of
defects is not larger than a prespecied number; otherwise, the lot is rejected.
If the lot is rejected, the lot may be handled in different ways, e.g., returning it to
the supplier or inspecting every item. The latter case is called the rectifying
inspection (or 100 % inspection). In the rectifying inspection, the defective items
will be removed or replaced with good ones.
Acceptance sampling can also be used by a manufacturer to inspect their own
products at various stages of production. The accepted lots are sent forward for
further processing, and the rejected lots may be reworked or scrapped.
13.2.2 Acceptance Sampling Plan
An acceptance sampling plan deals with the design of sampling scheme. Three typical
sampling plans are single-sampling, double-sampling, and sequential sampling. In a
single-sampling plan, one sample of items is randomly taken from the lot, and the
acceptance decision is made based on the information contained in the sample.
In a double-sampling plan, a decision based on the information in an initial
sample can be accept the lot, reject the lot, or take a second sample. If the
second sample is taken, the nal decision is made based on the information from the
initial and second samples.
In a sequential sampling, a decision is made after inspection of each item ran-
domly taken from the lot, and the decision can be accept, reject, or continue
the process by inspecting another item. The process ends when an accept or
reject decision is made. Sequential sampling can substantially reduce the
inspection costs. This is particularly true when the inspection is destructive and the
items are very expensive.
Depending on specic situations, there are other sampling plans (e.g., see
Ref. [5]). For example, two extreme sampling plans are (a) accepting the lot with no
inspection and (b) inspecting every item in the lot and removing all defective units.
For the concise purpose, we focus on the single-sampling plan in this chapter.
13.2.3 Operating-Characteristic Curve
In an acceptance sampling for attribute, let p0 denote a critical fraction defective,

called the acceptable quality level (AQL). It represents the level of quality that the
consumer would consider to be acceptable as a process average. Let p denote the
true fraction defective. Ideally, the lot should be accepted in probability 1 if p p0
and rejected in probability zero if p [ p0 . However, the estimated fraction defective
from the sample often deviates from the true value so that the acceptance
probability of the lot (denote it as Pa p) is usually larger than zero and smaller
than 1. The plot of Pa p versus p decreases with p and is called the
13.2 Acceptance Sampling for Attribute 237
Fig. 13.1 OC curve 1.2

Manufacturer's risk
associated with the binomial 1
distribution Ideal OC curve
0.8 A
Pa
0.6 OC curve
0.4
Customer's risk
0.2
B
0
0 0.01 0.02 0.03 0.04 0.05
p
operating-characteristic curve (or simply OC curve) of a sampling plan. The OC

curve represents the discriminatory power of a sampling plan. The closer to the
ideal case it is, the better the discriminatory power is. Figure 13.1 shows the ideal
and actual OC curves for a given sampling plan.
13.2.4 Average Outgoing Quality
Consider the rectifying inspection where all defective items are replaced with good
ones. Let N and n denote the lot size and the sample size, respectively. The average
fraction defective obtained over a long sequence of lots is p and the acceptance
probability is Pa p. When a lot is accepted, the total inspection number is n and the
outgoing lot has pN n defective items. When a lot is rejected, the total
inspection number is N and the outgoing lot has zero defective items. As a result,
the average fraction defective (or average outgoing quality or AOQ) of all the
outgoing lots is given by
AOQ pN nPa p=N: 13:1
When p ! 0 or 1, AOQ ! 0. This implies that the plot of AOQ versus p is uni-
modal with a maximum value, which is called the average outgoing quality limit.
The average total inspection number per lot is given by
na nPa p N1 Pa p N N nPa p: 13:2
Since Pa p decreases with p and hence na increases from n to N as p increases

from 0 to 1.
13.2.5 Acceptance Sampling Based on Binomial Distribution
Let n denote the sample size, nd denote the number of defective items in the sample,
and c is a critical defective number to be specied. Clearly, nd =n can be viewed as
an estimate of fraction defective p and c=n can be viewed as an estimate of p0 . As

such, p p0 is equivalent to nd c. The lot is accepted [rejected] if nd c
[nd [ c]. The acceptance sampling plan design is to specify the values of n and c.
When the lot size N is very large (e.g., N 10n), the acceptance sampling plan
can be determined based on the binomial distribution. Let X denote the number of
defective items in a sample with n items. If the lot fraction defective is p, the
probability of the event X x is given by
px Cn; xpx 1 pnx : 13:3
The acceptance probability is given by
X
c
Pa p px: 13:4
x0
For given values of c and n, Eq. (13.4) represents an OC curve.

As the sample size n increases, the actual OC curve can be closer to the ideal OC
curve but it will require more inspection time and cost. Two inappropriate
approaches that are sometimes used to design sampling plans are c 0 and a xed
n=N. In the case of c 0, the required sample size will be small since n c=p0 .
When n=N is a xed percentage, the required sample size can be small for small N,
and the required inspection efforts can be large for large N.
To achieve an appropriate tradeoff between the precision and test effort, common
approaches are to control the producers and customers risks. Referring to
Fig. 13.1, the risks are represented by the difference between the ideal and actual
OC curves for a given p. Specically, for p p0 , 1 Pa p represents the pro-
ducers risk; and for p [ p0 , Pa p represents the customers risk. Figure 13.2
displays a risk curve as a function of p. It is noted that the risk achieves its
maximum at p p0 , where the risk curve is generally discontinuous.
In practice, the risks are represented by two specic points in the OC curve. For
example, the manufacturers risk is specied by Point A in Fig. 13.1 with coor-
dinates p1 ; Pa p1 , and the customers risk is specied by Point B with coordi-
nates p2 ; Pa p2 . Since n and c must be integers, it is almost impossible to make
Fig. 13.2 Risk curve 0.7

(p0 0:01, n 295 and 0.6
c 2) 0.5 Producer's risk Customer's risk
0.4
Risk
0.3
0.2
0.1
p0
0
0 0.005 0.01 0.015 0.02 0.025 0.03
p
the OC curve exactly pass these two known points. As such, we nd (n, c) so that
the OC curve is closest to these two desired points and meets the inequalities:
Pa p1 1 a; Pa p2 b: 13:5
where a and b denote the risks of producer and customer, respectively.

An iterative procedure can be used to nd the value of n and c. We start from
c 0. For a xed value of c, we nd the value of n so that the following achieves
its minimum:
SSE 1 a Pa p1 2 b Pa p2 2 : 13:6
As c increases, n increases and the risks decrease. The process is repeated until
Eq. (13.5) can be met. We illustrate this approach as follows.
Example 13.1 Assume that p0 0:01, p1 0:005, p2 0:015 and a b 0:2.
The problem is to nd the values of c and n.
Using the approach outlined above, we obtained the results shown in Table 13.1.
As seen, when c = 2, the inequalities given in Eq. (13.5) can be met. It is noted that
the inequalities can be met for c [ 2 but more inspections are required.
For the sampling plan (c; n) = (2, 295) and N 5000, Fig. 13.3 shows the
average outgoing quality curve. As seen, the average outgoing quality limit equals
0:4372 %, which is achieved at p 0:7673 %. The acceptance probability when
p 0:01 is 0.4334. This implies that the risk of producer is larger than the risk of
customer when p p0 .
Discussion: The approach to specify the two risk points can be troublesome and
potentially unfair. In fact, the producers and customers risks are generally unequal
at p p0 (see Fig. 13.2 and Example 13.1). To improve, Jiang [2] presents an
Table 13.1 Computational c n a b

process for Example 13.1
0 80 0.3304 0.2985
1 186 0.2384 0.2305
2 295 0.1846 0.1800
Fig. 13.3 AOQ curve for 0.005

Example 13.1
0.004
0.003
AOQ
0.002
0.001
0
0 0.01 0.02 0.03 0.04
p
equal-risk approach. In this approach, two conditions are introduced to determine

the values of n and c. The rst condition is that the producers and customers risks
at p p0 are equal to 0.5, and the second condition is to control the producers
average risk given by
Zp0
1
r 1 Pa pdp: 13:7
p0
0
For Example 13.1, when c 2 and n 267, the producers average risk equals
0.1851, and the producers and customers risks at p p0 are 0.499829 and
0.500171, respectively, which are nearly equal. For this scheme, the producers risk
is a 0:1506 at p1 0:005, and the customers risk is b 0:2352 at p2 0:015.
This implies that the risk requirement given by the customer may be too high
relative to the risk requirement given by the producer.
13.2.6 Acceptance Sampling Based

on Hypergeometric Distribution
When the lot size N is not very large, the acceptance sampling should be based on
the hypergeometric distribution, which describes the probability of x failures in n
draws from N items. Let m denote the number of conforming items. The number of
defective items in the lot is N m. Table 13.2 shows possible cases among n, m
and N m, where xL and xU are the lower and upper limits of X, respectively.
The probability of the event that there are x defective items in n items drawn
from N items is given by
px Cmnx CNm
x
=CNn ; x 2 xL ; xU : 13:8
For the sampling plan n; c, the acceptance probability is given by
X
c
Pa p px: 13:9
xxL
Table 13.2 Range of X xL xU

nm 0
n[m nm
nN m n
n[N m Nm
Range of X max0; n m minn; N m
Fig. 13.4 OC curves for the 1

hypergeometric and binomial
0.8
distributions
0.6
Pa
0.4 BN (150)
BN (100)
0.2
HG (150)
0
0 0.01 0.02 0.03 0.04 0.05
p
For given N and p, we take
m int1 pN 0:5: 13:10
As such, the OC curve is dened by Eq. (13.9). The approach to specify n and c is
the same as that outlined in Sect. 13.2.5.
Figure 13.4 displays three OC curves:
(a) HG150, which is associated with the sampling plan based on the hyper-
geometric distribution with N; n; c = (800, 150, 2),
(b) BN100, which is associated with the sampling plans based on the binomial
distribution with n; c = (100, 2), and
(c) BN150, which is associated with the sampling plans based on the binomial
distribution with n; c = (150, 2).
From the gure, we have the following observations:
(a) The OC curve associated with the hypergeometric distribution is not smooth
due to the rounding operation in Eq. (13.10).
(b) The OC curves for the binomial and hypergeometric distributions with the
same n; c are close to each other when n=N is small.
(c) For the same n; c, the discriminatory power of the plan based on the hyper-
geometric distribution is slightly better than the plan based on the binomial
distribution.
13.3 Acceptance Sampling for a Normally

Distributed Variable
Let X denote the quality characteristic with sample being (xi ; 1 i n), and Y
denote the sample mean. The quality characteristic can be nominal-the-best,
smaller-the-better, and larger-the-better. For the larger-the-better case, we set a
lower limit yL . If the sample average is less than the lower limit, the lot is rejected;
otherwise accepted. Similarly, we set an upper limit yU for the smaller-the-better
case; and set both the upper and lower limits for the nominal-the-best case. Since
the smaller-the-better and larger-the-better cases can be viewed as special cases of

the nominal-the-best case, we only consider the nominal-the-best case in the fol-
lowing discussion.
Assume that X approximately follows the normal distribution with mean l and
standard deviation r. As such, Y can be approximated by the normal distribution
p
with mean ly l and standard deviation ry r= n.
The sampling scheme is described by two parameters: n and k, where k is the
difference between the nominal value l0 and the lower or upper limit, i.e.,
k l0 yL yu l0 . Let d0 denote the acceptable quality limit and d1 denote the
rejectable quality level. These imply that the normal item is dened by
jx l0 j d0 , and the defective item is dened by jx l0 j d1 .
For a given ly , the acceptance probability is given by
p p
Ply UyU ; ly ; r= n UyL ; ly ; r= n: 13:11
Letting d jl0 ly j, Eq. (13.11) can be written as below:

p p
Pd Ud k; 0; r= n Ud k; 0; r= n: 13:12
The risk at d d0 is given by 1 a 1 Pd0 , and the risk at d d1 is given by

b Pd1 . For given values of (l0 ; d0 ; d1 ; r; a; b), the initial values of n and k can
be obtained by minimizing the following:
SSE Pd0 a2 Pd1 b2 : 13:13
After rounding the initial value of n to an integer, we recalculate the value of k

through minimizing SSE given by Eq. (13.13).
Example 13.2 Assume that (l0 ; d0 ; d1 ; r; a; b) = (100, 0.2, 0.5, 0.3, 0.95, 0.1). The
problem is to nd n and k.
Using the approach outlined above, we rst nd the initial values of the sam-
pling plan parameters, which are n 8:572 and k 0:3685. We take n 9 and
recalculate the value of k, which is now 0.3701. The actual risks are 1 a
4:447 % and b 9:697 %. As a result, the lower and upper limits of Y are yL
99:63 and yU 100:37, respectively.
13.4 Acceptance Sampling for Lifetime
The lifetime of the product is an important quality characteristic. The sampling plan
for lifetime is usually to control the mean life and deals with a statistical hypothesis
test. The hypothesis can be tested either based on the observed lifetimes or based on
the observed number of failures. For the former case, we let l denote the average
13.4 Acceptance Sampling for Lifetime 243
life of a product and l0 denote the acceptable lot average life. A product is accepted
if the sample information supports the hypothesis:
l l0 : 13:14
For the latter case, the observed number of failures (m) is compared with the
acceptable failure number c. The lot is rejected if m [ c; otherwise, accepted.
Since the life tests are expensive, it is desired to shorten the test time. As such,
lifetime tests are commonly truncated. Fixed time truncated test (type-I) and xed
number truncated test (type-II) are two conventional truncated test methods. Many
testing schemes can be viewed as extensions or mixtures of these two truncated
schemes. Choice among testing methods mainly depends on testing equipment and
environment.
Suppose that a tester can simultaneously test r items, which are called a group. If
g groups of items are tested, the sample size will be n rg. A group acceptance
sampling plan is based on the information obtained from testing these groups of
items. Since r is usually known, the sample size depends on the number of groups g.
Sudden death testing is a special group acceptance sampling plan that can
considerably reduce testing time. Here, each group is tested simultaneously until the
rst failure occurs. Clearly, this is a xed number truncated test for each group.
There are several approaches for designing a sudden death testing, and we focus
on the approach presented in Ref. [4]. Assume that the product life T follows the
Weibull distribution with the shape parameter b and scale parameter g. Let Tj
(1 j g) denote the time to the rst failure for the jth group (termed as the group
failure time). Since Tj minTji ; 1 i r, Tj follows the Weibull distribution with
shape parameter b and scale parameter s g=r 1=b . It is noted that Tjb is also a
random variable and follows the exponential distribution with scale parameter gb =r.
Similarly, Hj Tj =gb is a random variable and follows the exponential distri-
bution with scale parameter (or mean) 1=r. The sum of g independent and identi-
cally distributed exponential random variables (with mean l) follows the Erlang
distribution, which is a special gamma distribution with shape P parameter g (an
integer) and the scale parameter l. This implies that V gj1 Hj follows the
gamma distribution with the shape parameter g and the scale parameter 1=r.
There is a close relation between the gamma distribution and the chi-square
distribution. The chi-square pdf is given by
1
fchi x xq=21 ex=2 13:15
2q=2 Cq=2
where q is a positive integer and known as degree of freedom. It is actually a

gamma distribution with shape parameter q=2 and scale parameter 2. As such,
Q 2rV follows the chi-squared distribution with degree of freedom 2g.
Let tL denote the lower limit of the lifetime. The quality of the product can be
dened as
p 1 exptL =gb 13:16
or
tL =gb ln1 p: 13:17
Noting that Hj Tj =gb Tj =tL b tL =gb , from Eq. (13.17) we have
Hj ln1 pTj =tL b : 13:18
P
g
Letting H Tj =tL b , from Eq. (13.18) we have
j1
Q 2r ln1 pH: 13:19
Clearly, a large H implies a large T. As such, T tL is equivalent to H c, where c

is a parameter to be specied. The lot is accepted if H c; otherwise, rejected.
According to Eq. (13.19), H c is equivalent to Q 2r ln1 pc q: The
acceptance probability at the quality level p is given by
Pa p; g; c PrQ q 1 Fchi q; 2g 13:20
where Fchi is the chi-squared distribution function. For given g and c, Eq. (13.20)
species the OC curve of the sampling plan.
Let a denote the producers risk at the acceptable reliability level p1 and br
denote the consumers risk at the lot tolerance reliability level p2 . The parameters g
and c can be determined by solving the following inequalities:
Fchi qp1 ; 2g a; Fchi qp2 ; 2g 1 br 13:21
or
1 1
qp1 Fchi a; 2g; qp2 Fchi 1 br ; 2g: 13:22
From Eq. (13.22) and noting q 2r ln1 pc, we have
ln1 p1 F 1 a; 2g
1 chi : 13:23
ln1 p2 Fchi 1 br ; 2g
13.4 Acceptance Sampling for Lifetime 245
Fig. 13.5 OC curve for the 1

sampling plan with r 10,
0.8
g 6 and c 25:77
0.6
Pa
0.4
0.2
0
0 0.01 0.02 0.03 0.04 0.05
p
As such, g is the smallest integer that meets Eq. (13.23). Once g is specied, we can
nd the value of c by minimizing the following:
SSE Fchi qp1 ; 2g a2 Fchi qp2 ; 2g 1 br 2 : 13:24
Excel function chidistx; q returns the probability of X [ x; and chidistp; q

returns the value of x for equation p chidistx; q.
Example 13.3 Assume that r 10 and the risks are dened as (p1 ; a) = (0.01, 0.05)
and (p2 ; br ) = (0.04, 0.05), respectively. The problem is to design a sudden death
test scheme (g; c).
From Eq. (13.23), we have g 6. Minimizing the SSE given by Eq. (13.24)
yields c 25:77. As a result, we have qp1 5:1794, which is smaller than
1 1
Fchi a; 2g (=5:2260); and qp2 21:0375, which is larger than Fchi 1 br ; 2g
(=21:0261). This implies that Eq. (13.22) can be met. Figure 13.5 shows the cor-
responding OC curve.
Suppose b 2:35, tL 100 and the group failure times are 117, 290, 260, 63,
284, and 121, respectively. Using these parameters and data yields
H 36:62 [ c 25:77, implying that the lot should be accepted.
It is noted that the lot will be rejected if b 1:97 in this example. Generally,
overestimating b results in a larger customers risk, and hence it is important to
appropriately specify the value of b. For example, tting the test observations to the
Weibull distribution yields b ^ 2:2544, which is very close to the given value.
13.5 Acceptance Sampling for Variable Based

on the Binomial Distribution
A variable acceptance sampling problem can be converted to an attribute accep-

tance sampling problem using the binomial distribution. The main advantage of the
approach to be presented in this section is that it is applicable for any distribution
family; and the main disadvantages are that it requires a larger sample size and can
lose some useful information for the same AQL.
Without loss of generality, we consider the case that the quality characteristic is
the lifetime. Let Ft; h1 ; h2 denote the life distribution, where h1 is a shape
parameter (e.g., b for the Weibull distribution or rl for the lognormal distribution)
and h2 is a scale parameter proportional to mean l (i.e., h2 al). The value of h1 is
usually specied based on experience but can be updated when sufcient data are
available to estimate a new shape parameter. Let l0 denote the critical value of the
mean so that the lot is accepted when l l0 .
Consider the xed time test plan with truncation time s. The failure probability
before s is given by p Fs; h1 ; al. For a sample with size n, the probability with
x failures (0 x n) is given by the binomial distribution with the probability
mass function given by Eq. (13.3). Let c denote the critical failure number. The
acceptable probability is given by Eq. (13.4). Given one of parameters n, s and c,
the other two parameters can be determined through minimizing the following:
SSE Pa p1 1 a2 Pa p2 br 2 13:25
where a and br are the risks of the producer and customer, respectively, and
p1 Fs; h1 ; al1 ; p2 Fs; h1 ; al2 ; l2 \l0 \l1 : 13:26
Example 13.4 Consider the Weibull lifetime with h1 b 2:35, h2 g al and

a 1=C1 1=b 1:1285. Assume that l0 100, l1 150, l2 80 and
a br 0:05. The problem is to design the sampling plan for two cases: (a)
n 10 and (b) s 50.
Using the approach outlined above, we obtain the results shown in Table 13.3.
As seen, Plan (a) is superior to Plan (b) in terms of the required test effort (ns); and
Plan (b) is superior to Plan (a) in terms of the time to complete the test (s).
The OC curve for Plan (a) is shown in Fig. 13.6. The OC curve for Plan (b)
almost overlaps with the OC curve for Plan (a), and hence is not displayed.
Table 13.3 Results for Case c s n a br ns

