You are on page 1of 243

IN

i%

AD-A143 E103
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
'.4,

"

A ero Prc:!.v*.-;!n Ln"zý:7,,c
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.

S

NANCY A. N•iITT Project Engineer MTurbine T FOR THE COMMANDER

/ACK RICHENS Chief, Components Branch Engine Division

H. I. BUSH Director Turbine Engine Division Aero Propulsion Laboratory

is no longer employed by your organization please notify

"I1f your address has changed, if you wish to be removed

f

m omr mailing ILqt. or if the addressee
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.

I" a

*

i

SECURITY CLASSIFICATION Of THIS PAGE (Whent Dat. Entered)

REPRTDOUMNTAIO
1. REPORT NUMBER

REPOT DCUMNTATON AGEREAD

PGEBEFORECOMPLETINGFORM
3.GOVT ACCESSION NO. 3.
______________

INSTRUCTIONS

RECIPIENT'S CATALOG NUMBER

*
4.

~~~AFWAI.-TWKI.2(J79
TITLE (and Sw.btillo)

A~4q/D

S. TYPE Of REPORT & PERIOD COVERED
Final_________

WEIBULL ANALYSIS HANDBOOK
______________________________________________

.Iulv 1982 to 31 August I1J83
S. PERFORMING ORG. REPORT NUMBER

7.

AUTNOR(s)

P&WA/GlIDl

FR- 17579

G. CONTRACT OR GRANT NUMBER(@)

Dr. R. B. Abernethy; .1. E. Breneman; C. H. Medlin, G'. L. Reinman
9. PERFORMING ORGANIZATION NAME AND ADDRESS

3fl-2C24
10. PRC..RAM ELEMENT, PROJECT, TASK & WORK UNIT N4UMBERS

U~nited Technologies Corporation Pratt & Whitney Aircraft GDroup
Chwernment Pri4NIIVIJ Divis~ion

IAREA

30660649
_ _ __ _ _ _

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

y

11.

*

Aero Propulsion LaihOFat(Iy (AFWAL/PO1TC) ~~Air Force Wright. Aeronautiald Labo~ratories (AFSC) WrgtPatterson Air Force Base, Ohio 4543:3
tMNITORING AGENCY NAME & #4ODRESS(ff Re'Htern~t from Controlling Office)

CON4TROLLING OFFICE NMAE AND ADDRESS

I2.

.NMUROPAE

Nov'ember 1983

REPORT DATE

228
IS. SECURITY CLASS. (ofthi
rep. toi)

5
16.

Unclassified
IS&. DECLASSIFICATION/DOWNGRADING

SCHEDULE

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

MWd Ides!4', IV block member)

Weiluall. Weihlmyes, Risk Analysis, Reliability Testing, Thorndike (Charts, Fatigue Testing. Life Testing. Expornential D~istribution

'

RI~~2. ABSTRACT (Gintms~ame n

eese. Old Un .e..ooryp

i

daj)

~

lhan.

U.

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 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
8/0 0102-014-4" d1

11SOLET19

UNCLASSIFIED
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
SAecess~n
TTS
OU'D@

For

G;RA&-l

.1:~TAB U 'ouxiced i

I£c if V•a

IlWI
LAvn

codes 'n1)t1!VtY
1

S~LU4a

.........................I.............................10 3....................................6 2...................... Weibull Paper and Its Construction ............. Changing Weibulis Establishing the Weibull Line .................................15 1........... Scope ...... ............ No-Failure Weibulls .................. L[w-Time Failures ................8 2............................. Case Study 3: System Risk Analysis Utilizing a Simulation M odel ............................................... ....... Engineering Changes and Maintenance Plan Evaluation ....16 1.............1 1...........13 Foreword ... Weibulls with Cusps or Curves ..........................................................11 3........... 3..11 Foreword ................................12 3.... Present Risk ......11 1....................... Failure Data Analysis Exemple ...................... Curved Weibulls ............. 2................... Calculating Risk .......4 3...................................... Case Studieb .......8 3....... Problems ..............................7 2.................5 2... ........................1 2.............. 13 13 14 14 14 17 IP 22 24 24 27 27 32 35 38 38 38 38 .......... Future Risk When Failed Units Are Not Fixed ................... 11)gleg nd.............................. The Use of Simulation in Risk Analysis ..........................9 2....6 3.Nonfailures ........... Close Serial Numbers .............5 3............... Forecasting Techniques .............................. Mathematical Models ...................... .............................4 2. Examples ......... Advantages of Weibull Analysis ............................. Stspended Test Items .......... Data Inconsistencies and Muhtimode Failures .......... Summary I I 2 2 5 ..................................2 1.... - 3 WEIBULL RISK AND FOREiCAST ANALYSIS ................................3 1.............. .. .......... Risk Predictions ..........................................................................3 2.. System Weibulls ..............9 3...............17 5 5 7 7 10 It) 10 10..10 2.... Risk Analysis Defi~iition ..2 3........................5 1.........................TABLE OF CONTENTS C(haptrr Page • I ...3 3.. Aging Time or Cycles ........................................7 1.......2 2.................................. :... Case Study 2: Bleed System Failures ...........13 1.... .. .......................................... Small Failure Sample Weibulls ........... 1... Case 93tudy 1: Bearing Cage Fracture ................................. V Ni ................. 1.........10 1.. Problems ............................ ANALYSIS ...................9 1................... .................................14 1............................ ..F L....................12 1. ................................ . Future Risk When Failed Units Are Repaired .......1 3.......................................7 3.................................................... Failure Distribution ....... INTI'IDOI))('i'ION TO W' Il'II............ .... Background .................. ............................8 1............38 38 39 40 40 43 43 44 52 60 "2 PERFORMING A WEIBULL ANALYSIS ............. Weibull Curve Interpretation ................................................................................. ...............................4 1........... _.........6 Ohje 'tiiv............

................................... .....................11 How Good 'Were the Forecasts? ...................... ...No F-ilure...........................8 5.2 4......5 5......... Foreword W eibayes M ethod ..2 5... . Initial Analysis -7 Small Sample ... ...10 Determining the Fix ..... 5.................... .....................................7 Example 2: Main Gearbox Housing Cracks .. Example 1: Turbopump Bearing Failures ..........................7 4. ......6 4...... Total Test Time ........................ 96 6-..................4 Two Months Later ..2 6..........................................................1A 4............. .... 99 6. ............ . 99 6... 107 6... ...16 Role of the Weibuil ................ ...........................15 Example 4: Support Cost Model ............... ...................... 103 6...Final Weibull Plot ......19 Concluding Remarks ...........5 4........................ 67 67 67 67 68 68 68 68 75 77 77 77 80 87 .......... Advantages and Limitations of the Zero-Failure Test Plans ......-.......... Zero-Failure Test Plans for Substantiation 'resting ............... 103 6.............. Unknown Failure Times ........No Failures .. 