Weibull Analysis Handbook

See
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/235105600
Weibull Analysis Handbook
Article November 1983
CITATIONS READS
35 1,588
4 authors, including:
James Breneman Grant Reinman

Tri-County Technical College Pratt & Whitney
8 PUBLICATIONS 35 CITATIONS 3 PUBLICATIONS 40 CITATIONS
SEE PROFILE SEE PROFILE
All content following this page was uploaded by James Breneman on 27 December 2013.
The user has requested enhancement of the downloaded file.

IN AD-A143 E103
i%
Uf
AFWAL-Tf-83-2073
- 2'1D"-1 \,
Dr. R. 0. Abernethy
J. E. 3reneman
C. H. Medlin
G.L Flainman
Pratt & Whitney Aircraft

Government Products Divzi.on
United Technolojis Corporation
P.O. Box 2C31
West Palm Beach, Florida 3-4Z2
November 1S33
Final Report for Period 1 July 1- 2 to 31 AL;zt 1::3
Approved for Pub!c 1.Cole-e, G;,,,,4,,z ,.. 1 ELECTE_

; JUL181984 "
A ero Prc:!.v*.-;!n Ln"z:7,,c '.4,
A~r Fron CT~-

7 t
Reproduced From
Best Available Copy
ALL *,-.v., t 4: -.- _.,
WU1fAJ *
N,: .;i:I'):.;, . : ::::,.: . ./', , . . . . ... . . . . . .. . . . . . .

NOTICE
When Government drawings, specifications, or other data are used for any purpose other than in
comnection -with a definitely related Government procurement operation. the United States Government
thereby incurs no responsibility nor any obligation whosoever; and the fact that the government may
have formulated, furnished, or in any ,-ay suppiivd the said drawings, specifications, or other data. is
not to be regarded by implication or otherwise as in any manner licensing the holder or any other
person or corporation, or conveying any rights or permission to manufacture use, or sell any patented
invention that may in any way be related thereto.
This report has been reviewed by the Office of Public Affairs (ASD/PA) and is releasable to the
National Technical Information Service (NTIS). At NTIS, it will be available to the general public.
including foreign nations.
This technical report has been reviewed and is approved for publication.
NANCY A. NiITT /ACK RICHENS