Example 13.4
(a) 5 108 10 0.0431 0.0458 1080
(b) 5 50 46 0.0403 0.0415 2300
Fig. 13.6 OC curve for case 1

of n 10
0.8
0.6
Pa
0.4
0.2
0
0 50 100 150 200

13.6 Supplier Selection 247
13.6 Supplier Selection
There are two kinds of supplier selection problem (SSP). One deals with specic
purchasing decision and the other deals with establishing a strategic partnership
with suppliers. The purchasing decision problem is relatively simple, and the
strategic partnership with suppliers is a much more complicated problem since it
involves many qualitative and quantitative factors, which are often conflicting with
each other. We separately discuss these two kinds of SSP as follows.
13.6.1 A Mathematical Model for Component

Purchasing Decision
Often, a manufacturer needs to select the component supplier from several sup-
pliers. The reliability and price of the components differ across component suppliers
and the problem is to select the best component supplier. Jiang and Murthy [3] deal
with this problem for the situation where the other conditions are similar and the
main concern is reliability. Here, we extend their model to the situation where the
main concerns are reliability and cost, and the other conditions are similar.
Suppose a key component is used in a system with a known design life (e.g.,
preventive replacement age) L. For a given supplier, assume that the life distribution
of its component is Ft. If the actual life of a component is larger than L, then the
associated life cycle cost is cp ; otherwise, the cost is cf 1 dcp with d [ 0. The
selection decision will be made based on the expected cost rate. Namely, the
selection will give the supplier with the smallest expected cost rate. We derive the
expected cost rate as follows.
The expected life is given by:
ZL ZL
EL L1 FL tf t dt 1 Ft dt: 13:27
0 0
The expected life cycle cost is given by:
EC cp 1 FL 1 dcp FL 1 dFLcp : 13:28
The expected cost rate is given by
J EC=EL: 13:29
Example 13.5 Suppose a certain component in a system has a design life L 50.
The components can be purchased from three suppliers: A1 , A2 and A3 . The lifetime
follows the Weibull distribution with the parameters shown in the second and third
Table 13.4 Lifetimes and costs of components

Alternative b g cp l cp l EC EL J
A1 2.1 115 130 101.9 1.28 140.38 47.33 2.97
A2 2.4 100 120 88.6 1.35 130.36 47.36 2.75
A3 3.2 80 110 71.7 1.54 120.96 47.51 2.55
columns of Table 13.4. The cost parameters are shown in the fourth column of the
table. We assume that d 0:5 for all the suppliers.
The expected cost, expected lifetime, and cost rate are shown in the last three
columns of the table, respectively. As seen, the expected lifetimes of the alterna-
tives are almost indifferent though the mean lifetimes are quite different. This
results from the fact that Alternatives 2 and 3 have larger shape parameters. Based
on the cost rate criterion, Alternative 1 is theworst and Alternative 3 is the best.
If selection criterion is the value of cp l, Alternative 3 is the worst and
Alternative 1 is the best. This reflects that b has a considerable influence on the
purchasing decision.
13.6.2 Supplier Selection Problem Involving Strategic

Partnership with Suppliers
Supplier selection problem involving strategic partnership with suppliers is a typical

MCDM problem. There are many approaches to handle this problem, and the AHP
has been widely used due to its simplicity and flexibility.
The implementation of AHP involves a multi-step procedure. The starting point
is to set up a team with the members from different departments such as material
planning, purchasing, stores, and quality control. The team members will involve in
the selection process. The main tasks of the team include identication of the
criteria (or attributes) and the sub-criteria (or characteristics), and making com-
parative judgments. The interview and questionnaire survey can be used to collect
the required data and information.
The hierarchical structure of SSP is generally composed of four hierarchical
levels: main goal, criteria, sub-criteria, and alternatives. The supplier evaluation
criteria depend on specic situations and hence it is not possible to identify a set of
generic criteria that are suitable for all SSPs. To identify the criteria for supplier
evaluation, the decision maker can provide a list of initial criteria for the team
members to discuss. During the process, the initial criteria allow being eliminated
and new criteria can be introduced. According to Ref. [1], price, quality, and
delivery are the three most important criteria. The other criteria can be manufac-
turing capability, service, flexibility, research and development, and so on.
For each criterion, a set of measurable characteristics will be identied by the
team. The characteristics for quality criterion include acceptable parts per million,
13.6 Supplier Selection 249
total quality management program, corrective and preventive action system, process
control capability, and so on. The characteristics for delivery criterion include
delivery lead time, delivery performance, and so on. The characteristics for cost
criterion include competitiveness of cost, logistics cost, manufacturing cost,
ordering cost, fluctuation on costs, and so on.
The priorities of the criteria will be derived through pairwise comparisons.
Supplier scores for each characteristic can be derived based on the indicators of the
characteristic or based on pairwise comparisons. The pairwise comparisons require
a signicant effort and hence the voting and ranking methods can be used to
determine the relative importance ratings of alternatives.
Once the above tasks are completed, it is relatively simple to calculate the global
scores of alternatives and make relevant decision. For more details about AHP, see
Online Appendix A.
The supplier evaluation method based on AHP is useful for both manufacturers
and suppliers. The manufacturer may use this approach for managing the entire
supply system and adopt specic actions to support suppliers; and through the
evaluation process suppliers may identify their strengths and weaknesses, and adopt
corrective actions to improve their performance.
References
1. Ha SH, Krishnan R (2008) A hybrid approach to supplier selection for the maintenance of a
competitive supply chain. Expert Syst Appl 34(2):13031311
2. Jiang R (2013) Equal-risk acceptance sampling plan. Appl Mech Mater 401403:22342237
4. Jun CH, Balamurali S, Lee SH (2006) Variables sampling plans for Weibull distributed
lifetimes under sudden death testing. IEEE Trans Reliab 55(1):5358
5. Montgomery DC (2007) Introduction to statistical quality control, 4th edn. Wiley Sons, New
York
Chapter 14
Statistical Process Control
14.1 Introduction
The items of a product should be produced by a stable process so that the variability
of the products quality characteristics is sufciently small. Statistical process
control (SPC) is a tool to achieve process stability and improve process capability
through the reduction of variability. There are several graphical tools for the pur-
pose of SPC. They are histograms, check sheets, Pareto charts, cause-and-effect
diagrams, defect concentration diagrams, scatter diagrams, and control charts [3]. In
this chapter, we focus on control charts. Typical control charts are presented;
process capability indices and multivariate statistical process control methods are
also discussed.
The outline of the chapter is as follows. Section 14.2 deals with control charts for
variable and Sect. 14.3 with design and use of the Shewhart control chart. Process
capability indices are presented in Sect. 14.4. Multivariate statistical process control
methods are discussed in Sect. 14.5. Finally, typical control charts for attribute are
presented in Sect. 14.6.
14.2 Control Charts for Variable
14.2.1 Concepts of Control Charts
In continuous production, the process begins with an in-control state. When some
of the controllable factors signicantly deviate from their nominal values, the state
of production process changes from in-control to out of control. If the change is
detected, then the state can be brought back to in-control in order to avoid the
situation where many nonconforming items are produced. Control charts can be
used to detect the state change of a process.

252 14 Statistical Process Control
The basic principle of a control chart is to take small samples periodically and to
plot the sample statistics of one or more quality characteristics (e.g., mean, spread,
number, or fraction of defective items) on a chart. A signicant deviation in the
statistics is more likely to be the result of a change in the process state. When this
occurs, the process is stopped and the controllable factors that have deviated are
restored back to their nominal values. As such, the process is monitored, the out-of-
control cases can be detected, and the number of defectives gets reduced.
Let X denote the quality characteristic and x denote its realization. A sample of
items with size n is taken per h hours and the quality characteristic of each sample
item is measured. Let sj ux denote the sample statistic at tj jh j 1; 2; . . ..

The horizontal axis of a control chart is t or j and the vertical axis is s. Point tj ; sj
on a control chart is called a sample point.
Usually, a control chart has a center line and two control lines (or control limits).
The center line represents the average value of the quality characteristic corre-
sponding to the in-control state, and the two control lines are parallel to the center
line and called the upper control limit (UCL) and the lower control limit (LCL),
respectively. The control limits are chosen so that nearly all of the sample points
will fall between them in a random way if the process is in-control. In this case, the
process is assumed to be in-control and no corrective action is needed. If a sample
point falls outside the control limits or several successive sample points exhibit a
nonrandom pattern, this can be viewed as an indicator of a change in the process
state from in-control to out of control. In this case, investigation and corrective
action are required to nd and eliminate the assignable causes.
14.2.2 Shewhart Mean Control Charts
Most of quality characteristics are continuously valued, and the statistic used to
represent a quality characteristic can be usually approximated by the normal dis-
tribution. Assume that the distribution parameters of X are l0 and r0 when the
process is in-control. Consider a random sample with size n. Let Xji denote the
observed value for the ith item at time instant tj . The sample mean is given by
j 1X n
sj X Xji 14:1
n i1
and the sample range is given by
Rj maxXji minXji : 14:2

i i
The control chart based on the sample average is called an X control chart, which
monitors the process mean; and the control chart based on the sample range is
called a range chart or R chart, which monitors the process variability.
14.2 Control Charts for Variable 253
Consider the mean control chart. When the process is in-control, X approxi-
mately follows a normal distribution with parameters lx l0 and rx given by
p
rx r0 = n: 14:3
is given by:
The interval estimates of 1001 a % of X

lx z1a=2 rx ; lx z1a=2 rx 14:4
where z1a=2 is the 1 a=2-fractile of the standard normal distribution.

Generally, let S be a sample statistic that measures some quality characteristic.
The mean of S is ls and the standard deviation is rs . The center line, the upper
control limit, and the lower control limit are dened as
CL ls ; LCL ls Lrs ; UCL ls Lrs : 14:5
The control charts developed according to Eq. (14.5) are called the Shewhart
control charts. The L in Eq. (14.5) is similar to the z1a=2 in Eq. (14.4). When L = 3,
the corresponding control limits are called the three sigma control limits.
14.2.3 Range Chart
and R charts are used simultaneously. Let R

In practice, the X denote the average
range estimated from m samples observed in the in-control condition. The center
and the control limits are given by:
line is equal to R
LCL D3 R
UCL D4 R; 14:6
where D4 and D3 can be calculated by
D4 1:2529 2:0156=n 10:6124 ; D3 max2 D4 ; 0: 14:7
The maximum relative error between the value of D4 calculated from Eq. (14.7) and
the value obtained from Appendix VI of Ref. [3] is 0.1469 %, which is achieved
when n = 3.
14.2.4 Errors of a Control Chart
A control chart can give two types of error. A Type I error occurs when the process
is actually in-control but the control chart gives an out-of-control signal. This
false alarm leads to a stoppage of the production when the process is in-control.
Fig. 14.1 Operating-

Probability in the limits

characteristic curves for an X
chart
n =5
n =10
n =15
Process mean shift
A Type II error occurs when the process is actually out of control but the control
chart gives an in-control signal. This type of error leads to a delay to initiate a
corrective action. When this occurs, more nonconforming items will be produced
due to the process being out of control.
to fall outside the
The probability of Type I error equals the probability for X
control limits and is given by
P1 ULCL; l0 ; rs 1 UUCL; l0 ; rs : 14:8
When L z1a=2 , we have P1 a; when L = 3, P1 = 0.0027. This implies that the

control limits are specied based on the Type I error. Usually, we take L = 3, which
corresponds to = 0.0027, or = 0.002 which corresponds to L = 3.09.
We now look at the probability of Type II error. Suppose that there is a mean
shift from l0 to ld l0 d. The probability for X to fall between the control
limits is given by
P2 UUCL; ld ; rs ULCL; ld ; rs : 14:9
The plot of P2 versus ld (or d) is called the operating-characteristic curve of a

control chart. For a given value of d > 0, a large value of P2 implies a poor detecting
ability.
Figure 14.1 shows the operating-characteristic curves as a function of sample
chart. As seen, the ability of detecting a shift increases as n and shift
size for an X
increase.
14.2.5 Average Run Length and Average Time to Signal
The basic performance of a control chart is the average time to signal (ATS). This
includes the two cases, which correspond to the concepts of Type I error and Type
II error, respectively. We use ATS0 to denote the ATS associated with Type I error,
and use ATS1 to denote the ATS associated with Type II error. ATS1 is an indicator
of the power (or effectiveness) of the control chart. A large ATS0 and a small ATS1
are desired.
14.2 Control Charts for Variable 255
We rst look at ATS0. For a given combination of n and h, the average number of
points (or samples) before a point wrongly indicates an out-of-control condition is
called the average run length (ARL) of the control chart. Let p denote the probability
that a single point falls outside the control limits when the process is in-control.
Clearly, p = = P1. Each sampling can be viewed as an independent Bernoulli trial
so that the number of samples (or run length) to give an out-of-control signal follows
a geometric distribution with mean 1/p. As such, the average run length is given by
ARL0 1=p: 14:10
When samples are taken at a xed time interval h, the average time to have a false
alarm signal is given by
ATS0 h ARL0 : 14:11

p
The geometric distribution has a large dispersion (with r=l 1 p 1) and
is very skewed so that the mean is not a good representative value of the run length.
In other words, the run length observed in practice can be considerably different
from the mean. This is illustrated by Fig. 14.2, which shows the geometric distri-
bution with p = 0.0027.
We now look at ATS1. Assume that the mean shift is d krs , i.e.,
ld l0 krs . In this case, the probability to detect the mean shift equals 1 P2 .
As such, the average run length to detect the out-of-control is given by
ARL1 1=1 P2 : 14:12
The required average time is given by
ATS1 h ARL1 : 14:13
Figure 14.3 shows the plot of ARL1 versus k for the X chart with 3-sigma limits.
As seen, for a xed k, ARL1 decreases as n increases. Since ATS1 is proportional to
h, a small value of ATS1 can be achieved using a small value of h.
Fig. 14.2 Pmf of the 0.003

geometric distribution with
p 0:0027
0.002
p (x )
0.001
x 0.05 x 0.5
0
0 200 400 600 800 1000 x 0.951200
Number of samples x
Fig. 14.3 Plot of ARL1 15

versus k
n =5
10
ARL 1
n =10
n =15
5
0
0 1 2 3 4
k
14.3 Construction and Implementation

of the Shewhart Control Chart
The design of a control chart involves two phases. In the rst phase, a trial control
chart is obtained based on the data obtained from pilot runs. In the second phase,
the trial control chart is used to monitor the actual process, and the control chart can
be periodically revised using the latest information.
14.3.1 Construction of Trial Control Chart
The data for estimating l0 and r0 should contain at least 2025 samples with a
sample size between 3 and 6. The estimate of l0 based on the observations obtained
during the pilot runs should be consistent with the following:
LSL USL
l0 14:14
2
where LSL and USL are the lower and upper specication limits of the quality
characteristic, respectively.
If the quality characteristic can be approximated by the normal distribution, the
p p
control limits are determined by l0 3r0 = n or l0 z1a=2 r0 = n with = 0.002.
The usual value of sample size n is 4, 5, or 6. A large value of n will decrease the
probability of Type II error and increase the inspection cost. When the quality char-
acteristic of the product changes relatively slowly, a small sample size should be used.
Sampling frequency is represented by inspection interval h. A small value of
h implies better detection ability and more sampling effort. The sampling effort can be
represented by the inspection rate (number of items inspected per unit time) given by
r n=h: 14:15
14.3 Construction and Implementation of the Shewhart Control Chart 257
Fig. 14.4 Mean control chart

80.01 USL
for Example 14.1
Sample mean
UCL
0
80
LCL
79.99 LSL
79.98
0 5 10 15 20
t
It depends on the available resources (e.g., operators and measuring instruments).

Let rm denote the maximum allowable inspection rate. Since r rm , we have
h n=rm : 14:16
Generally, the value of h should be as small as possible, and hence we usually take
r rm .
Example 14.1 A manufacturing factory produces a type of bearing. The diameter of
the bearing is a key quality characteristic and specied as 80 0.008 mm. The
process mean can be easily adjusted to the nominal value. The pilot runs yields
r0 0:002 mm and R 0:0021 mm. The maximum allowable inspection rate is
rm 4 items per hour, and the minimum allowable ATS0 is 400 h. The problem is
chart and a R chart.
to design a X
Clearly, we have l0 80. Taking n = 5 yields the sampling interval

h n=rm 1:25 h. Letting ARL0 1=a ATS0 =h yields a h=ATS0
0:003125. This implies z1a=2 2:9552. As a result, the control limits are given by
r0
UCL l0 z1a=2 p 80:0026; LCL 79:9974:
n
The mean control chart is shown in Fig. 14.4.