6............................... .12 Example 3" Opportunistic Maintenance Screening Intervals 103 6................. Weibayes Worries and Concerns ...........-.............4 5.......... 99 6..4..... 1 107 6..........17 Example 5: Vane and Case Field Cracks .............. .........4 4......................................... Weibest ........................ Recommended Method oor Solving Equations . .. ...6 5...... ............. ...7 5........................... 6............... I 5 SUBSTANTIATION AND RELIABILITY TESTING ...3 5....... 104 6......87 88 90 92 92 94 94 94 94 ..................3 4............ .... . ........ Weibayes . .8 Information Available for Analysis ................:lisk Prediction . Problem s ..... Non-Zero-Failure Test Plans . 106 6.... ........ . . ...........6 Four Months Later ...............................3 FOREWORD ............... .. 6......................... ..... ......1 5................. ............ ....... .............. ................8 WtIEN WEIIIJLIS ARE IMPOSSIIlIE ......1 6..14 Finding the Optimum Interva .........18 Resolving the Questions ......N TABLE OF CONTENTS (Continued) 4 WEIIIAYIS 4.............................9 Foreword ...... ....................................... .............. ...... ...... 107 vi .................Batch Identified ...104 6.................. Zero-Failure Test Plans for Reliability Testing .......13 Structuring the Problem .................................... ...... 102 6.......9 Risk Anal)'is .......... ... Designing Test Plans ........... 5 ............................................................. Examples of Problems/Analytical Solutions ....................6. ... Problems 6 CASE HISTORIES WITH WEIBULL APPLICATIONS .... ................. . .............

.......... The Likelihood Function ................ Example .......... Ranks.............................................. .4 7.. 155 155 155 159 159 159 160 161 16:1 163 163 Example and Step-by-Step Procedure APPENDIX D 1... 4.3 7........8 7....... "B............* - * ..........11 Foreword ... Weibull "Thorndike" Charts ........-rrrr~r ..... Risk Forecast Accuracy ........................................ APPENDIX C ............ "- ....................7 7.................................. and 95'........1 7....................... ........... vii • -.... 1......... Maximizing the Likelihood Function .................. 2........ 4... Confidence Intervals for Reliability .............. 2................. Risk Forpcast Precision ............. Weibayes Methods .......... 4... APP NDIX F ........ ...................... Maximum Likelihood Method of Weibull Analysis Foreword ............. 5......................... Foreword Derivation of the Weibayes Equation ..................... Eta (qL)Estimates ............. Risk Forecasts .................................... Confidence Intervals About a Failure Time ....... 1............. i 2...................................................•.. Confidence Intervals for 1 and 'p ............. 3....... 2..............................g .....................................10 7.....Rank Regression (Weibull Plot) Method of Weibull Analysis .......... Acruracy of ....... ................................................2 7..........................5 7....................... Confidence Bands on the Weibull Line ......................................................1 5............................ PROBLEM S .......... 5"...................................2 6..............w r r>f• ........ ........... Page 1 h It 110 110 113 113 114 116 128 129 129 113 141 143 APPENDIX A APPENDIX B - Ranks ............**........................................................... 3............ ......................... 7 CONFIDFNCE LIMITS AND OTHER ASPECTS OF THE W EIBU I..........................• TABLE OF CONTENTS (Continued) Chapter %..... ......... M onte Carlo Simulation .... ........................................ Confidence Intervals ...........-........ Method ............ Comparing the Weibull to Other Distributions .... 166 166 168 168 17:68 176 176 178 179 Foreword ................................ Shifting W eibulls ........................Monte Carlo Simulation Study Weib 11 Analysis Methods ....................................[ .... Glossary ........ AP'iNDIX E I.... 7..*..................... Weibull Goodness of Fit .......... Median Ranks........................................................... -i............ .6 7......1 Life Estimates ...................................... ] sBeta (Fo)Estimates .......9 7....................

........... ............. Chapter C hapler Chapter C hapive Chaipter Chapter Chapter -- ...)ILEMS I 2 :1 4 .....................195 199 199 199 206 2 220 222 222 * I.. Answer .......................... ..... ..........0 h~d vs vt Maximmn Rank IMetrehosidn A1P'IE.................. ............................... ........ Answers ... 7.. ..................... .......... ....NI)IX (G th d of \ ei u l m i ..............................................TABLE OF CONTENTS (Continued) ('hooter Pa•e'e •"M ILikv-!i~h. 18........ 4......... ......... A nsw ers ....... I..... A nswers ..........................raphs ..... 4.........inat ion Computer Pr....... ...... ..........-t AP'PENDIX i A'PPENDIX ...... Ii Vilii a im• ..............................1 ..................... ...................1... Aaswers .. An:swer .r 6 7 A nsw er.)gjamrns ..... 5... ........................... 2.......... A1'PENI)IX H -\Veilull Parameter W eiuill (............ 3................. -- ANSWERS TO IP•..... .......................................J I...

......................--- ......................... 26 28 29 ............ M ain Oil Pumps ...5 1. ...... Compressor Start Bleed System...............6 To Correction to Curved Weibull ........ ix --..................................................... .......2 1.......... Failure Dintribution Charecteristic .4 2.......................... 36 41 42 42 45 46 48 First Pass .................. Second Pass ......... Compressor Start Bleed System ................. ........................................... 30 31 33 Small Sample Beta Estimates Are Too Steep ......................................7 2....... Rivet Failures With Suspensions .....................8 2...... EEC ...... Bearing Population .6 1.......................... Mixing Failure Modes........ 1......................................................1 2..................LIST OF ILLUSTRATIONS AMU" Page 1.6 2...........................3 3.4 1... Ocean Base ....3 2............... Correction Applied Risk Analysis With Weibulls Simulation Logic Simulation Logic ........................... 8 9 11 12 16 19 23 25 .........7 2............................1 1.......... Construction of Weibull Paper Rivet Failures ............... Bleed System Failure Distribution Excluding Air Base D .........10 2.12 3..... Excluding Ocean Base ...................... MFP ......... ................ ............................... ............. 4 .......................:" lPerforming a Weilull Analysis Understanding the Results .....11 2..................... A Curved Weibull Needs to Correction Plotting to Correction t................................5 2........ Bearing Cage Fracture ........................ ...................9 2.... Weibull Plot for Augmentor Pump ..2 2. ......... Mixing Failure Modes.......................... Compressor Start Bleed System............................4 3.1 3.............. .............................2 3... 6 11) Louver Craicking .....................5 3..............................