Project Engineer Chief, Components Branch
T MTurbine Engine Division
S FOR THE COMMANDER
H. I. BUSH
Director
Turbine Engine Division
Aero Propulsion Laboratory
"I1f your address has changed, if you wish to be removed f m omr mailing ILqt. or if the addressee
is no longer employed by your organization please notify AFWAI./POTC
W-PAFB OH 45433 to help us maintain a current mailing liUs.
Copies of this report should not be returned umless retu*rn is required by security considerations.
contractual obligations, or notice on a specific document.
aI" * i *
SECURITY CLASSIFICATION Of THIS PAGE (Whent Dat. Entered)
REPOT DCUMNTATON
AGEREAD INSTRUCTIONS
REPRTDOUMNTAIO PGEBEFORECOMPLETINGFORM
1. REPORT NUMBER 3.GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBER
* ~~~AFWAI.-TWKI.2(J79 A~4q/D ______________
4. TITLE (and Sw.btillo) S. TYPE Of REPORT & PERIOD COVERED

Final_________
WEIBULL ANALYSIS HANDBOOK .Iulv 1982 to 31 August I1J83

S. PERFORMING ORG. REPORT NUMBER
______________________________________________ P&WA/GlIDl FR- 17579
7. AUTNOR(s) G. CONTRACT OR GRANT NUMBER(@)
Dr. R. B. Abernethy; .1. E. Breneman; 3fl-2C24

C. H. Medlin, G'. L. Reinman
9. PERFORMING ORGANIZATION NAME AND ADDRESS 10. PRC..RAM ELEMENT, PROJECT, TASK
U~nited Technologies Corporation IAREA & WORK UNIT N4UMBERS
Pratt & Whitney Aircraft GDroup 30660649
Chwernment Pri4NIIVIJ Divis~ion
P.O. Box 2%91. West Palm Beach. Fl._ _ _ __ _ _ _
y 11. CON4TROLLING OFFICE NMAE AND ADDRESS

Aero Propulsion LaihOFat(Iy (AFWAL/PO1TC)
I2. REPORT DATE
Nov'ember 1983
* ~~Air Force Wright. Aeronautiald Labo~ratories (AFSC) .NMUROPAE
WrgtPatterson Air Force Base, Ohio 4543:3 228

tMNITORING AGENCY NAME & #4ODRESS(ff Re'Htern~t from Controlling Office) IS. SECURITY CLASS. (ofthi toi)
5
rep.
Unclassified
IS&. DECLASSIFICATION/DOWNGRADING
SCHEDULE
16. DIST RIBUTION STATEMENT (l this Rope"
Approved for public release, distribution unlimited
17. DISTRIBUTION STATEMENT (of th abstract eatershi Bl.ck 20. if l'nU bronikofp~eft)
IS. SUPPLEMENTARY NOTES
19. REY WORDS (Coothbue an roeres. *fee it rM4e....r MWdIdes!4', IV block member)
Weiluall. Weihlmyes, Risk Analysis, Reliability Testing, Thorndike (Charts, Fatigue Testing. Life
Testing. Expornential D~istribution
' ABSTRACT (Gintms~amen

RI~~2. eese. Old Un .e..ooryp i daj) ~ lhan.
This handbook is intended to provide instructions or. how to do Weibull analysis. It will provide an]
understanding of Weibull analysis that is common between the military and industry. The handbook
U. contains seven chapters plus an appendix. The chapters are written containing a minimum of
mathematics with proofs given in the appendix.
DD 1473 90ITIOw OF I NOV 6SI08 11SOLET19 UNCLASSIFIED

8/0 0102-014-4"
d1
SECURITY CLASSIFICATION OF THIS PAGE (When. Date FAIMrn4
FOREWORD
Advanced Weibull methods have been developed at Pratt & Whitney Aircraft in a joint
effort between the Governmetit Products Division and the Commercial Products Division.
Although these methods have been used in aircraft engine projects in both Divisions. the
advanced technologies have never been published, even though they have been presented and
used by the LT. S. Air Force (WPAFB), U. S. Navy (NAVAIR) and several component
manufacturers.
The authors would like to acknowledge the contribution to this work made by several other
Pratt & Whitney Aircraft employees: D. E. Andress, F. E. l)auser, J1. W. Grdenick,
J. H. Isiminger, B. J. Kracunas, R. Morin, M. E. Obernesser, M. A. Proschan, and
B. G. Ringhiser.
The key Air Force personnel that encouraged publication were: Gary Adams, Dr. Tom
Curran, Jim Day, Bill Troha and Don Zabierek (the USAF Program Manager), all at W1'AFB.
The following members of the Ameriean Institute of Aeronautics and Astronautics

Systems Effectiveness and Safety Committee provided valuable constructive reviews for which
the authors are indebed: M. Berssenbrugge, R. Cosgrove, P. Dick, T. P. Enright, J. F. Kent, L.
Knight, T. Prasinos, B. F. Shelley, and K. L Wong.
The authors would be pleased to review constructive comments for future revisions.
I
For
SAecess~n
TTS G;RA&-l
OU'D@ .1:~TAB
iU 'ouxiced
if VaIc
codes
IlWI 'n1)t1!VtY
LAvn
1
S~LU4a
TABLE OF CONTENTS
C(haptrr Page
I . INTI'IDOI))('i'ION TO W' Il'II,.I, ANALYSIS ................................ I
1.1 Ohje 'tiiv,. ............................................ ....................... I

1.2 Background .. .............................................
1.3 Examples .......................... ......................
.F 1.4 Scope ....................................................................... 2
1.5 Advantages of Weibull Analysis ........................................ 2
1.6 Aging Time or Cycles ....................................... 5
L.7 Failure Distribution ............................................... ........ .
1.8 Risk Predictions ............................................................. 5
1.9 Engineering Changes and Maintenance Plan Evaluation ......... 5
1.10 Mathematical Models ..................... _.; .... 1. 7
1.11 Weibulls with Cusps or Curves .............................. 7
1.12 System Weibulls .......................................... 10
1.13 No-Failure Weibulls ....... ................................ It)
1.14 Small Failure Sample Weibulls .............. ................. 10
1.15 Changing Weibulis 10......................................
1.16 Establishing the Weibull Line .......................................... 13
1.17 Summary ............................................... 13
"2 PERFORMING A WEIBULL ANALYSIS ..................................... 14
2.1 Foreword ...................................................................... 14

2.2 Weibull Paper and Its Construction ................................... 14
2.3 Failure Data Analysis - Exemple ..................................... 17
2.4 Stspended Test Items - Nonfailures ........... :............... IP
2.5 Weibull Curve Interpretation ............................................ 22
2.6 Data Inconsistencies and Muhtimode Failures .............. ........ 24
2.7 L[w-Time Failures ......................................................... 24
2.8 Close Serial Numbers ..................................... 27
2.9 11)gleg nd.............................................. 27
2.10 Curved Weibulls ......................................... 32
2.11 Problems ....................................................................... 35
3 WEIBULL RISK AND FOREiCAST ANALYSIS ............................. 38
3.1 Foreword ............... ................................. 38

3.2 Risk Analysis Defi~iition ............... ..................... 38
3.3 Forecasting Techniques ................................................... 38
3.4 Calculating Risk ............................................................. .38
3.5 Present Risk ............................................ 38
3.6 Future Risk When Failed Units Are Not Fixed .................... 39
3.7 Future Risk When Failed Units Are Repaired ...................... 40
3.8 The Use of Simulation in Risk Analysis .............................. 40
3.9 Case Studieb ............................................................... . 43
3.10 Case 93tudy 1: Bearing Cage Fracture .......................... 43
3.11 Case Study 2: Bleed System Failures ................................. 44
3.12 Case Study 3: System Risk Analysis Utilizing a Simulation
M odel .......................................................................... 52
3.13 Problems ......................................... ..................... 60
Ni
.N TABLE OF CONTENTS (Continued)
4 WEIIIAYIS WtIEN WEIIIJLIS ARE IMPOSSIIlIE ................. 67
4.1A Foreword .............. 67

4.2 W eibayes M ethod .................... ................................... 67
4.3 Weibayes - No Failures ................................................. 67
4.4 Weibest - No F-ilure...................... ........ 68
4.5 Unknown Failure Times .............................................. 68
4.6 Weibayes Worries and Concerns ........................................ . 68
4.7 Examples of Problems/Analytical Solutions ................... 68
4.8 Problem s ........................................................................ 75
I 5 SUBSTANTIATION AND RELIABILITY TESTING ......................
5.1
5.2
5.3
Foreword ........................................................................
Zero-Failure Test Plans for Substantiation 'resting ................
Zero-Failure Test Plans for Reliability Testing .....................
77
77
77
80
5.4 Total Test Time ...... .......................................... 87
5.5 Advantages and Limitations of the Zero-Failure Test Plans . ,87
5.6 Non-Zero-Failure Test Plans ............................................. 88
5.7 Designing Test Plans ...................................................... 90
5.8 Recommended Method oor Solving Equations ....... ............ . 92
5.9 Problems ......................... ................. 92
6 CASE HISTORIES WITH WEIBULL APPLICATIONS .................. 94
6.1 FOREWORD ............................................. 94

6.2 Example 1: Turbopump Bearing Failures ............................. 94
6.3 Initial Analysis -7 Small Sample ............................. 94
6.4 Two Months Later - Batch Identified ............................... 96
6-..
5 . ..:lisk Prediction ........... ............. ...... ..-- -- -
6.6 Four Months Later - Final Weibull Plot ........................... 99
6.7 Example 2: Main Gearbox Housing Cracks .......................... 99
6.8 Information Available for Analysis ...................................... 99
6.9 Risk Anal)'is ................................................................. 102
6.10 Determining the Fix ........................................................ 103
6.11 How Good 'Were the Forecasts? ............ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.12 Example 3" Opportunistic Maintenance Screening Intervals 103
6.13 Structuring the Problem .4...................................
6.14 Finding the Optimum Interva .......................................... 104
6.15 Example 4: Support Cost Model ........................................ .104
6.16 Role of the Weibuil ........................................................ 106
6.17 Example 5: Vane and Case Field Cracks ............................. 107
6.18 Resolving the Questions ................ .................... 1
107
6.19 Concluding Remarks ...................................... 107
vi
-rrrr~r .**.,.*,*- * .g .w r .-.
r>f -i,.
TABLE OF CONTENTS (Continued)
Chapter Page
%; 7 CONFIDFNCE LIMITS AND OTHER ASPECTS OF THE 1

W EIBU I.[. ............................................... ................
7.1 Foreword ........................... It

7.2 Confidence Intervals ........................................................ 110
7.3 Confidence Intervals for 1 and 'p .......... .......................... 110
7.4 Confidence Intervals for Reliability ..................................... 113
h
7.5 Confidence Intervals About a Failure Time ......................... 113
7.6 Confidence Bands on the Weibull Line .............................. 114
7.7 Weibull "Thorndike" Charts ................................ 116
7.8 Shifting W eibulls ............................................................ 128
7.9 Weibull Goodness of Fit ................................... 129
7.10 Comparing the Weibull to Other Distributions .................. 129
7.11 PROBLEM S .................................................................. 113
APPENDIX A - Glossary ....................................... 141

APPENDIX B - Median Ranks, 5"; Ranks, and 95', Ranks .......... 143
APPENDIX C - Rank Regression (Weibull Plot) Method of Weibull

Analysis .. .................... ...................................................... 155
1. Method .......................................... 155
2. Example and Step-by-Step Procedure ......................... 155
APPENDIX D - Maximum Likelihood Method of Weibull Analysis . 159
1. Foreword ....................................................................... 159

2. The Likelihood Function ................................... 159
3. Maximizing the Likelihood Function ................................... 160
4. Example ................................................ 161
AP'iNDIX E - Weibayes Methods ............................... 16:1
I. Foreword 163
i
2. Derivation of the Weibayes Equation ................................ . 163
APP NDIX F - Monte Carlo Simulation Study - Acruracy of

Weib 11 Analysis Methods ................. ....................... 166
1. Foreword ............................................... 166

2. ] sBeta(Fo)Estimates ........................................................ 168
3. "B.1 Life Estimates .......................................................... 168
4. Eta (qL)Estimates ......................................................... 17:68
4. Risk Forecasts ................................................................ 176
5.1 Risk Forecast Accuracy .................................................... 176
5.2 Risk Forpcast Precision .................................................... 178
6. M onte Carlo Simulation ................................................... 179
vii
-......... ........... "-

TABLE OF CONTENTS (Continued)
('hooter Pae'e
A1P'IE.NI)IX (G Rank IMetrehosidn h~d vsvt Maximmn ILikv-!i~h.0

"M th d of \ ei u l m i .,...................... ............. I........ .... ......
A1'PENI)IX H -- \Veilull Parameter ;,inat ion Computer Pr,)gjamrns 18.1
* AP'PENDIX i -- W eiuill (;raphs ................ ........................ .. .195

I.-t
A'PPENDIX .J -- ANSWERS TO IP.)ILEMS .............................. 199
I. Chapter I A nsw er.; .......................................................... . 199

2. C hapler 2 A nswers ..................................................... 199
3. Chapter :1 Aaswers .......................................................... 206
4. C hapive 4 A nsw ers ................................. . . ............. ... .. 2
5. Chaipter .r Answers ....................................................... 220
4.1. Chapter 6 Answer . ........................................................ 222
7. Chapter 7 An:swer ....................................................... 222
Ii
Vilii
a im
LIST OF ILLUSTRATIONS
AMU" Page
1.1 lPerforming a Weilull Analysis ..
1.2 Understanding the Results . 4
1.:" 11) Louver Craicking ...................... ............ 6
1.4 To Correction to Curved Weibull ..................................... .......... 8
1.5 Mixing Failure Modes, MFP ....................................................... 9
1.6 Mixing Failure Modes, EEC ....................................... 11
1.7 Small Sample Beta Estimates Are Too Steep ....... ........ ............ 12
2.1 Construction of Weibull Paper ..................................... 16
2.2 Rivet Failures ................................................. 19
2.3 Rivet Failures With Suspensions .................................................. 23
2.4 Failure Dintribution Charecteristic ................................. 25
2.5 M ain Oil Pumps ....................................................................... . 26
2.6 Weibull Plot for Augmentor Pump .............................................. 28
2.7 Compressor Start Bleed System ......................................... .......... 29
2.8 Compressor Start Bleed System, Excluding Ocean Base ................... . 30
2.9 Compressor Start Bleed System, Ocean Base ................................. 31
2.10 A Curved Weibull Needs to Correction .............................. 33
2.11 Plotting to Correction ......................................
2.12 t. Correction Applied ............................................ 36
3.1 Risk Analysis With Weibulls ...................................... 41
3.2 Simulation Logic - First Pass ................................................... 42
3.3 Simulation Logic - Second Pass ................................................. 42
3.4 Bearing Population .................................................................... 45
3.5 Bearing Cage Fracture .......................................... 46
3.6 Bleed System Failure Distribution Excluding Air Base D ................... 48
ix
--.---
N I %
UST OF ILLUSTRATIONS (Continued)
:1.7 Bleed System Failure Distribution At Air Bas.e 1) ........................... 49
3.8 Bleed System Population - Air Base D ...................................... 50
3.9 Failure Distribution Input to Simulation ....................................... 53
:1.10 Risk Analysis Simulation Outline ................................................. 55
3.11 Module Population ................................................................. h57
:1.12 Risk Analysis Comparison ........................................................... . 61
:1.1.1 Overail Population .................................................................... . 63
:1.14 Location A Population 64
.015 Control Population ................................................................... C5
4.; Compressor Vane and Case ......................................................... 69
4.2 Weibayes Evaluation of New Design in Accelerated Test .................. 71
4.3 Weibull Evaluation of B/M Design ............................
4.4 Weilayes Evaluation of New Design in Accelerated Test .................. 73
4.5 Hydraulic Pump Failures ........................................................... 74
1.6 When Shoul2' We Pull the Suspect Batch? .................................... 76
5.A B ll and Roller Bearing Unbalance Distribution ............................. 78
5.2 Weibull Plot for 0.99 Reliability at 1000 hr ................................... . 81
5.3 Illustration of B.1, BI, and BI0 Lives ......................................... 82
5.4 Illustration of 7000 hr BIO Life ................................................... 83
5.5 A Reliability of 0.99 at 1800 Cycles Is Equivalent to an 8340.9 Cycle

Characteristic Life When 0=3 ............................... ...................... 85
l,6 A 2000 hr B10 Life Is Equivalent to a 6161.6 hr Characteristic Life

When 0=2 .............................................................................. 86
5.7 Probability of Passing the Zero-Failure Tests for X's Between

0.5 and 5 ..................................................... 89
6.1 Weibull Plot for Augmentor Pump Bearing .................................... 9.5
II
it s .- ..
7.- -
LIST OF ILLUSTRATIONS (Continued)
Figure Page
6.2 Augmentor 650 on Up ............................................................... 97
6.3 Ri.ik Analysis .......................................................................... 98
6.4 Weibull Ploi for Augmentor Pump ................................. 100
6.5 Main Gearbox Housing Cracks .................................................... 101
6.6 Cumulative Main Gearbox Housing Cracks .................................... 102
0.7 Approach To Optimizing Scheduled Maintenarnce ............................ 105
6.8 Sheduled and Unscheduled Maintenance Interaction .................. 10,5
6.9 Unscheduled Maintenance Input via Weibulls ................................ 106
6.10 1th Vane and Case Cracking ..................................................... 108
6.11 Expected UER's Due to 12th Vane and Case Cracking .................... 109
7.1 Weibull Test Case .............................................. 112
7.2 Weibill Plot Where/t = 2.0 and q 100 for 10 Failures ............... 115
7.3. Example of Confidence Bands on a Weibull Line ........................... 117
7.4 Cumulative Sums of Poisson (Thorndike Chart) .............................. 118
7.5 Weibull Thorndike Chert for ffi 0.50 .......................................... 120
7.6 Weibull Thorndike Chart for = 1.0

. ............................... 121
7.7 Weibull Thorndike Chart for ji = 1.5 ............................... 122
7.8 Weibull Thorndike Chart for t = 2.3 ........................................... 123
7.9 Weibull Thornd4-e Chart for = 2.5 ...........................................

ff 124
7.10 Weibull Thorndike Chart for / - 3.0 ........................................... 125
7.11 Weibull Thorndike Chart for - 4.0 ........................................... 126
7.12 Weibull Thorndike Chart for/ = 5.0 ............................... 127
7.13 Example of Shifting a Weibull .................................... 130
7.14 Flange Cracking ................................................ 131
7.15 Flange Cracking Estimated DNstribution ........................................ 132
xi
7.16 Hyptothesized Weihull ............................................................... I:I.
7.17 Picking the Best Distribution ..................................................... !34
7.18 Comparing Weibull and Log-normalDistributions for Coverpiate

Failures ................................................. ................................. .. 135
7.19 Weibull of Nonserialized Parts, Problem No. 5 .............................. 138
7.20 True Weibull, Problem No. 6 ...................................................... 139
7.21 Suspect Failure Mode, Problem No. 6 ....................................... 140
('.I Example Data With Least Squares Line ........................................ 158
1.'.1 Beta I"stimates: True Beta :3, Five Failures .............. .............. 167
F.2 W-1. Aecmacy ................................................. 169
.3 Befai Precision ......................................................................... [70
F.4 B.l Life Affuracy ..................................................................... 171
F.5 B.A Life Precision ........................... 172

2.........................................
F.6 Eta Accuracy Median Characteristics Life Estimates - A ................ 174
F.7 Eta Accuracy - Median. Characteristics Life Estimates - B ........... 175
F.8 Monte Carlo Simulotion Procedure ............................................... 181
F,'.9 Initial Conditions of Simulator .................................................... 182
.1.1 Problem 2.1 .................................................................. ........ 2200
.1.2 Problem 2-2, Infant Mortality ..................................... 202
.A1. Problem 2-3, Curve-. Weibull ..................................... 204
J.4 Problem 2-3, Overall Pnpulation ................................... 205
.1.5 Problem 3.2, Overall Population .................................................. 208
.1.6 Problem 3-2, Location A Only .... 209
J.7 Problem 3-2 ....................................................... ............... 2
J.8 Problem 7-15 ............................................................................ 225
xii
Figu re" Page
I
- .1.9 I'ruh'mb 7-5 .................... .................................... 226
I
.1.10 IProbhl m 7-1; ............................................................................. 227
.1.11 i P rolermC~l 7 6 ............................... ............................................. 228
I/
*1
LIST OF TABLES
Table Page
1.1 Wt'ibull Risk Forcs .............................. t.... 7
2.1 Coist ructi~on of Ordinate (Y) .............................................. t5
2.2 Construction of Abscissa W( .................................................. 15
2.A Baseline .................................................................... 17.
2.4 Weibull Coordinates........................................................ 18
2.5 Adjusted Rank ...... i....................................................... 290
'42.6 Median Rank ............................................................. 2
3.1 Present Risk................................................................ 39
3.2 Future Risk................................................................. 42
:3.3 Hearing Risk After7 12 Months ............................................. 47
3.4 Bleed System Failures by Air Base........................................ 47
3.5 Bleed System Risk After 18 Months ....................................... 51
:1.6 Bleed System Risk After 18 Months....................................... 52
:3.7 Simulation Output for 1000 Hour Inspection .............................. 58
:3.8 Simulation Output for 1200 Hour Inspection .............................. 59
:1.9 Table of Uniform Random Numbers from 0. to 1.0........................ 66
5.1. Characteristic Life Multipliers for Zero-Failure Test Plans Confidence

Level: 0.90 .................................................................. 79
* 5.2 Required Sample Sizes for Zero-Failure Test Plans Confidence Levek
0.90 ......................................................................... 80
6.1 Projected Pump Failures ................................................... 98
6.2 Follow-Up Analysis Results ................................................ 103
7.1 Confidence Levels.......................................................... 110
7.2 Confidence Bounds on the Weibull Line .................................. 116
7.3 Critical Values for Testing the Difference Between Log Normal and
Weibull (Favoring the Weibuill) ............................................ 136
xiv
LUST OF TABLES t. Untinued)
Table Page
7.4 Critical Values for Testing the Difference Between Log Normal and
Weibull (Favoring the Log-Normal) ............................................... 1:36
B.1 M edian Ranks .......................................................................... 143
B.2 . Five Perce..t Ranks ............ ... .................................. 147
B.3 Ninety-Five Percent Ranks .......................................................... 151
F.i Standard Deviations of the Characteristic Life Estimates .................. 173
F.2 Risk Forecast Accuracy 17.........................................
F.A Risk Forecast Standard Deviations . 178
i".
FA Componcnts of1 Monte Carlo Simulator ......................................... 18(0
xv
A.
.a *fe t.
Sil .a S.r.. S.. . '. ". " ..
CHAPTER 1
INTRODUCTICN TO WEIBULL ANALYSIS
1.1 OBJECTIVE
The objective of this handbook is to provide an understanding of both the standard and
advanced Weibull techniques that have been developed for failure analysis. The authors intend
that their presentation bee sech that a novice engineer can perform Weibull analysis after
studying this document.
1.2 BACKGROUND
Waloddi Weibull delivered his hallmark paper on this subjectl in 1951. He claimed that
his distribution, or more specifically his family of distributions, applied to a wide range of
problems. He illustrated this point with seven examples ranging from the yield strength of steel
to the size of adult males born in the British Isles. He clairied that the function "-.may
sometimes render good service". He did not claim that it always worked or even that it was
always the best choice.
Time has shown that Waloddi Weibull was correct in all of those statements and
particularly within the aerospace industry. The initial reaction to his paper in the 1950's and
even the early 1960's was negative, varying from skepticism to outright rejection. Only after
pioneers in the field experimented with the method and verified its wide application did it
become popular. Today it has many applications in many industries and in particular the
aerospace industry. There are special problems in aerospace and unusual arrays of data. Special
methods had to be developed to apply the Weibull distribution. The authors believe there is a
need for a standard reference for these newer methods as applied within the aerospace industry,
and to industry in general.
1A EXAMPLES
The following are examples of aerospace problems that may be solved with Weibull
analysis. It is the intent of this document to illustrate how to answer these and many similar
questions through Weibull analysis.
* A project engineer reports three failures of his component in service

operations in a'six week period. Questions asked by the Program Manager
are, "How many failures are predicted for the next three months, six
months and one year?"
* "To order spare parts that may have a two to three year lead time, how
* may the number of engine modules that will be returned to a depot be
forecast for three to five years hence month by month?"
* "What effect on maintainability support costs would the addition of the

new split compressor case feature have relative to a full case?"
* "If the new Engineering Change eliminates an existing failure mode, how
many units must be tested for how many hours without any failures to
demonstrate with 90% confidence that the old failure mode has either
"beeneliminated or significantly improved?"
W92bu7l. Waloddi (1951). A Statistical Distibution Function of Wide Applicability. Joural ofApplied fechanicu. pg.
293-297.
P. 1A SCOPE
As treated herein. Weibull analysis application to failure analysis includes:
0 Plottirg the data

" Interpreting the plot
Predicting future failures
* Evaluating various plans for corrective actions
* Substantiating engirneering changes that correct failure modes.
Data problems and deficiencies are discussed with recommendations to overcome

deficiencies such as:
Censored data
* Mixtures of failure modes
* Nonzero time origin (to cor. action)
No failures
Extremely small samples
* Strengths ,und weaknesses of he method.
Statistical and mathematical derivationi, are presented in Appendices to supplement the

main body of the handbook. There are brie, discussions of alternative distributions such as the
log normal. Actual case studies of aircraft engine problems are used for illustration. Where
problems are presented for the reader to solve, answers are supplied. The use of Weibull
distributions in mathematical models and simulations is also dOscribed.
1.5 ADVANTAGES OF WEIBULL ANALYSIS
One advantage of Weibull analysis is that it provides i simple graphical solution. The
process consists of plotting a curve and analyzing it. (Figure U.1). The horizontal scale is some --
measure of life, perhaps start/stop cycles, operating time, or as tu:bine engine mission cycles.
The vertical Peale is the probability of the occurrence of the 4vent. The slope of the line (ti) is
particularly significant and may provide a clue to the physics of the failure in question. The
relationrhip between various values of the slope and typical failure modes is shown in Figure 1.2.
This type of analysis relating the slope to possible failure modes can be expanded by inspecting
libraries of past Weibull curves.
Another advantage of Weibull analysis is that it may be useful even with inadequacies in
the data, as will be irdicated later in the section. For example, the technique works with small
samples. Methods will be described for identifying mixtures of failures, classes or modes,
problems with the origin being at other than zero time, investigations of alternative. scales other
than time, non-seriamized parts and ccmponents where the time on the part cannot be clearly
identified, and even the construction of a Weibull curve when there are no failures at all, only
success data.
In addition, as there are only a few alternatives to the Weibull, it is not difficult to make
graphic comparisons to determine which distribution best fits the data. Further, if there is
engineering evidence supporting another distribution, this should be considered and weighted
heavily against the Weibull. However, it has been the writers' experience that the Weibull
distribution most frequently provides the best fit of the type of data experienced in the gas
turbine industry.
II
.. - *ri-
-% .1
II I
bt~ a-
.iI
$1 "
" ', *N ( Z,
a
CLI
C 0 0
0 0r 0 0 A
E aEE v E-a C
ca=
0 0 CO() cc &
w. 'a 0 0 0
0ro w..c..l
401- L .
CP
1.6 AGING TIME OR CYCLES
Most applications of Weibull analysis are based on a single failure class or mode from a
single pt.,t or component. An ideal application would consist. of a sample of 20 to 30 failures.
Except for material characterization laboratory tests, ideal data are rare; usually the analysis is
staited with a few failures embedded in a large number of successful, unfailed or censored units.
The age of each part is required. The units of age depend on the part usage and the failure
F mode. For example, low and high cycle fatigue may produce cracks leading to rupture. The age
units would be fatigue cycles. The age unit Gf a jet starter may be the number of engine starts.
Burner and turbine parts -nay fail as a function of time at high temperature or as the number of
excursions from cold to hot and return. In most cases, knowkdge of the physics-of-failure will
"provide the age scale Whmn the units ot age are unknown, several age scales must be tried to
determine the best fit.
"1.7 FAILURE DISTRIBUTION
The first use of the Weibull plot will be to determine the parameter d, which is known as
the slope, or ;hmne parameter. Beta determines which member of the family of Weibull failure
distributions best fits or describes the data. The failure mode may be any one of the types
L. represented by the familiar reliability bathtub curve, infant mortality with slopes less than one,
random with slopes of one, and wearout with slopes greater than one. See Figure 1.2. The
Weibull plot is also inspected to determine the onset of the failure. For eximple, it may be of
interest to determine the time at which 1"( of the population will have failed. This is called BI
"life.Akernativeiy, it may be of int arest in determining the time at which one tenth of I " of the
population will have failed, which is called B.1 life. These values can be read from the curve by
- - inspection. See Figure 1.3.
1.8 RISK PREDICTIONS
I ;If the failure occurred in service operations, the responsible engineer will be interested in a
prediction of the number of failures that might be expected over the next three m,.tlths. six
months, a year, or two years. Methods for making these predictions are treated in Chapter ". A
"ty.icalrisk prediction is shown in Table 1.1. This p.-ocess may provide information on .dhether
'ornot the failure mode applies to the entire fleet or to only- one portion of t1e fleet, which is
often called a batch. After the responsible engineer develops alternative plans for corrective
action. including production rates and retrofit dates, the risk predictions wil be repeated. 'Me
decision maker will require these risk predictions in order to select the best course of action.
". 1.9 ENGINEERING CHANGES AND MAINTENANCE PLAN EVALUATION
VWeibull analysis is used to evaluate engineering changes as to their effect (n the entire
. fleet of engines. Maintenance schedules and plans are also evaluated using Weibull analysis.
These techniques are illustrated in Chapter 6 - Case Histories with Weibul! Applications. In
each case the baseline Weibull analysis is conducted without the engineering change or
maintenance change. The study is then repeated with the estimated effect of the change
mo)difying the Weibull curve. The difference in the two risk predictions represents the net effect
(if the change. The risk parameters may be the predicted number of failures, life cycle cost,
depot loading, spare parts unage, hazard rate, or aircraft availability.
L'
L:
F 5
-6"
; "- / .. ,,
PL
5.

"- . -' -k
- .
EIBULL DISTRIBUTION
.
0I =5 .544.
1886.037
SAMPLE SIZE = 11
FAILURES = 11 -
_ 4-
a. .- _
.....
Q..-
4 . - -III
"
'I. .
4%, . t
4 C ao.
IT ? ~ * . .... .. . i
TOTAL OPERATING CYCLES
Fgsre 1 ID Loouver Cracking
6
TABLE 1.1. WEIBULl RISK
FORECAST N
Risk Pred.ction 'for. 12 Months

Beginning July 1978
11.77 0.00 more failures in 0 m nths
15.12 3.35 more failures in I month
19.18 7.41 more failures in 2 months
53.68 41.91 more failures in 7 taonths
75.53 6.3.76 more failures in 9 months
102.42 90.65 more failures in 1i months
What if? Corrective action next month, next

._-@
! 1.10 MATHEMATICAL MODELS
Mathematical models of an ertire engine system including its control system mayy be
producet by combining the effects of several hundred failure modes. The (ombination may be
done by Monte Carlo simulation or by analytical methods. These models have been use-ful for
predicting spare parts usage, availability, module returns to depot, and maintainability support
coda. Generally, these models are updated with the latest Weibulls once or twice a year and ./
predictions an renerated for review.
1011 WEIBULLS WITH CUSPS OR CURVES

The Weibull plot should be inspected to determine how well the failure data fit the
straight line. The scatter should be evenly distributed about the line. However. sometimes the
failure points will not fall on a straight line on the Weibull plot, and modification of the simple
Weibull approach may be required. The bad fit may relate to the physicsof the failure or to the
quality of the data. There are at least two reasons why a bad fit may occur. First, the origin - If
the points fall on gf',-le curves, it may be that the origin of the age scale is not located at zero
See Figure 1.4. There may be physical reasons why this will be true. For example, with roller
bearing unbalance, it may take a minimum amount of time for the wobbling roller to destroy the
L cage. This would lead to an origin correction equal to the minimum time. The origin correction
may be either positive or negative. A procedure for determining the origin correction is given in
Chapter 2.
Second, a mixture of failure modes -- Sometimes the plot of the failure points will showy
cusps in sharp comers. This is an indication that there is more than one failure mode, i.e. a
mixture of failure modes. See Figure 1.5. In this case it is necessary to conduct a laborator-'
failure analyr'a of each failure to determine if separate failure modes are present. If this is found
to be the case, then separate Weibull plots are made for each set of data for each failure mode. If
the laboratory analysis successfully categorized thejfailures into separate failure modes, the
separate Weibull plots will show straight line fits, that is, very little data scatter. On each plot
"thefailure data points from the other failure modes are treated as successful (censored or non-
failure) units.
S: t I . ,7
/
IIBULL DISTR IBUT ION -"--

WEL "
0 = 1.330 ("'
I = W.2`52'
m..pL SIZE = 261-
LURES
Im. 71
i
. * t "t " ... ..
00 e
__________T =I-
F MM
CURVE
--
* 0
Fiur 1.4 to Cuve We
T. Corecio
, \,
... .. ... 6.7..111.131.0
TOTAL OPERATING TIME (t4RS)

-- - - - - - - - - - - - - - - -- --. ,..--. -.. . .-.-- .- , w o
. EIBULL
S=1.611DISTRIBUTION 4 - -
. LESIZE = 676 S=
MP 1160.911 0(
.- FAILLUES = 41
.. [---
U. - -7. . - ... ........... - ,
as-
U. . ....... J, .j : .~ 6...O
M /u
Fiw .5. Mxn Fa --~ ModaM "
/ 'I
1 A
1.12 SYSTEM WEIBULLS
If the data from a system such as a jet engine are not adequate to plot individual failure
modes, it is tempting to plot a single Weibull for the system based on mean-time-between-
, failuires (MFBF), assuming 4 1. This approach is fraught with difficulties and should be
avoided if possible. However. there may he no alternative if the system doq not have serialized
*! part identification or the data do not identify the type of failure for each failure time. Some
years ago it was popular to produce system Weibulls for the useful life period (Figure 1.6)
assuming constant failure rate (4 = 1.0). Electronic systems that do not have wearout modes
were often analyzed in this manner. More recently, some studies indicate electronics may have a
decreasing failure rate, i.e. a # of less than one.' Although data deficiencies may force the use of
system Weibull analysis, a math model combining individual Weibull modes is preferred
because it will be more useful and accurate.
1.13 NO-FAILURE WEIBULLS
In some cases, there is a need for a Weibull plot even when no failures have occurred. For
example, if an engineering change or a maintenance plan modification is made to correct a
failure mode experienced in service, how much success time is required before it can be stated
(with some level of confidence) that the problem has been corrected. When parts approach or
exceed their predicted design life, it may be possible to extend their predicted life by
constructing a Weibull for evaluati3n even though no failures have occurred. A method called
Weibayes 3nalysis has been developed for this purpose and is presented in Chapter 4. Methods
to design experiments to substantiate new designs using Weibayes theory are presented in
Chapter 5 -- Substantiation and Reliability Testing.
1.14 SMALL FAILURE SAMPLE WEIBULLS
Flight saiety considerations may require using samples as small as two or three units.
Weibull analysis, like any statistical analysis, is less precise with ;mall samples. To evaluate
t Iese small-sample prol)Iems, extensive Monte Carlo and analytical studies have been; made and
will be presented in Appendix F. In general, small sample estimates of tend to be too high (or
sleep) and the characteristic life, V7,
tends to be low. See Figure 1.7.
1.15 CHANGING WEIBULLS
After the initial Weibull plot is made, later plots will be based on larger failure samples
and more time on successful units. Each plot will be slightly different, but gradually the Weibull
parameters will stabilize as the data sample increases. The important inferences about B.1 life
and the risk predictions are that they should not change significantly with a moderate size
sample.
""Unified Field (Failure) Theory-Demise of the Bathtub Curve", Kam LiWong, 1981 Proceedings Annual Reliability
and Maintainability Symposium.
10
tla0N
. EIBLLL DISTRIBUTION
= 1.330
=696.4S68
S._LESIZE = 188G
FAILURES = 1880
W.- -
I
30..--
0.
0 12S
0Zi l *4 431.00. 1. 4 . 6. 00.-

low . 4. 6-1s..m.sao.
TOTRL OPERATING TIME (H-R.)
FO 272262
Figure 1.6. Mixing Failure Modes, EEC
11
'C. O
C13
LL
LO0
A CCO
0 ~ ~ ~ ~ .a coc0 U 'j V
> cn)
0212
1.16 ESTABLISHING THE WEIBULL UNE
The standard approach for constructing Weibull plots is to plot the time-to-failure data on
Weibull probability graphs using median rank plotting positions as described in Chapter 2. A
straight line is then fit to the data to obtain estimates of 0 and q/. This approach, has some
deficiencies as noted above for small samples I is simple and graphical. Maximum likelihood
estimates may be more accurate, but require ci.-.iAex computer routines. The advantages and
disadvantages of these methods are discussed in Appendices C and D.
1.17 SUMMARY
The authors' intent is that the material in this handbook will provide an understanding of
this valuable tool for aerospace engineers in industry and Government. Constructive comments
would be appreciated for future revisions of this handbook.
v--
131
/ *
I- /
':7-- * 1
4''
CHAPTER 2
PERFORMING A WEIBULL ANALYSIS
2.1 FOREWORD
This section describes how to construct Weibull paper and how to plot the data. Since
interpretation of the data is the most important part of doing an analysis, an extensive
discussion is given on how to interpret a Weibull plot. Examples are used to illustrate
interpretation problems.
The first question to be answered is whether or not the data can be described by a Weibull
distribution. If the data plots on a straight line on Weibull paper, the data can be approximated
by a Weibull distribution.
2.2 WEIBULL PAPER AND ITS CONSTRUCTION
11e Weibull distribution may be defined mathematically as follows:
F(t) -I -
where:
F(t) fraction failing

t = failure time
to : starting point or origin of the distribution
= characteristic life or scale parameter
$ = slope or shape parameter
e exponential.
F(t) thus defines !he cumulative fraction of a group of parts which will fail by a time t.
Therefore, the fraction of parts which have not failed u,) to time t is I - F(t). This is often
called reliability at time t and is denoted by R(t). By reaimnging the distribution function, the
following can be noted:
I - F(t) e - (a -- /
let t, -0
then
S-- F(t) e- -
14
/
/
V
,. . / ~V ,-,j~ ~ ~~./
:.
S1I F(t)
G t Y~
Iy~ '1t
n n t i 21-
nt -- n
Y=BX+A
* The expression Y = BX + A is the familiar equation for a straight lie. By choosing Rn t as

X, the scale on the abscissa, and
Rnn(
Q n
as y, the scale on the ordinate, the cumulative Weibull distribution can be represented as a
straight line. As noted in Tables 2.1, 2.2, and Figure 2.1, Weibull paper can be constructed as
follows:
TABLE 2.1. CONSTRUCTION OF ORDINATE (Y)
CWl 2Vein.-
:P() I - F(t) mui CW Vei. (.-#91)
0.001 -6.91 0 unis
0.01 -4.60
0.1 -2.25 4.-
4 4
0.6 -03
0.9 0.83 - 7.74
0.999 1.93 84
TABLE 2.2. CONSTRUCTION

OF ABSCISSA (t)
(Ar
Mnt -W ------
1 0 unit
2 O6
I5
3 1.10
4 1.39
1.61
15 2.71
20 3.00
100 4.61
1000 6.91
15
, \ - i,- . ' .~. ,,-- , .

/ / '-".p .4 .
, . ;. .. ',
ca
0
0L
1m
ID~
L
oCo. cv.
' ______5~
c!c
CM-
CdC
LAI I I Coo
'I"
I" I ... - //'.r0

LL....L..J.P.J" I II I ; / -
16
I' If the units used are common for the abscissa and the ordinate (i.e.. nches to inches or
eeitimeters Ito centimeters), the paper will have a one-to-one relationship for establishing the
* slope of the Weibull. (The Weibull parameter if is established by simply measuring the-slope of
the line on Weibull paper.) Of course, the scales can be made in any relationship. That .s2-'o-l,
10-to-I, I1)0-to-1, or any other combination to best depict the data. Throughot thishhandhook
data has ieen plotted on I-to-I paper wherever possible. However, the slopes will be displayed
on the charts. Sample Weibill paper has been included in AppendixL. (At first glance, this
paper may appear to be common log or log-log paper. Looks are deceiving be.cas it is not and
should not be used as such; nor can common log paper be used as Weihull p )
2.3 FAILURE DATA ANALYSIS - EXAMPLE.*
During the deveopment, testing, and field operation of gas turbine engines, items
sometimes fail. If the failure does not affect the performance of the aircraft, it will go unnoticed
until the engine is removed and inspected. This was the case for the compressar inlet airseal
rivets in the following example. The flare part of the rivet was found missing from one or more
of the rivets during inspection.
A program was put into operation to replace the rivets with rivets of a new design. A
fatigue comparison was to be used to verify the improvement in the new rivet. A baseline using
the old rivets war established by an acce!erated laboratory test. The rsults are presented in
Table 2.3.
TABLE 2.3. BASELINE
Rivet Serial Failure Time

Number (SN) (in,) Remar. .
3 90 Failure
2 96 Failure
3 100 Rivet flare loo d without Is
4 30 Failure
5 49 Failure
6 45 Rivet fare loomened without falow
7 10 Lug failed at rivet attachment,
. 82 Failure
Since rivet numberm 3, 6, and 7 were considered nonrepresentative fa thesedata will_

be ignored for the first analysis Thatlves five data points. The first step in establiahing a
Weibull plot is to order the data from low time to high time failure. This failitates establishing
the plottiaig positions on the time axis. It is also needed to establish the corresponding ordinate
F(t) values. Each failure in a group of tested units will have a certain percentage of the total
population failing before it. These true values are seldom known. Studies -have been made-as to
how best to account for this inaccuracy, especially with small samples. However, most of these
-" studies are limited, and more detailed discussion is beyond the scope of this handbook. It has
been the convention at P&WA to use "Median Ranks" for establishing F(t) plotting positions,
and tables can be found in Appendix B.
With five failures, the column in Appendix B headed with sample: size 5 is -used.. The
resulting coordinates for plotting the Weibull are shown in Table 2.4 and plotted in Figure 2.2.
One additional item should be noted. Points with the same time should be plotted at that time
at separate median rank values.
1
1(npur and Lambenmn, Reliability in Enineering Design, Wiley, pp 297-303.
17
MORMON
MORM-O..-
T!ABILE 2.4. WEIBiIIJ, COORDINATEs7
Failure
Order Time Median
Number SIN (Min) Rank
1 4 30 12.9
E
2 5 49 -31.3
3 8 82 50.0
8 2 96 87.0
A line is
drawn through the data points. Formal methods of rank regression and maximum
likelihood for establishing the line are discussed in Appendices C and D respectively. The slv)pe -
of the line is measured by taking the ratio of rise over run. Select a starting point and measure
4)ne inch in the horizontal direction (run). Then, measure vertically (rise), until the line is
itersected. In Figure 2.2, the rise is two inches. Therefore, the slope represented by Greek
sybo f, (fl) -rise/run -2/1 = 2. One needs two parameters to describe a Weibull distribution
when discussing or reproducing the curve. The first is P, and the other is the characteristic life
et~t (denoted by qv). Eta occurs at the 63.2 percentile of the distribution and is indicated on most
characteristic life q = 80 min.
1'he unique feature of the characteristic life is that it occurs at the 63.2'~ Point regardless
of the Weibull distribution (i.e., slope). By examining the Weibull equation it will become clear
why this is true. When time, t, is equal to q~ it does not matter what is; F(t) is always 63.2%/;:
Fit) = 4 00
= -- 00 when t=
=I--0.368
FMt = 0.6.32 regardless of the value of I
M. SUSPENDED TEST rrEMS - NONFAILURES
In t le example in Section 2.3, some rivets failed by causes other than the failure mode of
in' cremt. A rivet that failed by a different mode cannot be plotted on the same Weibull chart in
the 9 ~e manner as a rivet which fractured because the rivets do not belong to the same failure
hin.t ril'utiurn. These data points are referred to as suspended or censored points. There are
evrldefiiios of suspensions, but for Weibull analysis, they are always treated the same
wily. i~hev ca.inot be ignored when establishing the Weibull. The argument for including them
in the anuilys's is that if their failure had occurred in the same fashion as other failures, the rank
order o'f the other failures would have been influenced. Therefore, something needs to be done to
accoun for he potential influence of these points. To illustrate the adjustment of the rank
order n ipbers for the influence of thepe suspended items, the rivet test results will be used
again.
1
Type 1: Te-0, tarninetad after a fixed time haa elapsed.
Tvpe HI: Test terminated after aaet number of failures hae" o.-curd.
Type HlE Test terminated for a cause other thani the one of intereat.
18
- I I I.,.-l t! - -
" EIBLLL DISTRIBUTION

0 =2.0
2 i
so----4FILL; SlZ 5
.=80 i
S~IV FRILL,..,ESIZE
.--. SAMPLE
go,, =S S - - -
-
-
",- i = CHARACTERISTIC LIFE -/- -
-i I N.
"- - o IN.: /
- .0000
U
St. -
0.1 1 f
II0.
, / .a4. F.S..
TOTAL OPERATI.NG TIME (MIN)
FD 272263
Fi.','ure 2.2. Rivet Failures
19
Tlhe genterail formnula for adjus~t ing i hie rank jH)sit ion, conisdering, all poss4ible ways thle
suspended item may have fauiled and potentially influenced the results, is given lby the following
equation:
2
hankIncemen + (N + 1) - (previous adjusted rank)
T7(n-umbFer ofitems beyond p~resenlt suspended item)(2)
wherc N is the total number of rivetsA tested regardlesit of whether it failed, was suspended
(Type I orTl~ype 11), or suspended by the wrong failure mode (Type 111).
Applying this equation to the rivet test data, the values in Table 2.5 are ohtained.
TABLE 2.5. ADJUSTED RANK

Rivet S/N Order Time (minutes) Adjusted Rank
7 1 10 suspenasion -
4 2 30 failure 1.125
6 3 45 suspension -
5 4 49 failure 2.438
8 6 82 failure 3.751.
1 6 90 fdilure 5.064
2 7 96 failure 6.377
3 8 100 suspension -
The adjusted ranks were calculated in the following manner:
Rivet No. 7 is a suspension; therefore, it does not need a rank -Yalue because
it will not be p~lotted on the Weibull chart.
Rank Increment for Rivet No. 4 ( 1+1)70 .2
where:
8 is the total number of rivets tested whether they failed -)r not
o isthe previous adjusted rank (in this case there was none)
7 is the total number of items beyond the first suspension starting the count
with the first failure as illustrated below:
Rivet Time Items Beyond Suspension
47 3 saiuspenso I Starting here and counting forwardI!

6 45 suspension 2
5 49 failure 3 1
8 82 failure 4
1 90 failure 5
2 96) failure 6
31 100 suspension 7
2
Johnson, Leonard G. (1959). The Statistical Treatment of Fatiguve Experiments. Research Laboratories, General
Motora Corporation, pp. 44.50.
20
The adjusted rank is the previous rank (in this case 0) plus the rank increment of 1.125.
Therefore. t he adjusted rank is:
Adjusted Riank for Rivel No.4 4)1I 1.125 1.125
, Rivet No. 6 is a suspension and receives no rank value.
Rivet No. 5 is a failure and the formula has to be employed agami to identify
the new rank increment to use between failures.
Rank Increment (8 + 11)+5

- 1.125-- 1.313
"where
1.125 is the previous adjusted rank
/
5 is the number of items beyond the last Suspension starting
with the tailure following that suspension.
Rivet Time Items Beyond Suspension

7 10 suspension
4 40) failure
6 45 suspension
5 49 failure I Starting here and counting forward
8 82 failure 2
1 90 failure 3
2 96 failure 4
3 100 suspension 5
The adjusted rank, therefore, is the previous adjusted rank plus the new rank
increment.
Adjusted Rank No. 5 = 1.125 + 1.313 2.438
*"Rivets No. 5, Nc. 8, No. 1 and 2 are failures without any additional
suspensions between them and the previous failures. Therefore, no new rank V
increment needs to be calculated. The last value calculated (1.313) is still
- valid. Therefore, the adjusted ranks for these rivets are:
* Adjusted Rank No. 8 = 2.438 + 1.313 = 3.751

* Adjusted Rank No. I = 3.751 + 1.313 = 5.064
* Adjusted Rank No. 2 5.064 + 1.313 = 6.377.
=-
P
21
With these adjusted ranks, the median ranks tan be established. The sample size to be
used when entering the median rank table would be 8. While interpolation could be used for
determining the appropriate median rank, a -good approximation is provided by Benard's
N ~formiala t :
Pam - i- 0. X (X"
= N+0()4 x(2.4)
where N sample size
iadjusted rank value
t Joe of this formula in illustrated in Table 2.6.
(1.125 - 0.3) x lO0~'f .8"

=v 8 +0(.4 =98c
(2.438 - 0.3) x 100"t

P6. 8+0.4 95.45~
etc.
TABLE 2.6. MEDIAN RANK

Adjuated Rank Median
Order No. Rank
1.125 9.82%
2.438 25.45%
3.751 41.08%
5.064 56.71%
6.377 72.35%
timing the cialculated median ranks and the fatilure times, Figure 2.3 is derived. The slope of
the line.f 2.0. is the amue at; the earlier Weibull. This will generally be true if the suspensions
-
.-rt- randomilydispersed with the data. N'ote, however, the effect on the. characteristic life, q~.It
went from FA) minutes wit hout suspensions to 100) minutes with suspension's. The analysis
reoilt ing in I(X)minutes is the correct method.
-2.5_ WEIBULL CURVE INTERPRETATION -
Weibull curves may reveal clues about the failure mechanism, since different slopes imply
different failure mechanisms.
If the slope is less than 1.0, reliability increases as the unit ages. This is often referred to as
an infant mortality failure mode. Quality control and assembly rroblems may produce infant
mortality failures. For instance, some gas turbine failures having slopes less than 1.0 are:
a. Improper augmentor liner repair - quality

b. Improper installation of temperature probes - misassembly
C. Fuel pump leaks due to installation problems - misassembly
d. Overhaul-related failures of various components - quality/misasembly -~
e. Electronic control failures - quality.

'Kau and Lembweomn, (1977). Reliability in Engimeeting Design. John Wiley and Sone, Inc., op. 300.
22
./
KENGILL DISTRIBUTION __------

D =2.0
9 100 MIN -f
SAMPLE SIZE=8 -
*.._ FAILLIES
- ..-..- --
4-4
t1 111 P.
11,. 4 l
1..* 4..III 4 . 44. * 43 0 9
AVM?
Figure 2.:: Rivet Failures With Suspensions~
23
! 'M1.. '1
'Ilie ex poneliai ;l 6iitriblmio ispe ;IIsicial cast-(I* the We'j)ibll disi riluitition when 41 1.0. [he
ivl)onti~ial al retleits original (li-iu
jl45i- IIiideicienies, iiisiitficieit redunda~ncy, tinuexgpetedl
laiiiures (itloji jg.s, 1441, o~r vveil p~ro~iiit iilisiis. li'his wouild result ini a conustanlt halulure rate
(~iht Inii. Soiiiix on~il ~itIIiIsw~it Ih sloptis ii I or*near I are
tiliso WeiI
;I.B~earing cae inilurv

1).Ieilljvratiire probe ilr
C. IFLiel('(lit rod solenoid fanilure
dI. Ftiel oiil cool()kr f'ailure
e. Ellect rnmic engine c( nit rid failhre.
Slopesgreaiter than I represent wearout mlodles. For shallow slopes like 1.8 to 3.0 there is
moire scatter in tIhe fajtln re (data and therefore fi-iilurepredict ions will cover long t imespans
reflecting this uincertainty. As the slopes get sl eep~er, failuire times b~ecolme more predictable.
Some examples of*Weibnills wit h slopes greater t han I are: A
a. hirbine vane wearout

1). Augruentor liner loirnthrboigh
V. 1'eniperat tire prob~e boss fatigue
dl. (earhiix housing cracks
C. A uignient or fiamehol(Ier cracks
1'. Oil tIn b chaite thlrouglh.
A slope. jI, of' 3.44 would approximate the f'amiliar hell-shaped or normal curve, as
indicated in Fig-tire 2.4.
2.6 DATA INCONSISTENCIES AND MULTIMODE FAILURES
'111Ve~eoher
We sui itlet ies in We ibull analysis which might signal problems. Examples are
rate the f'ollowing:
gixen t hat illuist
a. Failuires aire most lY low-time p~arts

1). Serial numbers of lailed part s are close toget her
C. '[he dlata has a -(dogleg" bend or cusp when p~lot ted on Weib'ill paper
d 'I'he data has a gradual convex or concave bend on Weiliull paper.
2.7. LOW-TIMAE FAILURES
Figure 2.5 is an example of* tow-time part failuhres on main oil 1,uinps. (Gas turbine engiiles
-ire tested b~ef~ore being ship'ied to the customer, and since there were over 100) of these engines
in the field with no problems, what was goingr wrong? Upon examining the f'ailed oil pumps it.
was found t hat t hey containedI oversized p~arts. Somet hieg had changed in the manuifacturing
process which created ihis p~roblem. '[lie oversized plarts caused an interf'erence with the gears in
the lpuinp which resulted in finiluire. T'his was traced to a machining operation and corrected.
'I1lw point here is t hat low-t ime failuires often indicate wearout (abnormal in this case) by
having a slope greater than one when plot1ted. Low-time fiailures provide a clue to a production
or assembly p~rocess chan~ge, especially when there are many su-ce.isftjl high-time units in the
field.
24
-N -. - -,.
oo
Z CL
0 L0
I * .. A h-..?' ,~
0 '~ 0 " .,...

I.-
. r-.....&
C CO_
II .. r-
0 c
Q11
a.?
. .- . :.'.:" ,. .. .... _... .

. '.-i\ .<.., " ....... .. ... .- - - , i ..,. ...-. . .
:.x ," ; ,' ."' . - " -"a. " . .
p . .... . ' "- . . . . .
25 /'7"
. \
.4,
t*.",...jI' I I I I IIH
WEEIBULL DISTRIUTIOI N -
' = 1.5
"v= 3618.
S.- SRIH'LESIZE= 1090
w.._ FAILURES = -3 -
I .---- - -4-4----
'I
-C3
iiI
OERAINGTIME (HR)
TOTL
4FD 272265
Figure 2..5. Main Oil Pumpm
20
A1,A/
2.8 CLOSE SERIAL NUMBERS
The ,tinne reasoning can Ihe extendled tho ( her peculiar Cailure gr(i~pings. For example, if
lfailures occur in Ihe middle (i' I he time experienc'e. hat is. low-I ime units have no failures, mid-
time- untits hlave' failures, and high-tinw units have no failures, then a batch problem is
*!,,, smusplcletd. Somet hing may have changed in tlie anulfacturing process fior a short period olt ime
and then changed back. The closeness of the serial numbers of the pares are a very definite clue
to this type of problein.
Figure 2.6 is a prime example of a process change which happened midstream in

1production. Bearings were failing in the augmentor pump. The failures had occurred in the 2W.,
Sto4WX) hour time frame. At least 650 units had more time than the highest time failure. These
failures were traced to a process chai.ge that was incorporated as- a cost reduction for
manuiacturing the bearing cages.
2.9 DOGLEG BEND
A Weibull plot containing a "dogleg hend" is a clue to the potential of multiple failure
modes (see Figure 2.7). This was the case for a compressor start bleed system binding problem.
Upon examination of the data, 10 out of 19 failures had occurred at one base. It was concluded
that this hase's location was contributing to the problem. The base was located on the ocean and
the salt air was the contributing factor.
"I,
The data were broken apart and the two resultant Weibull. charts are presented in Figures
2.8 and 2.9. Note that the fleet Weibull presented in Figure 2.8 is less than one, j 0.837.
X This 7'
could be considered an infant mortality problem, while the ocean base Weibull, Figure 2.9, is
more of a wearout failure mechanism with / 5.223. This problem was related to lack of
maintenance. More attention was given to this area and the problem was resolved.
The failures do not have to be associated with an environmental factor to cause a dogleg
Weibull. In fact, they are usually associated with more than one failure mode. For instance, fuel
pump failures could be due to bearings, housing cracks, leaks, etc. If these different failure
modes are plotted on one Weibull plot, several dogleg bends will result. In cases where this
.Occurs without prior knowledge, a close examination of the failures will have to take place for
potential separation into different failure modes.
OP
27
-'%\ 27'.
...... ."....:. .,.. .. .....

. ... tV -,#./"-
/ -.-. 7 .. *""/.' "\ ... _
, 7
1,EIBULL DISTRIBUTION
0 = 4.994
4 -
= 478.210
SAMPLE SIZE = 387
ao. FAILURES = 9
. -. 49 ,, -4-.- -- 4 - _--- l-- b
Iil / .
1.. __. -0 U8- . . i.U .. 100
----- A - I TOTAL OPERATING TIME (HR)

/ v
FD 27226
F~igure 26. Weibull Plot for Augmentor Pump
28
S-, .. -
. WEIBLL DISTRIBUTION
-. _i 9. ==/ 63601.62
iP 1.122l 2 i1 .. .
S*FLE SIZE = 2%~
-.. FAILURES = 19 " :-
I. I, *1I "
52 3 ---
I18.', ....-
* 6.. -d- 438-- I
__4,6 g I ]
-o -- _ .'77 _
10. * mt 0. 14. 1?.. -t-..S.0SI.

TOTAL OPERATING TIME (HR) * -'
i 7
4 FO 272287
I --
Figure 2.7. Compressor Start Bleed System
29
7-'..
:
*/
W.8r
Ns IEIBULL DISTRIBUTION
1 =O.84
= 404444
SAMPLE SIZE = 2054
0._ F IFILURES 9
me..
.... - - -
C,
S.-
i - .
. , -.-- -.
q1 2 - 312 304
TOTAL OPERATING TIME (HR)

-.
I, L
Fll) 2n2268
Base
Start Bleed 'yst e, Excluding Ocean
.0
Fgur e2.8.
.
Compressor
.. _ . ..
4. . 3. 0. 1 . * ,,o . . ,, . , / ".
30
-AA-
/V
S.. I",EIBLLL DISTRIBUTION__ -. : - -
S. 202006..
4/LESIZE =202 - :::: -
u.- FAILt.GES= 10 - :
III-II-
w
LI6
_. ' .. . ".
0.5- i -
LIJ
2 3-
F S. . - - - ' O Base
--- AUWm i. - -Mm . -.
S. . .. 3IU . 313 4 .....
opexrSatBedS.tm
iiue29 ca
. a...'
AI MI- -
, U,,,
n I.i
2.10 CURVED WEIBULLS
In Sect ion 2.1, lhe cumnulaht ive (list rihution u'wntion F(t) is pr,-sented for the Weibull. It
wis illust rated as:
F(t) ",I -- e I ..,

,1
where:
t failure time
StI
t starting point or origin of the distribution.
In (licussions so far, to,was asmumed to be zero. When data are plotted on Weibull paper it
quickly becomes otvious if the origin of time is not zero. The data will appear curved as
illustrated in Figure 2.10 if the zero time origin is not true.
There are other reasons for poor fit (i.e., the data do not form a straight line on Weibull
paper). For example, another distribution like a Normal, log Norma!, etc. may better describe
* the data. If this is true, the distribution which best describes the data should be used.
But the data displayed in Figure 2.10 was from engine controls and there was no reason to
suspect that the Weibull distribution could not be used to analyze it. There are a couple of ways
to determine what adjustment is needed to make the data appear straight. First, there is an
analytical method that can be used to establish to. The equation is:
4tt-1 0 (42 -- ti)
t. (t0:1,,-
- t2, - t -t) - "....
Where t1 is the first failure time. t2 is the time corresponding to the linear halfway distance
on the vertical axis between the first and last failure, and t 3 is the last failure time. This is
illustrated in Figure 2.11. The values for t1 , t2, and t3 are:
t, = 16.9 hours - first failure
S= 42.0 hours - halfway failure
t3 = 389.0 hours - last failure

t. = 42.0 - (389.0 - 42.0) (42.0 - 16.9)
(389.0 - 42.0)-(42.0 - 16.9)
t,, -42.0 - 27.1
to= 14.9 hours.
If to is positive, it implies that the origin star'ts after zero; if negative, before zero. In other
words, there is a time, in this c ie approximately 15.0 hours, in which the control would not he
expected to fail.
32
Morr
._EIBLLL DISTRIBUTION
0 = 1.330
v = 648.2,S2
SAMPLE SIZE = 261
a=. FAILURES =71
U .- - -
F(t) = 1-e
i, WEN to ,00
CURVE
--
I I., - 4 I -"
* "
. . 4O 0.1; 00' o. '. 4. .s.6.

*. ooo0. 2.3. ;..L
6. 7.9.9.10ooo.
S~TOTRIL OPER/ITINQ TIME (HR)
Ixi 1 FD271851
Figure 2.10. A Curved Weibult Needs t0 Correction
.1 * 33
7. " ,.
//
Ei'
IWEIBJ.LL. DISTRIBUTION-
D = 1.330
v= 648.252
U. SMP4LE SIZE =261 -
a. FAILURES =71--
U0..- . - -a I 1- k.-
D F (t) 1-.e+T
CIV
34
Itdo
-t =42S
1
Upl)on questioning ite vendor, this was not fouind to Ih true. 'hwe vendor was actually
w ltesting the units for ahotit 15 hours prior to shipment and discarding or repairing failed units.
T''his made the (list ribuit ion appe.,r to be truncated at 15 hours with a zero probability of failure
before that time. Subtracting 15 hours frorm each of the failure times will adjust thecurve forthe
absence of this time. The resultant. curve is plotted in Figure 2.12.
The corrected curve provides a more accurate prediction of the probability of failure.
However, to determine distribution percentiles like the B.1 life or 111 life, one has to add the 15
hours to the time read from the Weibull plot.
For example, to determine the time to failure for the 1/1(M unit (often referred to as the BI
life), one would read the I percentage point of 8 hours from Figure 2.12 and then add 15 hours to
it. That is, the BI life is estimated to be 2.3 hours.
The second way to correct a curved Weibull uses a simplistic approach. Wheit the curved
t ,Weimbull becomes fairly perpendicular to the horizontal scale, extend the curved Weibull
vertically through the time scale. Where it intersects, simply read the curve and subtract or add
the time. Trying this technique on Figure 2.11 confirms that 15 hours is a good estimate. vyeye,
t his curve would be considered convex; therefore, a subtraction of time would be required. Data
._plotted on Weibull paper that curves in the other direction (concave) would require addingtime
; to each point.. The amount of time to he added w(,uld be found with either of the above
procedures.
2.11 PROBLEMS
;Problem 2-1:
Fatigue specimens were put on test. They were all tested to failure and the failure times were
150, W.5, 250, 240, 115, 2(A), 240, 1.50, 2C0, and 190 hours.
a. Construct a Weibull and determine its slope, d, and characteristic life, ,..
b. Would you have expected the derived slope for fatigue specimens?
c. If you were quoting the B 1 0 life, what would the value be?
!7
. 4
4 4464
35
i~i N
/
-
'
' P IIBLLL DISTRIB3UTION-

=00.970 i'
9 944.867 -
. SIZE = SPRE
261
FILURES = 71
To' CORRECTION = .0O
-1
0/
0.1<
.~ U.k. o. L 4. 11.11.6.10. * 000..

TOTRL OPERATING TIME-IHR) ,'
F0271853
Figure 2.12. to Correction Applied
36/
x.4
// , . \.-. , ..
".....:..; ___,_._..,: .
Pro~blemi 21-2:
Thlere were live failures .of a part in service. Thle informatiton on these parts is,
Se'rial Numiber Time (hours~) C'omm ent

KI1 9.0 Failure
W12 6.0) Failure
8*1 14.6 Suspension
*8:14 1.1 Failure
L-8:15 20.0 Failure
836 7.0 Suspension
8:17 6.5.o) Failure
8M8 8.0 Suspension
a. Construct. a Weibull with suspensions included and determine its slope, 0,

and characteristic life, q~.
1). What is the failure mode?
C. Are there other clues which n...y lead to an answer to the problem?
P~roblem 2-3:
The following set of failure points will result in a curved Weibull: 90, 130, 165, 220, 275,
370,52.5, and 1200 hours.
a. What value is needed to straighten the Weibull?

b. Will the value found in "a" be added or subtracted from the failure values?
Solutions to these problems are in Appendix J.
37
CHAPTER 3
WEIBULL RISK AND FORECAST ANALYSIS
3.1 'FOREWORD
One of the major uses of Weibull analysi.; is to predict the number of'occarrences of a
failure mode as a function o" time. This projection is important because it gives ma.agement a
clear view of the potential magnitude of a pro)blemI. In addition, if this prediction is made ftr
different failure modes, management is able to set the l)riority for the solution o" each problem.
In this chapter the use of the Weibuil probability d(ist ribution fiunction in predicting the
occurences of a failure mode is explained. The additional input needed for risk analysis will be
covered, and several examples are presented to explain further the techniques involved.
It should he emphasized that. the f'orecast analysis is only as good as the failure data. The .
data should be examined closely to ensure that they are from a single failure mode and will 'ihe
fit a
Weibull distributio-.
3.2 RISK ANALYSIS DEFINITION
A risk analysis calculates the number of incidents projected to occur over some future
period.
3.3 FORECASTING TECHNIQUES
The observed failures and the population of units that have not failed are used to obtain
the Weibull failure distribution, as discussed in Chapter 2. The tollowing additional input is
needed for torecasting:
a Usage rate per uni' per mont h (or year, day, etc.)
h) Inltroduction rate of new units (if they are subject to this same failure
mode)
With this inl'ormat ion a risk analysis can he p)roduced. The tech'niques used to produce the
risk analysis can vary from simple calculations to those involving Monte Carlo simulation.
Monte Carlo simulation is only required when comn)lications arise in the risk analysis. These will
he explained in the following sections.
3.4 CALCULATING DISK
Risk calculat ions are described in three sect ions:
"* Present risk -

"* Future risk when failed units are not fixed
"* Future risk when failed units are fixed.
3.5 PRESENT RISK
The simplest case arises when there are no new units (no production) and no replacement
of tailed units. 1f" there is a population of N items and each has accumulated t hours or cycles,
:. : ~/ ,. . ,.
the expected numtl-r of faltures from this iwpulathion is the p)rolbaility of failure by time t
multiplied by tile numilr of units. N. Therefore, Ifor a Weihull (listribution this becomes
Ixl)ecled numixer of Ifailures F(t). N

(0 e ut/yj) N. ;I:
lquuit.ion :3.I can he use(d immediately to calculate the following:
There are 25 units in a population: 5 units have accumulated I1KM) hours of operational
time, 5 units have accumulated 2000 hours, 5 units have accumulated 3000 hours, 5 units have
accumulated 4000 hours, and 5 units have accumulated 5000 hours. Assume that the population
is subject to a Weibull failure mode with = 3.0 and ? = 10000 hours. The question is, 'WhaL is
the cumulative expected number of failures from time 0 to now for this population?" Figure 3.1.
is the Weibull failure distribution with the cumulative probability of failure by each time on the
unitE as illustrated. Table 3.1 summarizes the calculations involved.
TABLE 3.1. PRESENT RISK
Number(N) Time (t)

of on Each-
Units Unit F(t) F(s) N Exampe of Calculation:
FMt) = 1 - 0 -wt-
5 1000 0.001 0.005 F(1000) =*I- e -0Q,
5 2000 0.008 0.040 - 1- e -tal
5 3000 0.027 0.135 1- a-o"'
5 4000 0.062 0.310 = 1- 0.999
5 &M00 0.117 0.58v) F(IO00) - 0.001
Sum - 1.075
The value of F(t) can also be read directly from the Weibull Cumulative Probability Plot.
(See Figure 3.1.)
The cumulative expected number of failures in this case is 1.075. A
&S FUTURE RISK WHEN FAILED UNITS ARE NOT FIXED .,
Given the same 25 units a in Table 3.1, the expected number of failures over the next 12
months can be calculated. Assume that one of the 4000 hour units has just failed. Since it is
assumed that failed units will not he replaced, it will be omitted from the population for the
calculation of fture risk.
Yearly usage of each unit wi I he 300 hours. The future risk will be composed of the risk o&'
the 1000-hour units failing by 13 hours, plus the risk of the 2000 hcur units failing by 2300
hours, plua the risk of the 30 ho r units failing by 3300 hours, etc.
In general, it a unit has accu ulated t hours to date without failure, and will accumulate u
additional hours in a future peri, then that unit's contribution to the total future risk is:
F(t+u) - F(t)
1 - F(t) (3.2)
39
:LU.m
' 'K" -" - . 2

where F(0) 1 e lirt,
'~is tIn lie ii If- i Ii g in I lit- first I lh ours ofservice,
prhI Iiiiuv oif' It( assiumiing
it follotws it Weibull li-tiliire dlist ribut ioin. ItI F(l ) is miutch less tha 10I.0, equaiot p.2) is'
(31
Fi.t 1 11) lF(t)
Table 3t.2 summarizes the future risk calculat ions for the population of 25 units, with tine failed
unit. at. 4000) hours.%
Hence the expected number of failures from this pop~ulation over the next 12 months is:
*Failures -5(0.(X)12) + N(00(K41) + 5(0.(X)99) +- 4(0.01 54) + 5(0.02316) =0.2506
3.7 FUTURE RISK WHEN FAILED UNITS ARE REPAIRED
The ca'culation of' the number tf failures t~hat will occur over some future time interval
lien the failed units will be repaired and returned to service involves the Saime concepts as
wien units tire noit fixed. VWhen t~he Jprot bilit~y oIf failure (if it unit over the time interval in
quiest ion is smnall (on the order of 0.5 or less), the techniques of Paragraph 3.6 can be applied. In
cases where the probability of' failure is greater than about 0.5, the chance of more than one
failure over the soni time interval becomes significant. Then, the exp~ected number of failures
may be calculated using published tabllest , complex mathematical formulas, otr Monte Carlo
siitulitt itol met ho~ds.
3.8 THE USE OF SIMULATION IN RISK ANALYSIS
The calculat ion of risk is easy for the simple case of a population with no inspections, no
p~roduict~ion added. and nol ret rof its. Of course, even simple risk analysis can become complicated
by the volumec of calculatitons involved. In this case, a computer program automating the
calcti~lat ions is4useful.
IIn sOMe iinStlnCeSK a part's%service life will depend on decisions to be made in the future
which will be dependent. (n a Weihull distribution. Since only the probability of this outcome
may bx- known. ,a poiwerful tool known as Monte Carlo simulation is useful. Monte Carlo
simnulatiorn enables an analyst to build a computer model of the decision plan as it affects a
i arl'o service life. It may include scheduled part inspections, random events such as the
rxtensive wear If*a particular part and its replacement, as well as the addition of new units; into
tIheIfield.
The effect of scheduled inspect ion on risk is straightforward. If a part is inspected and
removed from service, it no longer contriblutes to the fleet's risk. If it continues in service, it
coIntinues to coIntrib~ute to the fleet's future risk.
As an example, thle methodology used in a Mont~e Carlo simulation is; described for the case
(if t hree failure modes and a scheduled inspection. In this case, the number (Itfailures; occurring
in each mode before the scheduled inspection i6 desired. However, the occurrence of' any one
failure mode will not affect any other mode.
'WHITE~, J. S. (1964), "Weibull Renewal Analysis," in Proceedings of tie Aerospace Resinbility and Maintainability
Ctonference, Washington, D. C., 29 June-I July 1964, New York: Society of Automotive Engineers. 639.657.
40
.*g iA
IWEIBtLLLOISTRIBUTION
P = 3.0
=10,000
uiiii- IX
- -
I!~l-kOl~ TOI~TAL OPRTN TIM (HR)
/./
'W. 12 -.
_
-.
I1 -. - - 1. L k
2i7
a 3..Rs
Figur --
Aayi it eb
4I
100. 2. * . .. .661000 3. . 10O I*. . ) ,

: TOThL. OPERATING TIME (HR)
SF~D 271854,
i " Figure 3.l. Risk Analysis' With Weibulls

q
II
[:... . ... ...... ... . . '. 41
: ,-I"\4 "\ . ,
" ; "- '" " "/ " '": "
mm m -,...
I '
TABLE 3.2. FlITURE RISK
Current Time on
"Number Time Each Each Unit's
of on Each Unit Risk
Units Unit at Year's End F(t+u)-F(t)
(N) (t) (hr) (t+u) (hr) F(t+u) F(t) l-Frt)
5 1000 1300 0.0022 0.0010 0.0012
5 2000 2300 0.0121 0.0080 0.0041
5 3000 3300 0.0353 0.0266 0.0089
4 4000 4300 0.0764 0.0620 0.0154
5 5000 5300 "0.183 0.1175 0.0236
The following procedure is performed for each unit in the population. Using random
numbers Ihat are evenly (uniformly) distributed between 0 and I and the three Weibull failure
distributions, generate a time-to-failure for each failure mode. See Figure 3.2. The following
equation is used to calculate the time to failure:
Iime to failure = LR
xln( 1 - random number (3.4)
Scheduled
Inspection
0Mode 2 Mode l
Time to Failure Time to Failure Mode 3
Time to Failure
//
FD 259948
igure 3.2. Simulation Logic F-irst Pass
Advance the simulator to the first event; if this event is a failure, note the cause, and
regenerate a new time to failure for this mode. See Figure 3.3.
. Scheduled
Inspection
0 Mode 1 Mode 2 Mode 3

Time to Failure Time to Time to
Failure Failure
0
FD 259949
Figure 3.3. Simulation Logic - Second Pass
42
,- ,-v *,,-
% .- / .- .." / ,-
- .
.ii iii
/ /, - / I I |7
.1/
Continue this protess until the scheduled inspe4 Iion is reachtd. The numlberof failures of
each mode is recorded and the simulation is repeated. After many repetitions of this process,
each using a different set of random numbers, the result, are averaged to give the expected risk.
A more detailed example utilizing these principles is given in Section 3.12. V
' 3.9 CASE STUDIES
Several case studies in the use of the ideas developed in the previous sections are now
presentd. The first two examples, Sections 3.10 and 3.11, illustrate the direct calculation of risk
without simulation. The case study in Section 3.12 uses Monte Carlo simulation.
3.10 CASE STUDY 1: BEARING CAGE FRACTURE
Bearing cage fracture times of 230, 334, 423, 990, 1009, and 1510 hours were observed. The
population of bearings within which the failures occurred is shown in Figure 3.4. A Weibull
analysis similar to those described in Chapter 2 was followed to obtain the Weibull failure
distribution for bearing cage fracture (Figure 3.5). From this distribution plot we call see that
*, the Blo life (time at which 10"I. of the population will have failed) is approximately 2430 hours.
This was much less than the Bil design life of 8000 hours, so a redesign was undertaken
immediately. Additionally, management wanted to know how many failures would be obsetred
before this redesign entered the field.
The risk questions and solutions are:
1. How many failures could be expected by the time units had reached 1000
hours?
Calculate the number of units that will fail by 1000 hours, assuming failed
units are not replaced. Enter the x-axis of the Weibull plot (Figure 3.5) and
read at 1000 hours that approximately 1.3% of the population is expected
to fail. That is, after the entire population of 1703 bearings reach 1000
hours each, 1703 (0.013) = 22 bearings would be expected to have failed.
2. How many failures could be expected in the next year?
Utilizing the methodology explained in Section 3.6 and applying Equationw

3.2 with a monthly utilization of 25 hours or 12(25) = 300 hours in one year
results in the calculations shown in Table 3.3. Thus about 12 more failures
can be expected in the next 12 months.
S:3. How many failures could be expected when 4000 hours had been
accumulated on each bearing if we instituted a 1000 hour inspection? A
200 hour inspection? No inspection?
From the answer to Question 1, the probability of a bearing failure by 1000

hours is 0.013. Therefore, if it is assumed that each 1000 hour inspection
makes the bearing "good as new" relative to cage fracture, there is a total
expectation of failure for each bearing by 4000 hours of approximately
0.013 + 0.013 + 0.013 + 0.013 = 0.052. So, if all 1703 bearings ran to 4000
hours with 1000 hour inspections, 0.052(1703)= 89 failures can be
expected.
43
A.
S ...
. . .. . . . .* ,_: -~//<.. . .. . ]
On the other hand, if there is a 2(00) hour inspection, the probability of
failure by 2(00 hours is 0.065. Using the same approach as in the previous
paragraph, by 4000 hours about 0.065 + .065 = 0.13 failures would be
expected for each bearing. Therefort, the expected number of failures with
a 2000 hour inspection would be 0.13(1703)=221.
Now suppose no inspections were made until 4000 hours, at which time the
bearing will be retired. Again utilizing the Weibull in Figure 3.5, the
probability of failure by 4(X)0 hours is 0.28. Therefore, by the time all 1703
of the bearings have been retired, 0.28(1703)=477 will have failed.
3.11 CASE STUDY 2: BLEED SYSTEM FAILURES
Nineteen bleed system failures have been noted and the times and geographical locations
of these failures are listed in Table "3.4. The high incidence at air base I) prompted a risk
analysis to determine the cumulative number of incidents to be expected over the next year at
air base 1).
A Weibull analysis of the fleet failures excluding air base D (Figure 3.6), indicates a
decreasing failure rate phenomenon, that is, # <1.0. But a Weibull analysis of the failures at air
base 1) (Figure 3.7) indicates a rapid wearout characteristic. From comparison of the plots it
seems that the bases are significantly d:fferent. It is shown in Chapter 7 that the two failure
distribut ions may be proven statistically to be significantly different.
Since the probability of failure, excluding air base D, was quite low by 4000 hours (the life
limit of t lie part) for the fleet, a risk analysis for air base D only was requested.
The risk questions are:
I) What is Ihe expected number of incidents in the next year and a half with
a usage of 25 hours per month?
Using the histogram of the times on each bleed system at air base D
(Figure 3.8), set up the calculation as before (Table 3.5). Over the next
18 months, 56 failures can be expected using a 25 hours per month
utilization rate.
.4 2) If the usage drops to 20 hours per month immediately, how many fewer
failures can be expected?
Changing the utilization rate to 20 hours per month will chang? the
calculation of expected risk. The new risk over the next 18 months is given
in Table 3.6. About 42 failures, or about 13 fewer than for a utilization rate
of 25 hours per month, are predicted.
44
S.
N. *- --
CD
ti N
N
M' 0
10
w ~0
NN
a_ 0
~ 1 (0
00
0~~~~C a) C 0 D ( I-N
oww
L- I
CL U
Ij 0
.45
. EIBULL DISTRIBUTION
D=2.3
, = 6487
" -. AS
F L S I ZE = 17 0 3"
. _ FRILLIRES = 6 . o
U..
40-
I w
--- 1T13 -- 1
W /
-I-
Wo.,, . ,
445
. 48- - -.
LIII l
2.30 4. . .0 . .. 4.. s. 2. .2..,..S0. 7.8.9.o0=
TOTRL OPERATING TIME (HR)
FO 271855
Figure 1..5. R1-aring ('age Fracture
46
":11,.1
I ; -...
Fit
TABLE 1.3. BEARING RISK AFTER 12 MONTHS
Number Cumrnt Time on Each Unit'a

of Time on Each Unit Risk Total Risk
Units Each Unit at Year's End F(t+u)-F(t) F(t+u)-F(t).
(N) (t) (t+u) F(t) F(t+u) I7-Xt). N
288 50 350 0.0000 0.0012 0.0012 0.3480

148 150 450 0.0002 0.0022 0.0020 0.2963
125 250 560 0.0006 0.0034 0.0029 0.3607
112 350 650 0.0012 0.0061 0.0038 0.4301
107 450 750 0.0022 0.0070 0.0049 0.5193
go 550 850 0.0034 0.0093 0.0050 0.5859
110 650 950 0.0061 0.0121 0.0070 0.7731
.114 750 1050 0.0070 0.0151 0.0062 0.9325
119 880 1150 0.0093 0.0186 0.0094 1.1148
128 960 1250 0.0121 0.0225 0.0106 1.3558
124 1060 1350 0.0151 0.0268 0.0118 1.4691
93 1150 1450 0.0186 0.0315 0.0131. 12214
47 1250 1550 0.0225 0.0366 0.0144 0.6790
41 1350 1650 0.0268 0.0422 0.0158 0.6473
27 1450 1750 0.0315 0.0481 0.0179. 0.4631
12 1550 1850 0.0366 0.0545 0.0185 0.2225
6 1650 1950 0.0422 0.0613 0.0200 0.1197
0 1750 2060 .0.0481 0.0685 0.0214 0.000
1 1850 2150 0.0545 0.0761 0.0228 0.0228
0 1960 2250 0.0613 0.0841 0.0243 0.0000
2 2050 2350 0.0685 0.0925 0.0258 0.0816
Sum - 11.613
TABLE 3.4. BLEED SYSTEM

FAILURES BY
AIR BASE
Air Base Hours at Failure

A 153
B 872
C 1568
A 212
D 1198
D 884
A 1428
C 806
D 1251
D 1249
C 1405
D 708
D 1082
D 884
D 1105
D 828
D 1013
9 64
p 32
- 47
,, .,.
IWEIGULI ISRBUI - F
-T-
0 =0.84-
v = 404444. ---
-
SAHUL SIZE = 2054

a. FAILURES 9 ---
--
a..
- -
- -- - - - - - - -- - - -
Z .
I a.
U.In
- - - -1
554
4044
*9.L 7.s * .3 W9 0. 2o . 2 4. V. VT.e9g.s

TOTAiL OPERATING TIME. (HR)
FD 27185i
Figure 3i.6. Bleed System Failure Distribution Excluding Air Base D
48
Ii
IBiLL DISTRIBUTION
EIf
0 =S. 22I-
= 2006. 5
SAFI.LE SIZE = 202
m.. FAILLRES= 10 - -
S.-
I.I
6 13
1 1
0.3.- - 0 -n -
2. .
i49 A 6. *1..10 .4. . 2. S.6.7.U.j00
FO 271858
\ .\
Figure 3.7. Bleed Syvstem Failure Distribution At Air Base D
N 0-
(D\ CV)
0 04
E0*
C'J
C/)
(Im CI,
*~ ClL
0)
C-4-
500
TAHI,I,:3.5. IBLEED) SYSTEM RISK AIT HR18 MONTHS
(lllilivzaion IbRu, 25 Hours Per Month)
Number Curre'nt Time on Each lnit's

*Units Each Unit in is Afonths F(t+u)-F(t) F(t+u)-F(t)
(N) W) (tfu) F(t) F(t+u) lT-Frt) TWIT-
--- N
0 50 500 0.0000 0.0007 0.0007 0.0000

0 150 600 0.0000 0.0018 0.0018 0.0000.:1
2 2.50 '00 0.0000 0.0041 0.0041 0.0061
0 IN 800 0.0001 0.0082 0.0061 0.0000
0 450 900 0.0004 0.015! 0.0147 0.0000
2 50 1000 0.0012 0.0260 0.0249 0.0497
2 650 1100 0.0028 0.0424 0.0397 0.0794/
10 750 1200 0.0058 0.0659 0.060 O.0
26 850 1300 0.0112 0.0984 0.0882 2.2939
27 950 1400 0.0199 0.1415 0.1241 3.3500
22 1050 1500 0.0334 0.1965 0.1688 3.7130
24 1150 1600 0.0532 0.2640 0.2227 5.3445
24 1250 1700 0.0810 0.34,4 0.2856 6.A540
11 13.50 1800 0.1186 0.4328 0.3565 3.9218
it 1450 1900 0.1675 0.5286 0.4338- 4.7719
20 1550 2000 0.2287 0.6259 0.5149 10.2990
8 1660 2100 0.3023 0.7188 0.5969 4.7752
4 1750 2200 0.387' 0.8016 0.6763 2.7052
2 1850 2300 0.4802 0.8700 0.7499 1.4998/ 'V
3 1950 2400 0.5774 0.9218 0.8149 2.4446
3 2050 250 0.6732 0.9573 0.8693 2.6080
1 2150 2600 0.7618 0.9792 0.9125 0.9125
Sum - 56.2352
vJ
51
51 / :
;tl.''I
N ,- ,
TABLE :1.6. BLEED SYSTEM RISK AFTER 18 MONTHS
(Utilization Rate 20 Hours Per Month)

Number Current Time on Each Unit's
Units Each Unit in 18 Months F(t+u)-F(t) F(t+u)-F(t) N
(N) (t) (t+u) F(t) F(t+u) I-FTt) Tl'-F)
0 .50 410 0.0000 0.0002 0.0002 0.0000
0 150 510 0.0000 0.0008 0.0008 0.0000
2 250 610 0.0000 0.0020 0.0020 0.0039
0 350 710 0.0001 0.0044 0.0043 0.0000
0 450 810 0.0004 0.0087 0.0083 0.0000
2 .550 910 0.0012 0.0160 0.0148 0.0296
2 650 1010 0.0028 0.0273 0.0246 0.0493
10 750 1110 0.0058 0.0444 0.03M8 0.3877
26 8150 1210 0.0112 0.0688 0.0582 1.5136
27 950 1310 0.0199 0.1022 0.084J 2.2676
22 10.50 1410 0.0334 0.1465 0.1170 2.5739
24 1150 1510 0.0532 0.2027 0.1580 3.7909
24 1250 1610 0.0810 0.2714 0.2072 4.9738
11 1350 1710 0.1186 0.3520 0.2648 2.9127
11 1450 1810 0.1675 0.4422 0.3300 3.6296
20 1550 1910 0.2287 0.5384 0.4015 8.0301
8 1650 2010 0.3023 0.6355" 0.4775 3.8201
4 .1750 2110 0.3871 0.7276 0.5556 2.2222
2 1850 2210 0.4802 0.8091 0.6328 1.2656
3 1950 2310 0.5774 0.8759 0.7064 2.1192
3 2050 2410 0.6732 0.9260 0.7736 2.3208
1 2150 2510 0.7613 0.9660 0.8323 0.8323
Sum = 42;7432
3.12 CASE STUDY : SYSTEM RISK ANALYSIS UTILIZING A SIMULATION MODEL
Assume a jet engine has four independent failore modes:
* Overtemperature
e V:mne and Case cracking
* Oil Tube cracking
- Combustion chamber cracking.
The failure distribution of each of these modes is illustrated in Figure 3.9. In addition,
there is a scheduled inspection at 1000 hours. At fail'ire or schieduled inspection the modes are
made "good-as-new."
1) How many failures can be expected in each mode over the next 2 years?
(Assuming a usage rate of 25 hours/month)
2) How will lengthening the intpection interval to 1200 hours change this
risk?
52
/ ... ' - ., ,,
,# , " . , - /\
I .
I ~II I I It1111YlI
u*'s WEIBLIL DISTRIBUTION ....
1--
osm
a. VANE AND CASE CRACKING . I -

* =4.57 '1 2336,4
a COMBUSTION
20 _CHAMBER CRACK [NG
= 4.03 '7 3149 ,VERTEE-ATURE
0 = 2.-S =10193
OIL TUBE CRxCiKrI

= 1.885 . =12050
s
: I # I
! iI:
loc.. 4..;...O. 4.. *.9. 1O 2.!. 4 f. *1.19 L .
TOTAL OPERATING TIME (FIR)

FD 271860
Figure 3.9. Failure Distribution Input to Simulation
53
-L
I I-
," /1 .- .. - -- - -- -,--
\
- z
'I'There is no easv solution to this problem willtout sin,ulation. A Monte Carlo simulation
baset on I hese groundrules is illustrated in Figure 3.10.
'I'l)rovide more det-ail, one engine starting with 0 hours, will he followed step l)y stel) to
'the first scheduled inspect.ion at 1000 hours.
Stop 1
(enierate random times to failure for each failure mode. First-, using a table of random
*
numbers. (Reference 1), four random n:umbers converted to the 0 to I range are 0.(X)7, 0.02$,
0.517. and 0.60'3.
Using Fquation :3.4:
F1 = Overtemperature = 10,193 [Qn (1 - 0.07

= 951 hours
F2= Vane and case crac=king = 2,336 [Qn (I - 0.028

= 1,072 hours
F- -Oil tube cracking 12,0150 [Qn ( - 0517

10,160 hours
F4 = Combust ion chamber cracking = 3,149[Qn (1 - 0.603

I = 3,088 hours
Steps 2 & 3
rhe minimum of the times-to-failure and inspection time is 951 hours; therefore, the
scheduled inspecl ion was not reached.
Stop 4
This failure was an overtemperature (F1 ) and is recorded as occurring 951/(25 hours usage)
:18 months in the future.
Generate another time to failure for F1, using the next random number, 0.442.
New F, 10,193 1Qn (1 - 0.442 7,876 hou

+ 951hours on F failure
8,827 hours
Ref. l A Million Random Digits With 100,000 Normal Deviates, The Free Press; Rand CorpoJration, 1955.
54
/
, * \* ~ * *
O'm'
00
Vi
Coto
0 NrU
L.)
E% 0
00
co
00
C:
0 ca
tIO
S E ca IO ( (
CL CM 0 0
~~U.
1iz
55
iI Now, the minimum of (FI, F2, F.:, F 4, 10X) hours) is 0XX0 hours, which is the scheduled
inspection. This prcess can he continued for as many inspection intervals as desired.
For engines with greater than zero hours initially, the Monte Carlo process must be
, modified. First, the time since last 1(XM)-hour inspection is calculated and used as the engine's
initial age (since engines are made "good as new" at each 1000-hour inspection). Then, note that
thethfirst. set of four random failure times must be greater than the engine's initial age (since all
of the engines in the histogram are suspensions). If any are less, other random numbers are
drawn until all four failure times are greater than the initial age.
The above procedure is followed for each engine in the population (Figure 3.11) and is
repeated several times so that an average risk can be calculated.
The simulation in Figure :3.10 was run, and the risk for the first 24 months is presented in
Table :1.7 for the 0XX) hour inspection, and in Table 3.8 for the 1200 hour inspection. A plot
comparing the two risks is presented in Figure 3.12, Increasing the inspection interval to 1200
hours increases the expected number of failures from 25 to 34, a delta of 9, by the end of 1981.
%Y
\ .- / "" \"
'/.4 t'
/ / _. . -.- '- . .. ,
56
N.
1~100
~.E 9
E 10
69 0 a
99 C-4 0
9*D 0
.01
* ae)
Fri
Co CD
COS
CC'
57S
- TABLE 3.7. SIMULATION OUTPUT FOR 1000 HOUR INSPECTION
* Cumulative Incidents
Month EFHO Curn EFH* Oil Tube Vane Case Over/Temp Comb. Chamber
S.#
SJan 29,22.5 29,225 0.00 0.00 0.00 0.00

SFeb 29,225 58,450 0.17 0.33 0.21 0.17
Mar 29,22.5 87,675 0.38 0.67 0.47 0.34
Apr 29,225 116,900 0.60 1.15 0.74 0.58
May 29,225 146,125 0.78 1.47 0.95 0.74
Jun 29,22.5 175,350 0.92 1.71 1.13 0.87
Jul 29,225 204,575 1.22 2.27 1.49 1.15
Aug 29,225 233,800 1.46 281 1.77 1.41
Sep 29,225 263.025 1.66 3.16 2.02 1.60
Oct 29,225 292,250 1.95 3.90 2.36 1.95
Nov 29,225 321,474 2.07 4.03 2.51 2.03
SDec 29,225 350,699 2.38 4.90 2.87 2.45
I **SS****19815S * *
Jan 29,225 379,924 2.66 5.55 3.19 2.77
Feb 29,225 409,149 2.77 5.83 3.32 2.91
Mar 29,225 438,374 2.87 6.13 3.44 .05
Apr 29,225 467,599 3.07 6.68 3.67 3.31
May 29,225 496,824 3.28 7.33 391 3.62
Jun 29,225 526,049 3.37 7.48 4.02 3.69
Jul 29,225 555,274 3.64 8.26 4.33 4.06
Aug 29,225 584,499 3.70 8.45 4.40 4.15
Sep 29,225 613,724 3.76 8.59 4.47 4.21
Oct 29,225 642,949 3.80 8.59 4.47 4.21
Nov 29,225 672,174 4.16 9.40 4.95 4.62
Dec 29,2215 701,399 4.44 9.96 5.29 4.90
KEH = engine flight hours
58
: I
'rTAIL 3.8. SIMULATION OuT'itT FOR 1200 HOUR INSPECTION
Cumulative Incidents
Month EFtII ('urn EkI* Oil Tuhbe Vane Case thver/r,.mmp Comb. Chamber
)980
Jan 29,225 29,225 0.00 0.00 0.00 0.00

Feb 29,225 58,450 0.17 0.53 0.21 0.24
Mar 29,225 87,675 0.40 1.28 0.47 0.58
Apr 29,225 116,900 0.69 2.44 0.79 1.08
May 29,225 146,125 0.91 3.21 1.05 1.43
Jun 29,225 175,350 1.21 4.02 1.40 1.81
Jul 29,22.5 204,575 1.40 4.86 1.62 2.17
Aug 29,225 233,800 1.57 5.25 1.82 2.36
Sep 29,225 263,025 1.84 6.26 2.13 2.81
Oct 29,225 292,250 1.93 6.64 2.43 2.98
Nov 29,225 321,474 2.12 7.39 2.55 3.31
Dec 29,225 35G,699 2.35 8.25 3.08 3.69
*eS*ee*' 1981******
Jan 29,225 379,924 2.61 9.26 3.21 4.14

Feb 29,225 409,149 2.72 9.73 3.34 4.34
Mar 29,226 438,374 3.01 10.99 3.46 4.89
Apr 29,225 467,599 3.31 12.16 3.79 5AO
May 29,225 496,824 3.31 12.16 3.93 5.40
Jun 29,225 526,049 3.65 13.72 4.04 6.07
Jul 29,225 555,274 3.65 15.72 4.35 6.07
"Aug 29,225 584,499 3.93 14.92 4.42 6.58
Sep 29,225 613,724 3.96 14.92 4.59 6.58
Oct 29,225 642,949 4.01 15.16 4.64 6.68
Nov 29,225 672,174 4.16 15.77 4.99 6.94
Dec 29,22.5 701,399 4.46 16.47 5.32 6.94
* EFH = engine flight hours
4
4
59
Mj"
3.13 PROBLEMS
Problem 3-1
A fleet of I() engines is subjected to a Weihull failure mode. The Weibull has a slope of 3
and a characteristic life of I(XX) hours. The current engine times are as follows:
Number of
Engines Engine Time
20 150 hrs
20 200
20 250
20 300
20 350
A.) What is the expected number of failures now? B.) How many additional
engines will be expected to fail in 6 months if the utilization rate is
25 hr/mo? Assume that failed units are not fixed.
Problem 3-2
A turbine airfoil has caused unscheduled engine removals at the following times and
locations.
Time at
Failure Location
684 (hours) A
821 A
812 A
701 A
770 A
845 A
855 B
850 C
806 E
756 G
755 H J
741 G
681 E
667 C A
649 B
603 B
600 C
596 G
57b D
,504 E ------
476 H
A) Generate a Weibull using the attached populations, overall (Figure 3.13)

and at Location A (Figure 3.14). How do these Weibulls compare?
60
'I.-
30.
:101>
~0.
0010
00.
CLL
oCM CO 0
00. 0)
0.-0
0-0
CcU
0*- 0
00)1
H) How many failures can be expected in the next 12 months?, the next 24
months? from each population? (Use 30 hours/mo.)
Problem 3-3
Given a control failure mode with = 1.26 and iq 19,735 total operating hours, and the
population of nonfailed units Figure 3.15, A) how many failures can be expected by the time
each unit has reached 1000 hours? B) 2000 hours? C) If the life of a control is 4000 hours, what is
the projected total number of failures in the life of the control if no further controls are added to
the population? D) If inspections "zero-time," or make the control units "good-as-new' how
many failures are projected to occur in this population by 4000 hours with a 1000 hour
inspection? E) with a 2000 hour inspection?
Problem 3-4
Using the table of 0-1 random numbers in Table 3.9, and the three Weibull failure modes:
a. = 0.76 Z
,q = 96,587 hours
b. /i =2.038
ii =4996. hours
c. i =7.4
= 1126. hours
Assume two scheduled inspections, at 1000 hours and 2000 hours, that make modes a and c
"good-as-new." while not helping mode b. A usage ra'e of 25 hours per month is assumed.
The following population of 5 engines is at risk:
I engine at (X)hours, I engine at 200 hours, I engine at 500 hours,
I engine at 7M) hours. and I engine at 900 hours.
A) How many failures will occur over the next 48 months?
Use the Monte Carlo simulation technique to solve this problem.
B) Would it be advisable to drop the 1000 hour inspection?
62
.Go.
0 rC-S
C,4I
00
oro.
co 'J
o 0m v m 0
II60
I-C.
a-,O
fl
0 co
UC)
00
Nq 0
CC,
co
CD4
cmJ
CJ
IC
a'a
0~0
0 c
64
Ci
>
C',
oDC
Ml, 0
st IC
coo4
co-
00 F)
cm'
co co C
0 a0 65
/ \
TABILE :.9. TABLE OF UNIFORM RANDOM NUMBERS FROM 0. TO 1.0
0.329 0.604 0.615 0.300 0.070 0.845 0.494 0.624 0.085 0.194
0.612 C.337 0.393 0.163 0.774 0.620 0.596 0.503 0.857 0.794
0.I45 0.945 0.3.r% 0.429 0.7f9 0.675 0.689 0.203 0.643 0.577
0.2.32 0.511 0.311 6.213 0.124 0.827 0.354 0.556 0.8!1 0.811
0.221 0.480 0.315 0.167 0.390 0.987 0.428 0.257 0.298 0.198
0.210 0.457 0.010 0.083 0.837 0.265 0.638 0.943 0.747 0.164
0.519 0.668 0.717 0.230 0.133 0.612 0.658 0.491 0.772 0.676
0.037 0.971 0.169 S0.166
0.815 0876 0.668 0.649 0.205 0.551
0.138 0.601 0.761 0.490 0.655 0.238 0.277 0.123 0.918 0.984
0.214 0.71,8 0.224 0.706 0.748 0.090 0.389 0.699 0.5A2 0.761-
0.418 0.422 0.402 0.270 0.928 0.982 0.365 0.933 0.323 1.367
0.950 0.469 0.709 0.431 0.854 0.363 0.574 0.630 0.521 0.974
0.202 0..%03 0.4.34 0.394 0.851 0.909 0.168 0.058 0.673 0.012
0.180 0.104 0.384 . 0.013 0.364 0.480 0.687 0.636 0.340 0.805
0.447 O..36 0.506 0.980 0.605 0.408 0.833 0.544 0.961 0.476
0.412 0.785 0.084 0.222 0.750 0.600 0.495 0.497 0.821 0.105
0.580 0..12 0.855 0.990 0.765 0.669 0.895 0.635 0.842 0.850
0.083 0.963 0.i34 0.847 0.717 0.054 0.420 0.249 0.041 0.502
0.609 0.996 0.793 0.526 0.159 0.861 0.507 0.826 0.249 0.688
3.551 0.198 0.701 0.376 0.932 0.888 0.655 0.608 0.838 0.703
n-rI.s.-tr..r~z. uve... ~ . e.t.....- .- .- * .--------
66
r ~I"
CHAPTER 4
WEIBAYES - WHEN WEIBULLS ARE IMPOSSIBLE
4.1 FOREWORD
At times a Weibull plot cannot be made because of deficiencies in the data. Typical
situations would be when:
(1) There are too few or no failures.

(2) The age of the units is unknown, and only the number of failures is
known.
(3) A test ),lanfor a new design is needed.
Weibayes analysis has been developed to solve problems when Weibull analysis cannot be
used. Weibayes is never preferred over Weibull analysis but is often required because of
weaknesses in the daLa. Weibayes is defined as Weibull analysis with an assumed # parameter.
Since the assumption requires judgment, this analysis is regarded as an informal Bayesian
procedure.
4.2 WEIBAYES METHOD
lit a Weibayes analysis, the slope/shape parameter t is assumed from historical failure data
or from engineering knowledge of the physics of the failure. Deperoding upon the situation, this
may be a strong or weak assumption. Given #, an equation may be derived (Appendix E) using
the method of maximum likelihood to determine the characteristic life, i.
N
; (4.1)
Where t1 is the time or cycles on uniti, r is the -"imber of failed units and V is the maximum
likelihood stimate of the characteristic life.
With #iassumed and q calculated from equa4tion t4.1), a Weibull equation is determine-". A
Weibayes line can be plotted on Weibull paper. The plot is used exactly like a Weibull
distribution. -
4.3 WEIBAYES - NO FAILURES
In many Weibayes problems no failure has occurred. In this case, a second assumption is
required. The first failure is assumed to bw ir-.minent; i.e. r = 1.0 (otherwise, the denominator in
equation (4.1) would be zero). The Weibayes line based on assuming one failure is conservative,
with at least 63 ', confidence that the true Weibull lies to the right of the Weibaye,. line. (See
Appendix E.)
The exact confidence level of the Weibayes lower bound is unknown because it depends on
the time to the first failure. If the Weibayes line is always constructcd immediately before the
first failure. the Weibayes confidence level is 631". If Weibayes analyses are consistently done
long before the first. failure, the confidence level is actually much higher than 63',. Therein,
Weibayes displays conservatism since the conridence level, while unknown, is at least 63',.
67
t -
.. , , . -1 t . + i"
tI
N:p
4.4 WEIBEST - NO FAILURES
In the early development and use of this analysis* 0.693 failures would be assumed instead
of 1.0. This is less conservative. The result was called a Weibest line. The Weibest line is a 504,
%-4 lower confidence bound on the true Weibull characteristic life versus 63', for Weibayes with r
..%4 1.0. In fact, Weibayes lines may be calculated for any confidence level (Appendix E).
4.5 UNKNOWN FAILURE TIMES
Sometimes the number of failures is known, but not the times to failure; again Weibayes
may provide a solution. For example, if the failed part is nonserialized and the component or
system has been through overhaul, it may be impossible to determine the time on the failed
unit(s) or the success unit(s). However, if the time on the components or systems is known, it
may be reasonable to assume that the same distribution of times applies to the nonserializcd
parts. In this application, there is more uncertainty in assuming a value for 0. If the physics of
failure are known, a library of Weibull failure modes may provide an estimate or a range of
estimates; the maximum and minimum # may each be used to determine the sensitivity of the
analysis to the assumption.
"If the times ton the failed units are known but the times on the successful units are
unknown, a Weibull shift method may be employed. (See Section 6.3.)
4.6 WEIBAYES WORRIES AND CONCERNS
The Weibayes method is required when there are deficiencies in the data or when the data
are not available. The Weibull method is always preferred over Weibayes, so it is appropriate to
A critically question the assumptions required by the Weibayes method in each case since the
answers to these questions vary for each application. Of course, the validity of the results
depends on the validity of the assumptions. Typical questions to be raised are:
4
(1) How valid is the assumed slope, e? If this assumption is shaky, should a
range of slopes be tried?
(2) With a redesign, what is the probability that a new failure mode is
present? A Weibayes test may not discover a new mode.
(3) With nonserialized parts, some assumption must be made to obtain

success or failure times. How valid is the assumption?
4.7 EXAMPLES OF PROBLEMS/ANALYTICAL SOLUTIONS
Problem 1) Fifteen vane and case failures have been experienced in a large fleet of engines.
Weibull analysis provides a d of 5.0 (see Figure 4.1). Three redesigned compressor cases have
been tested in engines to 1600, 2900 and 3100 hours without failure. Is this enough testing to
.substantiate the redesign?
-. 1.
Mr. Joseph W. Grderuck of Pratt & Whitney Aircraft/Commercial Products Division is credited for much of the
original development of the Weibest concept
68
'N
K.. - ",
. . :h . .,. _Kl
"-
/;:-
.. .. /
.
: . ...
A %.X':...,, tx
Y, ~ V
404
. EIBULL DISTRIBUTION - - - -\
0 = 4.971
T= 2132.247
. SAPJLE SIZE = 275S
"U.- FAILURES 15
U.li --- -- - - -- 4 , - -!/ ..
rMU
-4 ' IS1JJ/ 4 io
111
'BIN
TIME SINCE OVERHAUL (HR)
FD 271a61
Figure 4.1. Compressor Vane and Case
69
, -
*
Assunuing 1 5.0 aind given the tinies on the three redesigned units, equation (4.1) mav be
used to calculate the characteristic life for a Weibayes solution.
= 1
= L'(1600):' + (2900)) + (3(M)03'] .. = 3406 hr
", ~~(4.2) ...
The Weibayes line is plotted in Figure 4.2. We may state with 63"1 confidence that the
failure mode for the redesigned units is to the right of this line and, therefore, significantly
better than the bill-of- material vane and case. It is possible that the redesign has eliminated this
failure mode but that cannot be proven with this sample of data. As more time is put on these
units without failure, the Weibayes line will move further to the right and more confidence will
be gained that the failure mode has been eliminated. The assumption of slope, in this case, is
* based on an established Weibull failure mode and is valid.
P'roblem 2) There have been 38 turbopump failures in service (Figure 4.3). Based on the physics
of the failure, an accelerated bench test was designed and two more turbopumps failed in a
much shorter time (Figure 4.4). Notice that the bench test Weibull has the same slope as the
'. , field failure Weibulls. This provides some confidence that the accelerated test provides the same
failure mode experienced in service. The turbopump was redesigned to fix the problem and two
units were tested on the bench to 500 hours without failure under the same accelerated
conditions. Is the redesign successful? What service experience should be expected?
Using equation 4.1 and the slope from the Weibulls in Figure 4.3, the Weibayes
characteristic life is calculated, assuming the first failure is imminent.
[r
+1 50027]"
~~~. 5o 27.,. - 646 hr
L (4.3)
This Weibayes line is plotted on Figure 4.5. If we assume that the ratio of characteristic
lives Wi's) for the B/M pump in service to the B/M pump in the rig test is a measure of the
acceleration of the test, a Weibayes line can be estimated for the redesigned pump in service. /
This line is also plotrted in Figure 4.5.
11
Ieeignnediervice = ('B/M/SVC 113B/M/Rig) iiRedem.in/Rig
= (2186.2 hr + 140 hr) 646.3 hr
= (15.6) 646.3
Redesigred/Semce, = 10,082 hr
I'roblem 3) One batch of turbopumps (DF3) produced nine service failures involving fire in
flight. From a Weibull analysis, it was decided to replace these pumps after 175 hours of
operation. Two other batches of these pumps, DF1 and DF2, had more service time but no
failures. Teardown and inspection of some of these pumps showed that the failure mode
(swelling of the ball bearing plastic cage) was present but to a lesser degree. There were not
enough spare pumps to immediately replace the DF1 and DF2 units. How long can replacement
of DFI and DF2 be safely delayed?
70
: " , : " " " i'Z