From Eq. (14.7) we have D4 2:1153 and D3 0. As a result, the control
limits of the R chart are given by LCL 0 and UCL 0:0044. Figure 14.5 shows
the designed R chart.
14.3.2 Sampling Strategy
A sampling strategy deals with how the samples are taken. An appropriate sampling
strategy can obtain as much useful information as possible from the control chart
analysis. Two typical sampling strategies are consecutive sampling and random
sampling.
Fig. 14.5 R chart for 0.006

Example 14.1 0.005 UCL
0.004
0.003
R
Center line
0.002
0.001
LCL
0
0 5 10 15 20
t
The consecutive sampling strategy takes the sample items from those items that
were produced at almost the same time. Such selected samples have a small unit-to-
unit variability within a sample. This strategy is suitable to detect process mean
shifts.
The random sampling strategy randomly takes each sample from all items that
have been produced since the last sample was taken. If the process average drifts
between several levels during the inspection interval, the range of the observations
within the sample may be relatively large. In this case, the R chart tends to give
more false alarm signals actually due to the drifts in the process average rather than
in the process variability. This strategy is often used when the control chart is
employed to make decisions about the acceptance of all items of product that have
been produced since the last sample.
If a process consists of several machines and their outputs are pooled into a
common stream, control chart techniques should be applied to the output of each
machine so as to detect whether or not a certain machine is out of control.
14.3.3 Nonrandom Patterns on Control Charts
Variability of process data can be random or nonrandom. Typically, there are three
types of variability in the use of a control chart. They are stationary and uncorre-
lated, stationary but autocorrelated, and nonstationary.
The process data from an in-control process vary around a xed mean in a
random manner. This type of variability is stationary and uncorrelated. For this case
the Shewhart control charts can be used to effectively detect out-of-control
conditions.
If successive observations have a tendency to move on either side of the mean,
this type of variability is stationary but autocorrelated. The variability with this
phenomenon is nonrandom.
If the process does not have a stable mean, the variability with this phenomenon
is nonstationary. This kind of nonrandom pattern usually results from some external
factors such as environmental variables or properties of raw materials, and can be
avoided using engineering process control techniques such as feedback control.
14.3 Construction and Implementation of the Shewhart Control Chart 259
When the plotted points exhibit some nonrandom pattern, it may indicate an out-
of-control condition. Three typical nonrandom patterns are
(a) the number of points above the center line is signicantly different from the
number of points below the center line,
(b) several consecutive points increase or decreases in magnitude, and
(c) cyclic pattern. This occurs possibly due to some periodic cause (e.g., operator
fatigue) and signicantly affects the process standard deviation.
Several tests for randomness can be found in Sect. 6.6.
14.3.4 Warning Limits
To help identify the nonrandom patterns, warning limits and one-sigma lines can be
displayed on control charts. The warning limits are the 2-sigma limits for the quality
characteristic with the normal distribution, or the 0.025-fractile and 0.975-fractile
for the case where the control limits are dened as the 0.001 probability limits (i.e.,
= 0.002). All these limits and lines partition the control chart into three zones on
each side of the center line. The region between the control limit and the warning
limit is called Zone A; the region between the one-sigma line and the warning limit
is called Zone B; and the region between the one-sigma line and the center line is
called Zone C.
When a point falls outside the control limits, a search for an assignable cause is
made and corrective action is accordingly taken. If one or more points fall into Zone
A, one possible action is to increase the sampling frequency and/or the sample size
so that more information about the process can be obtained quickly. This adjusted
sample size and/or sampling frequency depend on the current sample value. The
process control schemes with variable sample size or sampling frequency are called
adaptive schemes.
The use of warning limits allows the control chart to signal a shift in the process
more quickly but can result in more false alarms. Therefore, it is not necessary to
use them if the process is reasonably stable.
14.3.5 Out-of-Control Action Plan
The control chart does not indicate the cause of the change in the process state.
Usually, FMEA is used to identify the assignable causes, and an out-of-control
action plan (OCAP) provides countermeasures to eliminate the causes. The OCAP
is a flowchart of activities when an out-of-control occurs, including checkpoints and
actions to eliminate the identied assignable cause. A control chart and an OCAP
should be jointly used and updated over time.
14.4 Process Capability Indices and Fraction

Nonconforming
The process capability can be represented in terms of process capability indices and
fraction nonconforming, which can be used to compare different processes that are
in a state of statistical control.
14.4.1 Process Capability Indices
The process capability is the ability of a process to produce the output that meets the
specication limits. A process capability index (PCI) is a measure for representing
the inherent variability of a quality characteristic relative to its specication limits.
It is useful for product and process design as well as for supplier selection and
control.
Consider a quality characteristic Y with the specication limits LSL and USL,
respectively. Assume that Y follows the normal distribution with the process mean l
and variance r2 . The fraction of nonconformance (or defect rate) is given by
p ULSL; l; r 1 UUSL; l; r: 14:17
Figure 14.6 shows the influence of r on p. As seen, a good process has a small r
and a small fraction of nonconformance.
Under the following assumptions:
the process is stable,
the quality characteristic follows the normal distribution,
the specication limits are two-sided and symmetrical, and
the process mean is at the center of the specication limits,
the process capability index Cp is dened as
USL LSL
Cp : 14:18
6r
Fig. 14.6 Good and poor Target

processes
Good process
f (y )
LSL USL
Poor process
y
14.4 Process Capability Indices and Fraction Nonconforming 261
If the specication limits are one-sided, the process capability index is dened as
l LSL USL l
Cpl or Cpu : 14:19
3r 3r
If the process mean is not at the center of the specication limits, the process
capability can be represented by index Cpk given by
minl LSL; USL l

Cpk : 14:20
3r
This index can be used to judge how reliable a process is. When Cpk 1:5, the defect
rate equals 3.4 parts per million, which corresponds to famous Six Sigma Quality.
If the process mean is not equal to the target value T, the process capability can
be represented by index Cpm given by
s

lT 2
Cpm Cp = 1 : 14:21
r
If the process mean is neither at the center of the specication limits nor equal to
the target value, the process capability can be represented by index Cpkm given by
s

lT 2
Cpkm Cpk = 1 : 14:22
r
More variants and details of the process capability indices can be found from
Refs. [1, 4, 5].
For a given process, a small PCI means a high variation. Therefore, a large PCI
(e.g., Cp [ 1:0) is desirable. In general, the minimum acceptable PCI for a new
process is larger than the one for an existing process; the PCI for the two-sided
specication limits is larger than the one for the one-sided specication limit; and a
large PCI is required for a safety-related or critical quality characteristic.
Example 14.2 Assume that a quality characteristic follows the normal distribution
with the standard deviation r 0:5, and the specication limits and target values
equal LSL = 48, USL = 52 and T = 50.5, respectively. The problem is to calculate
the values of process capability indices for different process means.
For a set of process means shown in the rst row of Table 14.1, the corre-
sponding values of Cp are shown in the second row and the fractions of noncon-
formance shown in the third row. As seen, Cp maintains unvarying but p varies with
l and achieves its minimum when the process mean is at the center of the speci-
cation limits (i.e., l 50).
The fourth row shows the values of Cpk . As seen, Cpk achieves its maximum when
the process mean is at the center of the specication limits. The fth row shows the
Table 14.1 Process 49 50 51 50.33 50.5

capability indices for
Cp 1.3333 1.3333 1.3333 1.3333 1.3333
Example 14.2
p, % 2.2750 0.0063 2.2750 0.0431 0.1350
Cpk 0.6667 1.3333 0.6667 1.1111 1
Cpm 0.4216 0.9428 0.9428 1.2649 1.3333
Cpkm 0.2108 0.9428 0.4714 1.0541 1
values of Cpm , which achieves its maximum when the process mean equals the target
value. The last row shows the values of Cpkm , which achieves its maximum when
l 50:33, a value between the target value and the center of the specication limits.
14.4.2 Fraction Nonconforming
The fraction nonconforming is the probability for the quality characteristic to fall
outside the specication limits, and can be estimated based on the information from

the control charts that exhibit statistical control. To illustrate, we consider the X

chart and R chart. From the in-control observations of the X chart, we may estimate
the process mean l and average range R. The process standard deviation rs asso-

ciated with X can be estimated by:
2
rs R=d 14:23
where d2 is given by
d2 7:9144 7:8425=n0:2101 : 14:24
The maximum relative error between the value of d2 calculated from Eq. (14.24)
and the one given in Appendix VI of Ref. [3] is 0.5910 %, which is achieved at
n = 2. The process standard deviation r associated with X is given by
p p
r nrs nR=d 2: 14:25
Assuming that the quality characteristic X is a normally distributed random

variable with l LSL USL=2, the fraction of nonconforming is given by
Eq. (14.17). For the 3rs control limits, we have 6rs UCL LCL. As such, the
process capability index can be written as
USL LSL
Cp p : 14:26
nUCL LCL
This implies that the process capability index can be estimated by the information
from control charts.
14.5 Multivariate Statistical Process Control Methods 263
14.5 Multivariate Statistical Process Control Methods
The Shewhart control charts deals with a single quality characteristic. A product can
have several key quality characteristics. In this case, several univariate control
charts can be used for separately monitoring these quality characteristics if they are
independent of each other. However, if the quality characteristics are correlated, the
univariate approach is no longer appropriate and we must use multivariate SPC
methods.
Two typical multivariate SPC methods are multivariate control charts and pro-
jection methods. The multivariate control charts only deal with product quality
variables, and the projection methods deal with both quality and process variables.
We briefly discuss them as follows.
14.5.1 Multivariate Control Charts
To be concise, we look at the multivariate Shewhart control charts with two cor-
related quality characteristics. Let Y1 and Y2 denote the quality characteristics,
which are normally distributed; l1 and l2 denote their means and aij ; i; j 1; 2
denote the elements of the inverse matrix of the covariance matrix between Y1 and
Y2. Let
v2 a11 y1 l1 2 2a12 y1 l1 y2 l2 a22 y2 l2 2 : 14:27
This statistic follows a central chi-squared distribution with 2 degrees of freedom. A

multivariate chi-squared control chart can be constructed by plotting v2 versus time
with a zero lower control limit and an upper control limit given by v2a , where a is an
appropriate level of signicance for performing the test (e.g., a 0:01).
14.5.2 Multivariate Statistical Projection Methods
Let Y denote the quality characteristics set and X denote the process variables set.
When the number of optional quality variables is large, it is necessary to reduce the
number of the quality variables. In practice, most of the variability in the data can
be captured in the few principal process variables, which can explain most of the
predictable variations in the product. The principal component analysis (PCA, see
Online Appendix C) and partial least squares (PLS, e.g., see Ref. [6]) are two useful
tools for this purpose. A PCA or PLS model is established based on historical data
collected in the in-control condition, and hence it represents the normal operating
conditions for a particular process. Then, a multivariate control chart (e.g., T2-chart)
can be developed based on the few principal variables.
Different from univariate control charts that can give out-of-control signal but
cannot diagnose the assignable cause, multivariate control charts based on PCA or
PLS can diagnose assignable causes using the underlying PCA or PLS model. More
details about multivariate SPC can be found in Ref. [2] and the literature cited
therein.
14.6 Control Charts for Attribute
Attribute control charts are based on integer-valued measurements. The basic

principles to construct an attribute control chart are similar to constructing a vari-
able control chart. Three typical statistics that are widely used in attribute control
charts are fraction nonconforming, number of defects that occur in an inspected
item, and average number of defects per the inspected item.
14.6.1 Control Chart for Fraction Nonconforming
Suppose that we inspect m samples with sample size n. Let Di denote the number of
defectives of the ith sample. Sample fraction nonconforming is given by
pi Di =n: 14:28
The fraction nonconforming pi follows the binomial distribution with mean

p and variance r2 p1 p=n. As such, the center line and control limits of a
mean control chart are given by
p p
l p; LCL p 3 p1 p=n; UCL p 3 p1 p=n: 14:29
If LCL\0, then take LCL 0. The control chart dened by Eq. (14.29) is called
the p chart.
Another control chart (called the np chart) can be established for D. This results
from Eq. (14.28). Clearly, D has mean np and variance r2 np1 p. As a result,
the center line and control limits of a mean control chart are given by
p p
l np; LCL np 3 np1 p; UCL np 3 np1 p: 14:30
If LCL\0, then take LCL 0.

Example 14.3 Suppose that there are 30 samples with sample size 50. The values of
Di are shown in Table 14.2, where the values of pi are also shown. The problem is
to design the p chart and the np chart.
14.6 Control Charts for Attribute 265
Table 14.2 Data for Di 0 1 0 3 3 2 2 0

Example 14.3
pi 0 0.02 0 0.06 0.06 0.04 0.04 0
Di 4 2 1 1 3 0 1 3
pi 0.08 0.04 0.02 0.02 0.06 0 0.02 0.06
Di 3 3 4 1 1 2 1 0
pi 0.06 0.06 0.08 0.02 0.02 0.04 0.02 0
Di 2 2 3 1 0 7
pi 0.04 0.04 0.06 0.02 0 0.14
The mean and standard deviation of p are 0.0373 and 0.0314, respectively. As
such, the center line and control limits of the p chart are 0.0373, 0 and 0.0942,
respectively.
The mean and standard deviation of D are 1.87 and 1.57, respectively. As such,
the center line and control limits of the np chart are 1.87, 0 and 6.58, respectively.
14.6.2 Control Chart for the Number of Defects Per

Inspected Item
Let d denote the number of defects in an inspected item, and c0 denote the maxi-
mum allowable number of defects of an item. If d [ c0 , the inspected item is
defective; otherwise, normal. The c chart is designed to control the number of
defects per inspected item. Here, the defects can be voids of a casting item, a
component that must be resoldered in a printed circuit board, and so on.
Let c denote the total number of defects in an inspected item. It follows the
Poisson distribution with mean and variance k. As such, the center line and control
limits of the c chart are given by
p p
l k; LCL max0; k 3 k; UCL k 3 k: 14:31
14.6.3 Control Chart for the Average Number

of Defects Per Item
Let u D=n denote the average number of defects per item. Then, u has mean u,
and standard deviation r u=n. The u chart is developed to control the value of
u. The center line and control limits of the u chart are given by
p p
l
u; LCL max0; u 3 u=n; UCL u 3 u=n: 14:32
Clearly, the u chart is somehow similar to the p chart, and the c chart is somehow
similar to the np chart.
References
1. Kotz S, Johnson NL (1993) Process capability indices. Chapman and Hall, New York, London
2. MacGregor JF, Kourti T (1995) Statistical process control of multivariate processes. Control
Eng Pract 3(3):403414
4. Pearn WL, Chen KS (1999) Making decisions in assessing process capability index Cpk. Qual
Reliab Eng Int 15(4):321326
5. Porter LJ, Oakland JS (1991) Process capability indicesan overview of theory and practice.
Qual Reliab Eng Int 7(6):437448
6. Vinzi VE, Russolillo G (2013) Partial least squares algorithms and methods. Wiley Interdiscip
Rev: Comput Stat 5(1):119
Chapter 15
Quality Control at Output
15.1 Introduction
The reliability of a manufactured product usually differs from its design reliability
due to various quality variations such as nonconforming components and assembly
errors. These variations lead to a relatively high early failure rate. Quality control at
output mainly deals with quality inspections and screening testing of components
and nal products. The purpose is to identify and reduce defective items before they
are released for sale.
An issue with product quality inspection is to classify the inspected product into
several grades based on the quality characteristics. The partitions between two adjacent
grades can be optimized to achieve an appropriate tradeoff between manufacturing cost
and quality cost. This problem is called the optimal screening limit problem.
Two widely used screening tests are burn-in and environmental stress screening.
Such tests are required for the products with high reliability requirements. The tests
are generally expensive; and the losses of eld failures incurred by defective items
are usually high. As such, the test costs and eld losses must achieve an appropriate
tradeoff. This involves optimization of the test scheme.
The outline of this chapter is as follows. Section 15.2 deals with the optimal
screening limit problem, and Sect. 15.3 deals with relevant concepts of screening
test. Optimization models for component-level burn-in and system-level burn-in are
discussed in Sects. 15.4 and 15.5, respectively.
15.2 Optimal Screening Limit Problem
15.2.1 Screening Limit Problem
Based on whether or not each item produced is subjected to inspection, the

quality conformance inspection can be 100 % inspection and sample inspection

268 15 Quality Control at Output
(if the inspection is destructive). The screening inspection is a type of 100 %

inspection.
There are two categories of screening limit problem. In the rst category, the
product items are classied into several grades based on one or more quality
characteristics. The partitions between two adjacent grades are called the screening
limits. The screening limits can be optimally determined by minimizing the
expected total cost. A number of models have been proposed in the literature for
determining optimal screening limits (e.g., see [5]). A feature of this category of
screening problem is that the inspected items can be either conforming or defective.
The second category of screening problem deals with the items whose quality
characteristics are within the specication limits. The items whose quality char-
acteristics measured during the production process show anomalies (or outliers) will
be screened out since they may contain concealed defects and hence have a high
risk of early product failure. The anomalies are detected by pre-set screening limits,
which are determined through a Part Average Analysis (PAA, [6]). PAA can be
used to detect pre-damage of units and components as well as problems with the
measuring equipment.
15.2.2 An Optimization Model
Consider the problem where the product items are classied into three grades
(acceptable, reworked, and scraped) based on a single variable Y, which is highly
correlated with the quality characteristic of interest. Assume that Y follows the
normal distribution with mean l and standard deviation r, and has a target value
T. It is easy to adjust the process mean to the target value so that we have l T.
Let d denote the screening limit, which is the decision variable. The manufactured
products are classied into the following three grades:
acceptable if y 2 T d,
scraped if y\T d, and
reworked if y [ T d.
Clearly, a small d results in more items being screened out as nonconforming. This
is why it is called the screening limit.
Consider two categories of costs: manufacturing-related cost before the sale and
quality loss after the product is delivered to the customer. As d decreases, the
manufacturing cost per sold item increases and the quality loss decreases. As such,
the optimum screening limit exists so that the expected total cost per sold item
achieves its minimum.
The manufacturing-related costs include three parts: raw material cost cm , pro-
duction cost cp , and inspection cost cI . Generally, these cost elements are constant
for a given manufacturing process. As such, the total manufacturing cost per
manufactured item is given by CM cm cp cI .
15.2 Optimal Screening Limit Problem 269
An acceptable product involves a quality loss given by
cq Kv 15:1
where v is the variance of the doubly truncated normal distribution with support
y 2 l d, and is given by
v r2 f1 2d=r/d=r; 0; 1=2Ud=r; 0; 1 1g: 15:2
A scraped product involves an additional scraping cost cs ; and a reworked

product involves additional costs cr and a reinspection cost cI . According to Ref.
[5], we approximately take cr cp , where cr includes both the reworking cost and
quality loss cost.
The probability for a produced item to be scraped is ps Ud; 0; r; the
probability for a produced item to be reworked is pr ps , and the probability for a
produced item to be acceptable is pa 1 2ps . As a result, the expected total cost
per manufactured product is given by
CT t CM ps cs pr cp cI pa Kv: 15:3
The probability for a produced item to be eventually shipped to the consumer is

equal to Pd pa pr . The expected total cost per sold product item is given by
Jd CT d=Pd: 15:4
The optimal value of d can be determined by minimizing Jd.