Weilayes Evaluation of New Design in Accelerated Test ................ 61 63 64 C5 69 71 Failure Distribution Input to Simulation Risk Analysis Simulation Outline ................................................................... Probability of Passing the Zero-Failure Tests for X's Between 0.........11 :1. Weibayes Evaluation of New Design in Accelerated Test ..5 6..........................................................14 ..8 3......... 49 50 53 55 h57 ..............................- . Module Population . Weibull Plot for Augmentor Pump Bearing ......6 hr Characteristic Life When 0=2 ....................015 4........................1 II it s 7..............................................6 5.....................1 :1.......................... Hydraulic Pump Failures .................A 5.........4 4...7 89 9..................10 3......3 5..1........................................ Weibull Evaluation of B/M Design ...........e 1) ... Illustration of B...... BI.4 5......................9 Cycle Characteristic Life When 0=3 ....................7 3....................... 4.. . Bleed System Population Air Base D ..........99 Reliability at 1000 hr ................... and BI0 Lives ..........5 Bleed System Failure Distribution At Air Bas................................ Weibull Plot for 0......................5 •1...N I % UST OF ILLUSTRATIONS (Continued) :1.. Risk Analysis Comparison ....6 86 5.... ..................................1...............................................3 4.......................9 :1..............2 4...... ........................- .................... A Reliability of 0....................... 73 74 76 78 81 82 83 85 l... A 2000 hr B10 Life Is Equivalent to a 6161.................................. Overail Population ...... Illustration of 7000 hr BIO Life .........2 5.......................................... .. Compressor Vane and Case ....... When Shoul2' We Pull the Suspect Batch? ......12 :1.....99 at 1800 Cycles Is Equivalent to an 8340.. B ll and Roller Bearing Unbalance Distribution ............................................................ Location A Population Control Population ..............................5 and 5 .... .................

13 7.........3 6..............................9 6.......................... ff - 3......14 7......50 .... Cumulative Sums of Poisson (Thorndike Chart) ......4... Unscheduled Maintenance Input via Weibulls . ....... Weibull Test Case Weibill Plot Where/t = ..5 Weibull Thorndike Chart for t = 2..............8 6...............11 7...............0 .......................... Example of Confidence Bands on a Weibull Line ................. = 1................................. Weibull Thorndike Chart for ji = 1.....2 6....0 ............................................................................7 6......10 6.... Weibull Thornd4-e Chart for Weibull Thorndike Chart for / Weibull Thorndike Chart for Weibull Thorndike Chart for/• Example of Shifting a Weibull Flange Cracking = 2......0 and q 100 for 10 Failures ..6 0..5 6.................10 7..11 7.......15 Augmentor 650 on Up ........... = 5........................... Cumulative Main Gearbox Housing Cracks ................ xi .............ik Analysis ..............7 7...............1 7.....5 106 108 109 112 115 117 118 120 121 122 123 124 125 126 127 130 131 132 Main Gearbox Housing Cracks .......................... Weibull Thorndike Chert for Weibull Thorndike Chart for ffi 0........................... ..........................6 7..... 97 98 100 101 102 105 10.... Flange Cracking Estimated DNstribution .......12 7................0 .......................................................3 ....................3......... Approach To Optimizing Scheduled Maintenarnce ........................................ 7........... Weibull Ploi for Augmentor Pump .........9 7................................ 1th Vane and Case Cracking ....0 .5 ...............LIST OF ILLUSTRATIONS (Continued) Figure Page 6..........................4 6.4 7...................5 7......................................................... Ri......... .....8 7..... .....................................2 7.......................... Sheduled and Unscheduled Maintenance Interaction ....................... ... Expected UER's Due to 12th Vane and Case Cracking ..... 2.......................

..4 Befai Precision . Curve-.....................7 J.1 ... Weibull of Nonserialized Parts.....A1.... ................................................. 174 175 181 182 2200 202 204 205 208 209 Median.....LIST OF ILLUSTRATIONS (Continued) 7.................... True Weibull.... Problem No................1 ........................................................l Life Affuracy ....I 1.....8 Problem 3.........1........................ Beta I"stimates: True Beta :3................ Problem 3-2.......7 F............4 F.........6 J...2 ....... ...............'........................................... B.... 2 225 Problem 7-15 .................................................'.......... Problem 3-2 ...... Problem 2-2..3 F............................. Problem No............................. Aecmacy Five Failures ....................18 Hyptothesized Weihull ............................................. Problem 2...... Suspect Failure Mode.......... Example Data With Least Squares Line .............5 F........................... ........ B.............1............................ B ....................A Life Precision .17 7..............2.......... 135 138 139 140 158 167 169 [70 7............21 ('... Overall Population ..2 ....... 6 ........19 7............5 ........... Overall Pnpulation ............... Comparing Weibull and Log-normalDistributions for Coverpiate ...... xii ............................ Failures .......... Eta Accuracy Median Characteristics Life Estimates Eta Accuracy A ... !34 . Problem No.... ........9 .... J.....8 F......1 F........... ..................... Weibull Problem 2-3...........1...... Characteristics Life Estimates - Monte Carlo Simulotion Procedure ...............................................6 F.................................... 171 172 2....... W-1.................................................................................16 7........20 7.............. 6 .. 5 ..... Picking the Best Distribution .. I:I......1..................................................... Infant Mortality Problem 2-3. Location A Only ......................................... Initial Conditions of Simulator .......................... .. .

........ 226 227 228 I ..LIST OF ILLUSTRATIONS (Continued) I - Figu re" Page ...... ............... ..................... I/ *1 ...................1................ ............................1.................................................. i P rolermC~l 7 6 .............10 ......9 I'ruh'mb 7-5 .....................1.........11 IProbhl m 7-1...