P =..9-
So. * P= 413.247
.. %.. 1,..:
SAPESIZE = 2755
... ,;50. go.- . FAILURES= 15S.- -
"W.-
d 40.. ,
. EB IREIBAYES DISTRIBUTION
= 4.971-
i.
'V1 = 3406.
w
.. _ -- = 4,971
.I: F-1-" .. -
,% q.{' _
41
1 12
i 116
OS.j 120 I"
Of
'

!t .q
: _
Oa
-o~itd
--. '", ., 86>
:::.-ac
S726
.....- "
TIM SIC OVRAL(R
.. ~~~ ~ ~ ~ ~ ~ ~~
~"2_-...............862 ~~~~F 1-:.x-:.....
.12\ ia(sEauto fNi ~sg nAclr dTs

UI.S.--.I
*;. '._ (EIBULL DISTRIBUTION -

0 = 2.624
""= 2186.2
.SIP LE SIZE = 800
-, F~AILURES =38.-
B/M FIELD
EXPERIENCE
1.. - - ...
U-I
'I,>
4
''4
o 0., , .7.. .
TOTPIL OPERATING TIME - (HR)
FD 271864
Figure 4.3. Weibull Evaluation of BIM Design
72
, - i . " - -- . -... :- ' " iL." '

I -
-, -
* m. WE
BULLDI
EIBU..L DISTRIBUTION
ISI BT ION -
-- I - --
B/M LOW LUBRICITY

TEST EXPERIENCE
(2 FAILURES)
..
B/M FIELD
EXPERIENCE
- (38 FAILURES)
. Pm. -1- A. - -

',,
/ PID271865 'N
S/-
! :
Figure 4.4. Weibayes Eval'uation of New Design in Accelerated Test
73
I
*Ii I 1I
I I I I II1. ..
.-EIBULL DISTRIBUTION -
I i
.46
B/M LOW LUBRICITY O0,0828

._ TEST EXPERIENCE
(2 FAILURES),-
ILU.,PI REDESIGN LOW

C ITY
,Z-vTEST EXP --
ii 0 Vi
M
1/
B/M FIELD
EXPERIENCE
I-V (38 FAILURES)
REDESIGN
PROJECTED
FIELD EXPERIENCE
tO. . . 4. 11. L. 0 2. 3. 4. S. 4. I.S.IO. 2. 3. 4. L 6. l, am.6,1,L
TOTAL OPERATING TIME (FIR)
FD 271867
Figure .1.5. Hydrzulic Pump Failures
74
-' Sn . _. -. _ . . -.. : .,
'lThere were no failures in ID)I and ID)12 even ihough symphins or the t;alure mode were
presenl. A Weilwyes analysis. using I he exist ijg Weilmll slope of 0i 4.4; and assuming the (0.(9!t3
ifailures were imminent, produced (he Weihest (5(0', confidence) line shown in Figure 4.6. The
I )1.'3 rofil al 175 hours (morresponds to a risk level of 1.7 as indicated in Figure 4.6. The same
risk level was applied to the Weihayes line and a 7W0 hr safe period was recommended. DFI. md
l)F2 pumps were replaced when they reached 700 hours. This did not create a supportability
problem as these pumps had acquired very little time. Weibest and Weibayes lines move to the
right with time as long as no failures are observed due to the increase in success time. In this
case, the Weibest B.7 time eventually exceeded the pumps' overhaul time of 1000 hours.
"I. Therefore, many pumps were utilized to their full life without premature replacement based on
the Weibest Analysis.
4.8 PROBLEMS
Problem 4-1
Two bolt failures due to low cycle fatigue have been observed in a flight test fleet of six
engines having the following times: 100, 110, 125, 150, 901 and 40 hr. The bolts are not serialized
and as the failures were discovered after the engines were overhauled, it is not known which
engines had the failed parts. If low cycle fatigue "iilure modes usually have slope parameters
between 2 and 5, and after rebuild the engine will accumulate 100 hours in the next year, predict
the number of expected failures. (Assume the two new replacement bolts are installed in the
rebuilds of the high time engines.)
Problem 4.2
The design system predicted B.1 life for the compressor disk is 1000 cycles. Five disks have
:. accumulated 1-510* cycles and rive have 2000 cycles wiirhout any failures. If most disk LCF
.failures have a 11of 3.0, is this success data sufficient to increase the predicted design life?
Solutions to these problems are in Appendix J1.
.97
I.o
'i 75
~amsana
~ *~wrnamnsnwn r~- .V.,..n -- - - - - o
' WEIBLLL DISTRIBUTION-
P = 4.915
= 488.S20
SAWILE SIZE =387
FAILLIRES =24
CONF LEVEL =0.90
70... _ __
40._ - 73
wU sag
ii fy 20
Z)
MO 2
no
low~
aK.
Of) m
cc
J.-
1.17S
0.5.-
10. 2. 3. 4. i. 6.17.8.19.300. 2. 3. 4. S. 1. .7.8..1000. 2. 2. 4. asw~

TOTP.L OPERATING TIM1E (HR)
FO 271868
* ~Filure 4.6. Whzen Should We Pull the Sus~pect Batch?
76
CHAPTER 5
SUBSTANTIATION AND RELIABILITY TESTING
5.1 FOREWORD
The objective of this chapter is to address the statistical requirements of substantiation

and reliability testing when the underlying failure distribution is Weibull. Substantiation
testing demonstrates that a redesigned part or system has eliminated or significantly improved a
known failure mode (0 and 17are assumed to be known). Reliability testing demonstrates that a
reliability requirement has been met.
It is assumed in the reliability testing section that the Weibull slope parameter, #, is
known. If the failure distribultion is known to be Weibull but 4 is unknown, test plans developed
by K. Fertig and N. MannI may be used.
A test plan gives the required number of units and the amount of time to be accumulated
on each to either substantiate a fix or meet a reliability goal. It also gives a success criterion,
where the test is passed if the success criterion is met. In a zero-failure test plan the success
criterion is no failure: the test is passed if every unit runs the prescribed amotunt of time and no
unit fails while on test.
Test plans can also be generated with a 1-failure success criterion, a two-failure success
criterion, etc. But all of these plans require more testing than the zero-failure plan.
A mwasure of confidence is usually built into statistically designed test plans, guaranteeing
thait if the failure mode in question has not been fixed or the reliability requirement has not
been achieved, there is a low probability that the test will be paseed. The zero-failure test plans
in this chapter guarantee with 90% confidence that the test will be failed if the required goal has
not been achieved. Thus, a part or system will have at most a 10'"1 chance of being accepted as
sat islactory when in fact it is not.
5.2 ZERO-FAILURE TEST PLANS FOR SUBSTANTIATION TESTING
A ball and roller bearing system has a Weibull failure mode, unbalance, with 0 (the
Weibull slope parameter) equal to 2, and q equal to 500 hours. The system is redesigned, and
three redesigned systems are available for testing. How many hours should each syi,.em be
tested to demonstrate that this mode of unbalance has been eliminated or significantly
improved?
The Weibull plot in Figure 5.1 illustrates the time-to-unbalance distribution.
Table 5.1 is used to answer this type of question. It is entered with the value of 4 and the
number of units to be tested. The conesponding table entry is multiplied by the characteristic
life to be demonstrated to find the test time required of each unit.
iMann, N. R. and K. W. Fertig, (1980) Life-Test Sampling Plans for Two.Parameter Weibull Populations.
Tehnometris, 22,165-177.
77
inmt- .-
. EIGULL DISTRIBUTION
o00 HOURS 1= 1
h "
U
B .. -
--.
- - --
- -
- -
1BFALL AND ROLLER

BERRING UNBALANCE
z DISTRIBUTION
Iii
I-
Y
'I
* 0 3I .
-I---
tO. 3. 4. . .. SJO.3. 2. 5.. '..65000. '2. 3. 4. * .0*.9.10m0.
17OTPL OPERATING TIME (HR)
FD 271869
SFigure 5.1. I-
Ball and Roller Rearing Unbalance Distribution
78
TABLE 5.1 CHARACTERISTIC LIFE MUILTIPLIIERS FOR ZMRO-FAIILJRE Tl'tSTr
PLANS /
CONFIDENCE LEVEL: 0.90
0
0.5 2.5 4.0 4.5 5.0
Sample Infant 1.0 1.5 2.0 Gradual 3.0 3.5 Rapid Wearout
Size Mortality Random + ...... Wearout ------ + + - (Brick Wall) - +
3 0.589 0.767 0.838 0 0.900 0.916 0.927 0.936 0.943 0.948
4 0.331 0.576 0.692 0.759 0.802 0.832 0.854 0.871 0.884 0.895
5 0.212 0.460 0.596 0.679 0.733 0.772 0.801 0.824 0.842 0.8.56
6 0.147 0.384 0.528 0.619 0.682 0.727 0.761 0.787 0.808 0.826
7 0.108 0.329 0.477 0.574 0.641 0.690 0.728 0.757 0.781 0.801
8 0.083 0.288 0.436 0.536 0.608 0.660 0.701 t, 732 0.758 0.780
9 0.065 0.256 0.403 0.506 0.580 0.635 0.677 0.711 0.739 0.761
10 0.053 0.230 0.376 0.480 0.556 0.613 0.657 0.693 0.722 0.745
12 0.037 0.192 0.333 0.438 0.517 0.577 0.624 0.662 0.693 0.719
14 0.027 0.164 0.300 0.406 '3.486 0.548 0.597 C.637 0.670 0.697
16 0.021 0.144 0.275 0.379 0.461 0.524 0.575 0.616 0.650 0.679
18 0.016 0.128 0.254 0.358 0.439 0.504 0.556 0.598 0.633 0.663
20 0.013 0.115 0.237 0.339 0.421 0.486 0.539 0.582 0.619 0.649
25 0.008 0.092 0.204 0.303 0.385 0.452 0.506 0.551 0.589 0.621
30 0.006 0.077 0.181 0.277 0.358 0.425 0.480 0.526 0.565 0.598
40 0.003 0.058 0.149 0.240 0.319 0.386 0.442 0.490 0.530 0.565
50 0.002 0.046 0.128 0.215 0.292 0.358 0.415 0.463 0,505 0.1540
In the balh and roller bearing example. Table 5.1 is entered with # equal to 2.0 and a sample
size of three. The corresponding table entry is 0.876. The characteristic life to be demonstrated
is 500 hours. The number of hours that each system should be tested is:
0.876 X 500 hours = 438 hours.
Thus, the zero-failure test plan to subrtantiate the ball and roller bearing system fix is: test
three systems for 438 hours each. If all three systems are in balance at the end of the test, then
the unbalance mode was either eliminated or significantly improved (with 90'( confidence).
If there is a constraint on the amount of test time accumtlated on each unit, Table 5.2 is
used to determine the number of units required for the test. For example, suppose in the
previous example that no more than 300 hours could be accumulated on any bearing system.
-Table 5.2 is entered with the known value of d and the ratio of the test time to the characteristic
life being substantiated. In the ball and roller bearing system example, Table 5.2 is entered with
# equal to 2.0 and the ratio
300 test hours per system
500 hour characteristic life =0.6
The corresponding entry in Table 5.2 is seven. The resulting test plan is: test seven systems for
..00 hours each. If all seven systems are in balance at the end of the test, then the unbalance
mode was either eliminated or significantly improved (with 90'; confidence).
79
... / 7
TABILE 5.2 REQUIRED SAMPLE SIZES FOR ZERO-FAIIURE TEST PI,ANS
CONFIDENCE LEVEL: 0.90
46
0.5 2.5 4.0 4.5 5.0
Infant 1.0 1.5 2.0 Gradual 3.0 3.5 Rapid Wearout
Ratio Mortality R.,ndom + ------Wcarout ----- -4- + - (Brick Wall) - +
0.01 24 231 2303 2302.5 * * ***** ***** **0**
01.02 17 116 815 5757 40703 * **** ***** ***** .$.*.
0.03 14 77 444 2559 1477i 8.1278 ..... ..... .....
0.04 12 58 288 1440 7196 35977 001"s ***** 0** *
0.05 1 47 206 922 4119 18420 82377 '* *
0.06 10 39 157 640 2612 10660 43519 **
0.07 9 33 125 470 1777 6713 25373 95898 *
0.08 9 29 102 M60 1272 4498 15900 66214 ** "
0.09 8 26 86 285 948 3159 35094 350A4 *
0.10 8 24 73 231 729 2303 7282 21025 72812 *
0.20 6 12 26 58 129 288 644 1440 3218 7196
0.30 5 8 15 26 47 86 156 28.5 519 948
4 6 10 15 23 36 57 90 143 225
0.40
0.50 4 5 7 10 14 19 27 37 53 74
0.60 3 4 5 [ 9 11 14 18 23 30
0.70 3 4 1 4 5 6 7 9 10 12 14
0.80 3 3 4 4 5 5 6 6 7 8
0.90 3 3 3 3 3 4 4 4 4 4
1.00 3 3 3 3 3 3 3 3 3 3
Indicates sample size exceeds 100,000
5.3 ZERO-FAILURE TEST PLANS FOR RELIABILITY TESTING
This section contains Z.ero-failure test plans for demonstrating a reliability goal when the
underlying failure distribut!ion is Weibull with known slope parameter #. A turbine engine
combustor's reliability goal',was 99,. reliability at 1800 cycles under service-like conditiots.
Success was defined as a comhbustor having no circumferential cracks longer than 20 inches (out
of a possible 53 inches). The number of cycles required to reach a 20-inch crack was known to1
follow a Weibull distribution with # equr.l to 3. How many combustors must be tested, and how
many cycles must each accumulate, to demonstrate this goal with a high level of confidence?
First, the reliability goal is re-expregsed as a characteristic life goal, and then the test plan
is designed.
Re-expreaion of Reliabilijy Goal
Reliability requirements generally assume one of the following forms:
Form 1: The reliability of the unit is required to be at least X'(, aftcr a certaiin
number of hours or cycles of life. (This is equivalent to the pereent
failing being at most 300.X`',). The Weibull plot in Figure 5.2
illustrates the requirement of at Icast 99"1 reliability (at most 1'
unreliability) at 1000 hours.
Form 2: The B10 life (or B1 life, or B.1 life, etc.) is required to be at least X
hours or cycles. By definition, the unit has a 10'. chance of failing
before reaching its 1310 life, a 1 chance of failing before reaching its
BI life, etc. See Figure 5.3.
80 J
~~.,,
, . . . ., , , .K /:
avu'No
UJ- -,- - -4 -
I -
Hil
I F
W. 3
10. . .. 6 .. 800.t . 1 ..tj.S.1c000. . 3 4.t