Example 15.1 Assume that the process mean is T 30 and the process standard
deviation is r 10. The manufacturing-related cost parameters are cm 500,
cp 1000, cI 10, and cs 0, respectively. When d 30, the quality loss is
5000, implying K 5:5556. The problem is to nd the optimal screening limit.
Using the approach outlined above, we obtained the optimal screening limit
d 16.0. The corresponding expected total cost per sold product item is J
1787:49 and the probability for a manufactured product item to be scraped is
ps 5:49 %.
If the scrap probability is considered to be too large, the manufacturing process
has to be improved to reduce the value of r by using high precision equipment and
machines. This will result in an increase of cp . Assume that r decreases from 10 to
8, and cp increases from 1000 to 1100. Then, the optimal solution is now d 14.9
with J 1774:46 and ps 3:11 %. Since both J and ps decrease, the improvement
is worthwhile.
15.3 Screening Tests
Burn-in and environmental stress screening (ESS) are two typical screening tests for
electronic products. An electronic product is usually decomposed into three hier-
archical levels, i.e., part (or component), unit (or assembly or subsystem), and
system (or product). We use item to represent any of them when we do not need
to differentiate the hierarchy level.
15.3.1 Types of Manufacturing Defects
According to Ref. [11], defect is weakness or flaw of an item due to substandard

materials or faulty processes. Defect can be patent or latent. A patent defect is a
condition which does not meet specications and hence is detectable by quality
inspection or functional testing. The parts with patent defects are likely to fail early
in life and can be removed by burn-in. Patent defects can be prevented by redesign
or/and process control.
A latent defect is a defect that generally cannot be detected by usual inspection
or functional testing. Examples of latent defect include microcracks caused by
mechanical shocks and partial damage due to electrostatic discharge or electrical
overstress (see Ref. [3]). The latent defects can be changed into patent defects by
external overstresses.
The strength of a part with latent defect is smaller than the strength of a normal
part (i.e., design strength). As such, when a part with latent defect is exposed to a
stress level that is larger than its strength, a failure occurs. We call such a failure the
latent failure. When the strength of a part with latent defect is not much smaller than
the strength of a normal part, it may take a relatively long time (relative to burn-in
period) for the latent failure to occur. In fact, the latent failure usually occurs in the
normal use period. As such, some of latent defects cannot be detected by a burn-in
procedure. Latent failures can be reduced by redesign for extreme conditions or by a
specic environmental stress screening test so as to transform a latent defect into a
patent defect.
The stressstrength interference model can be used to model the time to latent
failure. Let z denote the defect size and x denote the corresponding strength, which is a
monotonically decreasing function of z (denoted as x uz). Let Yt be time-
dependant stress. Assume that stresses occur at random points in time due to external
shocks and the shocks occur according to a point process Nt modeled by the Poisson
process with intensity k. The stresses resulting from shocks are random variables with
distribution Gy. The reliability function is given by (see Example 9.5)
Rt expf1 Gxktg: 15:5

15.3 Screening Tests 271
From Eq. (15.5), the mean time to failure is given by
1 1
ET; z : 15:6
k1 Gx kf1 Guzg
Equation (15.6) relates the defect size to the mean lifetime in the normal use
condition.
For a given screening test, the time to latent failure determines the detectability
of the part. Figure 15.1 shows relations between the defect size, strength, lifetime,
and detectability. As seen from the gure, a large defect is more probably patent
and can be detected by functional test; a small defect is more probably latent and
can be transformed into a patent defect by ESS; a defect with intermediate size can
be either patent or latent and can be detected by burn-in.
The latent defects affect the failure pattern. If there are no latent defects, the
failure rate of a population of items can be represented by the classic bathtub curve.
If some of the items contain latent defects, they may fail in the normal use phase
under excessive stress conditions. Such failures result in jumps in the failure rate
curve. Figure 15.2 shows the failure rate curve superposed by the classic bathtub
failure rate curve and the failure rate curve resulted from latent defects. In the
literature, this superposed failure rate curve is called the roller coaster curve (e.g.,
see Refs. [7, 8]).
Fig. 15.1 Relations between

defect size, strength, lifetime,
and detectability
x, E (T )
x = (z )
E (T )
ESS Burn-in Functional test
Burn-in period
Fig. 15.2 Roller coaster

failure rate curve
Latent failure rate

r (t )
Burn-in PM
ESS
Early failure Wear-out

Useful life period
period period
t
15.3.2 Burn-in
Burn-in is a kind of test to expose defects of items or their components and screen
out those items with defects in order to prevent product early failure. It is usually
applied to the items with high initial failure rate, which result from defective parts
and quality variations due to assembly-related problems. Typical assembly prob-
lems include components damage and component connection defects. As such, the
burn-in can be used at component level and system level. Component-level burn-in
is done often by component suppliers to identify and eliminate defective compo-
nents and system-level burn-in is done by the manufacturer to remove component
defects and assembly defects.
The test conditions are application-specic. To accelerate the process, burn-in
can be conducted under relatively harsh environments. Burn-in of electronic
components is usually conducted at elevated temperature and/or voltage.
Figure 15.3 shows a typical temperature stress cycle used in burn-in.
The tested items usually operate for a xed time period (called burn-in period).
Any item that survives the burn-in will be released for sale. If the product is
repairable, then failures during burn-in are rectied and tested again until it survives
the burn-in period. Burn-in spends cost and consumes a part of useful lifetime but
can lead to less eld cost due to enhanced product reliability after burn-in. One of
the major problems with burn-in is to optimally determine the burn-in period for a
given criterion such as cost, reliability or their combination.
Reliability measure of the burnt-in product item can be the survival probability
over a prespecied mission time (e.g., warranty period or planning horizon) or
mean residual life. An age-based preventive replacement policy for the burnt-in
product can be implemented to further reduce the total costs. Jiang and Jardine [4]
simultaneously optimize the burn-in duration and the preventive replacement age
based on the cost rate, which considers both cost and mean residual life.
Most products are sold with warranty. Breakdown of a burnt-in item within the
warranty period causes warranty claims, which incur warranty costs. A balance
between burn-in costs and warranty costs can be achieved by minimizing the sum
of burn-in and warranty costs.
Fig. 15.3 A typical

High temperature
temperature stress cycle used
condition
in burn-in
Temperature
Operational condition
Low temperature
condition
t
15.3 Screening Tests 273
It is noted that eliminating the root cause of early failures is better than doing a
burn-in if possible. As various root causes for failures are identied and eliminated,
burn-in may eventually be no longer needed. Block and Savits [1] present a liter-
ature review on burn-in.
15.3.3 Environmental Stress Screening
ESS is a process for accelerating the aging of latent defects by applying excessive
stress without damaging the items [2]. The intensity and magnitude of shocks
produced by ESS test can be sufciently large, implying that k is large and Guz
is small in Eq. (15.6) so that the time to latent failure will be small, that is, ESS can
be effective.
A key issue with ESS is to appropriately determine the types and ranges of
stresses (or shocks) to be applied. Typical stresses used in ESS are thermal stress,
vibration, and shock. The thermal stress tests include low temperature test, high
temperature test, temperature cycling tests, and thermal shock test. Temperature
cycling tests simulate varying temperature operating environment. Thermal shock
test quickly changes the temperature by moving the tested item from one temper-
ature environment to another temperature environment. Vibration testing can be
random vibration and sine vibration, and may be carried out on a single axis or three
mutually perpendicular axes. Random vibration testing can excite all resonant
frequencies throughout the entire test and hence is preferred. Shock tests include
mechanical shock test and power cycling. A typical shock test simulates the stresses
resulting from handling, transportation and operation by applying ve shock pulses
at a selected peak acceleration level in each of the six possible orientations. Power
cycling is implemented by turning on and off the power at predetermined intervals.
Other extreme environments that ESS tests can simulate include high altitude, high
voltage, humid, salt spray, sand, dust, and so on. Some ESS tests can simulate two
or more environments at a time.
ESS exposes defects by fatiguing weak or marginal mechanical interfaces [2].
Since fatigue is the result of repeated stress reversals, ESS usually applies stress
cycles (e.g., thermal cycling, onoff cycling, and random vibration) to produce such
stress reversals. Generally, temperature cycling, random vibration, and their com-
bination are the most effective screening processes for electronic assemblies.
15.3.4 Comparison of ESS and Burn-in
Both ESS and burn-in processes emphasize on reducing early eld failures.
Generally, burn-in takes much lengthier time to power a product at an operating or
accelerated stress condition. On the other hand, ESS is generally conducted under
accelerated conditions to stress a product for a limited number of stress cycles, and
functional testing is needed to verify that the product is functioning after ESS testing.
As such, main differences between them are as follows (e.g., see Refs. [9, 11]):
(a) the tested item is powered for burn-in and stressed for ESS,
(b) the stress levels used for burn-in are usually lower than the stress levels used
for ESS, and
(c) test duration is from several hours to a few days for burn-in and from several
minutes to a few hours for ESS.
Generally, ESS is more effective in screening out stress-dependent defects,
which result in overstress failure, but it is less effective in screening out the defects
caused by time- or usage-dependent failure modes. Conversely, burn-in can screen
out the time/usage-dependent defects and provides useful information for predicting
reliability performance of the product.
ESS and burn-in can be combined to reduce burn-in time. For example, a two-
level ESS-burn-in policy (see Ref. [10]) combines a part-level ESS and a unit-level
burn-in. Under this policy, all parts are subjected to an ESS and the parts passing
the part-level screen are used in the unit. Then, all units are burned-in, and the units
passing burn-in are used in the nal system, for which there is no burn-in or ESS.
15.4 Optimal Component-Level Burn-in Duration
Component-level burn-in aims to detect nonconforming parts or units. Consider a

burn-in test, in which items are operated in a normal operation condition for a time
period s so that the weak items are found and get repaired. If an item does not fail in
the burn-in period, it passes the test; if the item fails during the test, a good-as-new
repair is performed and the item is retested until it passes the test.
The basic assumption is that the items are from a mixture population. The test
will nd most of the weak items. If these are not found, the warranty cost will be
much larger than the test cost. The burn-in period s is a key parameter to be
determined. If s is too small, some of the weak items will be delivered to customers
and this can lead to a large warranty cost; if s is too large, both burn-in time and
cost are high. As such, the burn-in duration can be optimized through minimizing
the total cost. A cost model is presented as follows.
Let p q 1 p denote the probability or proportion that the item is conforming
[nonconforming] and Fc t Fn t denote the life distribution of a conforming
[nonconforming] item. The life distribution of the item population is given by
Ft qFn t pFc t: 15:7

15.4 Optimal Component-Level Burn-in Duration 275
After the burn-in, the reliability of a nonconforming item is given by
Rb;n t Rn t=Rn s; t s 15:8
and the reliability of a conforming item is given by
Rb;c t Rc t=Rc s: 15:9
The probability that an item is conforming after the burn-in is given by
pb pRc s=qRn s pRc s 15:10
As such, the reliability function of the burnt-in item population is given by
Rb t 1 pb Rb;n t pb Rb;c t: 15:11
We assume that the burnt-in item will be put under operation with a mission time
L, which can be a warranty period or a plan horizon. Let Rb L denote the survival
probability of this burnt-in item within the mission time. Assume that a breakdown
cost cf is incurred if the item fails within the mission time. As such, the eld failure
cost is given by
CL 1 Rb Lcf : 15:12
The probability that the item will pass the test is Rs. To be concise, we simply
write it as R and let F 1 R. Let K denote the number of repairs before the item
passes the test. The probability that the item passes the test after k repairs is given by:
pk F k R; k 0; 1; 2; . . .: 15:13
Clearly, K follows the geometric distribution. The expected number of repairs is

given by EK F=R; and the expected number of tests is given by n 1=R.
When the item fails during the test, the mean test time is given by
Zs Zs Zs
1 p q
b xf xdx xfc xdx xfn xdx: 15:14
F F F
0 0 0
When gx is the Weibull pdf with parameters b and g, we have
Zs " #
s b
xgxdx lGa ; 1 1=b; 1 15:15
g
0
where Ga : is the gamma cdf. Using Eq. (15.15), the mean test time b can be
expressed in terms of the gamma cdf. The expected test time for an item to pass the
test is given by
T bn 1 s: 15:16
Let c1 denote the test cost per unit test time and c2 denote mean repair cost for
each repair. The total burn-in cost is given by
CB c1 T c2 n 1: 15:17
The objective function is the sum of the eld failure cost and total burn-in cost,
given by
Js CL CB : 15:18
The optimal burn-in duration is determined through minimizing Js.

Example 15.2 Suppose the lifetimes of both conforming and nonconforming items
follow the Weibull distribution with parameters shown in the rst four columns of
Table 15.1. The other parameters are shown in the last ve columns of Table 15.1.
The problem is to nd the optimal burn-in duration.
Figure 15.4 displays the plot of failure rate function. As seen, it is bathtub-shaped,
implying that the burn-in can be effective. Using the approach outlined above, we
obtained the total cost curve shown in Fig. 15.4. As seen, it is also bathtub-shaped.
The minimum point is at s 29:42 and the corresponding cost is 9.77.
If the objective is mission reliability rather than the cost, the optimal burn-in
duration is determined through maximizing the mission reliability. For the current
example, Fig. 15.5 displays the plot of mission reliability versus burn-in duration.
As seen, the plot is unimodal with maximum at s 41:92. Therefore, the optimal
burn-in duration associated with the mission reliability objective is 41.92.
Table 15.1 Parameters for c gc n gn P L c1 c2 cf

Example 15.2
3.5 500 0.85 150 0.1 100 1 30 200
Fig. 15.4 Plots of failure rate 30

function and total cost versus 25
burn-in period
r (t ), C ( )
20
15
C ( )
10
r (t ), 10-4
5
0
0 50 100 150 200
t,
15.4 Optimal Component-Level Burn-in Duration 277
Fig. 15.5 Plot of mission 1

reliability as a function of
0.8
burn-in period
Rb (L , )
0.6
0.4
0.2
0
0 50 100 150 200

Table 15.2 Field failure 0 29.42 41.92

probabilities before and after
burn-in Rb(L) 0.9460 0.9547 0.9552
Fb(L) 0.0540 0.0453 0.0448
Reduction (%) 16.0 17.0
Table 15.2 shows the mission reliabilities and eld failure probabilities before
and after burn-in. The last row shows the relative reductions in eld failure after
burn-in. As seen, the reduction is signicant and the performances obtained from
two burn-in schemes are close to each other. This implies that the burn-in period
can be any value within (29.42, 41.92).
15.5 Optimal System-Level Burn-in Duration
Component-level burn-in focuses on component nonconformance, whose effect on

reliability is modeled by a mixture model. System-level burn-in involves multiple
components and assembly errors (for the effect of assembly errors on reliability, see
Sect. 12.2.2.2). Since the component-level burn-in cannot screen out the assembly
defects, system-level burn-in is necessary.
Different from component-level burn-in, where the total time on test (TTT) is a
random variable, the TTT of system-level burn-in is usually a constant s. When a
component fails during burn-in, it is replaced with a normal one; when a connection
failure, it is perfectly repaired. The burn-in process continues until the TTT reaches
the prespecied duration. As such, the age of the replaced component (or repaired
connection) at the test end is a random variable, as shown in Fig. 15.6, where
symbol indicates a component replacement or connection repair.
There are two approaches to deal with the system-level burn-in problem. One is
to view the system as a single item and assume that its failure follows a nonho-
mogeneous Poisson process with bathtub-shaped failure intensity; and the other is
to decompose the system into component level so that the lifetime distribution of
the system is a function of the reliability of each component position. Since the
reliability information of some components of a product is known, the latter
Fig. 15.6 Repair processes in
Component position no.

component positions during
n
the system-level burn-in
Age of component after burn-in
i

TTT
approach appears to be more practical. We consider the latter approach and present
the reliability and cost models as follows.
15.5.1 Reliability Model
Consider a system that consists of n component positions connected in series.

Possible defects in each component position include component defects and com-
ponent connection defects (i.e., assembly error). Assume that the failure times of all
components and their connection are mutually independent.
For Component i, let Fi t denote its life distribution before assembly. After
assembly, it becomes defective with probability qi , and let Gi t denote the life
distribution of a defective component. As such, the failure distribution of the
component after assembly is given by
FPi t 1 1 Fi t1 qi Gi t: 15:19
The reliability function of the product before burn-in is given by
Y
n
Rt RPi t; RPi t 1 FPi t: 15:20
i1
After the burn-in, the reliability function at the i-th position is given by
0
Ri x si
RBi x 0 f1 qi Gi x s Gi sg; x 0: 15:21
Ri si
Since the probability that there is a replacement or repair is usually small, we

0
approximately take si s. As such, the mission reliability after burn-in can be
approximated by
15.5 Optimal System-Level Burn-in Duration 279
Y
n
RL; s RBi L: 15:22
i1
15.5.2 Cost Model
The cost of an item consists of the burn-in cost and the eld operational cost. The
burn-in cost consists of component-level cost and system-level cost. The compo-
nent-level cost includes component replacement cost and connection repair cost.
For Component i, the replacement cost is given by
Cri cri Mi s 15:23
where cri is the cost per replacement and Mi t is the renewal function associated
with Fi t. Assume that the repair for connection failure is perfect (with a cost of
cmi ) so that the connection failure for each component occurs once at most. As such,
the connection repair cost is given by
Cmi cmi qi Gi s: 15:24
The total component-level cost is given by
Ci Cmi Cri : 15:25
The system-level cost deals with burn-in operational cost. Assume that the
operational cost per unit time is a constant c0 . As such, the system-level cost is
given by
Cs c0 s: 15:26
We now look at the eld failure cost, which is given by Eq. (15.12) with Rb L
given by Eq. (15.22). It usually assumes that the cost for a eld failure, cf , is four to six
times of the actual repair cost to reflect the intangible losses such as reputation cost.
As a result, the total cost for each item is given by
X
n
Js Cri Cmi c0 s cf 1 Rb L: 15:27
i1
The optimal burn-in duration is determined through minimizing Js or maximizing

Rb L.
References
1. Block HW, Savits TH (1997) Burn-in. Stat Sci 12(1):119

2. Caruso H (1989) Environmental stress screening: an integration of disciplines. In: Proceedings
of 1989 annual reliability and maintainability symposium, pp 479486
3. Chan SI, Kang JH, Jang JS (2013) Reliability improvement of automotive electronics based on
environmental stress screen methodology. Microelectron Reliab 53:12351238
4. Jiang R, Jardine AKS (2007) An optimal burn-in preventive-replacement model associated
with a mixture distribution. Qual Reliab Eng Int 23(1):8393
5. Naidu NVR (2008) Mathematical model for quality cost optimization. Rob Comput-Integr
Manuf 24(6):811815
6. Wagner M, Unger W, Wondrak W (2006) part average analysisA tool for reducing failure
rates in automotive electronics. Microelectron Reliab 46:14331438
7. Wong KL, Lindstrom DL (1988) Off the bathtub onto the roller-coaster curve. In: Proceedings
of 1988 annual reliability and maintainability symposium, pp. 356363
8. Wong KL (1990) Demonstrating reliability and reliability growth with environmental stress
screening data. In: Proceedings of 1990 annual reliability and maintainability symposium,
pp. 4752
9. Wu CL, Su CT (2002) Determination of the optimal burn-in time and cost using an environment
stress approach: a case study in switch mode rectier. Reliab Eng Syst Saf 76:5361
10. Yan L, English JR (1997) Economic cost modeling of environmental-stress- screening and
burn-in. IEEE Trans Reliab 46(2):275282
11. Yan L, English JR, Landers TL (1995) Modeling latent and patent failures of electronic
products. Microelectron Reliab 35(12):15011510
Part IV
Product Quality and Reliability in
Post-manufacturing Phase
Chapter 16
Product Warranty
16.1 Introduction
Product support (also known as customer support or after-sales support) deals with
product service, including installation, maintenance, repair, spare parts, warranty,
eld service, and so on. Product support plays a key role in the marketing of
products and the manufacturer can obtain prots through product servicing (e.g.,
provision of spare parts and maintenance servicing contracts).
Product warranty is a key part of product support. In this chapter we focus on
product warranty-related issues, including warranty policies, warranty cost analysis,
and warranty servicing.
The outline of the chapter is as follows. We start with a discussion of product
warranties in Sect. 16.2. Typical warranty policies are presented in Sect. 16.3.
Reliability models in warranty analysis are presented in Sect. 16.4, and warranty
cost analysis is dealt with in Sect. 16.5. Finally, related issues about warranty
servicing are discussed in Sect. 16.6.
16.2 Product Warranties
16.2.1 Concepts and Roles of Warranty
A warranty can be viewed as a contractual agreement between the manufacturer and

buyer of a product. It establishes responsibilities of buyer and liability of manu-
facturer when a failure occurs in the warranty period. As such, warranty provides a
protection for both the consumers and the manufacturers.
The manufacturer can use warranty as a marketing tool since a longer warranty
period usually attracts more customers. However, warranty involves additional