.................................................................... i...............................1 6......... * 5............................................................... Confidence Levels............... Bleed System Risk After 18 Months .............. Baseline ..................................................................1 7..... Projected Pump Failures ..................................................................................... Hearing Risk After7 12 Months .. Future Risk..4 2.............1................... Follow-Up Analysis Results ....8 :1..... Bleed System Risk After 18 Months................. xiv 79 80 98 103 110 116 6...... Weibull Coordinates.................5 :1...................................90 ............................. '42...................................3 136 ...........4 3...... Construction of Abscissa W( ............................1 2....2 Median Rank .......................................... Simulation Output for 1200 Hour Inspection ..... Required Sample Sizes for Zero-Failure Test Plans Confidence Levek 0..........................5 Wt'ibull Risk Forcs ................A 2............................LIST OF TABLES Table Page 1........................6 3.................9 5........... t...........2 :3...................... to 1...0.................... Bleed System Failures by Air Base..................90 . Present Risk....................... 7 t5 15 17.......... Critical Values for Testing the Difference Between Log Normal and Weibull (Favoring the Weibuill) .... 18 290 2 39 42 47 47 51 52 58 59 66 Coist ructi~on of Ordinate (Y) .... Confidence Bounds on the Weibull Line .....1 2.......... Simulation Output for 1000 Hour Inspection ..............................................2 7.....2 7.....................6 :3......................................................................3 3..............................1 3........ Adjusted Rank ..........................................2 2.... Table of Uniform Random Numbers from 0................. Characteristic Life Multipliers for Zero-Failure Test Plans Confidence Level: 0................7 :3........

......... Risk Forecast Accuracy Risk Forecast Standard Deviations ......... ............3 F.......i F.................. 17......... Standard Deviations of the Characteristic Life Estimates ............. Ninety-Five Percent Ranks ............A i"............................................................ .......2 F.................................... 178 Componcnts of1 Monte Carlo Simulator ................................ FA Five Perce.....4 Critical Values for Testing the Difference Between Log Normal and Weibull (Favoring the Log-Normal) ............... Untinued) Table Page 7...................... 1:36 143 147 151 173 B...........LUST OF TABLES t.........2 B........t Ranks .1 B.............. ............... M edian Ranks .. 18(0 xv A.......... .......

a *fe t..may sometimes render good service". He did not claim that it always worked or even that it was always the best choice. CHAPTER 1 INTRODUCTICN TO WEIBULL ANALYSIS 1. six months and one year?" "To order spare parts that may have a two to three year lead time. 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?" of Wide Applicability. . It is the intent of this document to illustrate how to answer these and many similar questions through Weibull analysis. The initial reaction to his paper in the 1950's and even the early 1960's was negative. The authors intend that their presentation bee sech that a novice engineer can perform Weibull analysis after studying this document. 1.r. He claimed that his distribution. . 1A EXAMPLES The following are examples of aerospace problems that may be solved with Weibull analysis. Only after pioneers in the field experimented with the method and verified its wide application did it become popular. Questions asked by the Program Manager are. The authors believe there is a need for a standard reference for these newer methods as applied within the aerospace industry. varying from skepticism to outright rejection. Sil . 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. "How many failures are predicted for the next three months. * A project engineer reports three failures of his component in service operations in a'six week period.2 BACKGROUND Waloddi Weibull delivered his hallmark paper on this subjectl in 1951. pg.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.°a S. * *• * * W92bu7l. Waloddi (1951). Time has shown that Waloddi Weibull was correct in all of those statements and particularly within the aerospace industry. 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. applied to a wide range of problems. '.. and to industry in general. He clairied that the function "-.. Joural ofApplied fechanicu. Today it has many applications in many industries and in particular the aerospace industry. " . A Statistical Distibution Function 293-297. or more specifically his family of distributions. ".. There are special problems in aerospace and unusual arrays of data. S. Special methods had to be developed to apply the Weibull distribution.

and even the construction of a Weibull curve when there are no failures at all. Methods will be described for identifying mixtures of failures.1). The relationrhip between various values of the slope and typical failure modes is shown in Figure 1.und weaknesses of he method. Plottirg the data Data problems and deficiencies are discussed with recommendations to overcome deficiencies such as: • * * • • * Censored data Mixtures of failure modes Nonzero time origin (to cor. operating time. 1A SCOPE As treated herein. The use of Weibull distributions in mathematical models and simulations is also dOscribed. action) No failures Extremely small samples Strengths . discussions of alternative distributions such as the log normal. scales other than time. this should be considered and weighted heavily against the Weibull. (Figure U. perhaps start/stop cycles. non-seriamized parts and ccmponents where the time on the part cannot be clearly identified. Actual case studies of aircraft engine problems are used for illustration. only success data.5 ADVANTAGES OF WEIBULL ANALYSIS process consists of plotting a curve and analyzing it. In addition. 1. The -- measure of life. There are brie. problems with the origin being at other than zero time. it is not difficult to make graphic comparisons to determine which distribution best fits the data.. Weibull analysis application to failure analysis includes: 0 * * " Interpreting the plot Predicting future failures Evaluating various plans for corrective actions Substantiating engirneering changes that correct failure modes. are presented in Appendices to supplement the main body of the handbook. classes or modes. 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. This type of analysis relating the slope to possible failure modes can be expanded by inspecting libraries of past Weibull curves. as there are only a few alternatives to the Weibull. For example. or ýas 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. *ri- . Further. answers are supplied.2. investigations of alternative.P. if there is engineering evidence supporting another distribution. However. Statistical and mathematical derivationi. Where problems are presented for the reader to solve. 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. the technique works with small samples. The horizontal scale is some One advantage of Weibull analysis is that it provides i simple graphical solution. 2 II .