TOTAL. OPERAiTING TIME (HR)
FO 271870
Figure .5.2. Weibull Plot for 0.99 Reliability at 1000 hr
81
/1
w
Of- - - - -
C)I
I-I
C-)
D.
I. I
2.2. .S7. i-Ai0

10D . 2- .3. .L769ico.S ...0
TOA PRTN
I
TIME(R
AV 2137
~ Il ~~srto
,~~~~ f.. B1 l n 1 ie I~~
82 I
F'ig ire .i1 il IIst rat,es I ht requiruiment tl'a 7" HJ) hf)ur B I0o life.
U.9_
a.,
U-4-
0.-
iA
- I.
I.I
U.--t. -T -.- 4
0., - - .
14
' BIO81= 7000 HFIl
FO 265617
Figure 5.4. Illustration of 70(X) hr BIO Life
83
!
Reliability requirements assuming either of these two fornis can be expressed as a
minimum characteristic life requirement. (iven that the time-to-failure dist ribution is Weibull,
with a known J1, reliability at time t is a function of q:
r(t e
(.5.1)
This expression can be rearranged algebraically, giving
t
'7 == I-In R(t)I"" (5.2)
Equation (5.2) can be used to express either form of reliability requirement in terms of rl. If the
requirement is, for example, that the reliability of the turbine engine combustor must be at least
0.99 at 1800 cycles (1 - 3), then substituting t = 1800 and R(t) = 0.99 into equation (5.2) gives
1800
' - i _in(0.99)),.
or,; = 8340.9
The 0.99 reliability requirement is equivalent to the requirement that '7 be at least &840.9
cycles.
See Figure 5.5.
Similarly, if the requirement is a B10 life of 2000 hours, then substituting t = 2000 and
R(t) 0.90 into equation (5.2) gives
17
2000
1-1n(0.90)1" 2
, assuming 8 2
or i; 6161.6
Thus, the B10 life requirement of 2000 hours with 8 = 2 is equivalent to the requirement
that q is at least 6161.6 hours. See Figure 5.6.
D..igninp Test Plans
Once the minimum characteristic life requirement has been calculated, Tables 5.1 and 5.2
can be used to design the test plan.
In the combustor reliability example, the 99% reliability goal at 1800 cycles was re-
expressed as an 8340.9 cycle characteristic life goal. Ten combust3rs were available for this
reliability demonstration test. To find the test cycles required of each combustor, enter Table
5.1 with # equal to 3.0 and a sample size of 10. The corresponding table entry is 0.613. Multiply
the table entry by the characteristic life requirement to find thc test time required of each unit.
84
I.'I
. EIBf.L DISTRIBU11T I
-- 3
S=8340.9
Dz3 J M
U./
8340.9340.9
U.. -91 I
TOTAL OF'-RWTING CYCLES
F=D 271872
F'igure, 5.5. A Reliability of (1.99 at 1800 Cycles Is Equivalent to On &140,9; Cch

- -_,
II II " "
CharacteristicLife When 8=3
.. ,' -- - -
. EIBLLL DISTRIBUTIONTI
S=2
6161.6
-.
so.. - -
s ,
0.1 . ....-
- ... -. .. .. .. .. ..
SI '
U.6.U -. loa mm.---
0., - -- - - - -
__ 7-
AV271873
Figure .5.6. A 2000 hr B10 Life Is Equivalent to a 6161.6 hr Characteristic Life When 0= 2
86
I I lhe ciniistu.hr exai plel, niillily i tig I Ie 'it-; I ile 5. I en I ry itf u.t;i :i I Py tfie chari-I eris. i("
life reillreleiiw l Pit 83,14.9 ('ycle's gives a ItstI ill, oft:
0.6;3:i KI.I()09 cyc'les ,511304 cycles.
'Thus., the zero- fiailure test plan o11

(lemnonstrate 99', relialili:.'at 18(K)cycles reqtuires. l:est:ing 10
c(ombustors for '511: cycles each. If no combtustor develops a circunlfereptial crack longer than
'20 i(nches, I hen Ihe test is passed.
How many combustors are required if each can accumulate at most 750 test cycles? To
answer this, enter Table 5.2 with the assumed value of 13,the Weibuli siope parameter, and the
ratio of the test time to the calculated characteristic, life requirement. In the comhustor
example, # was assumed to be 3.0, and the ratio of the test time to the calculated characteristic
life requirement is:
750 test cycles per combustor

8340.9 cycles characteristic life 0.09
The corresponding entry in Table 5.2 is 3159. The resulting test plaa requires testing
3159 combustors for 750 cycles each. If no combustor develops a circumferential crack 1rnger
than 20 inches, then the test is passed.
5.4 TOTAL TEs r TIME
Two reliability test plans were constructed in Section 5.3 to demonstrate that a
characteristic life was at least 8340.9 cycles, with 90'e confidence.
Number of Test Cycles.. Total Test

Combustors Per Combustor Cycles
Plan 1 3,159 7,50 3,159 X 750 = 2,369,250

Plan 2 10 5,113 10 x 5,113 = 51,130
Note that, in terms of total test cycles, it is more efficient to run the smaller number of
combustors for a greater number of cycles. Plan 2 demonstrates the same reliability as, Plan 1,
but requires fewer total test cycles.
This efficiency is realized for every test plan in this section where exceet7s 1.9.
The situation is reversed for # less than 1. In this case, the greater the number of units on
test, the lower is the total test time.
When 0 is 1, the total test time is constant, regardless of the number of items on test.
S5 ADVANTAGES AND LIMITATIONS OF THE ZERO-FAILURE TEST PLANS

The test plans introduced in Section 5.2 limit the probability that substandard reliability
units will pass the tests. This is generally the most important goal in reliability testing. Also., the
test plans are simple and easy to use.
However, they are only designed to limit the acceptance of substandard reliability items.
They do not control the probability that units from a high reliability design will pass the tests.
87
................. ................................
iN some inst ances, ithe experimenter is interested in guaranteeing a high probability (if
acceptance for high reliability units.
For example. to demonsi rate that a characteristic life is at least 2044) hours with W)',
confidence, assuming ji 2.5. requires that 14 units he tested I1000 hours wit hout failure. (Enter
Table 5.2 with I1 = 2.5 and the ratio I(XX) test hours per unit/200() hours = 0.5 to find 14 units.)
Another requirement might be that designs with characteristic lives greater than 4M(X) hours
pass the test with at least 90', probability. The zero-failure test gives these high-reliability
designs only a 65", chance of passing. (See Figure 5.7.)
There are two remedies for this problem. The minimum characteristic lite requirement can
he reduced enough to guarantee a suitably high probability of acceptance for high reliability
units, or the size of the test can be increased until both requirements are met. The second option
is c(enidered in Sections 5.6 through 5.8.
The curves in Figure 5.7 assist the experimenter in determining how much to reduce the
minimum characteristic life requirements to meet high reliability requirements. They give the
prlhabl)ility of passing the zero-failure test as a function of the Weibull parameter # and the ratio
of the characteristic life of interest to the minimum required characteristic life. For example, the
probability of successfully completing the zero-failure test mentioned earlier in this section for
units whose characteristic life is 4000 hours is 0.65. To see this, enter the J1 2.5 curve of Figure
5.7 with the x-axis ratio of:
R 4(X0 hours (characteristic life of interest)

R - 'P.ours (minimum characteristic life requirement)
or R - 2.0.
The probability of successfully completing the test, from Figure 5.7, is 0.6-5.
In the preceding example, suppose the characteristic life requirement were dropped from
2000 hours to 1250 hours. Only five units would havc to be tested 1000 hours, instead of 14.
(-nter Table 5.2 with # = 2.5 and the ratio (1000 tst hours per unit/1250 hour characteristic life
requirement) = 0.8, to get the five-unit requiremen.;. To find the probability of acceptance of
the 40(M)-hour characteristic life design, enter the I = 2.5 curve in Figure 5.7 with a ratio of 40000.
hours/1254) hours = 3.2. The chanmes of passing are 88% - close to the 901" requirement.
Reliability demonstration tests that terminate successfully with no failure have one other
advantage. Very high reliability often makes a demonstration test-to-fa'lure impractical. In this
case, a zero-failure test plan is desirable. The risk is that unless some units are run to failure, the
statistical assumptions inherent in the test design cannot be validated. (For example, with failed
units, the Weibull slope parameter # can be estimated and compared to the assumed value of jI.)
5.6 NON-ZERO-FAILURE "lST PLANS
In Section 5.2, test plans were introduced to demonstrate that a lower limit characteristic
life has been achieved, with 90'e confidence. The plans assume that the unit's time-to-failure
distribution is Weibull with known slope parameter t.
88
. - \ *i' .- ' .. " -' .,
-. - ,\ ,
CD
CD0
a00
CVJ
101
-a CO)
'1a .. 0
CLC
89c
.,a-- ".---. -. =-
aa.a,.r.SS.att.-
<. t -.-.- *' --. .A ., .%- ." " -
As discussed in Section 5.5, these test:. do not control the risk of rejecting units with
acceptably high characteristic lives (called producer's risk). They only control the risks of
passing the test with a characteristic life below the lower limit (called consumer's risk). This risk
was set at l()1 .
The methods for test plan construction introduced in this section provide control over both
forms of risk if the zero-failure plans do not adequately balance the two. The methods can be
Found in "'Methods for StatisticalAnalysis of Reliatilityand Life Data-1 .
5.7 DESIGNING TEST PLANS
These tests will have the following structure:
A. Put n items on test for t hours (cycles) each. -
B. When an item on test fails, it is not replaced.

C. If roor fewer failures occur, the test is passed.
This section describes methods for calculating r0 and n satisfying the two constraints:
A. The probability of passing the test with a characteristic life as low as no

should be no more than aw(minimum life requirement).
B. The Probability of passing the test with a characteristic life as high as q,

should be at least a 1.
is the characteristic life to be demonstrated. ill is sometimes referred to as t'e "design"

characterist;c life. vis usually set at 0.05 or O.1, and avis usually set at 0.9 or 0.95.
The equations to be introduced require the definition of some standard mathematical

notation.'
A. Summation
Xi X15 + X2 + + X.
i 'I
SMann, Nancy IL. Ray K Schafer. and Noser D. Singpuwwala (1974), Methods for StatistitLd Analysis of Reliability and
Life Data, John Wiley and Sons, New York, Chapter 6, pp 312-315 and p 328.
90
" " " - ; \ " ' / '+I,!",\

B. Factorial
n! n- (n . 1).(n 2) ......... 2.1
C. Number of suhsets of size r from a set of n
(n ' ~ n!
r / (n -r
Assuming that the time-to-failure distribution of the items on test is Weibull, with known
parameter #, the following equations should be solved for r0 and n to satisfy the two "'probability
of passing" requirements.
r 0) r (5.3)
"= r
I' r.)
)r~ ((P .)
(5..l
where p0 = I-e-(Uv00
Pi = I -e-(t/' 1)
t is the test time per unit

n is the number ofunits on test
r. is the allowable number of failures
1o is the demonstrated characteristic life
"91 is the design characteristic life

i is the assumed value of the Weibull slope parameter
"0ois the probebility of passing the test with, a characteristic life equal to P0 Gro
is set by the experimenter)
a, is the probability of passing the test with a characteristic life equal to 1, (Gil
is set by the experimenter)
Equations (5.3) and (5.4) genera!:y require a computer 'or their solution. Certain computer
packages are available that solve these equations. Dr. K. E. Case of the Oklahoma State
Uriversity School of Industrial Engineering and Management (Stillwater, Oklahoma) built an
interactive program that includes the ability to solve equations (5.3) and (5.4.)'
Eqtvtions (5.3) and (5.4) generally cannot be solved for a combination of n and r0 that
satisfy the target probabilities a 0 and a, exactly. Some authors recomm2nd solving the equations
so that the actual probability ao of passing the test with il = n is no greater Ihar. ,r0 . and the
corresponding true probability a,' is at least as great as a,. The next section discusses the
method recommended by Dr. Case' for solving equations (5.3) and (5.4).
'Case, Kenneth K and Lynn L Jones (1979), "An Interactive Computer Program for the Study of Attributeii
Acceptance Sampling. Final Technical Report," Oklahoma State University, School of Industrial Engineering ared
Management, Stillwater. Oklahoma
91
..
5.8 RECOMMENDED METHOD FO'" SOLVING EOUATIONS
Dr. Case's final technical report' describes the recommended mrethod. It consists of the
following steps:
1. ('alculate p = I e-(t/1
P0 ="1
I -- -t'
2. Set r0 (= 0
:1. Find the values of n that satisfy equations 5.3 and 5.4. Call them no and n 1 ,
respectively.
4. Calculate a =i--"P" and b = p0/pl.

nip,
5. If a is greater than b, increase r0 by 1, and repeat steps 3, 4, and 5.
6. Continue the process until two contiguous values of r0 are found whose
calculated a-ratios bound b.
7. Select as the final value of r0 that which has the a-ratio nearer the desired
ratio b = p0/pl.
8. For the selected value of r0 , there are two values of n,n 0 and nj, calculated in
step 3. Average no and n, to get the final sample size:
n + n,
n f= 2
&9 PROBLEMS
Problem 5-1
A turbine engine exhaust nozzle control bearing was failing prematurely due to fatigue.
Bearing failures followed a Weibull distribution with # equal to 1.5 (a common value for bearing
fatigue) and q equal to R000 hours. The bearing was redesigned, and the environment in which it
operated was improved to give the bearing a higher expected life. Twenty redesigned bearings
were available ior testing. How long should each be tested to demonstrate, with 90`' confidence,
that the fatigue mode was significantly improved?
Problem S-2
High pressurs: turbire vanes were eroding beyond allowable limits. A significant percentage
Of the engines in servict were being removed for vane repair or replacement prior to their
scheduled turbine maintenance. The time to failure - determined by the worst vane in the set
- followed a Weibull distribution with # = 3 and i = 1300 cycles.
Through redesign and material changes the vane's durability was improved. Design a test
to demonstrate the new vane's goal: no more than 5% of the engines should be removed by 2:3(W)
cycles for vane erosion (with 90`% confidence). During this test, assume that the turbines are
'Ct, "(enneth K and Lynn L Jones (1979), "An Interactive Computer Program for the Study of Attributn
Acceptance Sampling. Final Techunical Report," Oklahoma Stat. University, School of lndusirial Engineering and
Management, Stiliwater. Okdahoma.
92
limited to running at most 5(XX) cycles each. Also.,ssume that the time to engine removal for
excessive vane erosion would st ill follow a Weihu!l distribution with ji 3.
Problem 5-3:
In Sect!ion 5.5. the zero--faihlre test plan was given to dtnionst ratte that the charatert-6.tic
life of it Weilull distribution with ji 2.5 is at least 2M(X) hours, with W', confidene It
required that 14 units he lested 10X0 hours. The test is passed if none of the 14 units fails during
t he lXK) hours of testing.
The additional requirement was added that units with characteristic lives greater than
40(X) hours should pass the test with at least 90", probability. It was shown that the 7ero-failure
test plan could only guarantee a 65r, chance of passing.
Use the methods introduced in Section 5.8 to construct a test satisf.;ng all of the above
requirements.
Solhtions to these problems are in Appendix J.
93
CHAPTER 6
CASE HISTORIES WITH Va. BULL APPLICATIONS
"6.1 F0'EWORD
A
'Thischapter provides examples of Weibull analysis used in a variety of situations. The
(examples were chosen from studies which include the complete cycle of analysis, deduction,
recornmendat ion, and imnplenientation. The case studies selected are:
(I) Tturhopiump Bearing Failures

(2) (earbox Housing Cracks
(3 ()pportutnistic Maintenance Screening Intervals
.. (4) Support Cost Model
N (5) Vane andI Case Cracking.
6.2 EXAMPLE 1: TURBOPUMP BEARING FAILURES
When this study began, three failures of the augmentor turhopump of an aircraft fighter
, engine had occurred in the field. This was an urgent problem because the failure enabled fuel to
escape and ignite. Because of this harard, top priority was assigned to the analysis of data that
a nmight help resolve this problem.
6.3 INITIAL ANALYSIS - SMALL SAMPLE
T'he first analysis was the evaluation of the three failures through Weibull analysis. Note
"that this was an extremely sma!l sample from the 978 turbopumps that were operating in the
Stfleet. The data were ranked by turbopump operating time, treating the successful pumps as
censored units. The resulting Weibull plot is shown in Figure 6.1.
ti
Fven with this small sample some valuable observations could be made. First, the very
;teep slope, il1 = 10, indicates that the failure mode is one of ranid wearout preceded by a
rlatively salfe period. Inspection of Figure 6.1 shows that the probability of a turbopump failure
prior to 20X) hours is negligible, but after 2.50 hours the probability increases rapidly.
* A second infer.,nce can be made from the initial Weibull analysis. The very steep slope (/d
" I)0) along with the existence of many unfailed pumps with run times greater than the failed
punips .uggests that the failed pumps are part of a unique batch. The method used to determine
whet her or not a given failure made is a hatch problem is to evaluate the Weibuli equation with
the parameters calculated (Figure 6.1) for each successful and failed turhopump. For each
*. pump, the probahility of failure is determined from the Weibull equation and these probabilities
iCre theo summed. If the failure mode applies to the entire flett, the sum of the cumulative
. lprobabilities should approximate the number of failures observed, in this case 3. For example:
4,
Ft
'I
,. EIBLLL DISTRIBUTION :,
0 = 10.094
' = 520.963
SAML.E SIZE = 978
-. FAILURES= 3
U
.. - . " - I- I
.5--------
j 5il
FO 271875
Figure 6.1. Weibull Plot for Augmenter Pump Rearing

!;F( ) -'.I1 - e.
S~(6.1)
where:
"ZF(ti) sum of probabilities of each unit

tV= time on each unit (both failed and unfailed)
= 520.963 = characteristic life
10.094 = slope of Weibull S
e = exponential (base of natural logarithms).
However, with these data the answer was 117 failures, indicating that the failure mode applied
to less than the entire fleet of turbepumos. Recommendations were made to Project Eogineering
that the turbopump vendor and the bearing vendor should review their processes to determine it
anything had changed, either in the process, the material, or the assembly. Initially, no change
was found that supported the batch hypothesis.
6.4 TWO MONTHS LATER - BATCH IDENTIFIED
At this point in the analysis there were seven confirmed and two unconfirmed failures. It
was observed that the serial numbers of the failed pumps were all quite high, ranging from
No. 671 to No. 872 in the sample of approximately 1000 pumps. The closeness of serial numbers
supported the hypothesis that this was a batch problem. If it is assumed that the batch started
at the first failed part, Serial No. 671, and extended to the latest pumps produced, the Weibull
equation generated fewer than nine failures. By iterating, it was found that by starting at Serial
No. 650 nine failures were generated, corresponding to the seven observed and two unconfirmed
failures. (See Figure 6.2.) This indicated there were about 353 pumps in the batch.
6.5 RISK PREDICTION
With a serious problem involving approximately 350 pumps, the next step was to forecast
the number of failures which could be expected in the near future. The risk analysis was
performed using the methods described in Chapter 13,and was limited to 353 suspect pumps.
The total operating tim, -n engines is kept in a data system that is updated monthly. It is
also known that each pump .cumulates an qverage of 25 hours operating time per month. The
risk analysis is illustrated in Figure 6.3. With the 353 pump times for the Weibull curve in
Figure 6.2, a cumulative total of 9.17 failures can be calculated for the "now" time using the
method explained in Chapter 3. Increasing each pump's time by 25 hours and again
accumulating the probabilities of failure, the value of 12.26 was obtained. The delta between
9.17 and 12.26 indicated that approximately three more failures were expected in the next
month. This analysis covered 24 months of operation and the results are presented in Table 6.1.
As the forecast indicates, almost all of the suspect lot was expected to fail within a little
more than two years. This was obviously a serious problem if the analysis was correct.
96
USh-, 99%. FAIL By 600 TOT-
ofi_ WEIBLLL
0'= S.936DISTRIBUTION HOURS
9= 462.193
' !SAMPJLE SIZE =353 --
!2.I YEAR
S S/N 650 AND UP: j
128 FAILURES INF= - -
i ~~2 340 YEARS

FAILURES IN .
zo 9 FAILURES NOW
tO. . -. ."* 4 .. 10.

. . 4 * * ... 00
Fiur 2 AugS
S/Nn 650 ND UP:
q'u~
-i. .- r r - ;.';a J- -a-r -a-. - JC- -. r -
97
Failure
Distribution
cumn% Prob of
Failure Engine Fleet Size
,l-p
-- 1 Movement of Fleet
Aging 25 Hours/Mo
Time (Hr)
FO 236M3
Figure 6.3. Risk Analysis
TABILE 6.1 I"ROJ('ECTED PUMP FAIoIJRES
Cumulatim
Failures Forecast Future Failurrs
9.17 0.0 More Failures In 0 Months
12.26 &12 More Failures In I Months
1&.22 7.06 More Failures In 2 Months
8&.44 44.27 More Failures In 7 Months
65.24 86.07 Mote Failure. In 8 Months
78.65 ftt.l Moto Failure. In 9 Months
110.1- 100.97 More Failure. In 11 Months
128,01 118.85 More Failures In 12 Months
167.21 158.05 Motm Failures In 14 Months
18&08 17891 More Failures In 15 Months
.WO.40 200.24 More Failures In 16 Months
M.f8x 221.66 More Failures In 17 Months
P5189 242.73 More Failure. In 18 Months
&=27 62.90 More Failures In 19 Months
290.75 281.58 More Failure. In 20 Months
340.34 331.18 More Failum In 24 Months
&-6.94 ,-46Z2 N - 33
98
, ,/i /
.. o -
// .
BJased tin t h)is ii 'ulvsis.
it w'is recnommende1dhl lot P'rJeci lngin ,.ring that tourl ip)U)S
No. 6W.41 an(d up with more than 175 hours of* lime be repllaed i, t-he lleet. Fort unately, there
were sufficient spare turbopumips to allow this toIbe accotmplished withbout groajding aircraft.
In addit ion, this would not have been possible wit hout t he knowledge of the relatively low risk
between 0 time and 20W hours. This action was effective as there were no more field failures.
Laboratory analysis of the failed pumps indicated that the failure mode was caused by
swelling of the plastic ball bearing cage to the extent that the balls would skid, causing the
bearing to fail. Coordinating with the turbopump manufacturer, the bearing manufacturer, and.
the plastic manufacturer, a statistical factorial experiment was designed to determine the cause
of the swelling of the plastic cages for corrective action.
6.6 FOUR MONTHS LATER - FINAL WEIBULL PLOT
Inspection of the turbopumps replaced in service (Number 6,50 and up with 175 hours or
more) revealed 15 more bearings considered to be imminent failures. The addition of these
failures to those originally seen in the field produced the final Weibull plot with 24 failures in a
sample of 387 turbopumps (Figure 6.4). Note that the original three-failure curve is a good
approximation of the final plot, the only difference being that the earlier curve had a steeper
slope (1I0 rather than 4.6) as indicated on Figure 6.1. Although this slope difference sounds large,
in fact, the inference from either curve would be substantially the same. that is. a rapid wearout
problem. The second Weibull based on seven failures was also a good approximation of the final
Weibull (Figure 6.4).
By this time the mesults of the statistically designed factorial experiment were available. It
was found that a process change had been made in the manufacture of the plastic cage to reduce
costs. The change resulted in cages of lower density. When these lower density cages were
subjected to the combination of heat, fuel, and alcohol, the alcohol diffused through the plastic
causing it to swell and crack. All such cages were removed from service. (Alcohol is a d( icing
agent added to jet fuel.)
6.7 EXAMPLE 2: MAIN GEARBOX HOUSING CRACKS
The main gearbox housing on some engines develrped cracks in the field. This type of
crack would usually be discovered during an inspection for oil leaks. Cracked housings were- .. .
"beingdiscovered at a rate of 1/20,000 hours of operating flight time. This was a ruggedly built
gearbox housing, and it was questioned whether each crack was one of a kind or whether they
were related events. Also, this identical gearbox was being introduced into a new aircraft, and it
was questioned whether the same failure mode would appear in the new installation.
6.8 INFORMATION AVAILABLE FOR ANALYSIS
Once the field was alerted to cracked housings, a quick inspection revealed 27 cracked
units. Of the 27, four housings were from the new aircraft.
At the outset,, there was considerable discussion as to whether the data should be grouped
together or a separate analysis should be completed for each aircraft type. Because of the
different missions of the two aircraft, it was decided that separate analyses should be run. The
Weibull analysis is presented in Figure 6.5 for both Aircraft 1 and Aircraft 2.
99
.j7. .. . ,
"F. "/ -,
mA..
. EI.LLL DISTRIBUTION -.-
0 = 4.61S .... -
= 488.520
' SRI4LE SIZE = 387 .-
. FAILUES = 24
CONF LEVEL = 0.90
7 1.
oil
S* .
INSPECTION OF PUMPS RETURNED TO THE

VENDOR FROM THE ONE TIME INSPECTION
PRODUCED 1S MORE PUMPS WITH
BEARINGS CONSIDERED TO BE IMMINENT
FAILURES. NO ADDITIONAL FAILURES WERE
EXPERIENCED fiN SERVICE AFTER
0.1 THE ONE TIME INSPECTION.
"o. .6. * ,.. ,.1 . . ,.L,., ,m. . 1. ,. L2 0,.0,,,

.
FD 2716"7
igu'l..I Wriblill Ihol. f[r Atim.epintor I111plp
j I'
100
s rEIBLLL DISTRIBUTION -
P = 2.842
S= 711.4925
. SAPLE SIZE = 213 -
,u.. _ FA ILLURES = 4
CONF LEVEL = 0.9
AIRCRAFT 2
03
I-I
Wz
ii , I
WEIBILL DISTRIBUTION
el-ol 0= 1 .736
v= 757.93
,~~~AM-SIZE 15I26 ....... ...
-FAILURES
,,, = 23i
CONF LEVEL = 0.9
.. 4. S.
S .. .100. 2 . 4. *S. . J1OJ. . k 14. L . 1.9.10M.

FO 271878
Figure 6.5. Main Gearbox Housing Cracks
At the time of the analysis. there were 23 cracked housings out of 1,526 in the Aircraft, 1
fleet. From Aircraft 2, 4 out.of 213 engines in the field were found with cracks in the housings.
Both curves represented wearout modes, with Aircraft 2 having failures occurring earlier and at
a faster rate (i.e., steeper slope).
101
__
_ _ .-.. / . - -, . ..- ..
tl( eadditional it c I whic'h Elh 1 Idli, nIoetd. eslxeviallv in Aircraft I. is. lu t lI
h dala do iio!
Ifall -m a straighlI lite. Or:linarilIy, datla cf this niattire indi'ate I hat there may Iih more t han otne
mode of faihinre. However, leaks and cracks of this type do not usually result in engine Iatihure
and are not disco! !red until an inspection is performed. The run time on the component at the
time of the leak or crack is usually not well defined. and the Weibull is distorted by the
clustering of events discove.ed at inspection. This would be especially true if the time between
inspections is large but occurring at specific times. One way to correct for this type of analysis
problem is to correet the data back to a common crack length. However, the correction factor
often comes under question and the easiest way to avoid this argument is to present the data as
they are obtained.
Ui RISK ANALYSIS
A ris,!c analysis for forecasting future failures was requested. This analysis used methods
discussed in Chapter 3. The r-sults of the analysis are prsented in Figure 6.6 for both aircraft
through 1982. it can be seen that Aircraft 2 has a lower characteristic lie than Aircraft 1. This
finding led to an investigation to determine if there were differences between the two aircraft
which would account for the difference in characteristic life. Strain gages and vibration pickups
were phlaed on gearboxes of each Pircraft and data were obtained. It was concluded thait Aircraft
"2was subject to more vibratory, stresqes which shortened the fatigue life of the gearbox. This
wotuld explain the steeper Weibull slope for Aircraft 2.
100
Assumptieons
90 . sWp GiG Flee
- 27 hrIGIBIMOnth
- Webul Falre Distribution/
.0 2 Weebuis
70
'30
I..
-so
S40 ~ 40 Aircraft I
20 Aircraft 2
0
80 I
~ 8 82.
Calenar YearwI
FD 256642
Figure 6.6. Cumulative Main Gearbox Housing Cracks
102
--- -,.. .. - .; . . . .. . ... - - . .-

*I ., .=/ . --
S" / \ . .. . . .
6.10 DETERMINING THE FIX
The fix for this problem was a f'airl. simple one. The (racking originated in the coverplate ,
of" the gearbox. T'hle driving force was a coverlplate resonance at certna engine speeds. The
coverpiat e was redesigned notIoi resonate at these f'reqiceniies. I hus eliminating the prol)!eni.
6.11 HOW GOOD WERE THE FORECASTS?
BIecatise of the tinme lag fronm |problem definition to the i ncorpo ration ,f a fix, adldi- lonal
flailtires niaY occur. 'l'hi:; presents an opporttunity to evaluate how well a forecast did and to
moin itior thlie elfcltc s of sample size on the WeihIill p)arameters.
'I'his problem was tracked for t1%o years afier the origim:l analysis was complete. The
findinjgs tt the initial analysis considered 23 cracks for Aircraflt and tiutr cracks for Aircraft 2.
"'l'e a-alvsis was repeated several times for each aircraft as more information became availab)le.
Resulls l'thlttle follow-on analyses are presented in T'able 6.2.
TABILE 6.2 FOLLOW-UI) ANALYSIS RESIITS
'I
Date No. of Engines No. of Cracks 0(hours)
Aircraft I
Original 1526 23 1.736 7578 3

12 mo. later 1609 41 1.782 6552.8
24 mo. later 1949 62 1.715 7038.0
Aircraft 2
Original 2131 4 2.842 711.5

12 mo. later 500 10 2.348 153..7
24 moa. later 7:12 13 1.805 :1604.8
Witlh the large numnher of cracks associated with Aircrnt I. tile Weihbull is stable. Hofwever,
Aircraft 2's Weiiull has changed considerably. 'T'his is typical and is discussed extensi',eiy in
Appendix F. Th'le risk analysis reflects the type of conservatism that would IK expected from the
results of I he initial analysis. The sieepness in the slope would cause an overprediction of the
expected number of cracks. From a rikk vie-jwpoint this could be considered as safety margin.
However, beflore any action is taken to incorporate an e;,gineering change to correct the
problem, an analysis must also be performed to determine the cost effectiveness of the change.
6.12 EXAMPLE 3: OPPORTUNISTIC MAINTENANCE SCREENING INTERVALS
Often gas turbine engines are sent to the shop because of unexpected hardware failures or
foreign object damage. Although the primary concern is the repair of the engine, the qui.. %on
also arises should the engine undergo its next scheduled maintenance while it is availablc it. inn
shop. The answer is based on economic considerations and depends on how close the engine or
its modules are to the next scheduled inspection.
For example, if an engine is in the shop for. repair after 1340 cycles of operation and is due
for a scheduled inspection at 1350 cycles (one cycle being equal to about 0.8 hour of engine flight
time), there would be no question that it should be inspected before re-installation in the
aircraft. If, however, the engine is in the shop at 1150 cycles, it is not so obvious that the 1350(
cycle inspection shou!d be performed. If the engine is in the shop at 500 cycles, it obviously
103
-~l -
N i
I "-m.l
shotIld not be inspected (coslly part rep)lac'ements are involved). There is, therefore, a break-
even point to lxe determined.
6.13 STRUCTURING THE PROBLEM
The trade to he made considers avoiding a scheduled engine removal versus the scrapp'ing
of parl s whose life is not quite used up. The opportunistic use of an unscheduled engine removal
to perform the scheduled inspection and replacement of life-limited parts not only allows saving
the laixor involved in engine removal and replacement but also allows the purchase of fewer
"'pipeline'" spare engines and modules. These savings are weighed against tae added costs
incurred by replacing parts early.
6.14 FINDING THE OPTIMUM INTERVAL
Monte Carlo simt;lation is the met hod preferred for evaluating the range of opportunist ic
maintenance intervals. The U.S. Air Force has such a simulator which (with some modification)
could be used for determining the optimum opportunistic maintenance interval. The simula:tor
is structur,4 to perform scheduled maintenance whenever the Monte Carlo process selects an
unscheduled engine removal which falls within a predefined screening interval. The process is
repeated for various screening intervals, and the resultant total support cost is plotted against
the selected screening interval to determine the optimum. Figure 6.7 is an illustration of this
procedure.
The Weibulls are used to describe each of the engine modules' major failure modes (reason
for unscheduled removal). Where inmprovements have been incorporated, the Weibulls are
adjusted to reflect the improvements. Only with a valid representation of the way in which each
removal cause varies with time could a realistic assessment be made.
The simuulation analysis was performed and an opportunistic mai,,ntance interval of

"30)
cycles was determined. This provided the Air Force with an economio decision criterion for
perorming scheduled maintenance.
6.15 EXAMPLE 4: SUPPORT COST MODEL
The support cost model uses a Monte Carlo approach to simulate the interaction of
scheduled and unscheduled maintenance events. The unscheduled events are entered in the
form of Weibull curves relating event, probability, and time. Scheduled events are entered at
specific times. A screening interval is input to define a time period during which scheduled
events can be precipitated by unscheduled opportunities. (See Figure 6.8.) Labor and material
costs are input for each event. The model selects corresponding labor and material costs for each
event and cm..npiies totals for the number of events and for labor and material costs by report
period (year). Totals are divided by the number of flight hours for the report period to derive
rates per flight hour.
The model matkes a predesignated number of passes through the life cycle (20 years) and
Creplorts the average of the passes by report period. The number of events per year can therefore
appear as a non-:nteger.
104
,S!t*
aInputut~l Outputl
lioutiel
*S.Ppeline [teen mantea j
Transporl
Mamtenance
Parts
Unectheduled Engine
Remova:. (UE.A
O"Crl'NSO
Sceduiled ~oal $nw
Removals (SER)
Sa" nt~erwal
Reemm of Conputw Model "Oppofsmamei Maintenmae Engine Snuai'on" Peald by

John Madden AFLCIXRS
FD 256843
Figurr 6.7. Approac3h T'o Optimizing Scheduled Maintenancet
Inpection Interval
I
I
-Opportunistic Maintenanc Wdteval
I I Unec~bedWe
Engine ReMoval
-
,.,
TieLkw ------ " ""I " I

u " i
Flight Hours or LCF Cycles
FO 25d544
Figure 6.8. Scheduled and Unscheduled Maintenance Interaction
"105
- x.
II lns.he.duhl(4 laihtt ui. n evenis. i 1111 -nj.asWeihll.s. are 4f,1"lour basic tvp.s: I)
un.chedulled engine renuI.vals (I lols). 2) tins.heduhed intsuhle reit.vals which are coincidental
witlli an engine removal (couincidentals). A) un.scheduled part remnovals which are coincidental
with a moalule removal (part coincidentals), and 4) installed maintenance events. As stated
above, each of these event inputs is accompanied by a corresponding labor and material cost. It
is also accompanied by factors which designate those percentages of events which precipitate
engine o.- module depot visits and demands for spare modules.
Scheduled event input is also accompanied by labor and material costs per event. Material
is input as a total cost of parts involved in the inspection along with e percentage to be scrapped
at each event. This scrap rate can vary among the events of a particular sequence.
The input is then a combination of Weibulls, scheduled intervals, material and labor costs,
depot visit factors, and supply system demand factors. Output is reported at the module (failure
mode) level, by report period (year), in terms of total quantities and rates per flight hour.
Parameters reported include engine removal and depot visits, module removals and depot visits.
module demands, labor and material cost.9 broken down by depot and base, and scheduled vs
unscheduled maintenance ior each report period and for the total life cycle.
&.16 ROLE OF THE WEIBULL
Unscheduled engine maintenance, as indicated above, is driven by both scheduled and

unscheduled events. The unscheduled events are caused by som'. failure modes that occur
randomly and others that exhibit wearout characteristics, i.e., an increasing failure rate. The
Weibull is the most convenient method of introducing these increasing rates into the model.
The Weibull is described by only two parameters, the characteristic life, ., and the slope. l3.
Figure 6.9 illustrates the use of Weibulls with 0 > I for life limited parts and 0I = I for randomly
distributed failure modes. Infant mortality, although seldom encountered in an operational
engine, can also be simulated by 1 < 1.
S> I Randomn(P-1
0.632 //
AePlIeIment
Fal e I II
,~I Characteristic
Fght Hours or LCF Cycles
FD 25S45
Figure 6.9. Unscheduled Maintenance Input via Weibulls
106
6.17 EXAMPLE 5: VANE AND CASE FIELD CRACKS
Cracks !'ismld in the 12thlstage vane and case

,'ei 1a high prpsure conpre.sor precipitated
Ihis -it dY. There was concern of case rupture il the cracks grew large enough to weaken the
,ruct tire. 'Ile major quest ions Were:
(I) How will the problem affect engines?