284 16 Product Warranty
servicing costs to the manufacturers and hence reducing warranty cost becomes of
great importance to the manufacturers.
The expected warranty costs depend on warranty requirements and associated
maintenance actions, and can be reduced through reliability improvement, product
quality control, and making adequate maintenance decisions in the warranty period.
16.2.2 Maintenance-Related Concepts
Warranty servicing involves maintenance activities. Maintenance can be classied

into two main types: corrective and preventive. Corrective maintenance (CM)
occurs after items failure and restores the failed item to an operational state by
repair actions; and preventive maintenance (PM) is performed before items failure
and aims to reduce items degradation and its risk of failure.
Warranty servicing usually involves CM actions. However, effective PM actions
in the warranty period may reduce the number of failures and consequently reduce
warranty servicing costs. During the post-warranty period, PM actions have a
considerable impact on the life cycle costs of the product. Therefore, the manu-
facturer should develop an appropriate PM scheme for its product. The PM scheme
can be different for different use environment and operational conditions.
Depending on whether a product is repairable or non-repairable, two basic
maintenance actions are repair and replacement of failed components. For a non-
repairable product, any warranty claim leads to a product replacement. For a
repairable product, a failed product can be either repaired or replaced.
Depending on the maintenance degree (or maintenance level), repair can be
minimal, perfect, or imperfect. A minimal repair does not change the failure rate of
the repaired item, a perfect repair is equivalent to a renewal, and the effect of an
imperfect repair is in between the effects of minimal and perfect repairs. Different
maintenance degree leads to different expected warranty servicing cost. For
example, a minimal repair spends a minimum cost to rectify the current failure but
can have the risk of subsequent failures. Frequent failures can increase customers
dissatisfaction since the customer has to bear some negative consequences of
failures (e.g., loss due to downtime).
16.3 Warranty Policies
16.3.1 Classication of Warranty Policies
The warranty policies can be classied in different ways. According to the de-
nition of warranty period, a warranty policy can be one- or two-dimensional. One-
dimensional warranty policies are characterized by a warranty period, which is
usually a time interval on the items age. In contrast, the two-dimensional warranty
16.3 Warranty Policies 285
policies are characterized by a region on the two-dimensional plane, where the axes
represent the age and the usage of the item. For vehicles, the usage is represented in
terms of mileages.
According to whether or not the warranty period is xed, a warranty policy can
be nonrenewable or renewable. The renewable warranty policies are usually
associated with replacement of a failed item.
According to whether the warranty is an integral part of product sale, warranty
policies can be divided into base warranty (or standard warranty) and extended
warranty (also called service contract). An extended warranty is optional for the
customer and not free.
In terms of the cost structure of warranty, a warranty policy can be simple or of
combination, where two simple policies are combined.
Depending on the type of product, warranty policies can be for consumer
durables, commercial and industrial products or defense products. The buyers of
these products are individuals, organizations and government, respectively. When
the buyers are organizations and government, products are often sold in lots. This
leads to a type of special warranty policies: cumulative warranty policies. For the
defense products, a specic reliability performance may be required. In this case,
development effort is needed and this leads to another type of special warranty
policies: reliability improvement warranties.
16.3.2 Typical Warranty Policies
In this subsection we present several typical warranty policies, which are applicable
for all types of products.
16.3.2.1 One-Dimensional Nonrenewing Free Replacement Warranty
This policy is usually called free replacement warranty (FRW), which is widely
used for consumer products. Under this policy, the manufacturer agrees to repair or
provide replacements for failed items free of charge up to a time W (the warranty
period) from the time of the initial purchase.
This policy is one-dimensional and nonrenewing. The word replacement does
not imply that the failed items are always rectied by replacement. In fact, it is
common to restore the failed item to operational state by repair, especially by
minimal repair.
16.3.2.2 One-Dimensional Nonrenewing Pro-rata Rebate Warranty
This policy is usually called pro-rata rebate warranty (PRW). Under this policy, the
manufacturer agrees to refund a fraction of the purchase price when the item fails
before time W from the time of the initial purchase. The refund depends on the age
of the item at failure X and is a decreasing function of the remaining warranty time
W X. Let qx denote this refund function. A typical form of qx is
qx acb 1 x=W 16:1
where a 2 0; 1, cb is unit sale price and x is the age of failed item.
When the rst failure occurs and a fraction of the purchase price is refunded, the
warranty expires. In other words, this policy expires at the time when the rst
failure occurs within the warranty period or at W. This policy is applicable for non-
repairable products.
16.3.2.3 FRWPRW Combination Warranty
Under this policy, the warranty period is divided into two intervals: (0, W1 ) and
(W1 , W). If a failure occurs in the rst interval, a FRW policy is implemented; if the
failure occurs in the second interval, a PRW policy is implemented and the refund is
calculated by

x W1
qx acb 1 : 16:2
W W1
16.3.2.4 Two-Dimensional Nonrenewing FRW
A two-dimensional warranty is characterized by a region in an age-usage plane.

Under two-dimensional FRW, the manufacturer agrees to repair or provide a
replacement for failed items free of charge up to a time W or up to a usage U,
whichever occurs rst, from the time of the initial purchase. Here, W is called the
warranty period and U the usage limit. As such, the warranty region is dened by
the rectangle shown in Fig. 16.1. This policy is offered by nearly all auto
manufacturers.
The two-dimensional FRW policy has several variants. One such variant is the
policy, whose warranty region is the triangular region shown in Fig. 16.1. The
boundary of the region is given by
y U1 x=W: 16:3
Another variant is an extension of the one-dimensional FRWPRW combination

warranty with four parameter W1 ; W and U1 ; U. The refund function is given by

x W1 u U1
qx acb 1 max ; : 16:4
W W1 U U 1
16.3 Warranty Policies 287
Fig. 16.1 Two typical

two-dimensional warranties y = U (1-x/W )
y
x
16.3.3 Special Policies for Commercial

and Industrial Products
In addition to the policies discussed above, four special warranty policies that are
widely used for commercial and industrial products are one-dimensional cumulative
FRW, extended warranties, PM warranty, and reliability improvement warranties.
We briefly discuss them as follows.
16.3.3.1 One-Dimensional Cumulative FRW
Industrial and commercial products are bought either individually or as a batch.

Cumulative warranties (also termed as fleet warranties) are applied when items are
sold as a single lot and the warranty refers to the lot as a whole.
Under a cumulative warranty, the lot of n items is warranted for a total time of
nW, with no specic time P limit for any individual item. Let Xi denote the ith items
service life and Sn ni1 Xi denote the total lifetimes of all the n items. If
Sn \ nW, a FRW policy is implemented. The warranty expires when Sn nW.
For non-repairable products, the manufacturer guarantees that the mean life of a
population of items will meet or exceed some negotiated mean life lL . If the mean
life of the fleet, l0 , meets or exceeds lL , no compensation is given by the manu-
facturer; otherwise, compensation in terms of number of free replacement items is
given according to the value of l0 =lL . The method to estimate l0 is specied in the
sale and purchase agreement.
16.3.3.2 Extended Warranties
An extended warranty (sometimes called a service agreement, a service contract, or

a maintenance agreement) provides additional coverage in addition to the base
warranty as an integral part of product sale. Extended warranties are purchased by
customer, and particularly applicable for complex products (e.g., wind turbines), for
which the buyer may lack the expertise to maintain it after expiration of the base
warranty.
The customer population is heterogeneous in terms of usage intensity and

environment. Therefore, the population can be divided into several subpopulations
based on the customers locations, usage intensity, and other characteristics. The
manufacturer can develop different extended warranty policies for customers to
choose. Similarly, it can design different servicing contracts in post-warranty period
for different customer subpopulations.
16.3.3.3 Preventive Maintenance Warranty Policy
Under a PM warranty policy, any product failures are rectied by minimal repair
and additional PM actions are carried out within the warranty period.
When the warranty period is relatively long (e.g., the case where the warranty
covers the whole life of the product), the manufacturer needs to optimize PM
policies. Often, the burn-in and PM are jointly optimized to reduce total warranty
servicing costs (e.g., see Ref. [8]).
16.3.4 Reliability Improvement Warranties
All the policies discussed above are also applicable for defense products. A special
policy associated with defense products is reliability improvement warranties,
which provide guarantees on the reliability (e.g., MTBF) of the purchased equip-
ment. Under this policy, the manufacturer agrees to repair or provide replacements
free of charge for any failed parts or items until time W after purchase. In the
meantime, the manufacturer also guarantees the MTBF of the purchased equipment
to be at least a certain level M. If the evaluated or demonstrated MTBF is smaller
than M, the manufacturer will make design changes to meet the reliability
requirements at itself cost.
The terms of reliability improvement warranties are negotiated between the
manufacturer and buyer, and usually include an incentive for the manufacturer to
increase the reliability of the products after they are put into service. The incentive
is an increased fee paid to the manufacturer if the required reliability level has been
achieved.
16.4 Reliability Models in Warranty Analysis
In this section, we discuss the reliability models that are needed in warranty
analysis.
16.4 Reliability Models in Warranty Analysis 289
16.4.1 Reliability Characteristics of Renewal Process
When the component is non-repairable, the failed component is replaced by a new

component. Assume that the failures are detected immediately; the items are sta-
tistically similar with failure distribution Ft, the failures are statistically inde-
pendent; and the replacement times can be ignored. In this case, the failures over
time occur according to a renewal process associated with Ft. The expected
number of renewals in (0; t) is called the renewal function given by Eq. (6.1), and
the renewal intensity function is given by mt dMtdt, which is also called the
renewal density function. In general, it is not possible to obtain the value of Mt
analytically for most distribution models, including the Weibull distribution.
Therefore, the renewal function is usually computed using approximations or
numerical methods.
For large t (e.g., t 2g, where g is the characteristic life), Mt can be
approximated by Eq. (6.2). For small t (e.g., t g), Mt can be approximated by
Mt Ft Ht lnRt 16:5
where Ht is the cumulative hazard function. Generally, the renewal function can
be approximated by [4]:
X
N
t il
Mt Ft U p 16:6
i2 ir
where U: is standard normal cdf and N is a sufciently large integer, or
X 2
N
l 2 r
Mt Ft Ga t; i ; 16:7
i2
r l
where Ga t; u; v is the gamma cdf with shape parameter u and scale parameter v.
For the Weibull distribution with t g, we have the following approximation [5]:
Mt pbFt 1 pbHt; pb Fw b 1; 0:9269; 0:8731 16:8
where Fw x; a; b is the Weibull cdf (with shape parameter a and scale parameter b)
evaluated at x. In the warranty analysis, the warranty period W is usually smaller
than the characteristic life so that Eq. (16.8) is accurate enough.
16.4.2 Reliability Characteristics of Minimal Repair Process
The failure of a system is often due to the failure of one or more of its components.
At each system failure, the number of failed components is usually small relative to
the total number of components in the system. The system is restored back to its
working state by either repairing or replacing these failed components. Since most
of systems components are not repaired or replaced, this situation is equivalent to a
minimal repair.
Let Ft denote the distribution function of the time to the rst failure, Ti denote
the time to the ith failure and Fi t; t [ ti1 , denote the distribution of Ti for the
repaired item. When a failed item is subjected to a minimal repair, the failure rate of
the item after repair is the same as the failure rate of the item immediately before it
failed. In this case, we have
Ft ti1 Fti1
Fi t ; t [ ti1 : 16:9
1 Fti1
If the item is not subjected to any PM actions and all repairs are minimal, then
the system failures can be modeled by a point process. Let Nt denote the number
of minimal repairs in (0; t). Nt follows the Poisson distribution with the MCF
given by [6]
Mt ln1 Ft: 16:10
It is noted that the variance of the Poisson distribution equals its mean. Therefore,
the variance of Nt is equal to Mt.
Specially, when Ft is the Weibull cdf, we have
Mt t=gb : 16:11
This is the well-known power-law model.
16.4.3 Imperfect Repair Models for Modeling Effect

of Preventive Maintenance
PM actions can affect both the rst and subsequent failures, and a PM is usually
viewed as an imperfect repair. As such, the effect of a PM on the reliability
improvement can by modeled by an imperfect maintenance model. Several specic
imperfect maintenance models are outlined as follows.
16.4.3.1 Virtual Age Models
The virtual age models are widely used to model the effect of PM on the reliability
improvement [10]. Suppose that a periodic PM scheme is implemented at ti is,
where s is PM interval. The failure is rectied by minimal repair, which does not
change the failure rate. Let vi denote the virtual age after the ith PM. The virtual age
Model I assumes that each PM reduces the virtual age of the product by a fraction
of the previous PM interval length s, i.e., as, where a is a number between 0 and 1
and called the degree of restoration. When a 0, the PM can be viewed as a
minimal repair and when a 1 the PM is equivalent to a perfect repair. As such,
we have
vi ti ias is1 a: 16:12
It is noted that the actual age at the ith PM is ti is, the virtual age just before
the ith PM is v i vi1 s, and the virtual age just after the ith PM is vi . As a
result, the failure rate reduction due to the ith PM is given by Dri rv i rvi ;
and the growth of the failure rate is according to rt ti vi rather than according
to rt. In other words, the effect of PM is twofold:
(a) current failure rate gets reduced, and
(b) the growth of the failure rate gets slowed down.
Given the distribution of time to the rst failure, Ft, and the parameter of
virtual age model I (i.e., a), the conditional distribution function after the ith PM
performed at ti is given by
1 Ft ti vi
Fi t 1 ; t ti : 16:13
1 Fvi
The virtual age Model II assumes that each PM reduces the virtual age of the
product by a fraction of v
i , i.e., avi1 s. This implies that the PM in the virtual
age Model II has a larger reduction in virtual age than the reduction in Model I if
the value of a is the same. Therefore, it is often used to represent the effect of an
overhaul.
The virtual age just after the ith PM is given by
X
i
1 a 1 ai1
vi s 1 a j s : 16:14
j1
a
When i is large, vi sa1 1, which is nearly a constant.

The conditional distribution function after the ith PM is given by Eq. (16.13)
with vi given by Eq. (16.14).
16.4.3.2 Caneld Model
Caneld [3] introduces a PM model to optimize the PM policy during or after the
warranty period. Let s denote the PM interval and d 2 0; s denote the level of
restoration of each PM. A minimal repair has d 0, a perfect repair has d s and
an imperfect repair has 0 \ d \ s. Clearly, d has the same time unit as t.
The model assumes that the ith PM only gets the failure rate slowed down and
does not change the value of current failure rate. As such, the failure rate after the
ith PM is given by
ri t rt id ci 16:15
where ci is a constant to be determined. According to the model assumptions, we

have
ri1 ti ri ti : 16:16
Letting ti is and from Eqs. (16.15) and (16.16), we have
X
i
ci ci1 Di Dj 16:17
j1
where Di ris d d ris d.

Example 16.1 Assume that the time to the rst failure follows the Weibull distri-
bution with shape parameter 2.5 and scale parameter 10. A periodic PM scheme is
implemented with a PM interval of s 2. Any failure is rectied by a minimal
repair. The problem is to examine the expected cumulative failure number Mt
associated with the PM models discussed above. For the virtual age models I and II,
we assume a 0:5; and for the Caneld model, we assume d 1:5.
Figure 16.2 shows the plots of Mt versus t for the three PM models. As seen,
the improvement effect associated with the virtual age Model II is the largest; and
the improvement effect associated with the Caneld model is the smallest. As such,
a PM with a large maintenance effort (e.g., an overhaul) can be represented by the
Fig. 16.2 Mean cumulative 16

No PM
functions associated with 14
different PM models 12
10
M (t )
8 Canfield Model
6 Model I
4
2 Model II
0
0 5 10 15 20 25 30
t
virtual age Model II; a PM with an intermediate-level maintenance effort (e.g., a

type II PM for vehicles) can be represented by the virtual age Model I; and a PM
with a small maintenance effort (e.g., a type I PM for vehicles) can be represented
by the Caneld model.
16.4.4 Bivariate Reliability Models
To analyze the two-dimensional warranty policies discussed in Sect. 16.3.2.4, we

need a bivariate reliability model. There are three different approaches for this
purpose. The rst approach is to use a bivariate failure distribution. This approach is
not preferred due to its complexity.
The second approach is to combine two scales (i.e., time t and usage u) into a
composite scale. Combining multiple scales into a composite scale can improve
failure prediction capability (see Ref. [7] and the literature cited therein). Suppose
that there are a set of data pairs (ti ; ui ; 1 i n), which are the observations at
failures. Let y ut; u; h denote the composite scale, where h is the parameter set
to be determined. Let ly h and ry h denote the sample mean and sample standard
deviation of the dataset (yi uti ; ui ; h; 1 i n). The composite scale has the
best failure prediction capability when CVy h ry h=ly h achieves its mini-
mum. As such, h can be optimally determined by minimizing CVy h. Two typical
function forms for ut; x; h are as follows:
u1 t; u t au; u2 t; x tua ; 16:18
For a given problem, both can be used as candidates, and the selection is given to
the candidate with smaller CVy h. As such, the reliability model can be repre-
sented by the distribution of random variable Y.
The third approach is to combine two scales into usage rate given by
q u=t: 16:19
For the population of a product, the usage rate q is a random variable and can be
represented by a distribution Gx, where x represents q. For a specic item, it is
usually assumed that the usage rate is a constant.
Consider the two-dimensional nonrenewing FRW with parameters W and U. Let
q0 U=W and p0 Gq0 . It is clear that the warranty of a sold item with usage
rate q\q0 [q q0 ] will expire at [before] t W. The proportion of the items
whose warranty expire at t W is 100p0 %.
Fig. 16.3 A failure-repair 35

process generated from 30
a bi-failure-mode model 25
Major
N (t)
20
15
10
5 Minor
0
0 20 40 60 80 100 120
t
16.4.5 Bi-failure-Mode Models
For a repairable product, the time to the rst failure can be represented by a
distribution function Ft. Failure can be minor (type I failure) or major (type II
failure). A minor failure is rectied by a minimal repair and a major failure is
rectied by an imperfect or perfect repair.
Let pt [qt 1 pt] denote the conditional probability of the event that the
failure is minor [catastrophic] if a failure occurs at age t. Generally, it is more
possible for a failure to be minor [major] when the age is small [large]. This implies
that pt [qt] decreases [increases] with age. However, pt can be non-monotonic
if there exist early failures due to manufacturing quality problems. As such, the
failure and repair process is characterized by Ft and pt. The characteristics of
this process can be studied using simulation.
The simulation starts with an initial age a ( 0). Then, a random life x is
generated according to Ft. The age of the item is given by a x. The failure type
can be simulated according to pa x, and the initial age is accordingly updated.
We illustrate this approach as follows:
Example 16.2 Assume that the lifetime of a product follows the Weibull distri-
bution with shape parameter 2.5 and scale parameter 10, and pt et=8 . Further,
we assume that a minor [major] failure is rectied by a minimal repair [replace-
ment]. Using the approach outlined above, a failure-repair process is generated and
displayed in Fig. 16.3. As seen, 10 of 30 failures are major. The MTBF of this
process is 3.56, which is much smaller than MTTF 8:87.
16.5 Warranty Cost Analysis
The manufacturer incurs various costs for rectication actions of failed items under
warranty and is interested in forecasting the warranty cost of the product for a given
warranty policy. The warranty coverage can be changed since it can affect buying
decisions. In this case, the manufacturer needs to estimate the warranty cost under the
new warranty policy. These problems deal with warranty servicing cost analysis.
16.5 Warranty Cost Analysis 295
The outcomes of warranty servicing cost analysis include expected warranty cost
per unit sale, expected total cost over a given planning horizon L for the manu-
facturer and buyer, and the prot of the manufacturer. In this section, we analyze
these costs for three typical warranty policies: one-dimensional FRW and PRW as
well as two-dimensional FRW.
16.5.1 Cost Analysis for Non-repairable Product