1 II I bt~ a- .iI $1 " '. " .-% . *N ( Z.

c... .l L .a I C CLI 0 0 0 0r 0 0 A E aEE v ca= 0 0 CO() E-a C cc & w. 'a 0 0 0 401CýP 0ro w.

For example. Beta determines which member of the family of Weibull failure L. several age scales must be tried to determine the best fit. The difference in the two risk predictions represents the net effect (if the change. . The units of age depend on the part usage and the failure mode. including production rates and retrofit dates.8 RISK PREDICTIONS the failure occurred in service operations.dhether 'ornot the failure mode applies to the entire fleet or to only. A "ty. This p. hazard rate. it may be of int arest in determining the time at which one tenth of I " of the population will have failed. a year. See Figure 1. low and high cycle fatigue may produce cracks leading to rupture. An ideal application would consist. The failure mode may be any one of the types represented by the familiar reliability bathtub curve. The age of each part is required. Methods for making these predictions are treated in Chapter ". In most cases. The risk parameters may be the predicted number of failures. usually the analysis is staited with a few failures embedded in a large number of successful. the responsible engineer will be interested in a I . ideal data are rare.. These values can be read from the curve by - • - inspection.2. which is called B. This is called BI "life. which is often called a batch. "1.. infant mortality with slopes less than one.If prediction of the number of failures that might be expected over the next three m. of a sample of 20 to 30 failures. Maintenance schedules and plans are also evaluated using Weibull analysis.tlths. 1.Akernativeiy. • ".-ocess may provide information on .. which is known as the slope. or aircraft availability. After the responsible engineer develops alternative plans for corrective action.7 FAILURE DISTRIBUTION The first use of the Weibull plot will be to determine the parameter d. or .Case Histories with Weibul! Applications. In each case the baseline Weibull analysis is conducted without the engineering change or maintenance change. The Weibull plot is also inspected to determine the onset of the failure. and wearout with slopes greater than one. knowkdge of the physics-of-failure will "provide the age scale Whmn the units ot age are unknown. Except for material characterization laboratory tests..one portion of t1e fleet. The age unit Gf a jet starter may be the number of engine starts. it may be of interest to determine the time at which 1"( of the population will have failed. See Figure 1. random with slopes of one. 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. depot loading.3.icalrisk prediction is shown in Table 1. the risk predictions wil be repeated. The age units would be fatigue cycles. unfailed or censored units.1. six months./ . 'Me decision maker will require these risk predictions in order to select the best course of action. spare parts unage. . or two years. distributions best fits or describes the data.t or component.1. 1.1 life. life cycle cost. . These techniques are illustrated in Chapter 6 .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.hmne parameter. L' L: F 5 -6" • .6 AGING TIME OR CYCLES F • Most applications of Weibull analysis are based on a single failure class or mode from a single pt. ". For eximple. The study is then repeated with the estimated effect of the change mo)difying the Weibull curve.

. t . i TOTAL OPERATING CYCLES Fgsre 1 ID Loouver Cracking 6 .037 SAMPLE SIZE = 11 FAILURES = 11 _ . • - . 1886.. -III 'I. IT ? ~ * .a Q. "- ....- _ 4 " . . ..544. - . -' -k EIBULL DISTRIBUTION 0I =5 . ...L P 5. 4. 4%. 4C a o.. ..

91 52.12 19.20 6. If this is found to be the case. these models are updated with the latest Weibulls once or twice a year and predictions an renerated for review. Generally. First.69 What if? in in in in in in 7 taonths 8 months 9 months 10 months 1i months 12 months Corrective action next month. There may be physical reasons why this will be true. next .7 .53 88.4. Second.65 105. The origin correction may be either positive or negative.77 15.5. However.58 90.76 76. the origin ._-@ !• 1.18 24.92 more more more more more more more more more more more more failures failures failures failures failures failures failures failures failures failures failures failures in 0 m nths in I month in 2 months in 3 months in 4 months in 5 months 6 months 32. it may be that the origin of the age scale is not located at zero See Figure 1. WEIBULLS WITH CUSPS OR CURVES The Weibull plot should be inspected to determine how well the failure data fit the straight line.97 75.ction 'for.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.If the points fall on gf'.Sometimes the plot of the failure points will showy cusps in sharp comers.30 18. In this case it is necessary to conduct a laborator-' failure analyr'a of each failure to determine if separate failure modes are present.35 102. very little data scatter. the separate Weibull plots will show straight line fits.-le curves.69 44. A procedure for determining the origin correction is given in Chapter 2.35 7. it may take a minimum amount of time for the wobbling roller to destroy the cage. a mixture of failure modes.07 29.TABLE 1. This is an indication that there is more than one failure mode. For example. sometimes the failure points will not fall on a straight line on the Weibull plot. and maintainability support coda. with roller bearing unbalance.1./ L S: t I . a mixture of failure modes -. There are at least two reasons why a bad fit may occur.10 24.68 63.92 41.87 36. then separate Weibull plots are made for each set of data for each failure mode.41 12.33 more failures in 53. 1011 .3.60 0. On each plot "thefailure data points from the other failure modes are treated as successful (censored or nonfailure) units. and modification of the simple Weibull approach may be required.e. The bad fit may relate to the physicsof the failure or to the quality of the data. This would lead to an origin correction equal to the minimum time.00 3. See Figure 1. that is. The scatter should be evenly distributed about the line. module returns to depot.42 117. WEIBULl RISK FORECAST N Risk Pred. availability. These models have been use-ful for predicting spare parts usage. 12 Months Beginning July 1978 11. i. . If the laboratory analysis successfully categorized thejfailures into separate failure modes. The (ombination may be done by Monte Carlo simulation or by analytical methods.