(2) What can be done to fix the problem?
(3) Can the problem be detected through inspection?
(4) What recommendations should be made to the Air Force?
6.18 RESOLVING THE OUESTIONS
Seven cracked cases were iacntified. The cracks were all of different lengths but shorter
than considered critical. One question whicri often arises fromo this type of analysis is whether to
normalize tlie times to a constant crack length. !t was decidee" to proceed without the corrections
on the times and construct a Weibull from the available infirmation. Figure 6.10 is the Weibull
basud on seven cracked cases.
The rate at which the fleet would run into this problem was then examined. It was assumed
that all engines were susceptible and that the crack could be detected upon inspection. These
eases would be repaired by welding, and the units would be placed back into operation. This
inspection and repair could be continued until a fix was in place. The more permanent fix was to
hardcoti the area of the cracking with a plasma sprry. It was also assumed that the fleet would
accumulate an average of 27 hours per month on each .ngine.
The engine would normally undergo inspection at 1.50 cycles. This is equivalent to about
10), hourn of engine operation.
With these assumptions and assuming that the hardeopti fix would be app!ied to all new
engines, the number of unscheduled engine removals due to this problem was projected using
methods described in Chapter 3. The results are illustrated in Figure 6.11. The forecast of 10 or
more engines develol*"ig cracks by the end of the first year and the total reaching about 40 by
the end of the following year resulted in implementation of the hardcoat fix.
6.19 CONCLUDING REMARKS
The plasma spray hardcoat has been incoirporated into production units. In addition, as old
units are received for their normal overhaul, hardcoating is applied to these units as well. At this
writing, a total of only 15 engines have been identified with cracks over the critical limit where it
could be said from the forecast that 40 additional engines would have been expected without the
fix. The quick action by the Air Force to implement the fix resulted in correcting the condition
in the field.
107
M..t
- . r , .T,-
-7.7
r *
.~
""EIBtILL DISTRIBUTION -
A = 4.117
90._
Ii._
SAMR-E SIZE
FRILLRES 7
= 1338 -
70.-
.,, SO.- 2. 3. 4.5 6..8S..O .
a20
I1z
"2 I. ,
'1U
%
t"
u W. 2787
li41E
.- .0 11,V 1,( 'na rakn
Sto. 2. . 4. S. 6.7 t.q.jdoo I

2. J. 4 .6 . .9.1O
. .2S . 1 m
I TOTAL OPERATING TIME (HR)
S~FD 271879
SI~i'u ' t; I0 12t1h Vaine and Case C'ractimA,
i ~108
80
701350 so~
TAC'$
70
Tram of Fwill
.50
40
4 130
j20
to
0
1901981 1982 1983 1984
Calena Yew
FO 25664?
Figure 6.11. Expectced IIER's Due to 12th Vane and Case Cracking
1
It
II
/
CHAPTER 7
CONFIDENCE LIMITS AND OTHER ASPECTS OF THE WEIBULL
7.1 FOREWORD
Now that some familiarity has been developed with the Weibull distribution and its
app)lication in risk analysis and life testing, further applications will be discussed. First among
these will be confidence intervals about the Weibull parameters # and n and about the Weibull
line. Secondly, special applications in risk analysis will be discussed, namely Weibull
"T'horndike" charts. The next topic to be discussed will be shifting Weibulls in the case of
insufficient information about the underlying population. Lastly, other options available when
the Weibull distribution may not fit the failure data will be discussed.
7.2 CONFIDENCE INTERVALS
Confidence intervals are measurements of precision in estimating a parameter. A

confidence interval around an unknown parameter is an interval of numbers derived from
sample d;,ta that almost surely contains the parameter. The confidence level, usually 901t or
higher, is the frequency with which the interval calculation method could be expected to contain
the parameter if there were repeated applications of the method.
7.3 CONFIDENCE INTERVALS FOR 8 AND i
Often it is of interest to determine how far from the "true" value an estimate of 4 or ,4
might deviate. For example, if the times to failure of every bearing ever made and every bearing
to be made in the future were known, it would be p(ssible to calculate # and q exactly. But, of
course, this is never the case; only a sample of bearings is available. The question is: how much
variation can be expected in the estimates of # and , (4 and qi)from one sample size to the next?
If this variation is small, then the particular sample will yield estimates close to the true values.
The problem involving censoring witn very few failures is not dealt with here. Reference (1)
is recommended for this situation. However, for large, complete (no suspensions) samples of size
n. the confidence intervals for # and ,qcan be approximated by equations (7.1) and (7.2),
respect ivly.
eXl, ( -0.78Z,,/. 2 0.78Z,./. (7.
\d -.5,/21 ! x/ 1.05Z,,/2,'
( - nexp -- xp (L5Z0 2 (7.2)
where Z.,, the upper a/2 point of the standard normal distribution, depends on what
confidence level is chosen. Table 7.1 gives Z.4 for various (usual) confidence levels.
TABLE 7.1. CONFIDENCE LEVEILS
Confidence Level 4,12
99% 2.576
95% 1.960
90% 1.645
"'Apolied Life Data Analysis, Nelson, 1982.
110
These confidence intervals are only approximate since: (I) the estimates of 11 and i9 used are
linear regression estimates (from a theoretical standpoint maximum likelihood estimates would
be required, see Appendix DI, (2) these estimates are only approximately normally distributed.
Example 7.1
Figure 7.1 shows a fitted Wt'ibull distribution with 45 failures and no suspensions. A 901,
confidence interval for 0 is desired. The relevant information is as follows:
94r
1.84
Confidence level 90",
7,,12 :- 1.645 (.rom Table 7.1)
Substituting into equat ion 7.1,
1.84 exl) (0 .78(.64) - 1.84 exp (0.78(0.645))
which reduces to 1.52 :5 # : 2.23
Example 7.2
Using the Weibull from Figure 7.1, what is a 9V0% confidence intervul for q? The relevant
information is:
n = 45
95.98, it = 1.84
Confidence level = 90',
7,,/2 = 1.645 (from Table 7.1)
Substituting into equation 7.2,
98.88 exp (--1.05(1.645)) <1.05(l.645))

988,x-84r " 958.88 exp 1.840(.5
or, 833.7 S i-1102.9
111
"_.s WE ILLL
01 I -
0 = 1.84 -
S= 958.88
um.. SPORE SIZE=45
-. ,FAILLRES
=45 -4S -
I I'iv II III
.. 3 11 i 1.o. 2 1610M
FO 271O80
Figure 7.1. Weibu-! Test Ce
. a ... im. . F7 .gaa ; ,"-

TORLOERTIN YCE
112
P0 27188
/ igu 7...bulT/tCa
,
112i
* 7.4 CONFIDENCE INTERVALS FOR RELIABILITY
Anot her p~roblem t hat appears in Weibull analyses it; t hat of obtaining confidence intervals
for the reliability ait a given point in .1imie. Tlhe reliab~ility at the point. Cis the probability of a life
of ait. least .' units, and will he dlenoted bky [Wt). Again assume a large sample with no censored
oliservat ionls.
The procedure is as follows:
1.Compute u = (Qn(t) - 2 n(4n))# 7.a
2. Compute Var (u) = 1.168 + 1.10 (t2) -0.19131 u J.(7.3b)
:3. Compute u, = u Z.,2 IVar(uN/ (7.3c)

AA
U. u + Z,,12 I1Var(u)I1 1/2 (ZO, from Table 7.1)
4. Then the confidence interval is: exp (-exp (U2)) :5 R(tV) -exp (-exp (ul)) (7.3d)
Example 7.3
Again using Figure 7.1. a 90"f. confidence interval for the reliability at 700 cycles is desired.
The step-by-step procedure follows:
A.u (n70 Qn(9158.88)) 1.84 = -0.579
2. Var (^u) = 1.168 + ( -0.579) 2(.1.0) - 0.1913 (-0.579)J - .03

O-
:1. ul -0.579 - (1.645) NTOiO.6 = -0.890
U2=-0.579 + (1.645) V05 _06 = -0.268
4. exp (- exp (-0.268) ) :s R (700) 5 exp (-exp(--0.890))
__ -Therefore,the-confidence interval is 0.465 :5 R(700) :s 0.663
7.5 CONFIDENCE INTERVALS ABOUT A FAILURE TIME
Engineers are often interested in a confidence interval for the time associated with a given
failure. This confidence interval can be approximated by equation (7.4). Ninety percent
confidence intervals will be assumed for all confidence intervals about the Weibull line in this
section.
t~~, 00 [n1- F1(0.05) ]i' tI .5 Q - F.(0.95)]i (7A.)
where ti, 0.05 and ti, 0.95 are the failure times associated with the ith failure and F1 (0.05) isRthe
5";. rank associated with the it 1' failure, while F3 (0.95) is the 95 *'rank associated with the i1t'
failure. Tables for Fj (0.05) and F1 (0.95) are in Appendix B.
113
Example 7.4
S~)Ispvm we are given Ihe Weieilml int Figure 7.2. if 2.0. it Il(H himirs. prwd'ced Iroumt 10
Iaitilires. The ah'cidallhat
imi n he ,fideim'e iilherval a itiu I lie fiml f'ailihre (21. I hotirs)
hire fior In'
foI;fws:
FI (0.05) 0.00-5) (froti Appendix 1)
1[, 0.05 (IM)[n I 0.005[i' 1 10.J'5l1l;i/2
= 7.08 hours
F1 (0.95) 0.258 (from Appendix B)
t1 , 0.95= 100 [Qn 10258"] = 10010.298411/2
= 54.63 hours
7.6 CONFIDENCE BANDS ON THE WEIBULL LINE
In Section 7.4 the confidence bands about a single reliability were calculated. Simultaneous
confidence bands can also be placed on the Weibull distribution fir complete samples. 2
lel,,nce,(2) contains the basic information for their construction. The resuls in Reference( )
have been extended t) the Weibull Distribution. Equation (7.5) together with Table 7.2 can be
used to cah'ulate 90", confidence bands about. the Weibull Distribution.
(F(x) K(n),F(x) + K(n)),
where
F(x) = 1 - e
(7.5)
and F(x) is the estimate obtained by substituting maximum likelih!od estimates for the
parameters.
1'2"An Approah to the Cmnstruction of Parametric Confidence Bands on Cumulative ritribution Functions,"
Srinivasan and Kanotsky, Biometrika, Vol. 59,3, 197Z
114
S.. 4 " /.
-< -, 7'.o
wEIBULL DISTRIBUTION - / -
P = 2.0
S= 100 HOURS
"SAMPLE SIZE = 10
-,. FAILURES =10
s-o
__ i
. ..... .. . ...... --....
5: .. . ... ..-
O.302. ____
w
w
-
w sI.- _____
LI..
: : "J.:
/
I. 2. 3.4~S 0910. 2. .LSGO0* *. .s~~oo

,
TOTRLL OPERATINg.. TIME (I-RS)
FO 271881
Figure 7.2. Weihil/ IPlot Where( # : 2.0) and ,q - 1(M) fir 10 Failures
115
TABILE, 7.2. COI
)N ID'El)IN(CE'
BOUIINI)D ON TIl'II
WII llIJi1 LINE
Sample Size (n) K(n)

3 0.540
4 0.420
5 0.380
6 0.338
7 0.307
8 0.284
9 0.269
o10 0.246
11 0.237
12 0.222
13 0.213
14 0.204
15 0.197
20 0.169
25 0.152
30 0.141
35 0.125
40 0.119
45 0.117
50 0.106
75 0.086
1oo 0.074
Kxample 7.5
Consider the Weibull in Figure 7.3, 0 = 2.0 and qp= 2000 hours, with a sample size of 7.
From Table 7.2 the critical value of KM(7) = 0.307. Therefore,
1- (x/2)'"- 0.307 _sF(x) _<1 -e-(XM20 + 0.307
for all x, 0 _ x < ou, with 90", confidence.
These Imnds are illustrated in Figure 7.3.
7.7 WEIBULL "THORNDIKE" CHARTS
A graphical method often used to determine the cumulative probabilities of the Poisson
distribution was named for F. Thorndike and is illustrated in Figure 7.4.
A random variable x has a Poisson distribution with a parameter 1W if P(X = x) = exp (- p)

;j,/x !. x = 0, 1, 2, :1..... ). (The Poisson distribution also arises as the limiting form of th,!
binomial when the sami1 le size becomes large.) As an illustration, suppose it is necessary to make
the statement: the expected number of occurrences is 3.0, and the actuai number of occurrences
will be between x and y with 0.90 probability. To find x and v, use the Tiorndike chart in
Figure 7.4 enter the x - axis at 3.0 and read up to the point wi,-.e 3.0 intersectE the 0.05 and
43.95 lines extending from the y - ax,'s. The values for "C" are found to be about 0 and 6.0
respect ively. Therefore, with probabiity 0.9, if the expected numbpr from a Poisson distribution
is :1.0, less than 6.0 will occur.
116
/ ... . _
S.7"
W-EI_LL DISTRIBUTION
0 =2.0
= 2000.
._ SA4MPLE SII7Z = 7
.- FAILURES =7
S. -- -
CIA
3. 4 . 7.1. 100. 2. 3. 4.10 . 2. . .8..10000.

FD 271882
Figure 7.3. Example of Confidence Bands on a Weibuli Line
117
/
0999C=6 7 8 9 10
0.
I5
0 0.99 C,,
S0.8 -- C-0i
0.07-
: 0.500-
0.4-
0..
0.3 25
~0.2 - S
- S0.1
0.601ii!
0.01 -
- S~0.001
0.0001
0.1 0.2 0.3 Vau

0.4 0.5 0.6 0.8 2f o: 3uPie
Value of Mu Prime
FO 2W565
Figure 7.4. Cumulative Sums of Poisson (Thorndike Chart)
118
I' A 1+
III I his sionn waY a graphical lechnipieit sioniiitr so the Tlhorndik&e chart has been developed
for Iiw( Weihull 411 isriomi olt Thiwsi charts gieI*thw lirshabililIy of*Ii...ig "C or fewer failuires by
III?V-ivenl I111Wv. The1y v Ilt aI-,u Ii~
1Ks.
IS 41 IN1it, how Itihenls'I oII fi-ailsaro- EwCUrring b~y
lIIItlIs ai~,i
11giv4en Iiime..
Figuires 7.5 1hrotigh 7.12 are. Weibiall Thornd'ike charts for )i's of04.5. 1.0. 1.5. 2.0, 2.5. :'.0I
1.0. arid 5.01. To Ilse t hese chart.,. determine the time Mt of interest (liossibl) the inspect ion
time), calculate thy (q~ is the characteristic~ life from the Weihull). enter the x - axis of the chart
with the closest 0 to the one of* in terest, and then read the probability of having "C" or fewer
faiilures. Several examples of this technique and other uses follow.
The usefulness or this information can arise, for example, when the inspection interval is
two or more times the characterist Iic life of the Weibull failure mode of a part. When this
happens, the part fail, is replaced (made "good as new"), fails again, is replaced again, etc. How
often can this process continue? The Weibull Thorndike charts answer this question.
E'xample 7.6
(Given Weibull parameters 4 = 1.5 and qj= 3000 hours, the probability of having three or
fewer filtires per unit by 60MN hours is to be calculated.
It this case, t/j Ohi4NN)/:N --- 2.0, and 41.5, so using Figure 7.7, enter the x - axis at till
2.0) and proceed uip to the point where the line -C z-3" is intersected. Then the probability of
observing:1 or fewer failures can be read frobn the y - axis as 0.9:3.
Exam ple 7.7
Suppose 4-1.5 and now t/q - :1.0. A 0.90 probability band can be placed abomut the
number of failures occurring by ti 7 3:.0.
Again using Figure 7.7, ent -r the x - axis at 3.0 and proceed to find the "C" values where
0.0(A at d 0.95 probabilities intersect. This yields I and 5, respectively.
E'xample 7.8
In spare parts provisioning, suppose the number of spare parts to be provided are required
for a part having a #4 3.0 and an inspection time/characteristic life ratio it/q) = 2.0. The
menager wishes to boe 90', confident that ho will not run out of parts. Using Figure 7.10,
entering the x - axis at 2.0 sad proceeding to the point where 0.9 on the y - axis intersects the
"C",lines, no more than two spare parts are needed per delivered part.
E'xample 7.9
A new design rotor bearing has been testEd for 22,00M hours. The current rotor bearing has
a limiting failure mode whose # = 1.7 and q = 4,9317 hours. Six failures, have been observed in the
test of the new desigr due to this mode. Is this unusual? With a t:7 ratio =22,000/4,9.37 =4.45,
entering the Weibull Thornidike chart for 4 = 1.5 (Figure 7.7), the p~robability of having six or
more failures is approximately 1.0 - 0.90 =0.1. Entering the Weibull Thornidike chart for 4= 2.0
(Figure 7.8) the probability of having six or more failures is approximately ItC - 0.92 0.08.
Therefore, it can be stated that the probability of observing six failures by this time in the
redesigned rotor bearing is from 0.08 to 0.10. Hence the redesigned rotor bearing is not as good
as the current bearing.
119
p0~
coo
40
C4
csv
120
C 41
In U)
I 04
Im 0) 0. 0
09 jo0; oq~ 4.jw
121
LO
CC
77j
C!
040
I-D 41 V)
SS91 Jo 0 jo souejjfl-,)3 jo qoic -
122
coo
0
InZ
1123
a
0 0
U.
In
C)
0
U
In
C)
S
C.) R.
I-
a
C)
C'. C
0
0
In
a
'0
o
E.
L
* 0
o
Cl '4
a
a..
C
R
.0
0 . I-0In40C'. Y V. V. -
.0 0
a SW1JO 0 .oueJJnooo JO 0Jd
124
1/2
C,,I 8
na o0; ool~ oqj
12
(IN
ql ORp'-pU ItciC
0 0
swl m~mo
o 0jo oqpdwu i
126
ae
cmI
cs 0
cio
dI
owl ojoqo0
0 o suwmmo
127.
7.8 SHIFTING WEIB.ULLS
In nIlIIlInht has IlKeeii don(. utp to thlis pointI Weiln IlI nilutre (list rihu t ions have been est inlated
In im oblservedl ftiulures, oiiten iii coi~mbint 4ionWit Ill IhelINipiun I io ns of'1tilfi led IInits. W hat call he
donle if the I inles onl each finiluireare known hut the timies on Ilite ;spiiut Wuntiailed units are
iiiof
unknown?
Tlhis p~roblem arises in the failure analysis of data from jet engines because all parts irn an
engine are nilt serialized; that is. the time (In the individual p~arts cannlot be tracked. In many
engines only the most important 4(M-r-A() parts are serialized, while the others,, (possibly as many
ats10014N)) are not.
01f course, if* the pairt f'ailure times are knowni, the engineer call generate at Weihull
distribution from the failures only, using the methods in Chapter 2. Fred IDauser, Statistician.
Commercial Products Division. P~ratt & Whitney Aircraft Group, United Technologies
Corporation, developed a met hod t0 "adjust" this Weibull if the number of untailed units in the
lplpulation isknown.
Ani out line of the met hod is as follows:
1. P'lot the failure data oil Weibull probability paper.
2. Kit imat e the Weibull parameters 11and q..
:1. Calculate the mean time toI failure OXMTl'R
mT'rF - times to failure for each part

No. failures(75
(now refer t(I Figure 7.1:3).
4. D)raw a vertical line through the MTTF.
5. (Calculate the proportion failed in the toItal population, No. failures/(No.

failures + No. suspensions), calculate the cumulative ",failed point = (1 -
e-Ill't")x 100, and draw a horizontal line from this point.
6I. At the intersection of the vertical and horizontal lines draw a line parellel-
to the failure distribution. This is an estimate of the "true" Weibull
distribution.
Suppose there have been four flange failures with times of 1165, 1300, 1393, and 1493
cycles iria population of 2500; however, the times on the unfailed units are unknowvn. The
proce(Iur~ to estimate the "true" Weibull distribution can be used:
Sleps land 2: See*Figure 7.14, 9.531, q 1400.7 cycles
St p3::
MTTF 116.5 + 1300 + 1393 + 1493 137

MTTF ~4 13.
128
No. I'iilurvs/(No. faildures I No. s.u.ppi.slions) (.I/2500)) t0.00)1;
Th'iverf,re, cuml ', hailed II v, 0.Wll;1 0 I. I61

0ti)
S leps .I. 5 and I;:
See Figure 7.15
The estimated distribution has a 4= 9.53 (same as the four failure Weibull), but the
",haracteristiclife
is now = 2629 cycles.
7.9 WEIBULL GOODNES3 OF FIT
The procedure to test whether a sample is from a specified Weilull dictribution can be
given in terms of the confidence bounds about the Weibull line developed in Section 7.6. Again,
complete samples will be assumed.
Procedure:
1. Using the Weibull estimates of il and q for the failure distributions,

calculate and plot the confidence bounds using the techniques of Section
7.6.
2. Now place the hypothesized Weibull on this same plot, as a dotted line.
:1. If this dotted line does not. lie entirely within the confidence bands, then
consider the sample to be from a different Weibull distribution.
IExample 7.)!
Given a sample of seven failures with 4 2.0, 17 = 2000, as in Section 7.6, can the
hypothesis that the sample comes from a Weibull distribution with 4 = 4.29 and 4 = 1500 be
rejected?
Plotting the hypothesized Weibull as a dotted line on Figure 7.3 gives Figure 7.16. One
would have to reject the hypothesis that this sample comes from a Weibull distribution with 4 =
4.29 and i1 1500 since the dotted line does not lie entirely within the bands.
7.10 COMPARING THE WEIBULL TO OTHER DISTRIBUTIONS
Figure 7.17 shows what happens if failure data from log-normal, normal, and extreme-
value distributions are plotted on Weibull probability paper. Since the log-normal is the most
frequent alternative in failure analysis, this will cover most of the practical cases that arise. For
example, suppose two plots like those on Figure 7.18 are given. In this case, the eye is unable to
discern which is "best," the log-normal or the Weibul. The statistical test from Referencet 3 ) can
be used to discriminate between these two failure distributions. The test can be set up in two
ways: to favor the log-normal or to favor the Weibull.
"t "Discriminat:on between the log-normal and Weibull Distributions," Dumonceaux and Antle, Technometrics;.Vol.
15,4,1973.
129
/M
... -. - : - :FAILURES
-WEIBULL OF
, ONLY
ESTIMATED -TRUE-
, IWEIBULL
DISTRUBTION
8X
IA
MTI
.. . .... + - --
L . le ''o.a.arl 0 S
.1.
. - ." EF
M.W
TOTL
OERAINGTIME (HR)
Figure 7.137. Example of Shifting a Weibull
130
=' .__WW LJ LL_ ,-r-
-. *L"-:
--JIEIBLLL =4
DISTRIBUTION t - - -
1400.7
SP,=MLE SIZE4 .1
,.. FAILURES=44-
40....__ _ -_
-. .4:
.I -_I__
I-
a.-
, -- - -
UI
00. . . 4. I, UO, ! . I.SIOO

. . 4. 14. . 8 9.10.
TOTRL OPERATING CYCLES
FD 271884
Figure 7.14. Flange Cracking
131
_ EIBLLL DISTRIBUTION
Pg.- 9 :S3
0 = 9!3 "..
1400 .7
SAMPLE SIZE= 4 S
-.. FAILURES= 4
e,= =
I H
za . - -
QI.-
IIll
ESTIMATED
DISTRIBUTION
.=9.$3
9 .2629-
SAMPLE SIZE=2SO0
FAILURES=4
MTTF
w.. I. '..8...100m.
TOTAL OPERAT.IN CYCLES
FD 271S86
Figure 7.1.5. Flange Ciacking Estimated Distribution
132
J I I
/
nU.
. _.. - .HYPOTHESIZED HEIBULL

-(P=4.29, T = 1500) --
S -... ESTIMATED WEIBULL

is. FROM SAMPLE 000
(D= 2.0, 1 = 2000)
S' I
NOTE:CA
I's. - CONFIDENCE BANDS CRN
BE VISUALLY "TRIMMED"
AFTER PLOTTING
Iv
01. -. - . m.t-1 4 11
O.L * * * ~TOTAL OPERATING CYCES . .4. JOD
FO 271W87
4:m im
Figure 7.16. Hypothesized Weibull
133
True Appearance on
Distribution Weibull Paper
Cun.
Normal %
VI
Time
Cum
Log-Normal %
Time
Cum
Extreme-Value %
(Largest extreme)
Time
FO 2S6466
Figure 7.17. Picking the Best Distribution
134
COVERPLATE
99.9 ---
0
w
S 50.0 363.2
-- NEIBULL DISTRIBUTION
10.0 '= 4523.8

U SAMPLE SIZE = 20 S
ILURES 20
w
1.0
cc
-j
100 1,000 10,000
99.999 LOG-NORMAL DISTRIBUTION

0
W MEAN=8.250 (4104.109)
M STD DEV =0.373093
U 99.0 SAMPLE SIZE = 20
SFAILURES 20
z
w
w 0.0 -
a-
cc
-j 1.0
0.001 "
100 1,000 10.000 100,000
FD 271888
Figure 7 18. Comparing Weibull and Lognormal D)istributions for Coverplate Failures
135
Favorhig tIit- Wcilmll. 4ah'ualdat,
W - (2wet)+' 112 f(lllt,ll:'I.)I

I ... l'_1Il) ]
-. n is sample size
p -WnW, t, is the time of failure
an ~)e *the We'lbull probability density function
W is compared it the appropriate table value for the confidence level desired (Table 7.3) and, if
W "Wtabk the Weihull is rejected in favor of the log-normal.
TABLE 7.3. CRITICAL VALUS FOR TESTING

THE DIFFERENCE BETWEEN LOG
NORMAL AND WEIBULL
(FAVORING THE WEIBULL)
Nimuber of Confidence Level
Failures 80% 90% 95%
20 1.008 1.041 1.067
30 0.991 1.019 1.041
40 0.980 1.005 1.026
5') 0.974 0.995 1.016
Favoring Ihe log-normal, calculate
W (2 r e a) z " ( t ,f At ,) l l tf(t -2)

l ... It, f )(0t, 7 .7)
and compare to the appropriate table value for the confidence level desired (Table 7.4) and, if W
2 Wi,, the log-normal is rejected in favor of the Weibull.
TABLE 7.4.. CRITICAL VALUES FOR TESTING

THE DIFFERENCE BETWEEN LOG
NORMAL AND WEIBULL
(FAVORING THE LOG-NORMAL)
Number of Con.idence Level

Failures 80% 90% 95 %
20 1.015 1.038 1.082
30 0.993 1.020 1.044
40 0.984 1.007 1.028
s0 0.976 0.998 1.014
Example 7.12
The coverplate failures that went into the plots in Figure 7.18 occurred at 1989, 2160, 2569,
27.8, 2813, 2979, ,016, 3283, 3294, 3503, 3853, 3916, 4294, 4462,5178, 5716, 5984, 6378, 6556, and
70(X) cycles. The estimated Weibull parameters are = 3.26, = 4523.8. The estimated log.
136
'4 normal Ipratieters
with SO', cnfiheince?
Atre, A
p 8.250 11. , 0.373093. Which fiailtre (list riI it ion fits thelata I Ixecr
Mol h tests. favoring first the Weihull and then the log-normal, will be performed.
i! FFavoring the Weilull, usirg equation (7.6),
2 2 17(X)0 f (700)o]1 -1)20

W - 12(3.141592) (2.71828) 0.373093 1-1 (11989 f (1989)1 ...
:U-= 1.039
comlaretl to a table value of 1.008. Hence, reject the Weibull.
Favoring the log-normal, using equation (7.7),
W - [2(3.141592) (2.71828) 0.3730932]+1/2 (11989 f (1989)1 ... [700 f (7000)1 ) +120

-- 0.962
compared to a table value of 1.015. Therefore, since the value W in the test that favors the
Weibull a Wble, the Weibull can be rejected in favor of the log-normal. This same decision is
reached in the test favoring the log-normal; in this case, since W s Wtable, the log-normal
cannot be rejected in favor of the Weibull.
In conclusion, the log-normal failure distribution seems to describe this coverplate failure
mode better than the Weibil distribution.
7.11 PROBLEMS
I. (iven a Weibull derived from 40 data points with / = 1.5, V = 2000 hours; what
are the 90', confidence intervals for # and i?
2. What is the 90',, confidence interval for Reliability at 1500 hours in problem 1?
:1. What are the 90', confidence intervals about th3 first three failures in problem
S! 1?
4. Given a Weibull with parameters 0 = 1, V - 1000 hours, what is the 90%

probability band on the number of failures to be expected by 4000 hours?
b. A 10 point Weibull of failures only was generated and is illustrated in Figure

7.19. These failures are of a non-serialized part with a total population size of
2000. Adjust this 10 failure Weibull for the entire sample size. Note: failure times
are 51, 79, 116, 164, 197, 230, 232, 327, 414, and 451 hours.
6. Are the Weibulls in Figures 7.20 and 7.21 significantly different? Assume the
"Weibull in Figure 7.20 is true.
137
4%;_ HEIBULL DISTRIBUTION _
=1.59
0 = 2S8
.. FL SIZE = 10
... FRILLUES ='10
U.. - ---
SM ........---- - ... -
I
S. .
.. -. 11. . 11 At
0If. - .. - - - -

FD 271890
Figure 7.19. Weibull of Non.serialized Farts, Problem No. 5
138
' W EIBLLLDISTRIBUTION --
SAMPLE SIZE = 10 ---
N. FAILLRECZ=10
--. - -------
MTTF
0.. -.-. -i M L-1 4- .4 L
Figure 7.20. True WeibulU, Problem No. 6 F 71
139
1.3LIL DISTRI1UTION
0 =2.97
t=895.4
U.SSHPL SIZE = 10
-FAI 1
tiIl iiII I MII
Ga-----44
an. . . . .*.
6. . . . . *.u A 7 2. .L . .. 6g4OH

FD 271893
Figure 7.21. Suspect Failure. Mode, Problem No. 6
140
I 140
I
APPENDIX A
GLOSSARY
1. 11"Beta) The parvieher of thle Weilhill distribution that determines

itS sphae and Ihalt implies the failun'r nmnle clutra'terislic
(infant mortalily, random, or wearout). It is alsocealled tihe
slope parameter because it is estimated by the slope of the
straight line on Weibull probability paper.
2. Bias The difference between the true value of a population

parameter and the grand average of many parameter
estimates calculated from random samples drawn from the
parent population. Also called fixed error.
3. Censored data - Data that contain suspended units.
4. Confidence - Relative frequency that the (statistically derived) interval

contains the true value being estimated.
5. Distribution - A mathematical function giving the cumulative probability

that a random quantity (e.g. a component's life) will be
less than or equal to any given value.
6. P1 (Eta) - The characteristic life of the Weibull distribution. 63.2("%

of the lifetimes will be less than the characteristic life,
regardless of the value of tf, the W-ibull slope parameter.
7. Hazard Rate - The instantaneous failure rate.
8. Infant - A failure mode characterized by a hazard rate that

Mortality decreases with age, i.e., new units are more likely to fail
than old units.
9. Monte Carlo - A mathematical model of a system with random elements,

Simulation usually computer-adaipted, whose outcome depends on the
application of randomly generated numbers.
10. MTBF - Mean or average time between failures.
11. Parameter - An unknown constant associated with a population (such

as the characteristic liWe of a Weibull population or the
mean of a normal population).
12. Precision - The degree of agrtEment among estimates calculated from

random samples drawn from a parent population. The
precision is usually measured by the standard deviation of
the estimates.
13. Random A failure mode that is independent of time, in the sense

(failure mode) that an old unit is as likely to fail as a new unit.. In other
words, the hazard rate remains constant with age.
141
/
14. Itt'liallilitv 'Tlhe pr bjlumilily thai. whlo olkrating in the manve*r
inleldecd. a syslemn will iero4rnm its inlended funi'tion
wit isl'actoirily IFor a specif ied interval of1"ime.
15. Risk Analysis A prediction f ithe oumimber of fI'miures exiiected Iti otcur in
some future tine period.
16. Suspension - A test or operational unit. that has not. failed by the mode
under consideration at the time of the life data analysis.
17. t- 7-pro age for the failure mode. It is known as the minimum
life paramete7 in the three-parameter Weibull
distribution: units have, zero probability of failure prior to
t . ,
18. Wearout A failure mode characterized by a hazard rate that

increases with age, ie., old units are more likely to fail
than new units.
19. Weibayes/Weibest A method for constructing a Weibull distribution based on

assuming a value of #, the Weibull slope parameter. It is
used wht-n there are certain deficiencies in the data (for
instance, when operating time has accumulated, but no
failures have occurred).
20. Wei' ill Procedure for finding the Weibull distribution that best
Analysis. describes a sample of unit lifetimes, in order to estimate
reliability, determine failure mode characteristics, and
predict the occurrences of future failures.
21. Weibull Plot A plot of time-to-failure data on Weibull probability