Under One-Dimensional FRW
For a non-repairable product, any failure during the warranty period is rectied by
replacing the failed item with a new item. Failures over the warranty period occur
according to a renewal process.
Let cs denote the cost of replacing a failed item and cb denote the sale price. The
expected warranty cost per item to the manufacturer is given by
CW cs 1 MW: 16:20
MW can be evaluated by Eq. (16.8) under the assumption of W g. It is noted

that Eq. (16.20) includes the manufacturing cost of the initially sold item. The ratio
between the warranty servicing cost and sale price is given by
rw cs MW=cb : 16:21
The prot per sold item is given by
Cp cb CW: 16:22
We now look at the costs of the manufacturer and customer in a given planning
horizon L. The sold item will be renewed by the customer at the rst failure after W,
when the expected renewal number is MW 1. Under the assumption of l L,
we have Mt t=l. As such, the expected renewal cycle length is given by
ET lMW 1: 16:23
The required number of items in the planning horizon is given by
L
n 1: 16:24
ET
As such, the users total cost in the planning horizon is given by ncb and the
manufacturers total cost in the planning horizon is given by nCW, and its total
prot is given by nCp .
Table 16.1 Results for pb 0.8082 CW 877.75

Example 16.3
FW 0.0962 Cp 122.25
HW 0.1012 ncb 3054.45
MW 0.0972 nCW 2681.05
l 4.4363 nCp 373.40
n 3.0544 rw 7.8 %
Example 16.3 Assume that the lifetime of a product follows the Weibull distri-
bution with shape parameter 2.5 and scale parameter 5 years. The warranty period is
W 2 years. The servicing and sale costs are 800 and 1000, respectively. The
planning horizon is 10 years. The problem is to estimate related costs.
We compute the renewal function using Eq. (16.8). The results are shown in
Table 16.1. As seen, the warranty servicing cost is about 7.8 % of the sale price.
16.5.2 Cost Analysis for Repairable Product

Under One-Dimensional FRW
For a repairable product, all failures over the warranty period are usually minimally
repaired. In this case, the number of minimal repairs over the warranty period is
given by Eq. (16.10). Let cm denote the cost of each repair. The expected warranty
cost per item to the manufacturer is given by
cW cm ln1 FW: 16:25
The manufacturers cost of per sold item is given by
CW cs cW: 16:26
For a given planning horizon L, the expected length of a renewal cycle depends
on the replacement decision. The models for determine the optimal stopping time of
a minimal repair process can be found from Ref. [6].
16.5.3 Cost Analysis for One-Dimensional PRW Policy
Under the PRW policy, the time to the rst failure (i.e., warranty expiration time) is
random variable. Conditional on X x, the manufacturers cost is given by
cx cs acb 1 x=W: 16:27

The expected warranty cost per item to the manufacturer is given by
ZW
CW cxf xdx RWcs : 16:28
0
The expected cost for user is given by
ZW
Cu W cb cx cs f xdx: 16:29
0
The expected prot per item to the buyer is given by
Cp Cb CW: 16:30
For the Weibull distribution, we have
ZW
acb
cxf xdx cs acb FW l 16:31
W W
0
where
ZW
lW xdFx lGa HW; 1 b1 ; 1: 16:32
0
Assume that the planning horizon L l. The expected renewals in the planning
horizon is given by
n L=l 0:51 r=l2 : 16:33
As such, the users total cost in the planning horizon is given by nCu W; the
manufacturers total cost in the planning horizon is given by nCW, and its total
prot is given by nCp .
16.5.4 Cost Analysis for Two-Dimensional FRW Policy
We use the usage rate approach to analyze related costs under this policy. Assume
that the usage rate given by Eq. (16.19) is a random variable and can be represented
by a distribution Gx, x 2(0; 1). Further, assume that any failure is rectied by
minimal repair. Let
q0 U=W: 16:34
Assume that the life at usage rate q0 follows the Weibull distribution with shape
parameter b and scale parameter g0 . Since the usage rate is similar to an accelerated
factor, we assume that the life at usage rate q follows the Weibull distribution with
shape parameter b and scale parameter g given by
q0
g g : 16:35
q 0
For a given value of q, the warranty terminates at
sq minW; U=q: 16:36
As such, the conditional expected repair number is given by

b a1 qb ; q\q0
nq sq =g 16:37
a 2 ; q q0
where
b b
W U
a1 ; a2 : 16:38
q0 g0 q 0 g0
It is noted that the expected repair number for q [ q0 is unchanged. As such, the
usage limit U actually controls the total repair number.
Removing on the condition, we have expected repair number per sold item given
by
Z1
n nxdGx: 16:39
0
Specially, assume that the usage rate follows the lognormal distribution with
parameter ll and rl . We have

bll brl 2 =2 lnq0 ll
n a1 e U ; brl ; 1 a2 1 Ulnq0 ; ll ; rl : 16:40
rl
Fig. 16.4 Influence of b on 0.4

warranty cost
0.3
0.2
n
0.1
0
1 2 3 4 5 6

The expected warranty cost per item to the manufacturer is given by
CW cs ncm : 16:41
The manufacturers cost of per sold item is given by Eq. (16.26).

Example 16.4 Assume that the warranty period is W 2 years and the usage limit
is U 20 (1000 km). The lifetime follows the Weibull distribution with shape
parameter 2.5 and scale parameter g0 5 years when q q0 10. Assume that
the usage rate follows the lognormal distribution with parameters ll 2:90 and
rl 0:78. The replacement, repair, and sale costs are 45000, 500, and 50000,
respectively. The problem is to estimate related costs.
The probability of q\q0 (i.e., warranty expires at W) is 22.2 %. The expected

number of repairs per sold item is n 0:0887. The expected serving cost is 44.35,
the cost for the manufacturer is 45044.35, and the prot is 4955.65.
Figure 16.4 shows the plot of the expected number of repairs as a function of b.
As seen, the expected number of repairs quickly decreases as b increases. This
implies that a large b is desired, as pointed out by Jiang and Murthy [9].
16.6 Product Warranty Servicing
Product warranty servicing is important to reduce the warranty cost while ensuring
customer satisfaction. In this section we briefly discuss three warranty servicing-
related issues: spare part demand prediction, optimal repairreplacement decision,
and eld information collection and analysis.
16.6.1 Spare Part Demand Prediction
Delay in repairs due to awaiting spare parts is costly but maintaining a spare part
inventory costs money. As such, the inventory can be optimized based on the
estimation of spare part demand. The demand estimation deals with predicting the
number of replacements for a specic component in a given time interval. The
replacements over the warranty period occur according to a renewal process and the
demand prediction needs to evaluate the renewal function and to consider the
variance of number of replacements.
Spare part inventory optimization needs to consider the importance of a spare
part. The decision variables include inventory level, reordering time, and order
quantity. These are related to the sales over time and component reliability.
16.6.2 Optimal RepairReplacement Decision
When a repairable item fails under warranty, it can be rectied by repair or

replacement with a new item. The repair cost is usually less than the replacement
cost in most situations. On the other hand, the time to the next failure for a repaired
item is statistically shorter than the one for a replaced item. As such, for a specic
situation, the decision to repair or replacement needs to be optimally made.
Strategies for making this decision can be based on the age (and/or usage) at failure
or based on the repair cost (or time). The former is called the age-based approach
and the latter is called the repair limit approach.
In the age-based approach and the case of one-dimensional warranties, a
threshold value for the (remaining) age of the item at failure is set. If the age is
smaller than the threshold value, a replacement may be more appropriate; other-
wise, a repair can be applied. The threshold value can be optimally determined to
minimize the expected cost of servicing the warranty over the warranty period. In
the case of two-dimensional warranties, two threshold values for the (remaining)
age and usage are set.
In the repair-limit approach, the cost or time to repair a failed item is a random
variable, which can be characterized by a distribution function. A threshold value
for repair cost or time is set and can be optimally determined. If the estimated repair
cost or time is smaller than the threshold value, the failed item is repaired; otherwise
replaced. More details about this approach will be presented in Sect. 17.3.
16.6.3 Field Information Collection and Analysis
A lot of data is generated during the servicing of warranty, including [1, 2]:
Technical data such as modes of failures, times between failures, degradation
data, operating environment, use conditions, etc. This type of information can be
useful for reliability analysis and improvement (e.g., design changes).
Servicing data such as spare parts inventories, etc. This type of information is
important in the context of improving the product support.
16.6 Product Warranty Servicing 301
Customer related data (e.g., customer impressions for product and warranty
service) and nancial data (e.g., costs associated with different aspects of
warranty servicing). This type of information is useful for improving the overall
business performance.
To effectively implement warranty servicing, adequate information systems are
needed to collect data for detail analysis. Such systems include warranty man-
agement systems and FRACAS mentioned in Chap. 9.
References
1. Blischke WR, Murthy DNP (1994) Warranty cost analysis. Marcel Dekker, New York
2. Blischke WR, Murthy DNP (1996) Product warranty handbook. Marcel Dekker, New York
3. Caneld RV (1986) Cost optimization of periodic preventive maintenance. IEEE Trans Reliab
35(1):7881
4. Jiang R (2008) A gammanormal series truncation approximation for computing the Weibull
renewal function. Reliab Eng Syst Saf 93(4):616626
5. Jiang R (2010) A simple approximation for the renewal function with an increasing failure
rate. Reliab Eng Syst Saf 95(9):963969
6. Jiang R (2013) Life restoration degree of minimal repair and its applications. J Qual Maint Eng
19(4):13552511
7. Jiang R, Jardine AKS (2006) Composite scale modeling in the presence of censored data.
8. Jiang R, Jardine AKS (2007) An optimal burn-in preventive-replacement model associated
with a mixture distribution. Qual Reliab Eng Int 23(1):8393
10. Kijima M, Sumita U (1986) A useful generalization of renewal theory: counting processes
governed by nonnegative Markovian increments. J Appl Prob 23(1):7188
Chapter 17
Maintenance Decision Optimization
17.1 Introduction
Maintenance is the actions to restore the system to its operational state through
corrective actions after a failure or to control the deterioration process leading to
failure of a system. The phrase actions to restore means corrective maintenance
(CM) and the phrase actions to control means preventive maintenance (PM).
Maintenance management deals with many decision problems, including
maintenance type selection (i.e., CM or PM), maintenance action selection (e.g.,
repair or replacement), maintenance policy selection (e.g., age-based or condition-
based), and policy parameter optimization. In this chapter, we present an overview
for key issues in maintenance management decision. Our focus is on typical
maintenance policies and their optimization models. More contents about mainte-
nance can be found from Ref. [5].
This chapter is organized as follows. We rst discuss maintenance policy
optimization in Sect. 17.2. Typical CM policies are presented in Sect. 17.3. Typical
component-level PM policies are classied into three categories: time-based
replacement policies, time-based inspection policies, and condition-based mainte-
nance policies. They are discussed in Sects. 17.4 through 17.6, respectively.
Typical system-level PM policies are group and opportunistic maintenance policies,
and are discussed in Sect. 17.7. Finally, we present a simple maintenance float
system in Sect. 17.8.
17.2 Maintenance Policy Optimization
A maintenance policy denes or species when or in what situation a certain

maintenance action is implemented. As such, maintenance policy optimization
involves the following three issues:

304 17 Maintenance Decision Optimization
1. Specication of maintenance task or action,

2. Time to trigger the maintenance task, and
3. Optimization of policy parameters.
We discuss these issues in the following three subsections, respectively.
17.2.1 Maintenance Tasks
Two types of basic maintenance actions are CM and PM. Typical PM tasks can be
found from reliability-centered maintenance (RCM) and total productive mainte-
nance (TPM). We rst look at the choice problem between PM and CM before
introducing RCM and TPM, and then summarize typical PM tasks.
17.2.1.1 Corrective Maintenance and Preventive Maintenance
The choice problem between CM and PM deals with two aspects: applicability and
effectiveness of PM. The applicability deals with failure mechanism and the effec-
tiveness with economic sense, which is addressed by optimization and discussed later.
The failure mechanism can be roughly divided into two categories: overstress
mechanism and wear-out mechanism. The failures due to overstress are hard to be
predicted and hence have to be rectied by CM. If the consequence of failure is
unacceptable, redesign is a unique improvement strategy.
Wear-out is a phenomenon whereby the effect of damage accumulates with time.
The item fails when the accumulated damage reaches a certain critical level. As
such, the failure due to wear-out mechanism implies that the item experiences a
degradation process before its failure. Generally, an item failure can be the result of
interactions among two or more mechanisms (e.g., stress-assisted corrosion). In all
these cases, the item is aging, the failure is preventable and the PM is applicable.
17.2.1.2 Reliability-Centered Maintenance
RCM [8] is an engineering-oriented maintenance methodology used to determine

maintenance tasks for dominant causes of equipment failures based on two criteria:
technical feasibility and cost effectiveness. Through integrating these identied
maintenance tasks, a complete maintenance regime (or maintenance program) for the
system can be established. RCM is particularly suitable for safety-related systems.
As mentioned in Chap. 9, RCM is a systematic application of the machinery
FMEA. RCM works through addressing the following issues:
Understanding the operating context, in which the system functions are main-
tained by maintenance. The operating context is characterized by items func-
tions, associated performance standards and failure denition.
Carrying out failure cause, effect, and criticality analysis.
17.2 Maintenance Policy Optimization 305
Determining appropriate maintenance tasks for the identied failure modes,

including PM tasks to prevent potential failure and the measures to reduce the
consequences of the failure when a suitable PM task cannot be found. This is
done by applying a set of decision-logic rules.
RCM also addresses continuous improvement through periodic review and
adjustment.
Dominant failure causes and key items are identied based on criticality or
failure consequences. RCM classies failure consequences into three categories:
Safety and/or environmental consequences,
Operational consequences, and
Nonoperational consequences.
Operational and nonoperational consequences are also called economic
consequences.
If a function is not critical (i.e., its failure risk is acceptable), a run-to-failure (i.e.,
CM) strategy is recommended. When the risk of a unpredictable failure is high,
design change is recommended. This involves redesigning a component so that the
new component has better reliability characteristics. This is sometimes called
design-out maintenance.
If failure risk is neither acceptable nor high, an appropriate PM task can be
carried out. RCM considers two age-related preventive tasks and two condition-
related preventive tasks. They are
Scheduled restoration task, which deals with remanufacturing a component or
overhauling an assembly at or before a specied age limit.
Scheduled discard task, which deals with discarding an item or component at or
before a specied age limit.
Predictive maintenance task (i.e., condition-based maintenance or CBM), which
uses condition monitoring and failure prediction techniques to determine the PM
time.
Detective maintenance task (i.e., failure-nding task or inspection), which deals
with implementing an inspection scheme to nd hidden failure for the case
where the failure is not self-announced.
RCM does not deal with quantitative optimization of maintenance policies, which
is the focus of this chapter. In addition, its six failure patterns have been criticized in
the literature. It is unclear that those failure patterns are about failure rate or failure
intensity and associated with components or systems; and how they are obtained. In
fact, the shape of failure intensity of a complex system is usually roller coaster shaped.
17.2.1.3 Total Productive Maintenance
TPM (e.g., see [9]) is a management- or organization-oriented maintenance meth-

odology and has been widely applied in manufacturing enterprises. Since the full
support of the total workforce in all departments and levels is required to ensure
effective equipment operation, it is sometimes called the people-centered
maintenance.
TPM increases equipment efciency through eliminating six big losses:
breakdown losses caused by the equipment
setup and adjustment losses
idling and minor stoppage losses
speed losses
quality defect and rework losses, and
startup and yield losses.
These losses are combined into one measure of overall equipment effectiveness
(OEE) given by
OEE A P Y 17:1
where A is equipment availability, P is performance efciency, and Y is the rate of