. . LURES 71 i ../ WEL IBULL DISTR IBUT ION -"-I 0 = 1. Corecio • to Cuve We 0 -CURVE .2`52' SIZE = 261m. \...0 TOTAL OPERATING TIME (t4RS) .7.4 T.. * t "t " .111. 6..330 •("' " I = W.131. 00 e __________T =I- F MM * Fiur 1.. . .pL Im. ....

4 0( - - MP 1160..j : . w o . . .-- .....-- - - . . ..O M Fiw . - .. ... .. J. .- ........- .911 LESIZE = 676 S= FAILLUES = 41 .. EIBULL DISTRIBUTION S=1. as- U. -•.~ A• 6.611 .-•-..5.....• - -7.-. .. .- [--- U. Mxn Fa /u -~ ModaM " / 1• 'I ..- - - - - - - - - -- --. .

how much success time is required before it can be stated (with some level of confidence) that the problem has been corrected.15 CHANGING WEIBULLS After the initial Weibull plot is made. V7.1 life and the risk predictions are that they should not change significantly with a moderate size sample. Each plot will be slightly different. When parts approach or exceed their predicted design life.12 SYSTEM WEIBULLS . like any statistical analysis. Some years ago it was popular to produce system Weibulls for the useful life period (Figure 1. 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 -.1. extensive Monte Carlo and analytical studies have been. some studies indicate electronics may have a decreasing failure rate.' Although data deficiencies may force the use of system Weibull analysis. assuming 4 1.Substantiation and Reliability Testing. a # of less than one. 1981 Proceedings Annual Reliability and Maintainability Symposium. small sample estimates of tend to be too high (or sleep) and the characteristic life.6) assuming constant failure rate (4 = 1. This approach is fraught with difficulties and should be avoided if possible. there is a need for a Weibull plot even when no failures have occurred. later plots will be based on larger failure samples and more time on successful units. i. However.0). To evaluate t Iese small-sample prol)Iems.7. Electronic systems that do not have wearout modes were often analyzed in this manner. In general.14 SMALL FAILURE SAMPLE WEIBULLS Flight saiety considerations may require using samples as small as two or three units. but gradually the Weibull parameters will stabilize as the data sample increases. Weibull analysis. More recently. it may be possible to extend their predicted life by constructing a Weibull for evaluati3n even though no failures have occurred. it is tempting to plot a single Weibull for the system based on mean-time-betweenfailuires (MFBF).mall samples. made and will be presented in Appendix F. tends to be low. 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. if an engineering change or a maintenance plan modification is made to correct a failure mode experienced in service. For example.13 NO-FAILURE WEIBULLS In some cases. 1. See Figure 1. *! If the data from a system such as a jet engine are not adequate to plot individual failure modes. 1.e. Kam LiWong. a math model combining individual Weibull modes is preferred because it will be more useful and accurate. is less precise with . The important inferences about B. ""Unified Field (Failure) Theory-Demise of the Bathtub Curve". 1. 10 tla0N .

- . 4.6.4S68 S. 6-1s..) FO 272262 Figure 1. 6.-- 0. 0 12S 0Zi l *4 431.00. 1 . low 00. Mixing Failure Modes. TOTRL OPERATING TIME (H-R.330 - I 30.sao. EEC 11 . . 4 . EIBLLL DISTRIBUTION =696._LESIZE = 188G FAILURES = 1880 W.m..- = 1.

'C. O C13 LL LO 0 A CCO 0 ~ ~ ~ ~ .a coc 0 > U 'j V cn) 0212 .

v-- 131 / ':7-- * * 1 I. A straight line is then fit to the data to obtain estimates of 0 and q/.iAex computer routines. This approach./ . has some deficiencies as noted above for small samples I is simple and graphical. 1. Constructive comments would be appreciated for future revisions of this handbook.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.-. The advantages and disadvantages of these methods are discussed in Appendices C and D. but require ci. Maximum likelihood estimates may be more accurate.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 2.1.

Since interpretation of the data is the most important part of doing an analysis.2 WEIBULL PAPER AND ITS CONSTRUCTION 11e Weibull distribution may be defined mathematically as follows: F(t) -I where: F(t) t to $ e - = : = = fraction failing failure time starting point or origin of the distribution characteristic life or scale parameter slope or shape parameter exponential. an extensive discussion is given on how to interpret a Weibull plot./ :.4'' CHAPTER 2 PERFORMING A WEIBULL ANALYSIS 2. the following can be noted: I - F(t) e - (a -- / let t.j~ • / ~ ~~. / / ~¼V . The first question to be answered is whether or not the data can be described by a Weibull distribution. This is often called reliability at time t and is denoted by R(t).) to time t is I . Therefore.F(t). 2. -0 then S-. the data can be approximated by a Weibull distribution. F(t) thus defines !he cumulative fraction of a group of parts which will fail by a time t. . . If the data plots on a straight line on Weibull paper.-.F(t) e. Examples are used to illustrate interpretation problems.1 FOREWORD This section describes how to construct Weibull paper and how to plot the data.- 14 V •. the fraction of parts which have not failed u. By reaimnging the distribution function. .

S1I F(t)
G t Y~

Iy~

'1t

n n Y=BX+A

t

i 21nt --

n

* 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

Q Rnn( 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. :P()
0.001
0.01 0.1 0.6 0.9

CONSTRUCTION OF ORDINATE (Y)
CWl 2Vein.-

I - F(t)
-6.91
-4.60 -2.25 -03

mui CW

Vei. (.-#91)
0 unis
4.4 4

0.83

-

7.74

0.999

1.93 TABLE 2.2. CONSTRUCTION OF ABSCISSA (t)
(Ar Mnt 1
-W

84

------

0 unit

I5

2
3 4 15
20

O6
1.10 1.39

1.61
2.71
3.00

100
1000

4.61

6.91

15

•,
/

\

-

i,, .
/

.
;.

'

.~.
.. ',
'-". p .4

,,--

,

.
.

ca

0 0L

•1m

ID~

L oCo. cv.

'

______5~

c!c

CM-

0
CdC

LAI

I I

Coo

'I"

I"

I

...

-

//'.r0

LL....L..J.P.J" I

II
16

I

;

/

-

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 Number (SN) 3 Failure Time (in,) 90 Failure Remar. .

2 3 4 5 6
7

96 100
30

Failure Rivet flare loo

d without Is

Failure
Lug failed at rivet attachment, Failure

49 45
.

10 82

Failure Rivet fare loomened without falow

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 Time (Min) 30 49 82 96 Median Rank 12.9 -31.3 50.0 87.0

Order Number SIN 1 4

E

2 5 3 8 8 2

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.

A line is drawn through the data points. Formal methods of rank regression and maximum

-

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

FMt

=

= -- 00 when t= =I--0.368 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 accouný for ýhe 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

-l t! - - €• " EIBLLL DISTRIBUTION 0 =2.E =S S S~IV so----4FILL.2.I I I. SlZ 5 ".0000 St .a4..S. U - 0.--. / . F.. TOTAL OPERATI.=80 SIZE go. .1 1 f II0..0 2 .NG TIME (MIN) FD 272263 Fi.- i - i - - i = CHARA•CTERISTIC LIFE -/- - -i I "- N.. o IN. Rivet Failures 19 .'....: / - .'ure 2.. SAMPLE FRILL.° .