paper.
142
14744N
APPENDIX B
MECIAN RANKS, 5% RANKS, AND 95% RANKS
TABLE B.I. MEDIAN RANKS
Sample Size
Rank
Order 1 2 3 4 5 6 7 8 9 10
1 50.0 29.2 20.6 15.9 12.9 10.9 9.4 8.3 7.4 6.6
2 70.7 50.0 38.5 31.3 26.4 22.8 20.1 17.9 16.2
3 79.3 61.4 50.0 42 1 36.4 32.0 28.6 25.8
4 84.0 68.6 57.8 50.0 44.0 39.3 35.5
5 87.0 73.5 63.5 55.9 50.0 45.1
6 89.0 77.1 67.9 60.6 54.8
7 90.5 79.8 71.3 64.4
8 91.7 82.0 74.1
9 92.5 83.7
10 93.3
Sample Size
Rank
Order 11 12 13 14 15 16 17 18 19 20
i 6.1 5.6 5.1 4.8 4.5 4.2 3.9 3.7 3.5 3.4
2 14.7 13.5 12.5 11.7 10.9 10.2 9.6 9.1 8.6 8.2
3 %3.5 21.6 20.0 18.6 17.4 16.3 15.4 14.5 13.8 13.1
4 32.3 29.7 27.5 25.6 2L.9 22.4 21.1 20.0 18.9 18.0
5 41.1 37.b 35.0 32.5 30.4 28.5 26.9 25.4 24.1 22.9
6 50.0 45.9 42.5 39.5 36.9 34.7 32.'t 30.9 29.3 27.8
7 a8.8 54.0 50.0 46.5 43.4 40.8 38.4 36.? 34.4 32.7
8 67.6 62.1 57.4 53.4 50.0 46.9 44.2 41.8 39.6 37.7
9 76.4 70.2 64.9 60.4 56.5 5311 50.0 47.2 44.8 42.6
10 85.2 78.3 72.4 67.4 63.0 59.1 55.7 52.7 50.0 47.5
11 93.8 86.4 74.9 74.3 69.5 65.2 61.5 58.1 55.1 52.4
12 3.3 87.4 81.3 76.0 71.4 67.2 63.6 60.3 57.3
13 94.8 88.2 82.5 77.5 73.0 69.0 65.5 62.2
14 ,5.1 89.j 83.6 78.8 74.5 70.6 67.2
15 95.4 89.7 84.5 79.9 75.8 72.1
16 95.7 90.3 8.5.4 81.0 77.0
17 96.0 90.8 86.1 81.9
18 96.2 91.3 86.8
19 96.4 91.7
20 96.5
143
auw.-
/
Sample Size
Rank
Order 21 22 23 24 25 26 27 2 2.9 3)
1 3.2 3.1 7.9 2.8 2.7 2.6 2.5 2.4 2.3 2.2
2 7.8 7.5 7.1 6.8 6.6 6.3 6.1 5.9 5.7 5.5
3 12.5 11.9 11.4 10.9 10.5 10.1 9.7 9.4 9.1 8.8
4 17.2 16.4 15.7 15.0 14.4 13.9 13.4 12.9 12.5 12.1
5 21.8 20.9 20.0 19.1 18.4 17.7 17.0 16.4 15.9 15.3
6 26.5 25.3 24.2 23.2 22.3 21.5 20.7 20.0 19.3 18.6
7 30.2 29.8 28.5 27.4 26.3 25.3 24.3 23.5 22.7
8 35.9 34.3 21.9
32.8 31.5 30.2 29.1 28.0 27.0 26.1 25.2
9 40.8 38.8 37.1 35.8 34.2 32.9 31.7 .30.5 29.5 28.5
"10 45.3 43.2 41.4 39.7 38.1 36.7 35.3 34.1 32.9 31.8
11 50.0 47.7 45.7 43.8 42.1 40.5 39.0 37.6 36.3 35.1
12 54.6 52.2 50.0 47.9 46.0 44.3 42.6 41.1 39.7 38.4
13 59.3 56.7 54.2 52.0 50.0 48.1 46.3 44.7 4.3.1 41.7
14 64.0 61.1 58.5 56.1 53.9 51.8 50.0 48.2 46.5 415.0
15 68.7 65.8 62.8 60.2 57.8 55.6 5.3.6 51.7 50.0 48.3
16 73.4 70.1 67.1 64.3 61.8 59.4 57.3 55.2 5,3.4 51.6
17 78.1 74.6 71.4 68.4 65.7 63.2 60.9 58.8 56.8 54.9
18 82.7 79.0 75.7 72.5 69.7 67.0 64.6 62.3 60.2 58.2
19 87.4 83.5 79.9 76.7 73.6 70.8 68.2 65.8 63.6 61.5
20 92.1 88.0 84.2 80.8 77.6 74.6 71.9 F9.4 67.0 64.8
21 96.7 92.4 88.5 84.9 81.5 78.4 75.6 72.9 70.4 68.1
22 96.8 92.8 89.0 85.5 82.2 79.2 76.4 73.8 71.4
23 97.0 93.1 89.4 86.0 82.9 79.9 77.2 74.7
24 97.1 93.3 89.8 86.5 83.5 80.6 78.0
25 97.2 93.8 90.2 87.0 84.0 81.3
2S 97.3 93.8 90.5 -87.4 84.6
27 97.4 94.0 90.8 87.8
29 97.5 94.2
29 91.1
97.6 94.4
30
97.7
7144
Sample Size
Rank
Order 31 32 33 34 35 36 37 38 39 40
1 2.2 2.1 2.0 2.0 1.9 1.9 1.8 1.8 1.7 1.7
2 5.3 5.1 5.0 4.8 4.7 4.6 4.4 4.3 4.2: 4.1
3 8.5 8.2 8.0 7.7 7.5 7.3 7.1 6.9 6.7 6.6
4 11.7 11.3 11.0 10.6 10.3 10.1 9 8 9.5 9.3 9.1
5 14.9 14.4 14.0 13.6 13.2 12.8 12.5 12.1 11.8 11.5
6 18.0 17.5 17.0 16.5 16.0 15.6 15.1 14.7 14.4 14.0
7 21.2 20.6 20.0 19.4 18.8 18.3 17.8 17.3 16.9 16.5
8 24.4 237 23.0 22.3 21.7 21.1 20.5 20.0 19.4 19.0
9 27.6 26.8 26.0 25.2 24.5 23.8 23.2 22.6 22.0 21.4
10 30.8 29.9 29.0 28.1 27.3 26.6 25.8 252 24.5 23.9
11 34.0 32.9 32.0 31.0 30.1 29.3 28.5 27.8 27.1 26.4
12 37.2 36.0 35.0 33.9 33.0 32.1 31.2 30.4 29.6 28.9
13 40.4 39.1 38.0 36.8 35.8 34.8 33.9 33.0 32.2 31.4
14 43.6 42.2 41.0 39.8 38.6 37.6 36.6 35.6 34.7 33.8
15 46.8 45.3 44.0 42.7 41.5 40.3 39.2 38.2 37.2 36.3
h4 50.0 48.4 47.0 45.6 44.3 43.1 41.9 40.S 39.8 - 38.8
17 53.1 51.5 50.0 48.5 47.1 45.8 44.6 43.4 42.3 41.3
18 56.3 54.6 52.9 51.4 50.0 48.6 47.3 46.0 44.9 43.8
19 59.5 57.7 55.9 54.3 52.8 51.3 50.3 48.6 47.4 46.2
20 62.7 60.8 58.9 57.2 55.6 54.1 52.6 51.3 50.0 48.7
21 65.9 63.9 61.9 60.1 58.4 56.8 55.3 53.9 52.5 51.2
22 69.1 67.0 64.9 63.1 61.3 59.6 58.0 56.5 55.0 53.7
23 72.3 70.0 67.9 66.0 64.1 62.3 60.7 59.1 57.6 56.1
24 75.5 73.1 70.9 68.9 66.9 65.1 63.3 61.7 60.1 58.6
25 78.7 76.2 73.9 71.8 69.8 67.8 66.0 64.3 62.7 61.1
26 81.9 79.3 76.9 74.7 72.6 70.6 68.7 66.9 65.2 63.6
27 85.0 82.4 79.9 77.6 75.4 73.3 71.4 69.5 67.7 66.1
28 88.2 85.5 82.9 80.5 78.2 76.1 74.1 72.1 70.3 68.5
29 91.4 88.6 85.9 83.4 81.1 78.8 76.7 74.7 72.8 71.0
30 94.6 91.7 88.9 86.3 83.9 31.6 79.,4 77.3 75.4 73.5
31 97.7 94.8 91.9 89.3 86.7 84.3 82.1 79.9 77.9 76.0
32 97.8 94.9 92.2 89.6 87.1 84.8 82.6 80.5 78.5
S" 33
34 97.9 95.1
97.9 92.4
95.2. 89.8
92.6 87.4
90.1 85.2
97.8 83.0
85.5 80.9
83,4
35 98.0 96.3 92.8 90.4 88.1 85.9
36 98.0 95.5 93.0 90.8 88.4
37 98.1 96.0 93.21 90.8
38 98.1 95.7 93.3
39 96.2 95.8
40 98.2
S~145
/
'I'ABiE 11.1. MEI)IAN RANKS
SAample. Uize
ituatk
Order 41 42 1.1 .14 .1.5 .16 .17 .I .19 s50
I 1.6 1.6 1.5 1.5 1.5 1.4 1.4 1.4 1.4 1.3
2 4.0 3.9 3.8 3.7 3.7 3.6 3.4
:.5 3.4 3.
3 6.4 6.3 6.1 6.0 5.8 5.7 5.6 5.5 6.4 5.3
4 8.8 8.6 8.4 8.2 8.0 7.9 7.7 7.5 7.4 7.2
5 11.3 11.0 10.7 10.5 10.3 10.0 9.8 9.6 9.4 9.2
6 13.7 13.3 13.0 12.7 12.5 12.2 11.9 11.7 11.4 11.2
7 16.1 15.7 15.3 15.0 14.7 14.3 14.0 13.7 13.5 13.2
8 18.5 18.1 17.6 17.2 16.9 16.5 16.2 15.8 15.5 15.2
9 20.9 20.4 20.0 19.5 19.1 1.7 18.3 17.9 17.5 17.2
10 23.3 22.8 22.3 21.8 21.3 20.8 20.4 20.0 19.5 19.2
11 25.8 25.2 24.6 24.0 23.5 23.0 22.5 22.0 21.6 21.1
12 28.2 27.5 26.9 26.3 25.7 25.1 24.6 24.1 23.6 23.1
13 30.6 29.8 29.2 28.5 27.9 27.3 26.7 26.2 25.6 25.1
14 33.0 32.2 31.5 30.8 30.1 29.4 28.8 28.2 27.7 27.1
15 35.4 34.6 33.8 33.0 32.3 31.6 30.9 30.3 29.7 29.1
16 37.9 37.0 36.1 35.3 34.5 33.8 33.1 32.4 31.7 31.1
17 40.3 39.3 38.4 37.5 36.7 35.9 3-5.2 34.4 33.7 33.1
18 42.7 41.7 40.7 39.8 38.9 38.1 37.3 36.5 35.8 35.1
19 45.1 44.0 43.0 42.1 41.1 40.2 , 39.4 38.6 37.8 37.0
20 47.5 46.4 45.3 44.. 43.3 42.4 41.5 40.6 39.8 39.0
21 50.0 48.8 47.6 46.6 45.5 44.6 43.6 42.7 41.8 41.0
22 52.4 51.1 50.0 48.8 47.7 46.7 45.7 44.?. 43.9 43.0
23 54.8 53.5 52.3 51.1 &3.0 48.9 47.8 46.Z 45.9 45.0
24 57.2 55.9 54.6 53.3 52.2 51.0 50.0 48.9 47.9 47.0
25 59.6 56.2 56.9 55.6 54.4 53.2 52.1 51.0 50.0 49.0
26 62.0 60.6 59.2 57.8 56.6 55.3 54.2 53.1 52.0 50.9
27 64.5 62.9 61.5 60.1 58.8 57.5 56.3 55.1 54.0 52.9
28 66.9 5.3 63.8 62.4 61.0 59.7 58.4 57.2 56.0 54.9
29 69.3 67.7 66.1 64.6 63.2 61.8 60.5 59.3 58.1 56.9
30 71.7 70.0 68.4 66.9 65.4 64.0 62.6 61.3 6El 58.9
31 74.1 72.4 70.7 69.1 67.6 66.1 64.7 63.4 62.1 60.9
32 76.6 74.8 73.0 71.4 69.8 68.3 66.9 65.5 64.1 62.9
33 79.0 77.1 75.3 73.6 72.0 70.5 69.0 67.5 66.2 64.8
34 81.4 79.5 77.8 75.9 74.2 72.6 71.1 696A 68.2 66.8
.35 83.8 81.8 79.9 78.1 76.4 74.8 ' 73.2 71.7 70.2 68.8
36 86.2 84.2 82.3 80.4 78.6 76.9 75.3 73.7 72.2 70.8
37 88.7 86.6 84.6 82.7 80.8 79.1 77.4 75.8 74.3 72.8
38 91.1 88.9 86.9 84.9 83.0 81.2 79.5 77.9 76.3 74.8
39 93.5 91.3 89.2 87.2 85.2 83.4 81.6 79.9 78.3 76.8
40 95.9 93.6 91.5 89.4 87.4 86.6 83.7 82.0 80.4 78.8
41 98.3 96.0 93.8 91.7 69.6 87.7 85.9 84.1 82.4 80.7
42 96.3 98.1 93.9 91.9 89.9 86.0 86.2 84.4 82.7
43 98.4 96.2 94.1 92.0 90.1 88.2 86.4 54,84.7
44 98.4 96.2 94.P 92.2 90.3 885 86.7
45 98.4 96.3 94.3 92.4 90.5 88.7'
46 96.5 96.4 94.4 92.5 90.7
47 96.5 96.5 94.5 92.7
48 965 96.5 94.
49 985 96.6
50 98.6
146
.141t47H
TABLE B.2. FIVE PERCENT RANKS
Sample Size
Rain
Order 1 2 3 4 5 6 7 8 9 10
1 5.0 2.5 1.6 1.2 1.0 0.8 0.7 0.6 0.5 0.5
2 22.3 13.5 9.7 7.8 6.2 5.3 4.6 4.1 3.6
3 36.8 24.8 18.9 15.3 12.8 11.1 9.7 8.7
4 47.2 .34.2 27.1 22.5 19.2 16.8 15A.
5 54.9 41.8 34.1 28.9 25.1 22.2
6 66.6 47.9 40.0 34.4 30.3
7 65.1 52.9 45.0 39.3
8 61.7 57.0 49.3
9 71.6 60.5
10 74.1
Sample Size
Rank
Order 11 12 13 14 i5 16 17 18 19 20
1 0.4 0.4 0.3 0.3 0.3 0.3 0.3 0.2 0.2 0.2
2 3.3 3U1 2.8 2.6 2.4 22 2.1 2.0 1.9 1.8
3 1.8 7.1 6.6 6.1 5.6 5.3 4.9 4.7 4.4 4.2-
4 13.5 12.2 11.2 10.4 9.6 9.0 8.4 7.9 7.5 7.1
5 19.9 1&1 16.5 15.2 14.1 13.2 12.3 11.6 10.9 10.4
6 27.1 24.5 22.3 20.6 19.0 17.7 16.6 15.6 14.7 13.9
7 34.9 31.5 28.7 26.3 24.3 22.6 21.1 19.8 18.7 1.7.7
.8 43.5 39.0 35.4 32.5 29.9 27.8 26.0 24.3 22.9 2L7
9 52.9 47.2 42.7 39.0 35.9 33.3 31.0 29.1 27.3 25.9
10 635 56.1 50.5 45.9 42.2 39.1 36.4 34.0 32.0 30X1
11 76.1 66.1 58.9 53.4 48.9 45.1 41.9 39.2 36.8 34.6
12 77,9 68.8 61.4 56.0 b1.b 47.8 44.9 41.8 39.3
13 79.4 70.3 63.6 58.3 .5.3.9 50.2 47.0 44.1
14 80.7 72.0 65.6 60.4 56.1 52.4 49.2
15 81.8 73.6 67.3 62.3 58.0 54.4
16 82.9 74.9 68.9 64.0 59.8
17 83.8 76.2 70.4 65.6
18 84.6 77.3 71.7
19 86.4 78.3
20 80.0
147
U471C'
/
/.
'I'ABILE 1.2. i'1,, II'I.( ENT RAN ,KS
Itank
(Irder 21 22 2:: 22. 27 -, 2J
1 0.2 0.2 0.2 0.2 0.2 ). 1 ).I I I 0.1
o..).
2 1.7 1.6 1.5 1.5 1.4 1.3 1.3 1.2 1.2 1.1
3 4.0 3.8 3.6 3.4 "L3 :3.2 3.0 2.9 2.8 2.7
4 6.7 6.4 6.1 5.9 5A6 5.4 5.2 5.0 4.8 4.6
S5 .8 9.4 8.9 8.5 8.2 7.8 7.5 7.3 7.0 6.8
8 13.2 12.6 12.0 11.4 1M.3 10.5 10.1 9.7 9.4 9.)
7 16.8 15.9 15.2 11.5 13.9 13.3 12.8 12.3 11.9 11.4
8 20.5 19.5 18.6 17.7 17.0 16.3 15.6 15.0 14.5 14.C
9 24.4 23,2 22.1 21.1 20.2 19.3 18.6 17.9 17.2 16.6
10 28.5 27.1 25.8 24.6 23.5 22,5 21.6 20.8 20.0 19.3
I1 32.8 31.1 29.A 28.2 26.9 25.8 24.7 23.8 22.9 22.1
37.1 35.2 33.5 31.9 30.5 29.2
13 41.7 39.5 31.5 35.7 28.0 26.9 25.8 24.9
34.1 32.6 31.3 30.0 28.9
14 46.4 27.8
43.9 41.6 31.6 37.8 36.2 34.6 33.3 32.0 30.8
15 51.2 48.4 45.9 43.7 41.6 39.8 38.1 :16.6 35.2 33.8
16 56.3 53.1 50.3 47.8 45.6 435 41.7 40.0 38.4 36.9
17 61.5 58.0 54.9 52.1 49.6 47.3 45.3 43.1 41.7 40.1
18 87.0 63.0 59.6 56.5 53.7 51.3 49.0 47.0 45.1 43.3
19 72.9 68.4 64.5 61.0 58.0 55.3 52,8 50.6 48.5 46.6
20 79.3 74.0 69.6 65.8 62.4 59.4 56.7 54.:. 52.0 50.0
21 86.7 80.1 75.0 70.7 67.0 6:3.7 R).7 58.1 55.7 53.4
22 87.2 80,9 76.0 71.8 68.1 64.9 62.0 59.4 57.0
23 87.7 81.7 76.8 72.8 69.2 66.0 63.2 60.6
24 88.2 62.3 77.7 73.7 70.2 67.1 64.2
25 88.7 8.3.0 78A4 74.5 71.1 68.1
26
89.1 83.6 79.1 75.3 72.0
27
89.4 84.1 79.8 76.1
29
89.8 84.6 80.4
30
90.1 85.1
90.4
148
/
TABIb, B.2. FIVE ;ANFS
PER('ENT
Sample Size
Rank
Order 31 32 33 34 35 36 37 3M 39 40
1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 v.I
2 1.1 1.1 1.0 1.0 1.0 0.9 0.9 0.9 0.9 0.8.
3 2.6 2.6 2.5 2.4 2.3 2.3 2.2 :. 1 21 2!or
4 4.5 4.3 4.2 4.1 3.9 3.8 3.7 3.A 3.5 3.4:
5 6.5 6.3 6.1 5.9 5.8 5.6 5.4 3.3 5.A S.):
6 8.7 8.4 8.2 7.9 7.7 7.5 7.3 7.1 6.9. 6.7
7 11.1 10.1 10.4 10.0 97 9.4 9.2 8.9 8.7 8.5
8 13.5 13.0 12.6 12.2 11.9 11.5 1L.2 10.9 10.6 10.3
9 16.0 15. 15.0 14.5 14.1 13.7 1'1..3 12.9 116 12.2
.0 18.6 18.0 17.4 16.9 16.3 15.9 15.4 1!..O 14.6 14.2
11 21.3 20.6 19.9 19.3 18.7 18.1 ,!.G 1.1 116.6 16.2
12 24.0 23.2 22.5 21.7 21.1 204 1'1.8 19.2 18.8 18.3
13 26.8 25.9 25.1 24.3 23.5 22.8 22.1 21.5 20.9 2014
14 29.7 28.7 27.7 2f.8 26.0 25.2 24.r, L3.8 :3.1 "21'5
15 32.6 31.5 30.4 29.5 28.5 27.7 26. 26.1 25.4 21-7
16 35.6 34.4 33.2 32.1 31.1 30.2 ,9:1 2A.4 27.6 26 9
17 38.6 37.3 36.0 34.8 33.7 32.7 31.7 30.4 30.0 29.1
18 41.7 40.3 38.9 37.6 36.4 35.3 3.L2 33.2 32.3 31A4
19 44.9 43.3 41.8 40.4 39.1 37.9 ;u;.8 :15.7 34.7 33. -
20 48.1 46.4 44.8 43.3 41.9 406 39.3 38.2 37.1 36A
21 51.4 49.5 47.8 46.2 44.7 42.3 41.9 40.7 39.5 28$At
22 54.8 52.7 ,50.9 49.1 47.5 46.0 44.6 41..3 42.0 40;A
23 58.2 E6.0 54.0 52.1 50.4 48.8 47.3 459 44.5 43.3
24 61.7 59.3 57.2 56.2 53.3 51.6 500 48.5 47.1 45.7
25 65.3 62.8 60.4 58.3 56.3 54.5 52.8 51.2 49.6 48.2
26 69.0 66.3 63.8 61.5 59.4 57.4 5.5.6 53.9 52.3 50.8
27 72.8 69.9 67.2 64.7 62.5 60.4 58.4 .56.6 54.9 53.3
28 76.8 736 70.7 68.1 65.6 63.4 61.3 59.4 57.6 55.9
29 81.0 77.5 74.3 71.5 68.9 66.5 64.3 62.3 60.3 58.6
30 86.5 81.6 78.1 75.0 72.2 69.7 67.3 65.2 63.1 61.2
31 90.7 86.0 82.1 78.7 75.7 72.9 70.4 68.1 66.0 64.0
32 91.0 86.4 82.6 79.3 76.3 7:1.6 71.1 68.9 66-7
33 91.3 86.7 83.0 79.8 76.9 74,1 71.8 69.6
34 91.5 87.1 83.5 80.3 77.5 74.9 72.5
35 91.7 87.4 83.9 80.8 78.0 75.4
36 92.0 87.8 84.3 81.3 78.5
27 92.2 88.1 84.7 81-7
38 92.4 88.4 85.0
39 92.6 MA6
40
149
.141W
/
TAIILI 1.2. FIVE, I'PIECNT RANKS
SAmple. Size
SRank
Olrdaer 41 42 43 44 4.5 46 .17 .18 49 50
I 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 0.1
2 0.8 0.8 0.8 0.8 0.7 0.7 0.7 0.7 0.7 0.7
3 2.0 1.9 1.9 1.8 1.8 1.8 1.7 1.7 1.6 1.6
4 3.4 3.3 3.2 3.1 3.0 3.0 2.9 2.8 2.8 2.7
5 4.9 4.8 4.6 4.5 4.4 4.3 4.2 4.1 4.1 4.0
6 6.5 6.4 6.2 6.1 5.9 5.8 5.7 5.5 5.4 5.3
7 8.2 8.0 7.8 7.7 7.5 7.3 7.2 7.0 6.9 6.7
8 10.0 9.5 9.6 9.3 9.1 8.9 &7 85 8.3 8.2
9 11.9 11.6 1"1.3 11.1 10.8 10.6 10.3 10.1 9.9 9.7
10 13.8 13.5 .13.1 12.8 12.5 12.2 12.0 11.7 11.5 11.2
11 15.8 15.4 15.0 14.6 14.3 14.0 13.7 13.4 13.1 12.8
12 17.8 17.3 16.9 16.5 16.1 15.7 15.4 15.1 14.7 14.4
13 19.8 19.3 18.9 18.4 18.0 17.5 17.2 16.8 16.4 16.1
14 21.9 21.4 20.8 20.3 19.8 19.4 18.9 18.5 18.1 17.7
15 24.0 23.4 22.8 22.3 21.7 21.2 20.8 20.3 19.9 19.4
16 26.2 25.5 24.9 24.3 23.7 23.1 22.6 22.1 21.6 21.2
17 28.4 27.6 26.9 26.3 25.6 25.0 24.5 23.9 23.4 22.9
is 30.6 29.8 29.0 28.3 27.6 27.0 26.4 25.8 25.2 24.7
19 32.8 12.0 31.2 30.4 29.6 28.9 28.3 27.6 27.0 26.5
20 35.1 34.2 33.3 32.5 31.7 30.9 30.2 29.5 28.9 28.3
21 37.4 36.4 35.5 34.6 33.7 32.9 32.2 31.5 30.8 30.1
22 39.7 38.7 37.7 36.7 35.8 35.0 34.2 33.4 32.6 31.9
23 42.1 41.0 39.9 38.9 31.9 37.0 36.2 35.3 34.5 33.8
24 44.5 43.3 42.2 41.1 401 39.1 38.2 37.3 36.5 357
25 46.9 46.6 44.4 43.3 42.2 41.2 40.2 39.3 38.4 37.6
26 49.3 48.0 46.7 45.5 44.4 43.3 42.3 41.3 40.4 39.5
27 61.8 50.4 49.1 4:.8 46.6 45.5 44.4 43.3 42.4 41.4
2b 54.3 52.8 51.4 50.1 48.8 47.6 .46.5 45.4 44.4 43.4
29 569 55.3 53.8 52.4 51.1 49.8 48.6 47.5 46.4 45.3
30 59.5 57.8 56.2 54.8 53.4 52.0 50.8 49.6 48.4 47.3
31 62.1 60.3 58.7 57.1 b6.7 54.3 52.9 51.7 50.5 49.3
32 64.8 62.9 61.2 59.5 58.0 56.5 55.1 53.8 52.6 51.4
33 67.5 65.5 63.7 62.0 60.4 58.8 57.4 56.0 54.7 53.4
l4 70.3 68.2 66.3 64.5 62.7 61.1 59.6 58.2 56.8 55.5
35 73.1 70.9 68.9 67.0 65.2 63.5 61.9 60.4 58.9 57.6
36 76.0 73.7 71.5 69.5 67.6 65.9 64.2 62.6 61.1 59.7
37 79.0 76.5 74.3 72.1 70.2 68.3 66.5 64.9 63.3 61.8
38 82.1 79.5 77.0 74.8 72.7 70.8 68.9 67.2 65.5 64.0
39 85.4 82.5 79.9 77.5 75.3 73.3 71.3 69.5 67.8 66.2
40 8.9 85.7 82.9 80.3 78.0 75.8 73.8 71.9 70.1 68.4
41 92.9 89.1 86.0 83.3 80.8 7M4 76.3 74.3 72.4 70.6
42 93.1 89.4 86.8 83.6 81.1 78.9 76.8 74.8 72.9
43 93.2 89.6 86.6 83.9 81.5 79.3 77.2 75.3
44 93.4 89.8 86.9 84.3 81.9 79.7 77.6
45 93.5 90.0 87.2 84.6 82.2 80.1
46 93.6 90.3 87.4 64.9 82.6
47 93.8 90.4 87.7 85.2
46 93.9 90.6 87.9
49 94.0 90.8
50 94.1
150
TABLE B.3. NINETY-FIVE PERCENT RANKS
nSample Size
Rank
Order 1 2 4 .5 6 7 x 9
I 95.0 77.6 63. 1 52.7 45.0 39.3 :14.8 :1i.2 28.3 25.8
2 97.4 86.4 75.1 65.7 58.1 52.0 47.0 42.9 39.4
3 98.3 90.2 81.0 72.8 65.8 59.9 54.9 50.6
4 98.7 92.3 84.6 77.4 71.0 65.5 60.6
5 98.9 93.7 87.1 80.7 74.8 69.6
6 99.1 94.6 88.8 83.1 77.7
7 99.2 95.3 90.2 84.9
8 99.3 95.8 91.2
9 99.4 96.3
1o 99.4
Sample Size
Rank
Order 11 12 13 14 1i 16 17 18 19 20
1 23.8 22.0 20.5 19.2 18.1 17.0 16.1 15.3 14.5 13.9
2 36.4 33.8 31.6 29.6 27.9 26.3 2&0 23.7 22.6 21.6
3 47.0 43.8 41.0 38.5 36.A 34.3 32.6 31.0 29.5 28.2
4 56.4 52.7 49.4 46.5 43.9 41.6 39.5 37.6 35.9 34.3
5 65.0 60.9 57.2 54.0 51.0 48.4 46.0 43.8 41.9 40.1
6 72.8 68.4 64.5 60.9 57.7 54.8 52.1 49.7 47.5 45.6
7 80.0 75.4 71.2 67.4 64.0 60.8 58.0 55.4 52.9 59.7
8 86.4 81.8 77.6 73.6 70.0 66.6 63.5 60.7 58.1 55.8
9 92.1 87.7 83.4 79.3 75.6 72.1 68.9 66.9 63.1 60.6
10 96.6 92.8 88.7 84.7 80.9 77.3 73.9 70.8 67.9 65.3
11 99.5 96.9 93.3 89.5 85.8 82.2 78.8 75.6 72.6 69.8
12 99.5' 97.1 93.8 90.3 86.7 83.3 80.1 77.0 74.1
13 99.6 97.4 94.3 90.9 87.6 84.3 81.2 78.2
14 99.6 97.5 94.6 91.5 88.3 85.2 8K.2
15 99.6 97.7 95.0 92.0 89.0 86.0
16 99.6 97.8 9&2 92.4 89.5
17 99.6 97.9 95.5 92.8
18 99.7 980 95.7
19 99.7 98.1
20 99.7
151
'TABIX B.:i. NINITrY-IIVK IRKRCENT IANKS
S,a&mple ,Sizt
Rank
Order 21 22 23 24 25 26 27 2m 29 30
1 1:1.2 12.7 12.2 11.7 11.2 11.8 1HU5 10.1 9.8
2 20.6 19.8 9.5
19.0 18.2 17.6 16.9 16.:1 15.8 15.3
3 27.0 25.9 24.9 14.8
23.9 23.1 22.2 21.5 20.8 20.1 19.5
4 .32.9 31.5 30.3 29.2 28.1 27.1 26.2 25.4 24.6 2.3.8
5 38.4 36.9 35.4 34.1 32.9 31.8 30.7 29.7 28.8 27.9
6 43.6 41.9 40.3 38.9 37.5 36.2 35.0 .33.9 32.8 31.8
7 48.7 46.8 45.0 43.4 41.9 40.5 39.2 37.9 36.8 35.7
8 53.5 51.5 49.6 47.8 46.2 44.6 43.2 41.8 40.6 39.3
9 5b.2 56.0 54.0 52.1 50.3 48.7 47.1 45.6 44.2 42.9
10 62.8 60.4 58.3 W6.2 54.3 52.6 50.9 49.3 47.9 46.5
I1 67.1 64.7 62.4 60.3 58.3 564 54.6 52.9 51.4 49.9
12 71.4 68.8 66.4 64.2 62.1 60.1 58.2 56.5 54.8 53.3
13 75.5 72.8 70.3 68.0 65.8 63.7 61.8 59.9 58.2 56.6
14 79.4 76.7 74.1 71.7 69.4 67.3 65.3 63.3 61.5 59.8
i5 83.1 80.4 77.8 75.3 73.0 70.7 68.6 66.6
16 64.7 63.0
86.7 84.0 81.3 78.8 76.4 74.1 71.9 69.9 67.9 66.1
17 90.1 87.3 84.7 82.2 79.7 77.4 75.2 73.0 71.0 69.1
18 93.2 90.5 87.9 85.4 82.9 80.6 78.3 76.1 74.1 72.'
19 95.9 93.5 91.0 88.5 86.0 83.6 81.3 79.1 77.0 75.0
20 962 96.1 93.8 01.4 88.9 86.6 84.3 82.0 79.9 778
21 99.7 98.3 96.3 94.0 91.7 89.4 87.1 84.9 82.7 80.6
22 99.7 96.4 96.5 94.3 92.1 89.8 87.6 85.4 83.3
23 99.7 98.4 96.6 94.5 92.4 90.2 68.0 85.9
24 99.7 96.5 96.7 94.7 92.6 90.5 88.5
25 99.7 96.6 96.9 94.9 92.9 90.9
26 99.8 98.6 97.0 95.1 93.1
27 99.8 98.7 97.1 95.3
28 99.8 98.7 97.2
29
99.8 98.8
30
99.8
152
.141II"
Smple Size
Rank
Order 31 32 .33 34 :25 Js 37 .M 39 40
1 9.2 8.9 8.6 8.4 8.2 7.9 7.7 7.5 7.3 7.2
2 14.4 13.9 13.5 13.2 12.8 12.5 12.1 11.8 11.5 11.3
3 18.9 18.3 17.8 17.3 16.9 16.4 16.0 15.6 15.2 14.9
4 23.1 22.4 21.8 21.2 20.6 20.1 19.6 19.1 18.6 18.2
5 27.1 26.3 25.6 24.9 24.2 23.6 23.0 22.4 21.9 2L4
a 30.9 30.0 29.2 28.4 27.7 27.0 26.3 25.6 25.0 24.5
7 34.6 33.6 32.7 31.8 31.0 30.2 29.5 2&.8 28.1 27.4
8 38.2 37.1 36.1 35.2 34.3 33.4 32.6 31.8 31.0 30.3
9 41.7 40.6 39.5 38.4 37.4 36.5 35.6 34.7 33.9 33.2
10 46.1 43.9 42.7 41.6 40.5 39.5 38.6 37.6 36.8 35.9
11 48.5 47.2 45.9 44.7 43.6 42.5 41.5 40.5 39.8 38.7
12 51.8 50.4 49.0 47.8 46.6 45.4 44.3 43.3 42.3 41.3
13 85.0 53.5 52.1 Su.8 49.5 48.3 47.1 46.0 46.0 44.0
14 58.2 56.6 65.1 53.7 52.4 51.1 49.9 48.7 47.6 46.6
15 61.3 59.6 58.1 56.6 55.2 53.9 52.6 51.4 50.3 49.1
16 64.3 62.6 61.0 59.5 58.0 56.6 55.3 54.0 52.8 51.7
17 67.3 65.5 63.9 62.3 60.8 59.3 58.0 56.6 55.4 54.2
18 70.2 68.4 66.7 65.1 63.5 62.0 60.6 59.2 57.9 56.6
19 73.1 71.2 69.5 .67.8 66.2 64.6 63.1 61.7 60.4 59.1
20 75.9 74.0 72.2 70.4 68.8 67.2 65.7 64.2 62.8 61.5
21 7.6. 76.7 74.8 73.1 71.4 69.7 68.2 66.7 65.2 63.8
22 81.3 79.., 77.4 75.6 73.9 72.2 70.6 69.1 67.6 66.2
23 83.9 81.9 80.0 78.2 76.4 74.7 73.0 71.5 69.9 68.5
24 86.4 84.4 82.5 80.6 78.8 77.1 75.4 73.8 72.3 70.8
25 88.8 86.9 84.9 83.0 81.2 79.5 77.8 76.1 74.5 73.0
26 91.2 89.2 87.3 C5.4 83.6 81.8 80.1 78.4 76.8 75.2
27 93.4 91.5 89.5 87.7 86.8 84.0 82.3 80.6 79.0 77.4
28 95.4 93.6 91.7 89.9 88.0 86.2 84.5 82.8 81.1 79.5
29 97.3 95.6 93.8 92.0 90.2 88.4 86.6 84.9 83.3 81.6
30 98.8 97.3 95.7 94.0 92.2 90.5 88.7 87.6 85.3 83.7
31 99.8 98.8 97.4 96.8 94.1 92.4 90.7 89.0 87.3 85.7
32 99.8 98.9 97.5 96.0 94.3 92.6 91.0 89.3 87.7
33 99.8 98.9 97.6 96.1 94.5 92.8 91.2 89.6
34 99.8 98.9 97.6 96.2 94.6 93.0 91.4
35 99.8 99.0 97.7 96.3 94.8 93.2
36 99.8 99.0 97.8 96.4 94.9
37 99.8 9A0 97.8 96.5
38 99.8 99.0 97.9
39 99.8 99.1
40 99.8
153
1.
Sample Size
Rank
Order 41 42 43 44 45 46 47 48 49 50
1 7.0 6.8 6.7 6.5 6.4 6.3 6.1 6.0 5.9 5.8
2 11.0 10.8 10.5 10.3 10.1 9.9 9.7 9.5 9.3 9.1
3 14.5 14.2 13.9 13.6 13.3 13.0 12.7 12.5 12.2 12.0
4 17.8 17.4 17.0 16.6 16.3 16.0 15.6 15.3 15.0 14.7
5 20.9 20.4 20.0 19.6 19.1 18.8 18.4 18.0 17.7 17.3
6 23.9 23.4 22.9 22.4 21.9 21.5 21.0 20.6 20.2 19.8
7 26.8 26.2 25.6 25.1 24.6 24.1 2.3.6 23.1 22.7 22.3
8 29.6 29.0 28.4 27.8 27.2 26.6 26.1 25.6 25.1 24.6
9 .32.4 31.7 31.0 30.4 29.7 29.1 28.6 28.0 27.5 27.0
10 35.1 34.4 33.6 32.9 32.3 31.6 31.0 q0.4 29.8 29.3
11 37.8 37.0 36.2 35.4 34.7 34.0 33.4 32.7 32.1 31.5
12 40.4 39.6 38.7 37.9 37.2 36.4 35.7 35.0 34.4 33.7
13 43.0 42.1 41.2 40.4 39.5 38.8 38.0 37.3 36.6 35.9
14 45.6 44.6 43.7 42.8 41.9 41.1 40.3 39.5 38.8 38.1
15 48.1 47.1 46.1 45.1 44.2 43.4 42.5 41.7 41.0 40.2
16 50.6 49.5 48.5 47.5 46.5 45.6 44.8 43.9 43.1 42.3
17 53.0 51.9 50.8 49.8 48.8 47.9 47.0 46.1 45.2 44.4
18 55.4 54.3 53.2 52.1 51.1 50.1 49.1 48.2 47.3 46.5
19 57.8 56.6 55.5 54.4 53.3 52.3 51.3 50.3 49.4 48.5
20 60.2 58.9 57.7 56.6 55.5 54.4 5.3.4 52.4 51.5 50.6
21 62.5 61.2 60.0 58.8 57.7 56.d 55.5 54.5 53.5 52.6
22 64.8 63.5 62.2 61.0 59.8 58.7 b7.6 56.6 55.5 54.6
23 67.1 65.7 64.4 63.2 62.0 60.8 59.7 58.6 57.5 56.5
24 69.3 67.9 66.6 65.3 64.1 62.9 61.7 60.6 59.5 58.5
25 71.5 70.1 68.8 67.4 66.2 64.9 63.7 62.6 61.5 60.4
26 73.7 72.3 70.9 69.5 68.2 67.0 65.7 64.6 13.4 62.3
27 75.9 74.4 73.0 71.6 70,3 69.0 67.7 66.5 65.4 64.2
28 78.0 76.5 75.0 73.6 72.3 71.0 69.7 68.5 67.3 66.1
29 80.1 78.5 77.1 75.6 74.3 72.9 71.6 70.4 69.1 68.0
30 82.1 80.6 79.1 77.6 76.2 74.9 73.5 72.3 71.0 69.8
31 84.1 82.6 81.0 79.6 78.2 76.8 75.4 74.1 72.9 71.6
X2 86.1 84.5 83.0 81.5 80.1 78.7 77.3 76.0 74.7 73.4
&1 88.0 86.4 83.4
W4.9 81.9 80.5 79.1 77.8 76.5 75.2
.34 89.9 88.3 86.8 85.3 83.8 82.4 81.0 79.6 78.3 77.0
35 91.7 90.1 88.6 87.1 85.6 84.2 82.7 81.4 80.0 78.7
36 93.4 91.9 90.3 88.8 87.4 85.9 84.5 93.1 81.8 80.5
37 95.0 93.5 92.1 90.6 89.1 87.7 86.2 84.8 83.5 82.2
38 96.5 95.1 93.7 92.2 90.8 89.3 87.9 86.5 85.2 83.8
39 97.9 96.6 95.3 93.8 92.4 91.0 89.6 88.2 86.8 855
40 99.1 980 96.7 95.4 94.0 92.6 91.2 89.8 88.4 87.1
41 99.8 99.1 98.0 96.8 95.5 94.1 92.7 91.4 90.0 88.7
42 99.8 99.1 98.1 96.9 95.6 94.2 92.9 91.6 90.2
43 99.8 99.1 98.1 96.9 95.7 94.4 93.0 91.7
44 99.8 99.2 98.1 97.0 95.8 94.5 93.2
45 99.8 99.2 98.2 97.1 95.3 94.6
46 99.8 99.2 98.2 97.1 95.9
47 99.8 99.2 98.3 97.2
48 99.8 99.2 98.3
49 99.8 99.2
50 99.8
7 154
/
./
APPENDIX C
RANK REGRESSION (WEIBULL PLOT) METHOD OF WEIBULL ANALYSIS
1. METHOD
Median rank regression uses a best-fit. straight line, through the data plotted on Weibull
p.l)er, to estimate the Weibull parameters beta and eta. The "best-fit" line is found by the
method of least squares.
First, the failure times and median ranks (see Chapter 2 for the calculation of median
ranks) are "transformed", as follows:
Y - Qn (cycles-to-failure)
X = Qn(Qn ( I - Median Rank of Y
(The median -ank is expressed in decimal form.)
Least squares is then used to estimate A and B in the equation Y = A + BX. These
estimates will be referred to as A and 1, respectively. The median rank regression estimates of
the Weibull parameters are:
1A
eA
2. EXAMPLE AND STEP-BY-STEP PROCEDURE
The median rank regression method will be illustrated with the censored data listed below.
Cycles Status
1500 Failure
1750 Suspension
2250 Failuie
4000 Failure
4300 Failure
5000 Suspension
7001 Failure
Step 1: Calculate the median ranks of the failure times us; ig the methods oi Chapter 2.
Cycles to Failure Raitk Order Number Median Rank (decimal form)

1500 1.0o0o 0.0946
2250 2.1667 0.2523
4000 3.3333 0.4099
4300 4.5000 0.5676
7000 6.2500 0.8041
155
Step 2: For each failure, calculate the natural. (base e) logarithm of the cycles-to-failure
(Y = Rn (cycles-to-failure))
and X =Rn(Rn(1 - Median Rank of Y))
Cycles to Failure Median Rank Y X

1500 0.0946 7.3132 -2.3088
22,50 0.2523 7.7187 -1.2353
4000 0.4099 8.2940 -0.6397
4300 0.5676 8.3664 -0.1763
7000 0.8041 8.A537 0.4887
A A
S'tep 3: Calculate the least squares estimates A and B of A and B in the equation
Y. =.A + BX, where:
A
A= V - X where V is the average of the Y's
and X is the average of the X's,
and
1.-I
x -i)?y, n
In the abmve example, n
SXiYi = -28.8735 Yi
Y = 40.5460 Y = 8.1092
Xi = -3.8714 X, 2 = 7.5356 X -0.7743
-28.8735 - (-3.8714)(40.5460)
B3=
7.5356 (-3.8714)2
5
A 2.5205
B = 4.5381
A
B = 0.5,5,54
and
A= 8.1092 - (0.5554) (-0.7743)
A =8.5392
156
Step 4: Calculate the median rank regression estimates of # and q:
== 1.80
4=e4 eO = 5111.25
The Weibull equation used to calculate the probability of failure before t cycles is then:
F(t) = 1 - e-(t/51115)'
Figure C.1 shows the data plotted on Weibull paper with the least squares line overlaid.
157
i a4-?1
N.
I.EIBLLL DISTRIBUTION .
0 1.8
v 5111.25
S. SRMPLE SIZE = 5
._ FRILURES= 5
-- III
U.--- -- I -- -
g ....
I. .. . .. . ... .. . .. . -
1.L --- - -- ' -

LEAST SQUARES LINE
O. , - - _- -
wa . 4 .g .aa.
: a . *o ~~oa . 2. 2rooo . *S ..
TOTAL OPERA-TING CYCLES

FO 2V1MS
Figure C. I. li.xampie D~ata With Le'ast Squares Line
158
APPENDIX 0
MAXIMUM UKELIHOOD METHOD OF WEIBULL ANALYSIS
1. FOREWORD
Weibull analysis consists of "fitting" failure data to the Weibull distribution by estimating
-the parameters, beta (#) and eta (q). The rank regression (Weibull plot) method was presented
in Chapter 2; this Appendix presents an explanation and an example of maximum likelihood
Weibull analysis. Details of the maximum likelihood method may be ound in Wayne Nelson's
text, Applied Life Data Analysis (1982), John Wiley and Sons, New York.
The maximum likelihood Weibull analysis method consists of finding the values of 0 and 7
which maximize the "likelihood," of obtaining the observed data. The likelihood of obtaining
the observed data, expressed in mathemat-cal form, is a function of the Weibull parameters 6
and i. Maximum likelihood finds the values of P and -9 which maximize this mathematical
likelihood function.
2. THE LIKELIHOOD FUNCTION
The likelihood function is the mathematical expression of the prpbability of obtaining the
observed data.
When the sample is complete (all units are run to failure), the likelihood function is:
L - i f(xi) f(x) f(x 2) ..... )
where n = sample size
dF(x)
x) - -
and F(x) = 1 - e-(X/e)O
In reliability terms, F(x) is the probability that a unit will fail before it acquires x units of
operating time. F(x) is often called the "unreliability" at time x, and satisfies
F(x) = - R(x)
where R(x) = reliability at time x.
When the time-to-failure distribution is Weibull,
f(x) =I e wo
and L n= ) - e 4000
S17 71
Note that the "likelihood" of the sample failure data x1 , x2 ,.... x, is a function of the Weibull
parameters # and V.
159
X X .
The general form of the lFkelihood function for censored samples (where not every unit has been
run to failure) is:
k
l, f(x,)" (1- FC(T))
iI j I
where
r = number of units run to failure

k number of unfailed units
x 1, x2 ,..... xr = known failure times
Ti, 'r 2, ..... Tk = operating time on each unfailed unit.
When the time-to-failure distribution is Weibull,
L i e -r e
3. MAXIMIZING THE LIKELIHOOD FUNCTION
The maximum likelihood method finds the values of / and q which maximize the
likelihood function. To find the values of #i and il which maximize the Weibull likelihood
function, differentiate the logarithm of the likelihood function with respect to JRand r/, equate
the resulting expressions to zero, and simultaneously solve for 3 and 17.
A
In the complete sample case, the maximum likelihood estimate of 13,denoted i3, satisfies
A
x' IInx
i-n
l nl - Qnx, - - - -'0
SII i;Q - x
Given the failure times x1 , x 2. xn, the maximum likelihood estimate of/ is found using
iterative proced,res.
The maximum likelihood estimate of il is:
A
where 13is the maximum likelihood estimate of/i.
When the sample is censored, the maximum likelihood 3 estimate, ji, satisfies
N A
x-Q
NA n=nx0
r i
i=1
160
31734-
>i7
where N = total sample size
(N --number of failures (r) + number of suspensions (k)).
UInit. censored at times Ti are assigned the values xrti Ti. The second term in the
eluatnon, -- Qnxi, sums the logarithms of the failure times only.
A
/ is again found using iterative procedures.
Analogous to the complete sample case, the maximum likelihood estimate of ,/ is, in
censored samples,
N A) 1/
A i
4L EXAMPLE
The maximum likelihood method will be illustrated with the censored data listed below.
Cycles Status
1500 Failure
1750 Suspension
2250 Failure
4000 Failure
4300 Failure
5000 Suspension
7000 Failure
A
The maximum likelihood 8 estimate, f, is the root of the equation
i-i 1 -' i In
G(8)GO -r-- Zxnxi 5 ZQn 1 - "3 -"
0n
0
Zx?'
i-i
161
314T
_____ ____ _____ ___