quality products.
TPM achieves effective maintenance through PM programs implemented by
maintenance departments and autonomous maintenance program implemented by
production departments. The autonomous maintenance is a critical aspect of TPM.
The operators are systematically trained to implement thorough and routine
maintenance on a daily basis. Typical activities include precision checks, lubrica-
tion, parts replacement, simple repairs, and inspections.
TPM gives emphasis to early equipment management. This involves designing
and installing equipment that needs little or no maintenance and is mistake-
proong. A mistake-proong design makes mistakes impossible or at least easy to
detect and correct. This is achieved through prevention and detection devices. A
prevention device makes it impossible for a machine or machine operator to make a
mistake; and a detection device signals the user when a mistake has been made so
that the problem can be quickly corrected.
TPM addresses safety and environmental issues by continuously and system-
atically carrying out safety activities, including the development of safety check-
lists, the standardization of operations and coordinating nonrepetitive maintenance
tasks.
Similar to RCM, TPM does not deal with quantitative optimization of mainte-
nance policies.
17.2.1.4 Summary
According to the above discussion, typical maintenance actions or tasks are CM,
routine maintenance (or autonomous maintenance), replacement, overhaul,
inspection, CBM, and design-out maintenance.
17.2 Maintenance Policy Optimization 307
17.2.2 Timing of Maintenance Tasks
Timing of a specic maintenance task deals with under what condition the task is
triggered and implemented. Generally, there are three cases to trigger a maintenance
task. They are
1. Failure triggered: it leads to a CM action,
2. Age or calendar time triggered: it leads to a time-based PM action, and
3. Condition triggered: it leads to a condition-based PM action.
There are several CM policies that involve optimal selection between two
optional actions: repair and replacement. Two types of typical CM policies are
repair limit policy and failure counting policy.
There are many PM policies, and they can be divided into component-level
policies and system-level policies. A component-level policy is dened for a single
component, and a system-level policy is dened to simultaneously implement
several maintenance tasks for several components.
Most of PM policies are of component level. These policies fall into two cate-
gories: time-based maintenance (TBM) and CBM. Here, the time can be age,
calendar time, and usage; and the maintenance can be repair, replacement, and
inspection.
There are several system-level maintenance policies, and two typical policies are
group and opportunistic maintenance policies. In group maintenance, the PM
actions are combined into several groups. For each group, the tasks are simulta-
neously implemented in a periodic way. A main advantage with group maintenance
is that it can signicantly reduce maintenance interferences. A multi-level PM
program is usually implemented for complex systems, and group maintenance is the
basis of designing such a PM program.
A failure triggers a CM action. This provides an opportunity to simultaneously
perform some PM actions by delaying the CM action or advancing PM actions.
Such PM policies are called opportunity maintenance. A main advantage with
opportunity maintenance is that it can reduce both maintenance cost and mainte-
nance interferences.
According to the above discussion, we have the following classication for
maintenance policies:
CM policies (or repair-replacement policies) at both component level and sys-
tem level,
Component-level PM policies, including TBM and CBM policies,
Inspection policies at both component level and system level, and
System-level PM policies.
17.2.3 Optimization of Maintenance Policies
Basic elements of an optimization problem are decision variables, objective func-

tion, and constraint conditions. Decision variables depend on the policy. For
example, in the context of CBM, the decision variables can be PM threshold and
inspection interval. Most of maintenance decision optimization problems do not
deal with the constraint conditions, or the constraints are application-specic. The
decision objective depends on whether or not a failure has safety or environmental
consequences. If not, the objective can be cost or availability; otherwise, the
objective is risk-based.
Consider the case where only economic consequences are involved. The opti-
mization needs to evaluate and predict the stochastic failure and repair behavior of
the system and its components. Field data are collected for characterizing the failure
process. The properties (e.g., trend, randomness, etc.) of the collected data are
studied so as to select an appropriate model for tting the data (see Sects. 6.46.7).
Based on the tted model, a decision model (i.e., objective function) is developed to
optimize the maintenance policy. The optimal policy parameters can be obtained
through minimizing the maintenance cost or maximizing availability.
When a failure has safety and/or environmental consequences, a risk-based
approach can be used. Risk-based maintenance (RBM) is a maintenance approach
developed for plant and equipment whose failure can have serious safety and
environment consequences. RBM uses risk assessment techniques to identify and
quantify the occurrence probability of an undesired event and evaluate its loss.
Based on the outcomes of the risk assessment, the components with high risk are
given a higher priority in PM effort than the components with low risk. As such, a
PM plan that the safety requirements can be met is developed. Generally, a RBM
methodology requires designing an optimum inspection and maintenance program
and involves the following three steps:
Identify most probable failure scenarios, carry out detailed consequence analysis
for the selected scenarios, and compute their risks,
Compare the calculated risks with known acceptable criteria, and
Determine the frequencies of the maintenance tasks.
17.3 Repair-Replacement Policies
In this section, we look at a type of failure-driven policies, where a repair or

replacement always occurs at some failure. We focus on the following three repair-
replacement policies:
(a) Repair cost limit policy,
(b) Repair time limit policy, and
(c) Failure counting policy with a reference age.
17.3 Repair-Replacement Policies 309
17.3.1 Repair Cost Limit Policy and Its Optimization Model
Under this policy, the item runs to failure. When a failure occurs, the failed item is
inspected and the repair cost is estimated; the item undergoes minimal repair if the
estimated repair cost is less than a prespecied cost limit x0 ; otherwise it is replaced
by a new one. The repair cost limit is a decision variable. The policy reduces into
the renewal process when x0 0 and the minimal repair process when x0 1.
The appropriateness of the policy can be explained as follows. When the item
fails, the decision-maker has two options: minimal repair and failure replacement. If
the failure is rectied by a minimal repair, the direct repair cost may be smaller than
the cost of a failure replacement, but this may lead to more frequent failure and
hence spends more cost later.
Repair cost, X, is a random variable with cdf Gx and pdf gx, respectively. For
a specied cost limit x0 , the probability that a failed item will be repaired is
px0 Gx0 and the probability that it will be replaced is qx0 1 px0 .
After a minimal repair the failure rate remains unchanged. The replacement rate (as
opposed to failure rate) of the item at time t is ht; x0 qx0 rt. As such, the
intervals between failure replacements are independent and identically distributed
with the distribution function Fx t; x0 given by:
Fx t; x0 1 expqx0 Ht 1 Rqx0 t: 17:2
Letting U denote the time between two adjacent renewal points, then Fx t; x0
represents the distribution of U. Let Nt; x0 denote the expected number of failures
in 0; t. Then we have
Mt; x0 qx0 Nt; x0 or Nt; x0 Mt; x0 =qx0 : 17:3
Average repair cost is given by
Zx0
1
cm x0 ugudu: 17:4
px0
0
The expected cost per failure is given by
Ct; x0 cr qx0 cm x0 px0 17:5
where cr is the average cost of a replacement. The cost rate in (0, t) is given by
Ct; x0 Nt; x0 Ct; x0 Mt; x0

Jt; x0 : 17:6
t tqx0
Let lq denote the mean time between replacements. It is given by
Z1
lq Rqx0 tdt: 17:7
0
lq gC1 1=b=q1=b x0 : 17:8
When t ! 1, we have
Mt; x0 1
: 17:9
t lq
As a result, Eq. (17.6) can be written as below:
cr cm x0 px0 =qx0
Jx0 : 17:10
lq
The optimal policy is to select x0 to minimize Jx0 .
17.3.2 Repair Time Limit Policy and Its Optimization Model
Under this policy, repair time X is a random variable with cdf Gx and pdf gx.
When an item fails, the completion time of repair is estimated. The item is rectied
by minimal repair if the estimated repair time is smaller than a prespecied time
limit x0 ; otherwise, it is replaced and this involves ordering a spare item with a lead
time.
The appropriateness of the policy can be explained as follows. In the context of
product warranty, the repair cost is usually less than the replacement cost in most
situations, and hence the manufacturer prefers repairing the failed product initially
before providing a replacement service. If the failed item is unable to be xed
before the prespecied time limit set by the manufacturer, the failed item has to be
replaced by a new one so that the item can be returned back to the customer as soon
as possible.
Let cm denote the repair cost per unit time; cd denote the penalty cost per unit
time when the system is in the down state, c0 denote the xed cost (including the
price of item) associated with the ordering of a new item, and L denote the lead time
for delivery of a new item.
17.3 Repair-Replacement Policies 311
In a similar argument as deriving the cost rate for the repair cost limit policy, we
have
px0 Gx0 ; qx0 1 px0 ; Fx t; x0 1 Rqx0 t: 17:11
The sequence of failure replacements forms a renewal process and the expected
number of renewals in (0, t) is given by the renewal function Mt; x0 associated
with Fx t; x0 . The expected number of failures is given by
Nt; x0 Mt; x0 =qx0 : 17:12
Average repair cost is given by
Zt
cd cm
cm x0 ugudu: 17:13
px0
0
Failure replacement cost is given by
cr c0 cd L: 17:14
The expected cost per failure has the same expression as Eq. (17.5), and the
expected cost per unit time has the same expression as Eq. (17.10).
17.3.3 Failure Counting Policy with a Reference

Age and Its Optimization Model
Let T denote a reference age, and tk denote the time when the kth failure occurs.
Under this policy, the item is replaced at the kth failure if tk [ T or at the (k + 1)st
failure if tk \ T; and the failures before the replacement are rectied by minimal
repairs. It is noted that event tk \ T includes two cases: tk 1 \ T and tk 1 [ T.
This policy has two decision variables k and T. When T 0 [T 1], the item
is always replaced at tk [tk 1 ] (i.e., a failure counting policy without a reference
age); when k 1, it reduces into a minimal repair process without renewal.
A replacement cycle is the time between two successive failure replacements.
Let X denote the cycle length, nx and nT denote the number of failures in 0; x
and [0; T], respectively. Table 17.1 shows the relations among X, T, nx and nT.
Table 17.1 Relations among nT k 1 nT k nT k 1

X, T, nx and nT
XT Impossible Impossible nx k 1
X[T nx k nx k 1 Impossible
Let Fx and Rx denote the cdf and reliability function of X, respectively. It is

noted that Rx can be interpreted as the probability to conduct a minimal repair
(rather than a replacement) at a failure. Let mx denote the number of minimal
repairs in 0; x.
When X T (implying that nT nx k 1 or mx k), the reliability
function (i.e., the probability of minimal repairs) is given by
X
k
R1 x Prmx k Pk x pn x 17:15
n0
Hx
xe
n
where pn x HCn 1 .
When X [ T (implying that nx k and, nT k and nx k 1), the
reliability function is given by
R2 x Prfmx k 1 or mx mT kg
Pk x pk x pk Tpk x: 17:16
The expected cycle length of the policy is given by
ZT ZT Z1 Z1
WT; k R1 xdx R2 xdx Pk xdx 1 pk T pk xdx:
0 0 0 T
17:17
For the two-parameter Weibull distribution, we have

( )
g Xk
Cn 1=b Ck 1=b
WT; k 1 Ga HT; k 1=b; 1
b n0 Cn 1 Ck 1
17:18
where Ga : is the gamma cdf.

The expected number of minimal repairs is given by
nm k 1 Prtk [ T k1 Prtk [ T k Prtk [ T 17:19
where Prtk [ T PrnT \ k Pk T pk T. As such, the cost rate is

given by
c m nm c r
Jk; T : 17:20
WT; k
The optimal parameters of the policy are the values of k and T that minimize
Jk; T.
17.4 Time-Based Preventive Replacement Policies 313
17.4 Time-Based Preventive Replacement Policies
When a component fails in operation, it can take a high cost to rectify the failure,
and hence it can be much cheaper to preventively replace the item before the failure.
Such preventive replacement actions reduce the likelihood of failure and the
resulting cost, but increase the PM costs and sacrice the partial useful life of the
replaced item. This implies that the parameters characterizing the PM policy need to
be selected properly to achieve an appropriate tradeoff between preventive and
corrective costs.
Three preventive replacement policies that have been used extensively are the
age replacement policy, block replacement policy, and periodic replacement policy
with minimal repair. Each of them involves one single-decision variable T. In this
section, we look at these policies and their optimization decision models.
17.4.1 Age Replacement Policy and Its Optimization Model
Under the age replacement policy, the item is replaced either at failure or on
reaching a prespecied age T whichever occurs rst.
Let Ft denote the cdf of item life, and cf cp denote failure [preventive]
replacement cost. Preventive replacement of a component is appropriate only if the
components failure rate associated with Ft is (equivalently) increasing and cp \ cf .
A replacement cycle can be ended by a failure replacement with probability
FT or by a preventive replacement with probability RT. The expected cycle
length for a preventive replacement cycle is T and for a failure replacement cycle is
given by
ZT
1
Tc tdFt: 17:21
FT
0
As such, the expected operational time for a replacement cycle is given by
ZT
WT Tc FT TRT Rtdt 17:22
0

" #
T b 1
WT lGa ; ;1 : 17:23
g b
where Ga : is the gamma cdf. The expected total cost is given by
EC FTcf RTcp cp 1 q 1FT 17:24
where q cf =cp is called the cost ratio. The optimum replacement time interval T is
the time that minimizes the cost rate given by
EC
JT : 17:25
WT
The preventive replacement age can be viewed as a BX life with X 100FT.

When it is hard to specify the cost ratio, Jiang [3] suggests specifying the value of
T by maximizing the following
yt tRt: 17:26
The solution is called the tradeoff BX life.

Example 17.1 The lifetimes of several car components follow the Weibull distri-
bution. Their Weibull parameters and cost parameters are shown in Table 17.2. The
problem is to nd their optimal preventive replacement ages.
Using the tradeoff BX approach and the cost model, the optimal preventive
replacement ages can be obtained. The results are shown in the second and third
columns of Table 17.3, respectively. Since the cost ratios are large, the results
obtained from the two approaches are signicantly different. Generally, we take the
results from the cost model if the cost parameters can be appropriately specied.
Table 17.2 Reliability and Component b g 103 km cp q

cost parameters for Example
17.1 A 3.0 23 25 60
B 2.6 124 800 5
F 3.4 46 70 10
O 4.7 16 30 100
P 2.4 135 700 8
S 1.7 48 80 3
Table 17.3 Replacement Component BX TAge TPeriodic TBlock TGroup

intervals of car components
A 15.9 4.7 4.6 4.6
B 85.9 61.2 79.3 55.3 55.2
F 32.1 18.7 18.3 18.4
O 11.5 4.6 4.6 4.6
P 93.7 52.5 87.9 55.3 55.2
S 35.1 42.7 55.3 55.2
17.4 Time-Based Preventive Replacement Policies 315
17.4.2 Periodic Replacement Policy with Minimal Repair

and Its Optimization Model
Assume that component is repairable. Under the periodic replacement policy with
minimal repair, the item is preventively replaced at xed time instants kT and
failures are removed by minimal repair.
Let cp and cm denote the costs of a preventive replacement and a minimal repair,
respectively. The expected number of minimal repairs in a replacement cycle is
given by the cumulative hazard function HT. As a result, the cost rate function is
given by
cp cm HT
JT : 17:27
T
More generally, we can specify a common preventive replacement interval for

several components with similar reliability characteristics by minimizing the fol-
lowing cost function:
Pn
cp;i cm;i Hi T
JT i1
17:28
T
where cp;i , cm;i and Hi : are the preventive replacement cost, repair cost, and the
cumulative hazard function of the ith component, respectively.
Example 17.2 Consider Components B and P in Table 17.2. It is noted that their
reliability characteristics are similar. Assume that these two components are
repairable with cp =cm 2. The problem is to determine the individual preventive
replacement intervals and the common replacement interval.
From Eq. (17.27), we can obtain the individual preventive replacement intervals
shown in the fourth column of Table 17.3. From Eq. (17.28) we have that the
optimum common replacement interval is 82.8, which is close to the individual
replacement intervals.
17.4.3 Block Replacement Policy and Its Optimization Model
The block replacement policy is similar to the periodic replacement policy with
minimal repair. The difference is that the phrase minimal repair is revised as
failure replacement. The cost rate function for this policy can be obtained from
Eqs. (17.27) or (17.28) by replacing cm and HT by failure replacement cost cf and

renewal function Mt, respectively. In such a way, Eq. (17.27) is revised as
cp cf MT
JT 17:29
T
and Eq. (17.28) is revised as

Pn
cp;i cf ;i Mi T
JT i1
: 17:30
T
Example 17.3 Consider the components in Table 17.2. The problem is to nd the
preventive replacement interval of Component F, the common preventive
replacement interval of Component group (A, O), and the common preventive
replacement interval of Component group (B, P, S).
The renewal function is evaluated by Eq. (16.8). Using the cost models given by
Eqs. (17.29) and (17.30), we have the results shown in the fth column of Table 17.3.
As seen, the results are close to those obtained from the age replacement policy.
17.4.4 Discussion
1. A periodic PM policy is more convenient in implementation than an age-

dependent PM policy since it does not require keeping records on item age. The
block replacement policy is more wasteful than the age replacement policy since
a young item might be preventively replaced.
2. A simple maintenance policy is often generalized in several ways. A popular
way to extend a simple maintenance policy is to replace minimal repair with
imperfect repair. Such generalized policies may result in cost savings to some
extent but they become more complicated due to mathematical intractability and
may be inconvenient in implementation.
3. Reliability model and cost parameters used for maintenance decision analysis
may be updated if new information is available.
17.5 Inspection Policies
The state (working/failed) of an item is unknown if it is not monitored continu-

ously. Such examples include protective devices and stored items. To reduce the
risk of failure, an inspection scheme has to be implemented. There are two options
17.5 Inspection Policies 317
to detect the state: discrete inspection and continuous monitoring. The continuous
monitoring is often impossible or too costly so that the discrete inspection scheme is
often used. The key decision variables for a discrete inspection scheme are
inspection times. This is because over-inspections will lead to high inspection cost
and low availability while under-inspections increase the risk of failure. Thus,
inspection times should be optimized.
An inspection scheme can be periodic, quasi-periodic, or sequential. Under a
periodic scheme, inspections are conducted at time instants jT, where j 1; 2; . . .,
and T is called the inspection interval. Under a quasi-periodic scheme, the rst several
inspections are conducted in a nonperiodic way and then a periodic inspection
scheme is implemented. A simple quasi-periodic inspection scheme is dened as
tj t1 j 1T 17:31
where tj is the time to conduct the jth inspection. Under the sequential inspection
scheme, the inspection interval tj tj1 varies with j. To be simple, we focus on the
periodic inspection scheme in this section.
The inspection actions can have influence on the reliability characteristics of the
inspected item. Two typical cases are:
(a) a thorough PM action is carried out at each inspection so that the item can be
good-as-new after the inspection, and
(b) nothing is done for the inspected item when it is in working state so that the
inspection can be effectively viewed as a minimal repair.
In this section, we consider inspection polices associated with the above two
cases, and present the corresponding optimization models with the objective being
cost or availability.
17.5.1 Inspection Policy with Perfect Maintenance and Its