5. Applying this equation to the rivet test data.438 3. 20 .751. 44.5. is given lby the following equation: 2 hankIncemen (N + 1) .064 6. Leonard G. (1959). conisdering. or suspended by the wrong failure mode (Type 111). all poss4ible ways thle suspended item may have fauiled and potentially influenced the results.5 are ohtained. Rank Increment for Rivet No.Tlhe genterail formnula for adjus~t ing i hie rank jH)sit ion.2 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 47 6 Time 3 saiuspenso 45 suspension Items Beyond Suspension I Starting here and counting forwardI! 2 5 8 49 failure 82 failure 90 failure 96) failure 100 suspension 3 4 5 6 7 1 1 2 31 2 Johnson. it does not need a rank -Yalue because it will not be p~lotted on the Weibull chart. Rivet S/N 7 4 6 5 8 1 2 3 Order 1 2 3 4 6 6 7 8 ADJUSTED RANK Time (minutes) 10 suspenasion 30 failure 45 suspension 49 failure 82 failure 90 fdilure 96 failure 100 suspension Adjusted Rank - 1. General Motora Corporation. the values in Table 2. therefore. was suspended (Type I orTl~ype 11). TABLE 2.(previous adjusted rank) + Fer ofitems beyond p~resenlt suspended item)(2) T7(n-umb wherc N is the total number of rivetsA tested regardlesit of whether it failed.377 - The adjusted ranks were calculated in the following manner: Rivet No. pp.125 - 2. Research Laboratories.50. 7 is a suspension. 4 where: 8 is the total number of rivets tested whether they failed -)r not ( 1+1)70 . The Statistical Treatment of Fatiguve Experiments.

8 = 2.064 + 1.125 +5 -- 1. The last value calculated (1. Rank Increment (8 + 1 1) . 2 5. therefore. Therefore.125 1.751 + 1. no new rank increment needs to be calculated. I = 3. t he adjusted rank is: Adjusted Riank for Rivel No.438 - *"Rivets No.751 Adjusted Rank No.438 + 1.The adjusted rank is the previous rank (in this case 0) plus the rank increment of 1.125 + 1. Therefore.313 = 5.313 "where 1. 8. 5.1. Therefore.313) is still valid. 5 = 1.125 Rivet No. is the previous adjusted rank plus the new ranký increment.4 . =- V * P 21 . 5 is a failure and the formula has to be employed agami to identify the new rank increment to use between failures.313 2.377. Rivet 7 4 6 5 8 1 2 3 Time 10 suspension 40) failure 45 suspension 49 failure 82 failure 90 failure 96 failure 100 suspension Items Beyond Suspension I Starting here and counting forward 2 3 4 5 The adjusted rank.064 Adjusted Rank No. Rivet No. Adjusted Rank No. the adjusted ranks for these rivets are: * * Adjusted Rank No. 1 and 2 are failures without any additional suspensions between them and the previous failures. 6 is a suspension and receives no rank value. No.125.313 = 3. 4) 1I 1. Nc.313 = 6.125 is the previous adjusted rank / 5 is the number of items beyond the last Suspension starting with the tailure following that suspension.

6.quality.377 timing the cialculated median ranks and the fatilure times.0. a -good approximation is provided by Benard's ~formiala t : Pam = N+0()4 i- 0. For instance. If the slope is less than 1. the median ranks tan be established. John Wiley and Sone.438 3.5_ WEIBULL CURVE INTERPRETATION - Weibull curves may reveal clues about the failure mechanism. While interpolation could be used for determining the appropriate median rank. C.misassembly Overhaul-related failures of various components . X x(2. .4) (X" where N sample size iadjusted rank value t Joe of this formula in illustrated in Table 2.0 are: a.125 2. Quality control and assembly rroblems may produce infant mortality failures.438 0.randomilydispersed with the data. q~.751 5.. The slope of the line.4 - . (1977). -~ 'Kau and Lembweomn. since different slopes imply different failure mechanisms.f 2.quality Improper installation of temperature probes .125 =v 0.3 is derived. The analysis reoilt ing in I(X)minutes is the correct method.35% Adjuated Rank Order No.4 95. (1. Inc.3) x 100"t 8+0. -2.3) x lO0~'f 8 +0(. The sample size to be used when entering the median rank table would be 8. is the amue at. (2. characteristic life. 22 .82% 25. however. some gas turbine failures having slopes less than 1. the earlier Weibull. Improper augmentor liner repair . This is often referred to as an infant mortality failure mode.misassembly Fuel pump leaks due to installation problems . b. reliability increases as the unit ages. TABLE 2. d. 1. e. MEDIAN RANK Median Rank 9. Figure 2.8" =98c P6. Reliability in Engimeeting Design. This will generally be true if the suspensions -rt. the effect on the. It went from FA) minutes wit hout suspensions to 100) minutes with suspension's.0.064 6.6.quality/misasembly Electronic control failures .08% 56.71% 72.N With these adjusted ranks. N'ote. 300.45~ etc. op.45% 41.

../ KENGILL DISTRIBUTION *.0 -f - - .:: Rivet Failures With Suspensions~ 23 ! 'M1..- -4-4 t1 11. 111 P.III 4 .. '1 ._ __------ 9 100 MIN SAMPLE SIZE=8 FAILLIES D =2..-... 4 l •1.* 4.. * 43 0 9 AVM? Figure 2. 44.

2.44 would approximate the f'amiliar hell-shaped or normal curve. . 1). was found t hat t hey containedI oversized p~arts. jI.6 DATA INCONSISTENCIES AND MULTIMODE FAILURES '111Ve~eoher We sui itlet ies in We ibull analysis which might signal problems.5 is an example of* tow-time part failuhres on main oil 1. 1).0. Ftiel oiil cool()kr f'ailure e.I. hirbine vane wearout Augruentor liner loirnthrboigh 1'eniperat tire prob~e boss fatigue (earhiix housing cracks A uignient or fiamehol(Ier cracks Oil tIn b chaite thlrouglh. Soiiiix on~il tiliso WeiI ~itIIiIsw~it Ih sloptis ii I or*near I are ivl)onti~ial al jl45i- 'Ilie ex poneliai . (Gas turbine engiiles -ire tested b~ef~ore being ship'ied to the customer. dl. what was goingr wrong? Upon examining the f'ailed oil pumps it.IIsicial cast(I* the We'j)ibll disi riluitition when 41 1. IFLiel ('(lit rod solenoid fanilure dI. and since there were over 100) of these engines in the field with no problems. For shallow slopes like 1. LOW-TIMAE FAILURES Figure 2. of' 3. Low-time fiailures provide a clue to a production or assembly p~rocess chan~ge. tinuexgpetedl laiiiures (itloji jg. failuire times b~ecolme more predictable. especially when there are many suý-ce. 2. as indicated in Fig-tire 2. V. 1441.7. d Failuires aire most lY low-time p~arts Serial numbers of lailed part s are close toget her '[he dlata has a -(dogleg" bend or cusp when p~lot ted on Weib'ill paper 'I'he data has a gradual convex or concave bend on Weiliull paper. Examples are gixen t hat illuist rate the f'ollowing: a.0 there is moi re scatter in tIhe fajtln re (data and therefore fi-iilure predict ions will cover long t imespans reflecting this uincertainty.8 to 3.isftjl high-time units in the field.l B~earing cae inilurv Ieilljvratiire probe ilr C. Somet hieg had changed in the manuifacturing process which created ihis p~roblem. Some examples of*Weibnills wit h slopes greater t han I are: a. A slope. Ellect rn mic engine c( nit rid failhre. A 1). iiisiitficieit redunda~ncy.4.6iitriblmio ispe . '[lie oversized plarts caused an interf'erence with the gears in the lpuinp which resulted in finiluire.uinps. T'his was traced to a machining operation and corrected. o~r vveil p~ro~iiit iilisiis. Slopesgreaiter than I represent wearout mlodles. 24 . [he retleits original (li-iu IIiideicienies. C. C. 1'.s. li'his wouild result ini a conustanlt halulure rate (~iht Inii. '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. As the slopes get sl eep~er.