The Weibull plot estimate, 1.8. was used as
zh.? initial value of i. This and
l i.twL,,t;
estimates of # are listed below with the corresponding
value of G3). (A Fortran ul~hroutint,
using a modified Newton-Raphson procedure was
used to find the value of giving t(;) o.)
G('8)
1.800 -0.1754
1802 -0.1746
2.179 -0.0255
2.182 -0.0248
Maximurm 2.253 -0.0007
Likelih(od 2.256 -0.0005
Estimate of 2.257 -0.0000
Beta
The maximum likelihood estimate of , is
A
'1=4900.1.
162j
I
APPENDIX E
WEIBAYES METHODS
1. FOREWORD
Weibayves is a method for constructing a Weibull distribution based.(on as.uming a value of

ji, the Weibull slope parameter. It is used when there are certain deficiencies in the data tfor
instance, when operating time has been accumulated, but no failures have occurred). Chapter 4
describes several applications of this method.
The Weibayes equation for q is
r (E.1)
where / is the assumed value of the Weibull slope parameter,

n is the number of' sispensions or unfailed units in the fleet,
ti, t2 , - - -, t, are the operating times tecumulated by units 1, 2. - - -, n.
r is the number of failures.
If no failures have ocwurred, r is assumed to be one, i.e, the first failure is imminei,,. q" is
then a conservative 63:, lower confidence bound on the true value of q. If failures have occurred
and Weibayes is used, q* is the maximum likelihood estimator of the true value of 7.
L DERIVATION OF THE WEIBAYES EQUATION
If no tailures have occurred, the Weibayes equation with r I gives a conservative 63',
lower confidence bound on the true value of qi.
The lower bound is derived using two facts from statisticr:
1. If ti, t2 , .. , t, represent failure times drawn from a Weibull population

.vith slope parameter #i and characteristic life Ut, then t 1p, t2
represents a random sample from an exponential population with mean
life a = i?. The exponential cumulative distribution function is
F(tI - 1 e-/.
2. If no failures have occurred in a fleet with n units having operating times

t1 , t2. t, and the units are sosceptible to an, exponential failure mode,
then a conservative 100 (I - (0"'. one-sided lower confidence limit on the
mean life 0 is:
n
Zt,
-1I
(E.2)
163
M47'
~ ~ -i
""C(,onservat ive meatis t Irue confidence .levelis unkmiowui. limt is lit least
I hait he
"~ ~ IM o|
r)". .
'hu.s. if no failtires have oct'irred in it fleet withIi unifs having owrating times I . In
awlt I lte units lire susceptible to it Weiulll failure Inode with known ji (i!ind unknown 17),then a
,) ive I(W) (I
*nserva .) ' one.sided lower confidence limit on I# -j is:
or '
.'.The Weihayes lower bound on q (E.1) is eq-.jl to the lower confidence bound in (E.3) with
[the denominator, - Qn a = 1.. Solving for twr'find:
-Rn 1.0
R~na = j
-1.0
-
'
i .
, ~(V =
=0.368 e- 1.0
Thus, the Weihas.es lower b~ound on I is a 100 (1 -0.368)% = 63.2"t conservative lower
confidlnce bound for q. i
The confidence level can! be increased hy decreasing a in the denominator of the expression
on the right hand side of ndeouahty
(E.3). For example, a conservative 90V lower confidence
im~und on I can he calculated by setting (v = 0.10, giving
S>-(
" ,)""with at least 90% confidence
"enO.1I
onth rgthadsie f re~ahty E.).Fr xapea onevaiv 0%loercn6dec
""Not e that with # assumed, determining a lower bound f: r also determines a lower bound
,..lo~r I he Weibuli line.

a T If failures have occurred, and Weibayes is 1
usea , * is the maximum likelihood estimator of
*the in. This is shown by finding the value of that maximizes
true value of d the WeibuU likelihood
equations from Appendix D, whille assuming that 8,the Weibull slope parameter, is known.
"T'hesecalculat ions, similar to those discussed in Appendix D, result in the following equation for
the maximmn likelihood estimat or o1"q (assuming that tiis known):
Sr Y (E.4)
Equations E. I (Weibayes) and E.4 (maximum likelihood) are identical, demonstrating that
the Weibayes equation yields the maximum likelihood estimator of V, when failures have
occurred and i is known.
41
I"
''
ii
165
;4- It
APPENDIX F
MONTE CARLO SIMULATION STUDY

ACCURACY OF WEIBULL ANALYSIS METHODS
1. FOREWORD
Safety considerations in aerospace operations require corrective action based on very small
samples of failure data. As this is unusual compared to other Weibull applications, Monte Carlo
simulation was used to study-the accuracy of Weibull analysis when applied to data from a fleet
of several thousand successfully operating units and very few service failures (three to ten
failures). Of prime importance were the accuracy and precision of the risk forecasts, the 3
estimates, and the B.I life estimates. Two methods of Weibull analysis were considered and
comparedi: median rank regression and maximum likelihood. These two methods are among the
most commonly used for estimating ji and I with multiply censored data (i.e., data containing
hot h failures and suspensions, with the suspensions and failure times intermixed). Median rank
regression is covered in Chapter 2 and Appendix C. Maximum likelihood methods are
introduced in Appendix D. Details of the Monte Carlo simulation method are discussed later in
this appendix.
Figure F.1 illustrates the meaning of the accuracy and precision of Weibull parameter
estimates. A simulated fleet with 2000 operating units was introduced to a Weibull failure mode
with d = 3 and il = 13,000 hours. When the fleet experienced its 5th failure, a Weibull analysis
was performed. The estimated values of # and 17were stored, the fleet was re-created and re-
introduced to the Weibull failure mode. A Weibull analysis was again done at he time of the 5th
failure. 'rhe 11 and q estimates were stored, and the process was repeated many times, for a total
of 1000 estimates of # and jj under the simulated circumstances. Figure F.1 is a histogram uf the
ft estimates. It shows that 294 of the 1 000 estimates of # were between 0.5 and 2.5. (No estimates
were below 0.5). In 392 out of I000 simulation trihls, 0 was estimated between 2.5 and 4.5.
"Accuracy" refers to the difference between the "typical" # estimate and the true value of 11.In
this case, the median d estimate is 3.5 (150", of the estimates are below 3.5, 50% are above 3.5)
and the true value of d is 3.0: the # estimates are quite accurate with 5 failures. Precision refers
to the variability in the # estimates, and is normally measured by the standard deviation. The
more the # estimates deviate from their mean value, the higher is their standard deviation.
The standard deviation of the 1000 0 estimates f 1,#, .. 1000, is calculated as:
ton
S - I-,
S999
where
1 0j,g'A+
f2 + ... +0m 100o
166
40
35 - 1000
Sample Size
Average
30 True
25 -
Percent 20 -
ofI
SOccurrences I-
15 -Precdslon
10 -
5o
0.5 2.5 4.5 6.5 8.5 10.5 12.5 14.5 16.5 18.5 20.5 22.5 24.5
Beta Estimate
FD 256570
FiOcurerFre. Bcla Etimats: True Beta 3. Five Failures
167
4., 2. BETA (0) ESTIMATES
The accuracy of the maximum likelihood (ML) and median rank regression (MRR)
estimates, calcuated with a small number of failures and a large number of unfailed units, is
illustrated in Figure F.2.
The standard deviations of the small sample # estimates are shown in Figure F.3.
The following can be seen from Figures F.2 and F.3:

I. Both methods of analysis, median rank regression and maximum likeli-
hood, tend to overestimate 6. When the failure data are plotted on Weibull
paper, the slope is generally too steep.
2. The accuracy and precision of the # estimates improve as the number of
failures increases.
3. Both methods of analysis produce estimates of comparable accuracy.

Maximum likelihood estimates are noticeably more accurate for 0 = 5.0,
especially with very few failures.
4. Maximum likelihood # estimates are more precise than median rank
regression estimates, especially when the data contain as few as three to
five failures. In the three-failure case, maximum likelihood estimates have
Sbetween estimates.
44% and 68% less variability than median rank regression
3. B.1 LIFE ESTIMATES

The HI life is frequently used as a design criteria. Its estimate from field data is often
compared to the predicted or design BI life, so its accurate estimation is important.
Figure F.4 illustrates the accuracy of the B.1 life estimates over the range of the study.
Figure V.5 shows the B. life standard deviations as a function of sample size (3 to 10
failures, I0(X) to 20XX) unfailed units).
147W
168
U- 0
u* -6
9 E
I Cf
0
o co
IXI
Li) -6
4a)
co m
Co
169
0 0
cm 0
-00 0
I P
E co
a I.
Z; > IZ
0 I) ) I
I II170
II
Go OD
ogo
FC4
040
"Co C 71
| | 1! !|
171
.1 I
I 39c
caI
ccI
LUI
I I I*
Figures F.4 and F.5 indicate that:
1. Median rank regression estimates of the B.1 life are typically conservative:
the estimated B.I life is more often than not less than the true 3.1 life.
2. The accuracy of the median rank regression B.1 life estimates improves as
the number of failures increases.
:3. For = 0.5, 1, and 3, the maximum likelihood method typically

overestimated the B.1 life (ML overestimated the B.1 life in approximately
58% of the simulation trials, for = 0.5, 1, and 3.0 MRR overestimated the
B.1 life in - 40% of the simulation trials).
4. For # = 5, maxinum likelihood typically underestimated the B.- life (ML

underestimated the B.1 life in 52% of the simulation trials, vs. 58% for
MRR).
5. The precision of the B.1 life estimate does not necessarily improve as'the
number of failures increases from three to ten.
4. ETA (n) ESTIMATES
The medians of the characteristic life estimates from simulated fleets with few failures are
shown in Figures F.6 and F.7.
"TableF.1 contains the standard deviations of the characteristic life estimates.

TABLE F.I. STANDARD DEVIATIONS OF THE CHARACTERIkTIC.
LIFE ESTIMATES
Standard Deviations of the

Eta Estimates
Type True True 3 5 U0
Estimates Beta Eta Failures Failures Failures
MRR 0.5 100,003,000 4.8 x 10 2" 4.3 X l0' 6.5 X 1016
M11 2.7 x 1014 1.9 x 10"4 3.0 X 1010
MRR 1.0 500,000 &8 x 1013 1.3 X 1012

8
1.2 X 109
6
ML 4.6 X 10" 9.2 X 10 8.1 X 10
MRR 3.0 13,000 320,000 76,000 21,000

ML 27,000 15,000 6,300
MRR &.0 5,175 8,200 5,900 2,200

ML 3,300 3,000 1,900
MRR - Median Rank Regression

ML - Maximum Likelihood
173
/
True
lX10 8
- -= - Rank Regression
Maximum Likelihood =0
-
Eta (,,) 1X1o 7-'-I
- /
ixio 6 I I 1 I I
0 1 2 3 4 5 6 7 8 9 10
Number of Failures
FD 256548
Figure 1-76. inL AccuravY Alhdian ('harac'trisics Ife, .tima,.s -- A
174
N"
500,1000
sos-
100.000
True q
rado t m-. ao m m 3
ETA )
10,000 _ _ _ _ _ _ _ _
"5,175 -=True 5
-- - --- -- - -m - n
riS
, -w - , Rank Regression
- , - Maximum Ukelihood
1,000 5 6 7 8 9 10 11 12
Number of Failures
FD25669
Figure F.7. Eta Accuracy Median CharacteristicsLife Estimntes - B
175
Table F. I arid Figt:res F.,, and F.7 indicate that.
I. The median rank regression (MRR) and maximum likelihood (ML)

estimates are conservative: q is typicall", underestimated.
2. The accuracy of both types of estimates improves as the number of failures

increases from 3 to 10.
:1. The precision of the characteristic life estimates improves as the number
of failures increases.
4. The standard deviations of the r, estimates are extremely large, reflecting

the presence of several extremely large estimates of q at each simulated
condition.
S. RISK FORECASTS
This section addresses the accuracy and precision of risk forecasts made when there are
only 3 to 10 failures and 1000 to 2000 suspensions in a fleet (a risk forecast is a prediction of the
number of failures expected to occur over a period of calendar time). Methods fc- constructirg
risk forecasts from multiply-censored life data can be found in Chapter 3.
The accuracy and precision of the risk forecasta are assessed by advancing a simulated fleet
through a known Weibull failure mode to the time of its 3rd, 5th, and 10th failures. A Weibull
analysis is done anid the risk forecast is made, up to 12 months into the future. The fleet is then
advanced 12 months further through the Weibull failure mode using an average military aircraft
utilization rate. The risk forecast is then compared to the actual number of failures csused by
the fleet's additionas 12 month advance. This procedure is repeated 100 times with 100 different
simulated Ilvel s to assess the variability and accuracy of the risk estimates.
5.1 RISK FORECAST ACCURACY
Tah!e F.2 contains the medians of the maximum likelihood and rank regression forecasts
for 0, 6. and 12 months into the future. The entry under "0" months ahead indicates the
cumulative number of failures expected 'to date' (at the time the analysis is performed). Thus,
when 11 :1 and there are three failures in the fleet, there are 5 - 3 = 2 additional failures
expected over the 12 month interval following the occurrence of the third failure. The rank
regression method predicts an "average" of 11 - 5 = 6 additional failures, and maximum
likelihood predicts an "average" of 6 - 3 = 3 additional failures.
176 /
.117Mn,- .. /
TAPLE F.2. RISK FORECAST ACCURACY
= 0.5
Median Median
No. Months Median MRR ML
Faikures Ahead Risk Forecast Forecast
3 0 3 4 3
6 3 6
12 4 7 4
5 0 5 6 5
6 5 7 5
12 6 7 6
10 0 10 11 10
6 10 11 10
.12 10 11 i1
a='
3 0 3 4 3
6 3 5 4
12 4 6 4
5 0 5 7 5
6 5 7 6
12 6 8 6
10 0 10 11 10
6 11 il 11
12 11 12 11
1-3
Median Median
Net. Maonth. Median MR9 ML
FailureX Ahead Risk Forecast Forecast
0 3 3
6 4 8 4
12 5 if 6
0 5 7
6 6 c 6
12 7 11 8
10 0 10 12 10
6 12 14 12
12 14 17 15
a-s
Median Median
No. Months Median MRR ML
Failureq Ahead Risk Forecast Forecast
3 0 3 6 3
6 4 12 5
12 7 22 8
5 0 5 8 5
6 7 14 8
12 10 25 12
10 0 10 13 1'
6 14 19 14
12 19 28 21
177
',Np
From Table F.2. it is seen that:
1. Rank regression risk estimates tend ito be very conservrtive . they

overpredict the numlbr of failures. Maximum likelihoKo risk forecasts are
only slightly conservative.
2. Forecasts based on maximum likelihood Weihull analy.wes are always more

accurate than forecasts based on rank regression Weibull analyses.
3. The accuracy differences between the two methods increases as 0

increases. Maximum likelihood methods are much more accurate than
rank regression with # 5; they are only marginally more accurate when ji
=0.5.
5.2 RISK FORECAST PRECISION
Table F.3 presents the standard deviations of the risk forecasts 0, 6, and 12 months into
the "future". (The risk forecast 0 months into the future is the expected number of failures to
date.)
TABLE F.3. RISK FORECAST STANDARD DEVIATIONS

ii = o..5
"No. Months MRR ML
Failures Ahead Forecast Forecast
3 0 22.0 0.003
6 52.0 0.98
12 91.0 2.3
5.0 3.8 O.002

6 4.8 0.45
12 6.1 0.91
10 0 2.8 0.004
6 3.1 0.22
12 3.3 0.43
#- I
No. Months MRR ML
Failures Ahead Forecast Forecast
3 0 73.0 0.002
6 94.0 0.84
12 112.0 2.3
6 0 34.0 0.002
6 56.0 0.59
12 67.0 1.4
10 0 3.2 0.0o6
6 3.6 0.2,1
12 4.1 0.49
178
TABLE F.3. RISK FORECAST STANDARD DEVIATIONS
a=3
No. Montha MRR ML

* Failurta Aherd Foreefat Foremst
3 0 14.0 0.006
6 19.0 1.2
12 24.0 4.4
a 0 7.4 0.006
* 10.0 1.1
12 14.0 3.3
10 0 4.0 o.011
6 5,.1.79
12 6.5. I
No. Month* MRR IL

Failums Ahead Fomeast Forecast
3 0 11.0 0.00
6 37.0 U.4
12 88.0 5.U
0 4.7 O.011
6 14.0 V7
12 36.0 5e
10 0 4.4 0.016
6 12.0 L.5
12 29.0 4.4
"TableF.3 shows that:
I. The rank regression forecasts ,iary substantially (up to 50 times more than
forecasts based on maximum likelihood Weibull analyses).
2. Maximum likelihcod forecasts are far zhore precise than rank regression
forecasts, over the entire scope of this study.
3. As the number of failures used in the Weibul' 3nalysis increases, the
S. precision of theresulting risk forecasts im roves.
L MONTE CARLO SIMULATION
The Monte Carlo simulation consists of input, rocessing, and output segments. The
components of each segment are listed in Table F.4.
179
TAHBLE F.4. (COMP)ONENTiS OF MONTE (CARLOI.)1MUILATIOR
nput Frovensing Ouhtput

I. F*leet gist 1. Fleet -mstruction 1. Weibull parameter estimates
I Fleet age distribution 2. Fleet ag~ingl 2. One-year-shead ris k forecasts

.1 . Prodluctikon schedule 31. New member addition 3. Actual one-year ahead failure
.,4. Usage rate 4. Random failure time genera- c.ounts
"-"&5 True Weibull pararneters tion
'& Random numlh 5. Parameter estimation
6 Risk forecasting[
Monte Carlo S~The
simulation procedure is illustrated in Figure F.S.
Fleet characteristics and Weilull failure mode parameters arc input into the simod;|tor.
The fleet characteristics are (11, the number of units in the fleet; (2) the fleet age distribution. (3)
the production schedule; and (4) the per-unit monthly usage rate. The age distribution is
assumed to be normal, with IO(XX) to "-.)(XX) units in the fi.,-et. New members are brought into the
fleet ac~cording to a production schedule and utilized art a rate typically experienced with Pratt &
Whitney Aircraft military gas turbine engines. The simulator generates fikilares among thle units
in the fleet according to the Weihull failure modle input. Sliope paramete, fibl~s) of one-half, one,
three. and rive were chosen to represent infant mortality, random, wearout, and rapid wearout
failure modles. respectively. Se Figure F.9.
The processing segment includes constructing and aging the fleet, incorporating new
S~members, generating random failure times, estimating the Weihull parameters, and forecasting
the additional number of failures expected in the year ahead. Failure times are gen erated for
each member ofl the fleet using random numbers and the inpu~t WeibuUl paramneters. The fleet is
aged util :1. 5, and 10 1AH~
.4 (OM()I~TSOF
failures ON1:('RI()1I8LTO
occur. Weibull analyses are performed at these times. and the
year-ahead I'-recarts are made. This proc~essq continues until 10 failures occur in the fleet. The
fleet is i hen revonst ructed, and the process begin.s again. For the rank regres.sion estimators. this
p~rocess was repeated i0(X) i.ihes fopr each failitre modte considered. Covst considerations limited
this nui lhr to I(N) fr the nmaximum likelihlmd estimato~rs.
I'aranivihr tesliniates, year-ahead failure forecasts, and actual year-aheadifailure cou.nits are
oonlput. The error in the estimation method is reflected in the differen('e. hetween the parameter
e.tiniatp.4 and the actual Weibuli parameters input to the simulator. Tlhe forecasting error is
simply the difference between the actual and forecast
180' failure counts.
|.
0
U~(3
IW,
L +3
Cd
*ZIA
181
*. *..... .
* p
Fleeu
/
Wei~bull
Failure
Mode
M.
C-
FD 256451
Figure F.9. Initial Conditions of Simulator
.8
182
S
SIi
APPENDIX G
RANK REGRESSION METHOD VS MAXIMUM LIKELIHOOD

METHOD OF WEIBULL ANALYSIS
This appendix contains a summary of the strengths and weaknesses (and general
comments) of the rank regression method and the maximum likelihood method of Weibull
analysis.
1. Rank regression provides a graphical display of the d-.ta. This helps to

identify instances of:
a. The Weibull distributin not fitting ..-ell to the data
(suggesting perhaps another distribution, like the log-nor-

Mal),
b. More than one failure mode affecting the units,
c. Data needing a t. correction,
d. Outliers in the data,
e. Batch problems.
Maximum likelihood does not provide a graphical display of the data. . -
2. Rank regression provides accurate estimates of "low" percentiles like the

B.A life under the conditions simulated in Appendix. F, even for small
sample sizes. "Low" percentiles refer to percentiles close to the time of the
first failure. Maximum likelihood "lower" percentile estimates have a
slight positive bias with small numbers of failures and a large number of
suspended items.
3, Rank regression risk forecasts are conservative (overestimate the risk) and
less procise, when computed with few failures and a couple of thousand
suspensions. Maximum likelihood risk forecasts are more accurate and
preci3e than are rank regression risk forecasts with small failure samples.
4. Both rank regression and maximum likelihood tend to overestimate t with
small failure samples. (The slope on the Weibull plot is too steep.) This
positive bias decreases as the number of failures increases.
5. Confidence intervals on the Weibull parameters # and q based on rank

rekresaion estimates are not available. Statistical hypotheses about P and q,
a
(e.g. Is = 1, implying the failure mode is random or memoryless?) cannot
be tested using these estimates. Exact or approximate (large sample)
confidence intervals on the Weibull parameters #i and q, based on the
maximum likelihood estimates, are available for all commonly occurring
forms of censored data. Statistical hypotheses about t and q can be tested
using these estimateu.
183
-N
-. . 5-.. 5..-' .5 % . N1 . _ ,%
7- -'77-7' .77~-777-7 7 7577
".5,
...
"APPENDIX H
WEIBULL PARAMETER ESTIMATION

COMPUTER PROGRAMS
Enclosed are program listings for estimatinIg the Weibull (distrilnt ion 6ir data containing
- both complete and censored sample%. Two programns are provided, ont, in BASIf' and one in
FORTRAN. both ,)f which were developed I'm and widl run on a TRSS4 MNKslol I or Ill
microcomputer. They take approximately lOK bytes of RAM. hut. in t he case of the FORTRAN
program, will require a FORTRAN compiler, at least one d..;k drive. and 48K memory.
Botl; programs will, estimate the parav.ieters of a Weibull (listrihution by the rank
regression and maximum likelihood technique. Hence, they can be used to solve many of the
5problems and examples in this Handbook. Of course, use of the programs beyond this Handbook
.a can be male; however, first the programs should be checked thoroughly by the user.
*4
The programs run in the "immediate" mode, i.e.. the user is prompted for i.iput. with no
printout options. The programs can be used to generate estimates of the Weibuil distributior
parameters for samples of size less than or equal to 100. If more than 100 fiailures are to be input,
the dimension statements at *he front of both programs should be increased.
If a histogram of a suspension population is available, along with the failure times of

%
interest, a Weibull analysis can be produced for quite large samples. The histogram is input bar
by bar and the programs assign the midpoint time of each bar to all of the units in the bar.
Of p1 ecitl note is the maximum ,ikelihood parameter estimation capability in both

programno. These programs illustrate the solution of the maximum likelihood equation
(Appendix 1).:I by the use of the Newton-Raphson iteration procedure.
It should be noted that, for those example canes that ate run with histograms from this
Handbook, the same answers as given in the Handbooxk may not be achieved. The parameter
estimates should be close, however, and will only be different becautse of the histogram input in
Slie case of t hese programs, where the individual times on each suspended unit were used for the ...
examples in the Handbook.
184
. -:
.7w 7-76 1
Listing 1. .A.SIc(' Pro.,raim for Weihu!l I'arrnet'trE.s/itiatitpnn
10 I)IM ]SA(100),tIA(100),A(100),IO()100),'I'T(P10),XZ(100)
20 I)IM \'X( Iom,IN(100)),TII t100),V'(100),IZ(I 00),X( 100)
30 IP1)
40 IM 0
50 PRINT"ARE YOlU INPIUTTING A HISTOGIRAM OF S1JSIPENSIONST?"
60 PRINT"ANSWElR Y OR N"
7O INPtUT A$
80 IF A$="N" GOTO 310
90 IP=I
100 PRINT"AN INTERVAL SIZE OF 50 IS ASSUMED"
110 PRINT"0.K .... ANSWER Y Ol N"
120 INPUT A$
130 IF A$ < > "N" GOTC 170
- 140 PRINT"PLACE THE INTERVAL SIZE YOU WILL USE IN"
150 PRINT"CC 1-10 W/DECIMAL"
160 INPUT PE
170 PRINT"PI,ACE THE NUMBER OF 'ELEMENTS IN EACH INTERVAL OF"
180 PRINT"THE HISTOGRAM IN CC 1-10 W/DECIMAL"
190 PRINT"USE -99. TO INDICATE THE END"
200 M=0
210 M=M+I
220 INPUT XI
t 230 IF XI = -99. GOTC 250
240 IN(M)=XI: GOTO 210
250 MM=M-1
260 FOR I=1 TO MM: PRINT IN(I);: NEXT I:PRINT
270 IM=0
280 FOR ,1=1 TO MM
N4 290 TI(J)=PE/2.+(.J-1)*PE
300 IM=IM+IN(.I): NEXT J
310 PRINT"INPUT THE FAILURE DATA AND SUSPENSIONS IN"
320 PRINT"CC 1-10 W/DECIMAL... USE -99999. TO INDICATE"
330 PRINT'"THE END OF DATA (NEGATIVES INDICATE SUSPENSIONS,"
340 PRINT"UNLESS A HISTOGRAM WAS INPUT)"
350 1=0
360 I=I+l
370 INPUT A(M
380 IF A(I) =0.0 GOTO 410
390 IF A(I) = -99999. GOTO 430
400 GOTO 360
410 I=I-I
420 GOTO 360
430 N=I-1
440 BN-=N+IM
450 FOR ,I=1 TO N: AW=A(,J): V(.J)=ABS(AW): NEXT J
460 GOSUB 5000
470 rOR I=1 TO N: IU=IZ(i): X(I)=A(IU): NEXT I
480 FOR 1=1 TO N: A(I)=X(I): NENT I
490 BI=BN+I: DJ=1.0: 13J=0.0: M=0: SX=0.0: SY=0.0: XX=0.0: YY=0.0: XY-0.0
500 PRINT"PT. DATA ORDER MEDIAN RANK"
510 FOR K=l TO N
185
Ibj.ingji 1. B4Ii.1.1 I' n f~pr WlEjibUi/I / ira
grajr~ b.'.'
Fs irmiorl'n(( 'mjlimuv 4)
52 IM 1)
.50IF II' (110T() 180)
540) FOR V- I TO MMI
*550) IF TI(.i) < A(K) THEN IM~IMi IN(.J)
56CIA NEXT J1
570 IS(K)=!Mi
580' BK=IM+K
581 IF IP=1 THEN BK=BK-1.0
590 IF IPil AND K--I THEN D~J=(II1-BJ)/AH1-BK)
6W! K~lGOTO 6:30
610 IF IS(-K)=IS(K-1) GOTO 6:30
620 D-I=(B1-B.J)/(B1-BK)
6,30 IF A(K) < 0.0 COTO 660
640 IF A(K) =0.0 COTO 840
650 IF A(K) ->0.0 GOTO 670
660 D.J=(BI-13.J)/(BIl-BK): GOT() 720
*670 BJ=B.J+D.J: RO(K)-B.J: RA(K)=(RO(K)-.3)/(IlNE-4)
680 XI =LOG'(A(K)):YPz 1./(l.-RA(K)):Y=1.0OG(IOG(YIl)):Y'X(K) =Y
681 REM PRINT: "B.J=";BJI;"D.J=";D.1;"BK=";BK;"B1 =";B13
690 PRINT K.A(K),R.O(K),RA(K)
700 M=M-II
710 SX=SX+X1:XX=XX+X1*X1:SY=SY+Y:YY=YY+Y*Y:XY=XY+X1*Y
.720 NEXT K
7310 GM7=M
740 BE~= ((M*YY-SY*SY)/(GM*XY-SX*S;Y)
750) At,= (BE*SX-SY)/GM:AV=~AL/BE:AV=EXI)(AV):ST- LIE
760 PRINT "I3ETA=";ST;" ETA=";AV
770 RmI).(
780 IF (XX-sSX/(X;M) < 0.0 COTO 820
790 XN=~XY-SX*SY/GM
*800 DE- 'QR((XX-SX*SX/CM)*(YY-SY*SY/CM))
* -810 R=XN/DE
*820 RQ=R*R
*821 IF' RQ >1.0 THEN RQ=1.0
83% PRINT "R= ";R;" R SQtJARF,";RQ
8:11 IPRINT-DO YOU WISH TO DO MAXIMUM LIKELIHOOD ESTIMATION?"
9:32 PRINT"ANSWER Y(E.S) OR N(O)"
8413 INPUT A$
8414 IF A$="N" GOTO 840
815 NF=0: PRINT"PLEASE BE PATIENT..IT'S ITERATING"
&36 GOSUI3 2000
840 END
2000 FOR I=1 TO N
2020 IF A(I)<0.0 GOTO 2050

2030 NF=NF+I
2040 XZ(NFi=A(I)
2054) NEXT I
2064) OT=.0001: NL'=l00: XB=BE: YA=-.001: NC=0: DX=-.001: DY= .01
21:30 GOSUJ3 3000: YB=AU
186
7% 7C.-"-
Listing I. BASIC Program for Weibull Parameter Estimation (Continued)
21 :L5 GOS1B 40X)0: XB -BB#

2136 ON JK (oTro 2130,2150,2140,2150
21:17 GOTO 2130
2140 IPRINT"ITERATION FAILURE"
2145 PRINT"BETA=";XB;" !.N MAXIMUM ,IKELIHOOD=";YB
2147 RETURN
2150 BL=XB: SU=0.0: RN=NF: FOR I=l TO N: SU=SLJ+TT(I)[BL- NEXT I
2160 IF IP=O GOTO 2250
2170 FOR I=I TO MM:SU=SU+IN(I)*Tl(I)[BL: NEXT I
2250 SU=SU/RN:TL=SUI(I.0/BL)
2260 PRINT"MAXIMUM LIKELIHOOD ESTIMATES FOLLOW"
2270 PRINT'BETA=";BL;" ETA=";TL
2280 RETURN
3000 SI#=0.0 : S2#=0.0 : S3#=0.0
3010 IF XB>15. OR XB < =0.0 THEN XB=0.1
3020 FOR 1=1 TO N
3021 PO#=TT(I)IXB
3022 SI#=S1#+PO#
3023 S2#=S2#+LOG(TT(I))*PO#
3030 NEXT I
3040 IF IP=0 GOTO 3130
3050 FOR I=1 TO MM : PO#=TI(I)[XB: SI#=Sl#+IN(I)*PO#
3060 S2#=S2#+IN(I)*LOG(TI(I))*PO#: NEXT I
3130 FOR I-1=TO NF: S3#=S3#+LOG(XZ(I)): NEXT I
3140 AU=(S2#/Sl#)-(S3#/NF)-(1.0/XB)
3150 RETURN
4000 JK=1: BB#=XB
4010 IF (ABS((YA-YB)/YA)-OT) < = 0.0 GOTO 4290
4020 IF (NC-I) < = 0 GOTO 4040
4030 GOTO 4090
4040 DX#-BB#
4050 DY#-YA-YB
4060 NC=NC+1
40M0 BB#=BB#1*.02
4080 RETURN
409C IF NC > NL GOTO 4300
4100 X2#-BB#
4110 D2#fYA-YB
4120 IF ABS(D2#-DY#) < .00001 GOTO 4320
4130 BB#=X2#-D2Vt*(X2#-DX#)/(D2#-DY#)
4150 IF BB# < = 0.0 GOTO 4250
41G0 IF BB#<X2# GOTO 4190
4170 IF BB#=X2# GOTO 4240
4180 IF BB#>X2# GOTO 4210
4190 IF BB#/X2u > =.6 GOTO 4250
4200 BB#=X2#*.75: GOTO 4250
4210 IF BB#/X2# < 1.4 GOTO 4250
4220 BB#=X2#*1.25
4230 GOTO 4250
4240 BB#=X2#*1.02
4250 DX#=X2#
4260 DY#=D2#
187
,N, W . "M
Listing 1. BASIC' Iriogrntm loor Wrihull Parameter E.'lionation (Conttinued)
4270 NC NC, I
4280 r.TI'N
4290 .1K 2: NC 2: RKi 'IRN
43WN) IRIINT"FAIIEI) TO CONVEI{R;E"
43110 .1K -3: NC -1: RTUI'tRN
4320 JKA4: RETURN
50O) FOR .J=1 TO N: IZ(J)=J: NEXT J
5010 IF N=I RETURN
5020 NM=N1-
5030 FOR K=1 TO N
5040 FOR J=l TO NM
5050 NI=IZ(J)
53W60 N2=IZ(J+I)
5070 IF V(NI) < V(N2) GOT3 5090
5080 IZ(J+I)=Nl: IZJ)=N2
W90 NEXT J
5100 NEXT K
5110 RETURN
188
a-.
!*
i.isting 2. FY)R''RA.N Pr g#ram for Weihill Parameter Estimation (C'ontinued)
COMMON /BiO(CK I/INT.TIME,MM,IHIST

DIMENSION ISt I00).RANKMI)(Ii)),A( I0xU).)Rii)ERI(10H))
DIMENSION YX(IXE)).INT( 100),TIME(I00),V(I(X))
DIMENSION IZ(100),X(AGW)
DATA ANO'N'/,PERINT/50./
IPRNT=I
IPOP=0
ISUM=0
WRITE (6,4001)
4001 FORMAT (2X,' ARE YOU INPUT'IING A HISTOGRAM OF SUSPENSIONS?',
?/,2X,'ANSWER Y OR N')
READ (5,4002) ANS
4002 FORMAT (Al)
IF (ANS .EQ. ANO) GOTO 701
IPOP=1
WRITE (6,4003)
4003 FORMAT (2X,'AN INTERVAL SIZE OF 50 IS ASSUMED,',
?/,2X,'O.K.?......ANSWER Y OR N')
READ (5,4002) ANS
IF(ANS .NE. ANO) GOTO 8801
WR;TE (6,4004)
4004 FORMAT (2X,'PLACE THE INTERVAL SIZE YOU WILL USE IN',
?/,2X,'CC 1 - 10 W/DECIMAL')
READ (5,4005) PERINT
4005 FORMAT (F10.0)
8801 CONTINUE
WRITE (6,4006)
4006 FORMAT (2X,'PLACE THE NUMBER OF ELEMENTS IN EACH'
?J,2X,'INTERVAL OF THE HISTOGRAM IN CC 1 - 10,',/,2X,
?'W/DECIMAL. . USE -99. TO INDICATE THE END')
m=0
211 M=M+I
READ (5,1007XINT
1007 FORMAT (F1O.0)
IF (XINT .EQ. -99.) GOTO 212
INT(M) = XINT
GOTO 211
212 MM - M. 1
WR!TE (6,3090)(INT(KL),KL= 1,MM)
3090 FORMAT (1014)
ISUM = 0
DO 2001 J=i,MM
TIME(J)=PERINT/2.+(J-1)*PERINT
ISUM=ISIJM+INT(J)
2001 CONTINUE
C INTERMEDIATE PRINT
WRITE(6,798)(TIME(J),J= 1,MM)
798 FORMAT(8F10.1)
C INTERMEDIATE PRINT
701 WRITE(6,2000)
189
2tXX) FORMAT (2X,.'INPU'1T THE FAILUIRE D)ATA AND StUS PENS IONS'
?./.'-X.'IN COlS 1-10 WITH DECIMAL, -9999%.'
?,/.2X.'NDICATIES THE ENI) OF D)ATA (NEGATIVFES'
?./2XINI)ICATE SLt SIENS IONS).. tJ NLESS HISTOG RAM'
.I.X.W. INPUT'
READ) (.5,101) AMI

101 FORMAT(FIO.0)
IF WADI -EQ. 0.0)GOTO 3
IF (AMI .EQ. -99999.)GOTO 2
GOTO I
3 1=14i
GOTO 1
2 N=I-i
BN N+ISITM
DO 4 i-I=1,N
4 V(Jl)=ABS(AGJI))
(CALL. ORD(V,N,IZ)
DO 22 1I=1IN
1811B= IZ(I)
22 X(1)=A(ISUB)
1)0 23 1I=N
2:1 A~l)-X(I)
BNi =BN+i
D1=1.0
Ki 0.0
N; =0
SUIMX =0.0
StIMY ~-.0
SUNIXX-o0.
.SIJMYY--o.0
SUMXY- 0.0
WRITrE (i,990))
990 FORMAT (2X,'PT.',4X,'DATA',4X,'ORDER',4X, 'MEDIAN RANK')
DO 634) K IN
ISI IM=0
IF ([POP .EQ. 0)GOTO 6312
DO 631 .J=1I,MM
IF (TIME(.J .1.1T. A(K))ISUM=ISUM+INT(J)
631 CONTINUE
IS(K)=ISUM
632 13K=ISUM+K
IF (IPOP .EQ. I)BK=BK.1.0
IF(IPOP .EQ. 1 .AND. K .EQ. I)DJ=(BNI-BJ)/(BNi-BK)
IF'%K .EQ. 1) GOTO 3911
IF (IPOP .EQ. 0) GO TO 3911
IF(IS(K) .EQ. IS(K-1))GOTO 1911
3901 DJ=(BNI-BJ))/(jBNI-BK)
3911 IF WAK))390,900,40X)
390 DJ-(BNI1.BJ)/(BNI1-BK)
GOTO 630
190
Listing 2. FORTRAN Program for Weibull Parameter Estimation (Continued)
400 3J B.J+ DJ
ORDER(Kh13.J
RAN KMD (K) =(0RD ER(K) -.3)/AB N+4)
XXX=ALOG,(A(K))
YPRIME= ./(1-RANKMD(K))
Y=ALOG(ALOG(YPRIME))
YX(K)=Y
WRITE(6,300)K,A(K),ORDER(K),RANIVMD(K)
300 FORMAT (16,F1O. 1,Fl0.4,Fl2.5)
M=M+1
SUMX=SUMX+XXX
SUMXX = SUMXX + XXX * XXX
SUMY=SUMY+Y
SUMYY=SUMYY + Y * Y
SUMXY=SUMXY + XXX * Y
60CONTINUE
CWRITE (6,800O)SUMX,SUMY,SUMXX,SUMYY,SUMXY
800 FORMAT (2X,'SUMX=',E20l.7,'SUMY=',E20.7,/,2X,
?'SUMXX =",E20.7,'SUMYY=',E20.7,'SUMXY=',E20.7)
GM=M
BETA= (GM*SUMYY-SUMY*SUMY)/(GM*SUMXY-SUMX*SUMY)
*1 ALPLN=(BETA*SUMX-SUMY)/GM
AVED=ALPLN,/IETA
ETA=EXP(AVED)
WRITE(6,3101)
* 3101 FORMAT(2X,'THE FOLLOWING ESTIMATES ARE RANKED REGRESSION',
?'ESTIMATFS')
WRITE (6,3100)BETA,ETA
'3100 FORMAT (/,2X,'BETA=',F1O.4,' ETA =',E20.7)
R-0.0
IF((SUMXX-SUMX*SUMX/GM) .LT. 0.0)GOTO 7871
XNUM=SUMXY-SUMX*SUMY/GM
DENOM=SQRT((SUMXX-SUMX*SUMX/GM)*(SUMYY-SUMVV SUMY/GM))
R=XNUMJDENOM
781IF (R .GT. 1.0)R= 1.0
781RSQ=R-R
WRITE(6,3200)R,RSQ
3200 FORMAT (2X,'R-',Fl0.5,'R**2=',F10.5)
WRITE (6,5001)
500) FORMAT (2X,' DO YOU WISH TO DO MAXIMUM LIKELIHOOD ESTIMATION?',
* ?/,2X,'ANSWER Y OR N')
READ (5,4002) ANS
IF (ANS .EQ. ANO) GOTO 900
IHIST=IPOP
CALL MAXL(A,N,IPRNT,BML,TML,BETA)
90CONTINUE
STOP
END
SUBROUTINE ORD(A,N,IZ)
DIMENSION A(1),IZ(l)
.J.DO 1 J1,N
191
Listing 2. FORTRAN Program for Weibull Parameter Estimation (Continued)
4 1 IZ(.)' .j
IF (N .EQ. 1) RETURN
NM -N -I
D)0 2 K -1,N
DO 2 J= I,NM
N I -IZ(.)
N2rIZ(.+l)
IF (ANI) .LT. A(N2)) GOTO 2
.44 ~IZ(.J+ 1)
IZ(.J=kN2
=NI
2 CONTINUE
RETURN
END
SUBROUTINE MAXL(T,NUM,IPRNT,BMLTM!,,BETA)
DIMENSION 771 (l00),T(1),XX(100),INT(I00),TIME(IO0)
COMMON /BLOCKfl'T,NSAMP
COMMON 1131.OCKIIINT,TIME,M.M,'HIST
COMMON /BLOCK2/XX,NFAIL
4 IH=IHIST
NSAMP=NUM
NFA IL =0
* DO 1 I=I,NS.%MP
TT(I)=ABS(T(1,)
IF(T(1) .LT. 0.ODO)GO TO 1
NFAIL=NFAIL+!
XX(NFAIL)=T(I)
I CONTINUE~
TOL= .0OW00
NLIM=100
X=BETA
PB=0.001
NCT=0
DIELX=.001
DELY=.01
30 PRN= AUX(X)
IF (IPRNT .EQ. 1)WRITE(6,206)X,PRN
206 FOMT2,BETA=',E15.5,' LN MAXLIKELIHOOD= ',E20.7)
GO TO (30,50,40,50),ISIG
GO TO 30
40 -IF (IPRNT .EQ. I)WRITE(6,205)FPR,PB,PRN
205 FORMAT(2X,' ITERATION FAILURE ',3E20.7)
IFLAG=1
RETURN
*50O BML=X
SUM=0.ODO
- .* .RN=NFAIL
DO 110 I=1,NUM
110 SUM=SUM+TT(I)**BML
IF (IHIST .EQ. C) GO TO 112
DO Ill I=1,MM
111 SUM=SUM+FLOAT(INT(I))*TIME(I)**BML
192
Listing 2. FOKIl'RAN P~rogram for We'ibull P'arame'ter &Rstimation(Continued)
112 SUM 7-SUM/RN