Optimization Model
Under this policy, inspection actions are periodically performed and the item is
preventively maintained at each inspection. The PM is assumed to be perfect. The
decision variable is inspection interval T.
Since the inspection is perfect, an inspection ends a cycle and resets the time to
zero. The probability that an operating item survives until T is RT; and the
probability that the item fails before T is FT. The mean downtime from the
occurrence of the failure to the time when it is detected is given by:
ZT ZT
1 1
td T T tf tdt Ftdt: 17:32
FT FT
0 0
When Ft is the two-parameter Weibull cdf, we have
gC1 1=b
td T T Ga Ht; 1 1=b; 1: 17:33
FT
Let s1 c1 s2 c2 denote the mean time (cost) to perform an inspection [and

repair] if the item is working [failed], and c3 denote the mean penalty cost per unit
downtime. The availability is given by
T td TFT
AT : 17:34
T s1 RT s2 FT
The mean cost rate is given by:
JT c1 RT c2 FT c3 td T=T: 17:35
The optimal inspection interval corresponds to the maximum of AT or the min-

imum of JT.
17.5.2 Inspection Policy with Minimal Repair

and Its Optimization Model
Under this policy, inspection actions are periodically performed at tj jT and do

not influence the reliability of the inspected item. The decision variable is
inspection interval T.
Let c1 and c2 denote the cost per inspection and the cost per unit time of an item
being unavailable due to an undetected failure, respectively. The expected total cost
to nd a failure is given by:
X
1
JT c1 j c2 tj Ftj Ftj1 c2 l 17:36
j1
where l is the mean lifetime. The optimal solution corresponds to the minimum of
JT.
If c1 represents the time of an inspection and c2 = 1, then JT represents the
expected downtime. The optimal solution with availability objective also corre-
sponds to the minimum of JT given by Eq. (17.36).
17.6 Condition-Based Maintenance 319
17.6 Condition-Based Maintenance
Time- or usage-based maintenance decision is for a population of identical or

similar items. The health state among these items may be considerably different
since their degradation levels can be considerably different due to the variability in
unit-to-unit performances, operational conditions, and environments. As such, the
PM time specied by a TBM policy can be too early for some items and too late for
some of the other items. This is especially true when the reliability model is
obtained from the eld data from considerably different operating conditions and/or
from the data pooled from nominally identical components produced by different
suppliers [3]. CBM can avoid such problems since the PM time of CBM is
dependent on the state of a specic item and the decision is individual-oriented
rather than population-oriented.
CBM continuously or discretely monitors one or more condition variables of an
item. The condition variables vary with time, and can be represented by one or
more degradation process models. The extrapolation is used to estimate or predict
the failure time or residual life that is usually described by a distribution. The PM
action is scheduled before an upcoming failure occurs.
In simple cases, the degradation process of an item is represented by a single
condition variable, which is monitored through implementing a periodic inspection
scheme. The degradation process is usually represented by a gamma or Wiener
process model. The failure is dened by a xed and known degradation level,
which is referred as functional failure threshold. To facilitate the implementation, a
PM degradation threshold can be set to trigger a PM action. The PM threshold can
be optimally determined using a decision model such as cost model. If the degra-
dation level observed at some inspection is smaller than the PM threshold, the next
inspection is scheduled; if the observed degradation level is larger than the failure
threshold, a CM action is carried out; otherwise, a PM action is carried out.
The use of a xed failure threshold is reasonable if the monitored condition
variable directly relates to the state of the item (e.g., wear amount or drop amount in
a certain performance). However, in most of practical problems, the degradation
process is represented by several condition variables, which indirectly relate to the
state of the item. In this case, the condition variables are usually combined into a
composite condition variable and its failure threshold is no longer known and xed.
Instead, the failure threshold associated with the composite condition variable is
random and time dependent. Similarly, the PM threshold can be time dependent. In
this case, the item can fail at or before the PM threshold. Jiang [4] deals with this
kind of case. There, several condition variables are combined into a composite
condition variable using a weighted power model. The functional failure threshold
is represented by a Gaussian process model, and the PM threshold is age dependent.
CBM is applicable for key components of a system and effective only for
predictable failures (with a wear-out failure mechanism). The appropriate use of
CBM can improve system reliability and decrease maintenance costs. However,
CBM requires high initial investment cost and it is technically challenging to turn
the observed condition information into actionable knowledge about health of the
system.
Prognostics and health management (PHM) can be viewed as a systematic CBM
approach for engineering asset health management. It attempts to integrate various
knowledge and available information to optimize system-level maintenance
decision.
17.7 System-Level Preventive Maintenance Policies
In this section, we look at two important system-level PM policies: group main-

tenance and opportunistic maintenance.
The PM tasks and PM intervals of components of a complex system can be very
different. If these are separately implemented, the systems operation will be fre-
quently interrupted. To reduce frequent maintenance interferences, the components
with similar maintenance needs can be grouped into a category to share a common
PM interval. In the meantime, the PM intervals of different categories are set as
integer times of the minimum PM interval so that the PM task with a longer PM
interval and the PM tasks with shorter PM intervals can be simultaneously per-
formed. This is the idea of group maintenance. The key issues for group mainte-
nance include grouping components into categories (groups or packages) and
determining the common PM interval of components in each group.
When a failure occurs, the CM action can be delayed to a later time to combine
with a PM action that is coming soon or the PM action is advanced to undertake if
the CM cannot be delayed. This is the idea of the opportunity maintenance.
Advantages of opportunistic maintenance are that it can further reduce maintenance
interferences while reducing a maintenance setup cost. A key issue for opportunistic
maintenance is to determine an opportunistic maintenance window for each com-
ponent and each component group.
Specic details for these two policies are presented as follows.
17.7.1 Group Preventive Maintenance Policy
For a complex system with many components, group replacement is an effective

maintenance strategy to combine the preventive replacement activities for the dif-
ferent components of the system into packages for execution. The procedure to
determine these preventive replacement packages involves the following two main
steps: grouping components and determining the PM interval of each component
group.
We rst look at the rst step. The number of groups is usually determined based
on experiences. For example, many complex systems implement a three-level PM
regime, implying that the number of groups is three in these cases. The similarity
17.7 System-Level Preventive Maintenance Policies 321
among different components can be measured based on their optimal PM intervals.

According to Ref. [6], the method to determine the PM interval of a component
depends on whether or not the component is safety related. For the safety-related
component, it is determined based on the reliability or risk requirement; for the
other components, it is determined based on the age replacement model.
Let Ti denote the preventive replacement age or PM interval of component i. We
arrange them in an ascending order and denote the ordered PM intervals as
T1 T2 . . . Tn : 17:37
Let xi Ti 1 Ti . Let K denote the number of groups, and Xk

1 k K 1 denote the kth largest value of xi ; 1 i n 1 with
Xk Tik 1 Tik . Let bk Tik Tik 1 =2. Then bk ; 1 k K 1 divide

(Ti ; 1 i n) into K groups. We call the group that contains Tn T1 the rst
[Kth] group. The above approach is graphically illustrated by Fig. 17.1.
We now look at the second step. Let lk denote the mean of individual PM
intervals in the kth group. Clearly, lK \ lK1 \. . . \ l1 . Let
nk intlk =lk1 0:5: 17:38
Let sk denote the common PM interval of components in the kth group. These PM
intervals must meet the following relations:
sk1 nk1 sk ; 2 k K: 17:39
This implies that we just need to determine the value of sK . It can be optimally
determined based on a periodic replacement policy discussed in Sect. 17.4, e.g., the
model given by Eq. (17.28) if minimal repairs are allowed or the model given by
Eq. (17.30) if minimal repairs are not allowed.
Example 17.4 Consider the PM intervals shown in the third column of Table 17.3.
The problem is to divide the components into three groups and to determine the
value of n1 , n2 and s3 .
30
25
20
xi
15
10
5 b2 b1
0
0 10 20 30 40 50 60 70
T (i )
Fig. 17.1 Dividing the components into K 3 groups based on components PM intervals
We rst implement the rst step. Using the approach outlined above, we have
X1 42:7 18:7 24 with b1 30:7 and X2 18:7 4:7 14 with b2 11:7.
As such, the component with Ti b1 belongs the rst group; the component with
Ti b2 belongs the third group; and the other components belong the second
group. As a result, the components in the rst group are (B, P, S); the component in
the second group is (F); and the components in the third group are (A, O).
We now implement the second step. The groups mean of PM intervals are
(l3 ; l2 ; l1 ) = (4.65, 18.7, 52.13). This yields n1 ; n2 = (2.79, 4.02) (3, 4). As such,
the remaining problem is to adjust the value of s3 so that s2 4s3 and
s1 3s2 12s3 . Based on the total cost rate model given by Eq. (17.30), we have
s3 4:6.
The nal PM intervals of the components are shown in the last column of
Table 17.3. As seen, it is almost the same as those in the fth column, obtained
from the block replacement policy for each group.
17.7.2 Multi-level Preventive Maintenance Program
In Example 17.4, we actually deal with a three-level preventive replacement pro-

gram for a simplied example. Generally, a manufacturer needs to develop a multi-
level PM program (or regime) for its product. The PM program will include various
PM tasks for the components and assemblies of the product, and the idea of the
group maintenance plays a key role in the development of such a PM program.
17.7.3 Opportunistic Maintenance Policy
Consider two components (denoted as C1 and C2 , respectively). Figure 17.2 shows

the triggering event and time window for implementing an opportunistic mainte-
nance, where the solid line indicates the scheduled PM time, the box indicates the
opportunistic maintenance window, and sign indicates a triggering event.
We rst look at case (a) in Fig. 17.2. Suppose that component C1 has a failure at
T1 and component C2 is planned to replace at T. An opportunistic replacement
Fig. 17.2 Triggering event

and opportunistic
maintenance window Case (a) Case (b) Case (c)
T1
TL T TR t
17.7 System-Level Preventive Maintenance Policies 323
window given by TL ; TR is set for component C2 . If T1 falls into this window, the
replacement of C2 can be advanced to T1 .
We now look at cases (b) and (c) in the gure. An opportunistic PM window is
set for a group of components. In case (b), a failure triggers an opportunistic PM
action, which can be advanced. In case (c), the PM action cannot be advanced since
the failure time is smaller than the lower limit of the opportunistic window.
However, the CM action may be delayed to the PM opportunistic window to
complete if the delay is allowed.
According to the above discussion, we see that the key problem of opportunistic
maintenance is to set the opportunity maintenance window for key components and
all PM packages. We look at this issue as follows.
An opportunistic maintenance action can save a setup cost for the combined
maintenance actions. But, advancing the replacement time of a component reduces
the useful life of the component; and delaying a CM may have a negative influence
on production.
The opportunistic maintenance window can be derived through adjusting rele-
vant cost parameter. To be simple, we consider the age replacement policy for a
component. For the other cases, the method to determine the opportunistic main-
tenance window is similar but more complex.
Suppose that the preventive and failure replacement costs for a component are cp
and cf , respectively, and its preventive replacement interval T is determined by the
cost model of this policy. Let cs;p denote the setup cost for a preventive replace-
ment. Advancing the PM implies that the setup cost can be saved so that the
preventive replacement cost cp in normal condition is reduced to cp cp cs;p .
This results in an increase in the cost ratio and a decrease in the optimal PM
interval. As such, the optimal PM interval obtained for this case is set as the lower
limit of the opportunistic replacement window.
Similarly, let cs; f denote the setup cost for a failure replacement, which is usually
much larger than cs; p . Delaying a CM implies that the setup cost can be saved so
that failure replacement cost cf in normal condition is reduced to cf cf cs; f .
This can result in a signicant reduction in the cost ratio and an increase in the
optimal PM interval. As such, the optimal PM interval obtained for this case is set
as the upper limit of the opportunistic replacement window. If the downtime loss
must be considered, we can take cf cf cs; f cd Dt, where cd is the loss per
unit downtime and Dt is the expected delay time.
Example 17.5 For the data shown in Table 17.2, assume that cs;p cs; f 0:5cp .
We do not consider the downtime loss. The problem is to nd the opportunistic
replacement windows of those components.
Using the approach outlined above, we have the results shown in Table 17.4,
where w TR TL =T is the relative width of the opportunistic replacement
window. Figure 17.3 shows the plot of w versus b. As seen, a large b allows a small
opportunistic window.
Table 17.4 Opportunistic Component TL TR w

replacement windows of
components A 3.7 4.7 0.2128
B 44.6 64.5 0.3252
F 15.0 19.0 0.2139
O 3.9 4.6 0.1522
P 38.1 54.2 0.3067
S 23.7 52.1 0.6651
Fig. 17.3 Plot of w versus b 0.7

0.6
0.5
0.4
w
0.3
0.2
0.1
0
0 1 2 3 4 5

17.8 A Simple Maintenance Float System
A maintenance float system is characterized by (a) one or more standby or backup

items (equipment or machines) to assure the system reliability, availability, and
required production rate, and (b) a maintenance workshop with a certain mainte-
nance capability in terms of number of maintenance servers (crews or persons).
Two key problems with the maintenance float system are (a) system design to
determine the key parameters such as the number of standby items and the number
of maintenance servers (e.g., see Ref. [1]); and (b) performance evaluation for a
given system conguration (e.g., see Ref. [7]).
Figure 17.4 shows a simple maintenance float system, which is composed of a
working item, a backup item and a repair workshop. The working and backup items
are statistically identical and follow a known life distribution Fx. When the
working item fails, the backup item will replace it to work immediately if available.
In the meantime, the failed item will be repaired as soon as possible. When the
backup item is not available, the system has to wait until the backup item gets
repaired and begins to work. The repair is assumed to be perfect and the time to
repair (denoted as Y) follows a distribution Gy. The system fails when the
working item fails and the backup item is being repaired. The problem is to evaluate
the availability of this system, and sometimes called the machine interference
problem or machine repairman problem (e.g., see Ref. [2]).
A working cycle starts at the time when the current working item begins working
and ends with when the backup item begins to work. The time for the backup item
to work can be immediately after the working item fails or the time when the
17.8 A Simple Maintenance Float System 325
Fig. 17.4 A simple

maintenance float system Backup
item
Working Repair
item workshop
Item to be
repaired
backup item gets repaired. Let X denote the operating time of the working item and
Y denote the repair time of the backup item. The reliability of the system is the
probability of event X [ Y, and can be evaluated using the stress-strength model
(i.e., X is equivalent to strength and Y is equivalent to stress) given by
Z1
R PfX [ Yg 1 FzdGz: 17:40
0
The expected uptime per cycle is given by
Z1
EX 1 Fxdx: 17:41
0
If ignoring the item switch time, the cycle length is given by T maxX; Y.
This implies that T follows the twofold multiplicative model given by Eq. (4.33)
with F1 t being replaced by Fx and F2 t being replaced by Gy. The expected
cycle length is given by
Z1
ET 1 FzGzdz: 17:42
0
As a result, the availability of the system is given by
A EX=ET: 17:43
In complex maintenance float systems, the number of working items, the number
of backup items, or the number of repair workshops can be larger than one. The
items may be subjected to a multi-level PM program. In this case, Monte Carlo
simulation is an appropriate approach to analyze the characteristics of the system.
Table 17.5 Results for ll R EX ET A

Example 17.6
0.5 0.9473 8.8726 9.0049 0.9853
0.4 0.9567 8.8726 8.9729 0.9888
0.3 0.9646 8.8726 8.9482 0.9916
Example 17.6 Assume that Fx is the Weibull distribution with parameters b 2:5
and g 10, and Gy is the lognormal distribution with parameters ll 0:5 and
rl 0:8. Using numerical integration to evaluate the integrals of Eqs. (17.40) and
(17.42), we obtained the results shown in the second row of Table 17.5.
If the system reliability or/and availability is not acceptable, they can be

improved by increasing the maintenance resources to decrease the time to repair.
Assume that ll is decreased to 0.4 and 0.3, respectively, the corresponding per-
formances are shown in the last two rows of Table 17.5. This illustrates the
influence of maintenance on system reliability and availability.
References
1. Chen M, Tseng H (2003) An approach to design of maintenance float systems. Integr Manuf
Syst 14(5):458467
2. Haque L, Armstrong MJ (2007) A survey of the machine interference problem. Eur J Oper Res
179(2):469482
3. Jiang R (2013) A tradeoff BX life and its applications. Reliab Eng Syst Saf 113:16
4. Jiang R (2013) A multivariate CBM model with a random and time-dependent failure threshold.
Reliab Eng Syst Saf 119:178185
Beijing
7. Lopes IS, Leito ALF, Pereira GAB (2007) State probabilities of a float system. J Qual Maint
Eng 13(1):88102
8. Moubray J (1997) Reliability-centered maintenance. Industrial Press Inc, New York
9. Tajiri M, Got F (1992) TPM implementation, a Japanese approach. McGraw-Hill, New York

Introduction To Quality and Reliability Engineering - R-Jiang

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Introduction To Quality and Reliability Engineering - R-Jiang

Hochgeladen von

Copyright:

Verfügbare Formate

Springer Series in Reliability Engineering

Introduction to Quality and

Additional material to this book can be downloaded from http://extras.springer.com.

ISSN 1614-7839 ISSN 2196-999X (electronic)

Jointly published with Science Press, Beijing

Library of Congress Control Number: 2015939148

Springer Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer-Verlag GmbH Berlin Heidelberg is part of Springer Science+Business Media

Manufacturing businesses need to develop new products and improve current

Fig. 1 Smile curve

LIFE CYCLE STAGES

Part I Background Materials

2 Engineering Activities in Product Life Cycle . . . . . . . . . . . . . . . . 11

2.4 Engineering Activities in Post-manufacturing Phase. . . . . . . . . 22

5 Statistical Methods for Lifetime Data Analysis . . . . . . . . . . . . . . . 67

6 Reliability Modeling of Repairable Systems . . . . . . . . . . . . . . . . . 89

Part II Product Quality and Reliability in Pre-manufacturing Phase

7 Product Design and Design for X . . . . . . . . . . . . . . . . . . . . . . . . 113

7.3 Design for Several Overall Performances . . . . . . . . . . . . . . . . 116

8 Design Techniques for Quality. . . . . . . . . . . . . . . . . . . . . . . . . . . 129

9 Design Techniques for Reliability . . . . . . . . . . . . . . . . . . . . . . . . 147

9.4 Reliability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

10 Reliability Testing and Data Analysis . . . . . . . . . . . . . . ....... 169

10.5 Accelerated Degradation Testing Models . . . . . . . . . . . . . . . . 181

11 Reliability Growth Process and Data Analysis . . . . . . . . . . . . . . . 193

Part III Product Quality and Reliability in Manufacturing Phase

12 Product Quality Variations and Control Strategies . . . . . . . . . . . . 221

12.2.1 Variations of Quality Characteristics

13 Quality Control at Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

14 Statistical Process Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

15 Quality Control at Output. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

Part IV Product Quality and Reliability in Post-manufacturing Phase

16 Product Warranty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283

17 Maintenance Decision Optimization . . . . . . . . . . . . . . . . . . . . . . . 303

17.3 Repair-Replacement Policies. . . . . . . . . . . . . . . . . . . . ..... 308

5M1E Materials, Manufacture, Man, Machine, Measurements,

DFM Design for Manufacturability

MTTR Mean Time to Repair

TTFF Time to the First Failure

Manufacturing businesses need to come with new products in order to survive in a

1.2 Product and Product Life Cycle

From a marketing perspective, a product is anything that can be offered to a

Science Press, Beijing and Springer-Verlag Berlin Heidelberg 2015 3

structure. A typical decomposition of a product includes six hierarchical levels,

1.2.2 Product Life Cycle

1.2.3 Technology Life Cycle of a Product

Fig. 1.1 Technology life

1.3 Notions of Reliability and Quality

1.3.1 Product Reliability

Reliability of a product (or system) conveys information about the absence of

1.3.2 Product Quality

1.3.3 Link Between Quality and Reliability

Therefore, reliability is expressed as the proportion of the survived items to the

1.4 Objective, Scope, and Focus of this Book

1.5 Outline of the Book

Chapters 79 deal with product design relevant issues. Chapter 7 discusses

2.2 Engineering Activities in Pre-manufacturing Phase

2.2.1 Main Activities in Front-End Stage

Science Press, Beijing and Springer-Verlag Berlin Heidelberg 2015 11

2.2.2 Main Activities in Design and Development Stage

2.2.2.1 Quality and Reliability Activities in Detail Design Stage

2.2.2.2 Reliability Activities in Development Process