TIML ,SIM*-(I.0/BMI,)
* ~WRI'I'(6,996)
96FORMAT(2X.' MAXIMUM LIKELIHOOD ESTIMATES FOR THIS CASE FOLLOW')
WRITE(6,995)BIM,TML
995FORMAT(2X,' BETA=',F1O.3,' ETA=',F20.2)
RETURN
N * END
FUNCTION AUX(X)
DIMENSION T(100),XX(100),INT(100),TIME(100)
* COMMON /BLOCK/T,N
COMMON /BLOCK I/INT,TIMV,,MM,IHIST
COMMON /BLOCK2/XX,NFAIL
ZZZ=NFAIL
SUM 1 =0.0
SUM2=0.O
SUM3=0.0
IF(ABS(X) .GT. 15.0 .OR. X .LE. 0.0)X=0.1
DO 10 I=-1,N
SUMI=SUMI+T(I)**X
10 SUM2=SUM2+ALOG(T(J))*T(I) 4*X
IF (IHIST .EQ. 0)GO TO 11
DO 20 I'=1,MM
SUM1 =SUJM1+FLOAT(INT(I))*TIME(I)*-X
20 SUM2=SUM2+FLOAT(INT(I))*ALOG(TIME(I))-TIME(I)-'X
11 DO 15 I=1,NFAIL
15 SUM3=SUM3+ALOG(XX(I))
A AUX=SUM2/SUMI-SUM3/ZZZ-1.0/X
"I: RETURN
END
SUBROUTINE SLOPE( X , YA ,YB ,Xl,DEL1,TOLJK,NMTNTIMELOOP)
JK=l
IF((ABS((YA-YB),'YA))-TOL')6,6,3
-3 IF(NCT-1)1,1,2
1 Xl=X
DELI =YA-YB
NCT=NCT+1
X=X*1.02
GO TO 9
2 IF(NCT-NTIME)b,5,4
5 X2=X
DEL2'=YA-YB
IF (ABS(DEL2-DELI) .LT. 1.E-06) GO TO 20
* ~X=X2-DEL2*(X2-X1 )/DEL2-DELI)
IF(X)8,8,10
10 IF(X-X2)11,7,12
11 IF((X/X2)-.6)13,13,8
13 X=X2*.75
GO TO 8
12 IF((X/X2)-l.4)8,14,14
193
i ~ Listing 2. FORTRAN Program for Weibull iParaineierEstimation (Continued)
V.(% TO 4,r .W*V .. J
'14 X -X2*I1.25
7 X - X2"1.0}2
'8 XI=-X2
DELI =DEL2
NCT = NCT+ I
GOTO 9
6JK=2
NCT=2
TO 9 S(GO
" i4... WRITE(g,100)LOOPX,YA,YB
i 100 FORMAT(IH0,'CONVERGENCE FAILURE IN LOOP',I2/1H ,'X =',E14.8,4X,' YA=',
?El4 9,4X,'YB =',E14.8/
JK=3
NCT= I
9 CONTINUE
S~RETURN
S20 JK=4
I RETURN
END
.4
194
APPENDIX I
WEIBULL GRAPHS
Ii
4..
.1195
Op a
195 :
- 7
,
99.11WEIBULL DISTRIB3UTION
99.5
99.0 Beta
Eta-
sample size
06.0 Failures -
90 0Key: i o 1i
63.2
* 40.0
40.
* 30.0
20.0
/T
0.5
0.1 1 .2 4 556789 1.5 3 4689 1.5 35 78a911
D2M
196
7. .
.-.--. 7
H H+
ti t
I '-i t ---- I
iI I f I-IItf !
It ++
a
1 -tttt l -- --
CIA-
0 0
ID 00
C) )Q 0 cC
co q
OtCYOA
6Og
GO P o( C) 0N 0 C
WUBOJ9d GAfl9,flwfl
197
, 1 !/ /
WEIBULL DISTRIBUTION
4,' t .9 i , .. t .
99.0 i i
Beta=
" 95.0 Eta
"Sample Size =
" 90.0 -Falure3
-,Key: 1 to3
80.0 A~- ~j
4 70.0 _ -
63.2 - - --
60.0o
"50.0
40.0
30.0 -
20.0 ~4. I~ -
* "U
* 0
* a. 10.0 - I_ _ _ ____ _____._.i_..
05. - - - -
0 i1
- ii mz
Lzz . ... ....
-.. . .....
* , - - -_ -
0.5 -- 1-. 2--- . 3-.. -- 4- -
o.____ . .
_.. . - - --
01 -- ~1.5- - 3 4 5 6
FU 26WI
198
APPENDIX J
ANSWERS TO PROBLEM'S
i. CHAPTER 1 ANSWERS: None
K. 2. CHAPTER 2 ANSWERS
2.1 Probem 2-1
h , a. # =3.5 j=200 hours

b. Yes. Wearout or fatigue failures usually have steep slopes
c. B 1.0 54 hours.
Time Median*
Rank (hr) Ranks
1 85 6.6 %
2 135 16.2
3 150 25.8
4 150 35.5
5 190 45.1
6 200 54.8
7 200 64.
8 240 74.1
9 240 &3.7
10 250 93.3
From median rank table
&-tFigure J.1.
199
-. -. - \\ >.,
. 4
11 1
I J1
. EIBULL DISTRIBUTION I
4' Fj = 3.5 WERROUT -
"=200 HOURS
SSWLE SIZE1,= 10
a.. FAILtJRES = 10
B1.0 54 HOURS
U..-- 4 ..-....
II
Si' - - - I
Fiur. -
1. -ier -'- -1
20
A 'I -
'NK I
2.2 Problem 2-2: Answers
a. if 0.75 q = 29ghours
b. Infant Mortality
C. Serial numbers are very close. A batch problem mow be suspect.
Median
Rank, Comment New Rank Rank
1 Failure 1 8.3%
2 Failure 2 20.2
3 Suspension
4 S,..spension
5 Failure 3.4' 36.9
6 Sizapension ,
7 Fa.&ure 5.32 59.5
8 Fa&' ire 7.2' 82.1
Rank Increment - (N + 1) (previous rank order number)

1 + (number of items beyond present suspended items)
= 9-2= 7 1.4
I+4 5
=2 + 1.4 - 3.4
Cank Increment - 9 - U.4 - 5.6 - 1.87
1+2 3
, ew Rank - 3.4 + 1.87 - 5.27 a 5.3

SNew Rank - 5.3 + 1.87 - 7.17 7.2
See Figure J.2.
I-
S~201
_._ ; .... '. ,. - , - ',

7 I
w.- I,,EIBLLL DISTRIBUTION - - -
P = 0.75
= 29 HOURS
. SAMPLE.SIZE= S
u..: FAILURES =5 - - - - -
INFANT MORTAL I TY
IL.__ - - - -
THWE SEIA NUMBERS OF FIRST TO

I LAST FAILURE-- RRE VERY CLOSF
MAYBE A HEAT TREAT THAT
S- . . . . . . . ... . - . . . . . . ., ,
*/~!
FO 27189?
Figure' J.2. I'ruhlu'n, 2-2. Infari Mortality
202
2.3 Preblem 2-3: Answerm
Median *
Rank Time Ranks
1 90 hr 8.3
2 130 20.1
'1 165 32.0
4 220 44.0
5 275 55.9
6 370 67.9
7 525 79.8
8 1,200 91.7
"From m.edi-in renn table
Using the formit,la
(t3- t2) - 42 - ti)

where tJ = 90
t 2 - 185
t3 = 1,200
t, 185 -_1,200- 185) (185- 90)

"1,200 - 185) -(185 90)
-185 -(1,01,k) (95)

1,015 - 95
w 185 96,425
920
= 185- 104.8
8S
Median*
Time Time - 80.2 hr Rank
90 hr 9.8 8.3
130 49.8 20.1
165 84.8 32.0
S220 139.8 44.)
275 194.8 55.9
370 289.8 67.9
525 444.8 79.8
1,200 1,119.8 91.7
-SeeFigures J.3 and J.4.
i 203
A .
N'V
1--44-~ -4-4
W.IBULL DISTR I[BUT IONhT

SD0 = CAN'T BE ESTAB.
9 = CAN'T BE ESTAB.
"",SAP, 'LE S IZE =8
r
T, FAILURES = 8
'. .... low.. ..
St8U DSRIUTO
2204
TBE iES AB.T

=CI T , =10
FOATLUTIME TOTa. (H8
2O
'p.1
-~ -"
.EIBLOL DISTRIBUTION ,
=279.HR
SAMPLESIZE=88
S.. FAILURES =8
=80.2
ior
a.T
I- I
,o. , ;
/4 1
00.
isl
,. 'i .
'.7,
Figure J.4. Problem 2-3, Overall Population
205
& CHAPTER 3 ANSWERS
3.1 Problem 3-1
(a) 'T'oday
NumI7Lb(er Time on
Engines Each Engine F(t) F(t). N
20 1,50 0.0033 0.067
20 200 0.008 0.16
20 250 0.0155 0.31
20 300 0.0266
20 350 0.533
0.042 0.84
S=1.909
(b) In Six Months
Number Time on
Engines Each Engine F(t) F(t) ..N
20 300 0.0266 0.533
20 350 0.042 0.84
20 400 0.062 1.24
20 450 0.087 1.74
20 500 0.118 2.35
ff 6.704
Therefore, atdditional failures 6.704 -1.909 = 4.8
2W6
RM F.. .. .7
3.2 Problem 3-2: Turbine Airfoil Unscheduled Engine Removals
(a) First, overall using Figure 31.13 po:ihlation:
P"Ioint i)alta Mean Order Median Rank
"47 476.0M) 1.3:14 0.06582

48 504.(XX) 2.687 0.01330
58 576.0()0 4.128 0.02134
59 596.000 5.570 0.02937
60 600.000 7.011 0.03741
S61 60:3.000 8.453 0.04545
62 649.0M 9.894 0.05348
72 667.000 11.455 0.06218
7:3 681.000 13.016 0.07088
74 684.000 14.576 0.07958
75 701.000 16.137 0.08828
76 741.000 17.697 0.09698
85 755.000 19.388 0.10640
86 756.000 21.079 0.11582
87 770.000 22.769 0.12525
88 806.000 24.460 0.13467
89 812.000 26.151 0.14409
90 821.000 27.841 0.15.352
91 845.000 29.532 0.16294
93 850.000 31.242 0.17247
100 855.000 33078 0.18271
then, using the Figure 3.14 population
Point L,,ta Mean Order Median Rank
3 384.000 1.067 0.02442

4 701.000 2.133 0.05839
6 770.000 3.240 0.09361
7 812.000 4.346 0.12884
8 821.000 5.452 0.16407
9 845.000 6.558 0.19930
The associated Weibull Plots are in Figures J.5 and J.6; they "seem" different. However,
Figure J.7 illustrates the total population Weibull with confidence bounds (from Chapter 7);
since the Location A Weibull lies outside the All Locations Weibull, the two Weibulls are
significantly different.
20
. i/ . ,-. ..
4.
.s... IEIBULL DISTRIBUTION - -
5 = 5.634
V=1107.078
-- S*LE SIZE = 179
U. -L =
U..--- FAILURES 21 l l/ I
I I II I i l!I II l
25
448
1i148
- - .. - --.
TOTAL OPERRTINQ TIME (HR)
Figure J..5. Problem .-Overall

-V7 Population
208
P.....
1~..~' .S. ~ U I I I fl li I
E113LLL DISTRIBUTIONr
W
" 9.610
-.
= 981.8337
S*VHLE SIZE = 31
-- ,FAILUMS=66V
S.. "- - -- . .. .. .....

- . -; . . - .-. .. ..-
'I.I
0.01 10
hi
TOTAiL OPERATING TIME (HR)

i_ \
_ _ __ _ __ _ _ __ _ _ __
_ __ _ _ _ __ _ . iii___
FD 272253
Figure J.6. Problem 3-2. lwation A Only
209
. V,
.%,TWI S WZ W W. n.'P
a. FAILURES =21
V.
J.
S.S2
210
(b) For the entire popla)htion. in 12 months
Number of ('umulative H(t+360) -F(t

Units (N) IInits t I 1 360 t
FM) Pt 136o) I -- F(t) (;(t).N
9.0 9.0 50.0 410.0 0.0 0.0 0.0

7.0 16.0 150.0 510.0 0.0 0.0 0.0 0.
14.0 30.0 250.0 610.) 0.0 0.0 0.0 0.5
9.0 39.0 350.0 710.0 0.0 0.1 0.1 0.7
I 7.0 46.0 450.0 810.0 0.0 0.2 0.2 1.1
9.0 55.0 55".0 910.0 0.0 0.3 0.3 2.4
9.0 64.0 650.0 1,010.0 0.0 0.4 0.4 3.8
8.0 72.0 750.0 1,110.0 0.1 0.6 0.6 4.8
T7. 79.0 850.0 1,210.0 0.2 0.8 0.8 5.3
6.0 85.0 950.0 1,310.0 0.3 0.9 0.9 5.3
6.0 91.0 1,050.0 1,410.0 0.5 1.0 1.0 5.7
6.0 97.0 1,150.0 1,510.0 0.7 1.0 1.0 5.9
. 7.0 104.0 1,250.0 1,610.0 0.9 1.0 1.0 7.0
6.0 110.0 1,350.0 1,710.0 L.0 1.0 1.0 6.0
7.0 117.0 1,450.0 1,810.0 1.0 1.0 !.0 7.0
8.0 125.0 1,550.0 1,910.0 1.0 1.0 1.0 8.0
!. i
7.0I 1132.0 1,650.0 2,010.0 1.0 1.0 1.0 7.0
,% 8.0 140.0 1,750.0 2,110.0 1.0 1.0 1.0 8.0
8.0 148.0 1,850.0 2,110.0 1.0 1.0 1.0 8.0
4.0 152.0 i,950.0 2,310.0 1.0 1.0 1.0 4.0
3.0 155.0 2,050.0 2,410.0 1.0 1.0 1.0 3.0
2.0 157.0 2,150.0 2,510.0 1.0 1.0 1.0 2.0
1.0 158.0 2,250.9 2,610.0 1.0 1.0 1.0 1.0
- 96.6
211
, For the enwire p4pthilai i. in 2.1 monihs
.Number of Cumulativ.
nlits (N) IlInits (I f 720) Fi)
I1 720 Fit) H, 720) Hi,)
. ., .9.0 9.0 50.0 770 0 0.0 .1. 0.1I.I
7.0 16.0 150.0 870.0 0.0 0.2 0.2
14.0 30.0 250.0 970.0 0.0 0.4 0.4 5.:
9.0 39.0 350.0 1,070.0 0.0
7.0 46.0 450.0 1,170.0 0.6 0.6 5.1
9.0 55.0 550.0 1,270.0 0.0 0.7 0.7 5.2
0.0 0.9 0.9 7.9
9.0 64.0 650.0 1,370.0 0.0 1.0 1.0 8.7
8.0 72.0 750.0 1,470.0 0.1 1.0 1.0 7.9
7.0 79.0 850.0 1,570.0 0.2
6.0 85.0 1.0 1.0 7.0
950.0 1,670.0 0.3 1.0 1.0
6.0 91.0 1,050.0 1,770.0 0.5 6.0
1.0 1.0 6.0
6.0 97.0 1,150.0 1,870.0 0.7 1.0 1.0 6.0
7.0 104.0 1,250.0 1,970.0 0.9 1.0 1.0 7.0
6.0 110.0 1,350.0 2,070.0 1.0 1.0 1.0 6.0
7.0 117.0 1,450.0 2,170.0 1.0 1.0 1.0 7.0
8.0 125.0 1,550.0 2,270.0 1.0 1.0 1.0 8.0
7.0 132.0 1,6.)0.0 2,37(;.0 1.0 1.0 1.0 7.0
8.6 140.0 1,750.0 2,470.0 1.0 1.) 1.0 8.0
8.0 148.0 1,850.0 2,570.0 1.0 1.0 1.0 8.0
4.0 152.0 1,950.0 2,670.0 1.0 1.0 1.0 4.0
3.0 155.0 2,050.0 2,770.0 1.0 1.0 1.0 3.0
2.0 157.0 2,150.0 2,870.0 1.0 1.0 1.0 2.0
1(.0 158.0 2,250.0 2,970.0 1.0 1.0 1.0 1.0
= 128.8
212
212
II~uo
____
For the population at Ical.ion A, in 12 months:
Number of ('umulative
Units (N) llnu;ts tI3 F(t__ F(t 43.60) G(t) (;(t).N
0.0 0.0 50.0 410.0 9.0 0.0
,. 0.0 0.0 0.0
0.0 1Vro0.0 510.0 0.0 0.0 0.0 0.0
0.0 0.0 250.0 610.0 0.0 0.0 0.0 0.0
1.0 1.0 350.0 710.0 0.0 0.0 0.0 0.0
0.0 1.0 450.0 810.0 0.0 0.1 0.1 0.0
0.0 1.0 550.0 910.0 0.0 0.4 0.4 0.0
1.0 2.0 650.0 1,010.0 0.0 0.7 0.7 0.7
1.0 3.0 750.0 1,110.0 0.1 1.0 1.0 1.0
2.0 5.0 850.0 1,210.0 0.2 1.0 1.0 2.0
2.0 7.0 950.0 1,310.0 0.5 1.0 1.0 2.0
2.0 9.0 1,050.0 1,410.0 0.9 1.0 1.0 2.0
2.0 11.0 1,150.0 1,510.0 1.0 1.0 1.0 2.0
2.0 "3.0 1,250.0 1,610.0 1.0 1.0 1.0 2.0
1.0 14 9 1,350.0 1,710.0 1.0 1.0 1.0 1.0
1.0 15.0 1,450.0 1,810.0 1.0 1.0 1.0 1.0
2.0 17.0 1,550.0 1,010.0 1.0 1.0 1.0 2.0
1.0 18.0 1,650.0 2,010.0 1.0 1.0 1.0 1.0
1.0 19.0 1,750.0 2,110.0 L.0 1.0 1.0 1.0
2.0 21.0 1,850.0 2,210.0 1.0 1.0 1.0 2.0
2.0 23.0 1,950.0 2,310.0 1.0 1.0
2.0 1.0 2.0
25.0 2,050.0 2,410.0 1.0 1.0 1.0 2.0
1 23.7
IV
Ni
- g
213
- - - --
For the plpulation at iACation A. in 24 months:
' Number of Cumulative

UInits (N) Units t t t 720 FYt) F(t f 720) G(t) G(t).N
0.0 0.0 50.0 770.0 0.0 0. 1 0.1 0.0

0.0 0.0 I R).0 870.0 0.0 0.3 0.3 0.0
0.0 0.0 250.0 970.0 0.0 0.6 0.6 0.0
1.0 1.0 350.0 1,070.0 0.0 0.9 0.9 0.9
0.0 1.0 450.0 0.0 1.0 1.0 0.0 -1,170.0
0.0 1.0 550.0 1,270.0 0.0 1.0 1.0 0.0
1.0 2.0 650.0 1,370.0 0.0 1.0 1.0 1.0
1.0 3.0 750.0 1,470.0 0.1 1.0 1.0 1.0
2.0 5.0 850.0 1,570.0 0.2 1.0 1.0 2.0
2.0 7.0 950.0 1,670.0 0.5 1.0 1.0 2.0
2.0 9.0 1,050.0 1,770.0 0.9 1.0 1.0 2.0
2.0 11.0 1,150.0 1,870.0 1.0 1.0 1.0 2.0
2.0 13.0 1,250.0 1,970.0 1.0 1.0 1.0 2.0
1.0 14.0 1,350.0 2,070.0 1.0 1.0 1.0 I.0
1.0 15.0 1,450.0 2,170.0 1.0 1.0 1.0 1.0
2.0 17.0 1,550.0 2,270.0 1.0 1.0 1.0 2.0
1.0 18.0 1,650.0 2,370.0 1.0 1.0 1.0 1.0
1.0 19.0 1,750.0 2,470.0 1.0 1.0 1.0 1.0
2.0 21.0 1,850.0 2,570.0 1.0 1.0 1.0 2.0
2.0 23.0 1,950.0 2,670.0 1.0 1.0 1.0 2.0
2.0 25.0 2,050.0 2,770.0 1.0 1.0 1.0 2.0
M= 24.9
i.e., all of them will have failed and will have been fixed.
M3 Problem 3-3
(A) . = I.IXX),F(t) 1 -- e "/p' equal I - e
F(1,000) = 0.0231
Number of failures = (0.0231) (1,308) - 30
(B) t = 2,000
F(2,000) = 0.0544
Number of failures = (0.0544) (1,308) = 71 _-
(C) t = 4,000
F(4,000) = 0.1253
Number of failures = (0.1253) (1,308) = 164
(D) Inspection at 1,000 hours, makes units "good as new"

P (failure at 4,000 hours) = F(1,000) + F(1,000) + F(1,000) + F(I,000)
= 0.0231 + 0.0231 + 0.0231 + 0.0231
= 0.0924 N.
Number of failures = (0.0924) (1,308) 121

/21
214 ,
/. '~ ~' '
(R) lnsixe'tion at 2,(NN) hours
P (failure at 4,(NN) hours) IMA(2,00) i F(2,(XX0)

0.0544 0.0544 + 0,1188
Number of failures (0.1088) (1,308) 1142
U3 Problem 3-4
Start with the top row of random numbers, and note that one's answer will vary depending
on the random number string.
(1) Engine at 100 hours

I..
F, - 96AW (Inl(-6W~~)
28,8A1.2 hours - 1,149 months from present
4A5&82ourm - 190 months from present
IF,. - 1,126(tn( .66))r-

V
1.11.9bours - 40 months from preent
1,000 hour inspection is 900 hours or 36 months from present
Reset Mode A and C to "0"
((T
lpA- 96,8874 -)Y"'-4
24,877 hours - 996 months from lint inspection
789 hour - 31 months from lost inspection
2,000 hour inspection is 40 months from last inspection
Past 48 months pt. of interest
,- 96W(In )1471
219,214 bous - 8,760 mouths from present
-, 4,996 (in I ))h/&494
215
.,,.A.
. 7.4 >, ._
S*-4' . - ~.*. ... . -- - - '.
.47
4.319 hours 164 months from present
F.. - ,.126(n(.())
1 -
1.122 hours - 36 months from present
1,000 hour inspection is in 800 hours = :2 months from present
F, - 96,587 (In ) -
3,994 hours - 159 months from last inspection
F.- - 1.126 ~(In

916 hours - 36 months from last inspe-tion
F4 - 96,587 (fn 897'. hours -3,575 months from Iment
4,99-W-- -O )" 3.566 hour.

-- 88 122 monthsfrom prosent
K.. .325 (In- 1,025 hours - 21 months from present
1.000 hour inspection is in 500 hours = 20 months
FA " 96,587 (in. - 9,96S hours , 3D8 months fre"n last inspection
,,.- 1,12( (in (I- 774V))'- 1188hours 47 months from last inspection
216
Il*
e.
(4) Engine al 70X) hours
FA - 96$,587 (f.n (---- ))' ;- 92,48houos - 3,671 months from present
F,, 4,996 (In ( ))'" - 4,813 hours - 164 months frem present
F,. - 1,126 (n . - 1,072 hours - 14 mouths from present
1,000 hour inspection is ir, 300 hours = !2 months
Reset Mode A anc C to "0" )
V,- 96,1"( (I( ))"" 231776hour 9,270mo.ntba, ,.rto o
t 1.--~w~-)''
F,.- ii~(n -1,197 hours -47 months from 1I-t inspection
2,000 hour inspection -a 40 months from last inspection
,. - 96.587 11'"(- 70,530hous. - 2,78montsfrompresent
J. - 4.996 (in (-.L" - 7,480 hours - 263 months from present
F,. - 1,126 n( In .) - 1,008 hours - 4 months from preset
1,000 hour inspection is in 100 hours = 4 months
Inspection before failure - Fc = 4.33 months actually
Reset A and C to "0"
FA - 96,87 (9. - ))' - 45,O7hours 1- monthsfrom tinspectio

--
F,. ,2 ~ 1,l85hours.- 47 months from lastinspectioa

2,000 'our inspection is 40 months %iom last inspection
Inspection before failure

A) No Failures B) 5 engines fail without 1,000 hour inspection
I additional engine fails before the 2,000 hour inspection
217
", "* ' . . .. ... . . .. ..... , ! .

z ' _ _' " " : . . .. ..
. . b ,. , '5 5 ".5 . ..-. '."
4. CHAPTER 4 ANSWERS
4.1 Problem 4-1
Leti, ---2,12 = 5
Consider the case when 1 = 2: a

10o' + 110 + 1252 + 150' + 90' +40' 187.0
Let T,. = Prerent Time PIow = Prob of Failure at

Present
TIoD = Time in 100 Hours P 10o = Prob. of Failure at 100
Hours in Future
Engine TP
1 40 0.045 140 0.429

2 90 0.207 190 0.644
3 100 0.249 200 0.681
4 110 0.293 210 0.717
5 0 0 100 0.249
6 0 0 100 0.249
F7V4 2.969
Thus, one would expect 2.969 - 0.794 = 2.18 additional failures during the next year.
Consider the case when a2 = 5:

. u- 1001 + 110 + 125 + 950o9
+ 405 1 1,,,147.3
2 LI
Engine T,., P . T___

I o P__
1 40 0.0015 140 0.540

2 90 0.082 190 0.972
3 100 0.134 200 0.990
4 110 0.207 210 0.997
5 .0 0 100 0.134
6 0 0 100 0.134
Thua, one would expect 3.767 - 0.4245 = 3.34 additional failures during the next year.
So:
As 0 increases from 2 to 5, the number of expected failures during the next year increases from
2.18 to 3.34.
4.2 Problem 4-2
The predicted design B.1 life = 1000
%wsume a -3.
218
. o -
rT5hen(K))- + 5(2000)'
I i .7
Let L = B.1 life predicted from the Weibayes analysis of the data.
Then:
0.001 =1 - e-(L/3M'T)
e-(L/38.7)3 =fi0.999
3
-(L/3845.7) = In 0.999
r L = 3845.7 (-In 0.999)1/
L =384.6
Since the Weibayes analysis predicts B.1 life = 384.6 (which is less-than the predicted design
value of 1000), one can conclude that the present data is insufficient to increase the predicted
design life.
,1
i-
9.21
9.4, . .. o ;. .... . . ",..
9. .. .. ' /, . .X ,-:" '.

9... . .1 .. , ,- , ..
.' .. . ,,'"9" , " .. ", , .
9.,
5.1 Problem 5-1 Substantiation Testing
Enter Tlble 5.1 with a sample size o" 20 and l equal to 1.5. Mhe corresponding entry is 0.237.
The required test time per bearing is:
0.237 x "000 hours = 711 hours.
"' Thus, the zero failures test plan is: run 20 bearings for 711 hours each. If no bearing fatigue
failures occur during the test, then the failure mode has been significantly improved, with 90%.
confidence.
5.2 Problem 5-2
The reliability goal may be stated mathematically at R(2300) = 0.95, which means that the
reliability of the vane system is 0.95 (95% succeeding, 5% failing) at 2300 cycles. First, convert.
this reliability goal to a characteristic life gal: substitute t = 2300 cycles; R(t) = 0.95, and 3
into'equation 5.2. The results are:
2300cycles 1/3
I -f
I-n(0.95)I
or qi= 6190.2 cycles.
The number of test cycles per turbine was not fixed. The only constraint was that it should not
exceed 5000 cycles. The table below shows the number of turbines required, assuming 3000,
4000, and 5000 test cycles accumulated on each.
Test Cycles Ratio of Test Cycles Number of Turbines

per Turbine to V1= 6190.2 Required - Table 5.2
3000 0.48 22
4000 0.65 9
.5000 0.81 5
/
/
Either of these test plans will satisfy the requirements of the test. However, the plan to test
5 turbines for 5000 cycles each requires the fewest total test cycles. (The first plan requires
22 X 3000 = 66,000 cycles, the second plan requires 9 X 4000 = 36,000 cycles, and the third plan
requires 5 x 5000 = 25,000 cycles)
Therefore, the test plan that satisfies the test requirements and that requires the fewest total
test cycles is: test 5 turbines for 5000 cycles each. If all turbines complete the test, with vane
erosion within the allowable limits, then no more than 5% of the turbines will be rejected for
excessive erosion prior to 2300 cycles, with 90% confidence.
5.3 Problem 5-3
In the terminology of Section 5.9,
220
4'4
./ = .2.5 ,
".9 t W00O hours,
"40 2000 hours,
%- 4(4X) hours,
and t, 0.9.
Equations I':..7 . . ,.,! no,- ,aire fLw r nnd n, using the method of Section 5.9.
Step 1: p, = 1 - exp(-(lO00/2000)2-5) = 0.162

p1 = 1 - exp(-(1000/4000)2- 5 ) = 0.031
Step 2: Setting r0 = 0, no, the value of n satisfying equation 5.3, was found to be 14. nj, the
value of n satisfying equation 5.4, is 3.
=14
a 33XX 0.031
0.162 -2"24.4
0.16 5.23
Step 3: Since, for ro = 0, a is greater than b, r. is increased by 1.
For r f=1, n. = 23 and n= 18, giving a =6.677.
Step 4: a is still greater than b, for r. = 1. For r= 2, n, 31 and nj 36, giving a =4.5.
Step 5: For r,, = 2, a is less than b, so the process of increasing ro is stopped here. The a-ratio
for r. = 2 is 4.5 and is closer to b = 5.23 than the a-ratio for r. f 1. So, the final value of r. is
2.
Step 6: The final value of n is
n,, + n,
2
S..
......... ----... 2+36 .,for r.
.-31 = 2
n -33.5
or n = 33 (rounding 33.5)
The final test plan is: test 33 units for 1000 hours. If 2 or fewer units fail while on test, the test is
passed. Note that the additional requirement in this plan has more than doubled the number of
* #units and test time required. (The zero-failures test plan requircd that 14 units be tested
1000 hours each.)
221'
..- .7 *....,,, V. , .,
-- .- P J
- 6. CHAPTER 6 ANSWERS: None
7.1 Problem 7-1
Using equat ion 7. 1.
|.5et 4k7V. ,)140 d <1.5 e"'4 7 4.7V.)/4II,
1.22 :s :s 1.84 is the 90"c confidence interval for .
. Using equation 7.2,
20(Oe4 0.,AMIf)/Ih 0 _ 1 2000 e' 0 h.6 /,.U40
1667 _<V _<2.199 is the 90' confidence intervEI for n.
7.2 Problem 7.2
Using equation 7.3a, b, c, and d.
7.3a U = (In (1500) - In (2000)) . 1.5 = - 0.4315
7.3b Var,(U) = 1.168 = (-0.4315)2 (1.1) - (0.1913) (-0.4315) =.. 364

40
7.3c U, (-0.4315) - (1.645) (0.0364),/2 = -0.7454
U2 = (-0.4315) + (1.645) (0.0364)/2 = -0.1177
7.3d e (-e.1177)) _<R(1500) S e*(-'.7464))
Therefore, 0.411 _<R(1500) < 0.622
7.3 Problem 7.3
For n = 40, from Appendix Tables 11.2 and 11.3
FA,o.0 = 0.001 P',oM - 0.072
F2,o.05 = 0.008 F 2,ot. 0.113
F3. = 0.020 F3A, 0.149 -
"- =1.5 =2000D

"Usi-agequation 7.4
t 00 10.001 , 20.01
t,2. 200In -1O.OW .". 80.21
222
-~-o.
11I.4141. 21N4)i III II.7
I ;k54.:1 "-
* t .~.. :--" 2(X(jfIn I-0.11 = 48M.33:
. 200n1-0149 1 = 592.74 1
90' confidence intervals on first 3 failures:
Failure1: 20.01 5 Time _<354.81
Failure2: 80.21 : Time 5 486.33

Failure2: 148.36 :5 Time _5592.74.
7.4 Problem 7.4
Use Weibull-Thorndike Chart
Figure 7.6 1)
T/Y = 4000/1000 = 4.
II We want 0.9 probability bands on the number of failures occurring at T = 4000 hours.,
* Entering x-axis at T/9 = 4,
a. When p = 0.05, C = 0
b. When p = 0.95, C = 7.
Thus, a 0.9 probability band on number of failures by T 4000 hours is (0,7).
7.5 Problem 7.5

Steps I and 2: Completed on Figure J.8
Id 1.59 =258.0
Step 3: MTTF = 1/10 (51 +... + 451) 1/10 (2261). 226.1
Step 4. On graph.
Step 5: 10/2000 x 100 = 0.5% of the populc.tion failed
4 Step 6: On graph Figure J.9
q new - 6400 hours.
223
S.q,
. . ... . . . . . .o,. . .... ... n.- -. -. b- - *. ;. .- bh,.*
7.6 Problem 7.6
I r Use the techni(I'e of Section 7.6 to calculate hands on Weibull from F'igureI.10
n=iO k(n) 0.246
- Using F(x) 1 - e(-X/)O, Bands are:
] [F(x) - 0.246, F(x) + 0.246]
For
= 2.974 ,1= 895.42
- e x/895"4212.974_ 0.246 _ 1 - e -[X/895.4212.974+ 0.246
"x Upper Band Lower Band
100 0.247 0
1 200 0.253 0
300 0.284 0
400 0.333 0
500 0.408 0
600 0.508 0.016
800 0.757 0.265
1000 0.997 0.505
1200 1.0 0.662
2000 1.0 0.754
"Now,plot on Figure J. 11; since "true" Weibull lies partly outside the confidence bands, we must
conclude tha' the Weibulls are significantly different.
'1
22
122
i /.
?. . ..
- . - / _ ,.
-w!
Sr I I I
"-EIBULL DISTRIBUTIONJ I
/P = 1.59I
-
= 258.0
SA._ LE SIZE = 10
.u.- FAILURES =10
.3D.
w I22
V I
1.1-
" .. -- . ., . .. "- :. ... ....
is
. 2 '. 3. S. .' .00
-4 "
4-0- - Iq 1
I
Fi ure .. -rhe 7-5- -- - -- - 4
-A -
.EIBILL DISTRIBUTION _ -"
_ 9== 258.0
1.59
Sm.. "".--.S
- FA IlLU -E SIZE
.RES = 10
= 10
U. -
.1'..
-4 I
4.D
.4/
1'
- Jr~-- -
.4#
si-4
a d - - o./.... 2 . .. s .~jao
SO . * . *.J(.* 4
Figr .1. Prole 7-5
//
I t
WEIBLLL DISTRIBUTI ON
P = 2.97- --- --
v8=95.4
".' S*F SIZE = 10
" --. FAILURES =10
4 u.V I 1 I- I
,4,
11.. -. t. t. t. Joy -~t -. W -11 - -W.
1.. ......- --
TOT~AL OPFERATING TIME (HR)
FO 272257
Figure J.10. Problem 7-6
22V
9. _. ...
.
'
!b'- -
WEIBULLL DISTRIBUTION II A
/
"D= 2.97
q =89S .4
SAMR.E SIZE =- 0.
F0ILRSE.- =O j0
U.- - - -
Iif
/I1 ' ,
FiurJ.1. -rbe -6-

7 S, / J
,.. I
II..
I
1s10.4 . .
11 11,1 /
ii
eL,-I- -,- , - -I - - "--"
A.,
aoimn
Prntn -L& - 75-G/ -iis
I ,
-98
!; ~Figure J. II. Problem 7-6 -;
/:
S~~228
*tU.SGov ernment Pilntin g Office: 1984 -- 7 9i9-0 02 /918i ... .
i .. ...
.
.i
. S#
View publication stats

Weibull Analysis Handbook

Uploaded by

Copyright:

Available Formats

Weibull Analysis Handbook

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Weibull Analysis Handbook

Uploaded by

Copyright:

Available Formats

What is the document about?

What is the document about?

What is the Weibull distribution used for in the document?

What is the Weibull distribution used for in the document?

See

Weibull Analysis Handbook

Article November 1983

James Breneman Grant Reinman

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

Pratt & Whitney Aircraft

Final Report for Period 1 July 1- 2 to 31 AL;zt 1::3

Approved for Pub!c 1.Cole-e, G;,,,,4,,z ,.. 1 ELECTE_

A ero Prc:!.v*.-;!n Ln"z:7,,c '.4,

A~r Fron CT~-

Best Available Copy

ALL *,-.v., t 4: -.- _.,

N,: .;i:I'):.;, . : ::::,.: . ./', , . . . . ... . . . . . .. . . . . . .

NANCY A. NiITT /ACK RICHENS

4. TITLE (and Sw.btillo) S. TYPE Of REPORT & PERIOD COVERED

WEIBULL ANALYSIS HANDBOOK .Iulv 1982 to 31 August I1J83

Dr. R. B. Abernethy; .1. E. Breneman; 3fl-2C24

P.O. Box 2%91. West Palm Beach. Fl._ _ _ __ _ _ _

y 11. CON4TROLLING OFFICE NMAE AND ADDRESS

WrgtPatterson Air Force Base, Ohio 4543:3 228

16. DIST RIBUTION STATEMENT (l this Rope"

Approved for public release, distribution unlimited

IS. SUPPLEMENTARY NOTES

' ABSTRACT (Gintms~amen

DD 1473 90ITIOw OF I NOV 6SI08 11SOLET19 UNCLASSIFIED

The following members of the Ameriean Institute of Aeronautics and Astronautics

I . INTI'IDOI))('i'ION TO W' Il'II,.I, ANALYSIS ................................ I

1.1 Ohje 'tiiv,. ............................................ ....................... I

"2 PERFORMING A WEIBULL ANALYSIS ..................................... 14

2.1 Foreword ...................................................................... 14

3 WEIBULL RISK AND FOREiCAST ANALYSIS ............................. 38

3.1 Foreword ............... ................................. 38

4 WEIIIAYIS WtIEN WEIIIJLIS ARE IMPOSSIIlIE ................. 67

4.1A Foreword .............. 67

I 5 SUBSTANTIATION AND RELIABILITY TESTING ......................

6 CASE HISTORIES WITH WEIBULL APPLICATIONS .................. 94

6.1 FOREWORD ............................................. 94

TABLE OF CONTENTS (Continued)

%; 7 CONFIDFNCE LIMITS AND OTHER ASPECTS OF THE 1

7.1 Foreword ........................... It

APPENDIX A - Glossary ....................................... 141

APPENDIX C - Rank Regression (Weibull Plot) Method of Weibull

2. Example and Step-by-Step Procedure ......................... 155

APPENDIX D - Maximum Likelihood Method of Weibull Analysis . 159

1. Foreword ....................................................................... 159

AP'iNDIX E - Weibayes Methods ............................... 16:1

APP NDIX F - Monte Carlo Simulation Study - Acruracy of

1. Foreword ............................................... 166

-......... ........... "-

A1P'IE.NI)IX (G Rank IMetrehosidn h~d vsvt Maximmn ILikv-!i~h.0

A1'PENI)IX H -- \Veilull Parameter ;,inat ion Computer Pr,)gjamrns 18.1

* AP'PENDIX i -- W eiuill (;raphs ................ ........................ .. .195

A'PPENDIX .J -- ANSWERS TO IP.)ILEMS .............................. 199

I. Chapter I A nsw er.; .......................................................... . 199