. I. " ... r-.... .. 25 /'7" ." . _.~ 0 '~ 0 " . .-N -.. i ... .. . "• . \ ....... ... • ..:" .?' . .•<.&• C CO_ II 0 c .? . :.. ..•.. . - -... '... :.. .- . ... r- Q11 a..i\ .. A h-.• "• .. ..x p • .....' . • . . .. oo Z 0 L0 CL I * .'.. .."' . -"a. ' "... .. ...

.- "v= 3618. Main Oil Pumpm 20 A1.5 w..._ I S.jI' I I I I IIH - WE EIBULL DISTRIUTIO I N ' = 1. t*..5..---- - -4-4---- 'I -C3 iiI TOTL OERAINGTIME (HR) 4FD 272265 Figure 2.A/ . SRIH'LESIZE= 1090 FAILURES = -3 - .".4..

This was the case for a compressor start bleed system binding problem. In cases where this Occurs without prior knowledge.#. smusplcletd. .:.2.. "I.9. charts are presented in Figures 2..a cost reduction for manuiacturing the bearing cages.. For example. Upon examination of the data. is more of a wearout failure mechanism with / 5.. etc. More attention was given to this area and the problem was resolved.. In fact. 10 out of 19 failures had occurred at one base.... / . low-I ime units have no failures. they are usually associated with more than one failure mode. It was concluded that this hase's location was contributing to the problem..223. and high-tinw units have no failures. -.. Somet hing may have changed in tlie to this type of problein....8 CLOSE SERIAL NUMBERS The . anulfacturing process fior a short period olt ime and then changed back."... housing cracks. These failures were traced to a process chai./"- _ .6 is a prime example of a process change which happened midstream in 1production..•' "\ . then a batch problem is *!. . . several dogleg bends will result.. If these different failure modes are plotted on one Weibull plot. while the ocean base Weibull. j 0. -. 7' OP 27 -'%\ 27'..-... At least 650 units had more time than the highest time failure. Note that the fleet Weibull presented in Figure 2. leaks. Sto 2. • tV . a close examination of the failures will have to take place for .untits hlave' failures. midtime. Figure 2. The data were broken apart and the two resultant Weibull.tinne reasoning can Ihe extendled tho ( her peculiar Cailure gr(i~pings.9 DOGLEG BEND A Weibull plot containing a "dogleg hend" is a clue to the potential of multiple failure modes (see Figure 2.837.9. 4WX) hour time frame. hat is. ..ge that was incorporated as. The failures do not have to be associated with an environmental factor to cause a dogleg Weibull.8 and 2.. 7 . The failures had occurred in the 2W. For instance. if lfailures occur in Ihe middle (i' I he time experienc'e. fuel pump failures could be due to bearings.7).. X This could be considered an infant mortality problem. potential separation into different failure modes. The closeness of the serial numbers of the pares are a very definite clue Figure 2. Bearings were failing in the augmentor pump.. This problem was related to lack of maintenance.8 is less than one. . The base was located on the ocean and the salt air was the contributing factor. *""/•.

1. i.U . -4-. • FAILURES = 9 SAMPLE SIZE = 387 .- - ao. 49 -.210 4 . __. 7 1. 100 ----- A - I TOTAL OPERATING TIME (HR) / v FD 27226 F~igure 26.994 = 478.4 . Weibull Plot for Augmentor Pump 28 . -.EIBULL DISTRIBUTION 0 = 4.._--.l-- b Iil / . .. -0 U8- ..•.

62 S*FLE SIZE = 2%~ FAILURES = 19 " :- . -d438-- . _i WEIBLL DISTRIBUTION iP 1. I. .S. = =/ 63601. - .. -.122 i1 .'.7. * 1?. FO 272287 I -- TOTAL OPERATING TIME (HR) i 4 Figure 2.. *1I " --- I18.. * __ý4. 52 3 I.. 2 l 9. -. * mt 0.S-..I ] ý g I -o -- _ .6 6. 7 Compressor Start Bleed System 29 .. -' -t-...0SI.'77 _ 10. 14..

. Fgur e Compressor Start Bleed . 4. ._ . - . . 0...-- -. L Fll) 2n2268 -.. _ . / • ". 1 .- . 3. .- S. i - . */ W. .8r Ns IEIBULL DISTRIBUTION 1 =O. ..: 7-'. . -AA- . Excluding Ocean * 2. . . -.... . I. • q1 2 - 312 304 TOTAL OPERA•TING TIME (HR) Base 'yst e. C..0 .o 30 . me.8.84 = 404444 SAMPLE SIZE = 2054 9 F IFILURES 0.

/V S. 202006.EIBLLL DISTRIBUTION__ S.. : - - - :::: : - u... a. opexrSatBedS. I. I".. S. ".- III-II- w LI6 _.. ..5- i - LIJ 2 3- F S. - - - ' O Base --- AUWm i. 313 4 .. •.i . 0. AI MI- .. ' . n U.... 4/LESIZE =202 FAILt.' - . . 3IU . -..GES= 10 -. - -Mm .tm iiue29 ca .

in this c ie approximately 15. In (licussions so far. First.11. if negative.0 hours .halfway failure t3 = 389. there is a time.9 hours . The data will appear curved as illustrated in Figure 2.0 . log Norma!. This is illustrated in Figure 2..ti) (t 0:1. = 42.t2.9 hours.0) (42..9) t.-sented for the Weibull.42.t -t) -• ". There are a couple of ways to determine what adjustment is needed to make the data appear straight. = 16.2. etc. the distribution which best describes the data should be used.0 hours . But the data displayed in Figure 2.0 . it implies that the origin star'ts after zero. before zero.was asmumed to be zero. For example.10 CURVED WEIBULLS In Sect ion 2..1.. wis illust rated as: •F(t) ". The values for t1 . there is an analytical method that can be used to establish to.I lhe cumnulaht ive (list rihution u'wntion F(t) is pr..(389.. When data are plotted on Weibull paper it quickly becomes otvious if the origin of time is not zero.e.0)-(42.e I .10 was from engine controls and there was no reason to suspect that the Weibull distribution could not be used to analyze it. * Where t1 is the first failure time.. may better describe the data.1 14. t2. . There are other reasons for poor fit (i.0 hours. 32 Morr . where: t failure time StI t starting point or origin of the distribution.. In other words.0 . If to is positive.0 . to..10 if the zero time origin is not true.42. 4tt-1 0 (42 -. The equation is: t. and t3 are: t.16.9) (389. t2 is the time corresponding to the linear halfway distance on the vertical axis between the first and last failure. another distribution like a Normal.27. If this is true. It -.0 .16. -42.0 to= . and t 3 is the last failure time.1 . the data do not form a straight line on Weibull paper). in which the control would not he expected to fail.last failure t.first failure S= 42.

" * 33 .00 I I. '. 7... *.330 v = 648.L S~TOTRIL Ixi OPER/ITINQ TIME (HR) 1 Figure 2._EIBLLL DISTRIBUTION 0 = 1.2.s. - 4 I -" * " . . 4. . ooo0.- - - F(t) = 1-e i. • CURVE -- WEN to .. 00' o. .1..10ooo.6. // . 2. A Curved Weibult Needs t0 Correction FD271851 .9.10.9.1 7. 6.3. .S2 SAMPLE SIZE = 261 FAILURES =71 U .a=. 4O 0.

. SMP4LE SIZE =261 FAILURES =71-- - U0.- .252 U.LL. a.330 v= 648.- D F (t) 1-.e+T CIV 34 Itdo -t =42S . DISTRIBUTIOND = 1.Ei' IWEIBJ. - -a I 1- k.

T''his made the (list ribuit ion appe._•plotted • .50. The second way to correct a curved Weibull uses a simplistic approach.11 PROBLEMS .uld be found with either of the above procedures.11 confirms that 15 hours is a good estimate. Trying this technique on Figure 2. 1. They were all tested to failure and the failure times were 150. Construct a Weibull and determine its slope. That is. the BI life is estimated to be 2. one would read the I percentage point of 8 hours from Figure 2. 250. t his curve would be considered convex. what would the value be? . 'hwe vendor was actually ltesting the units for ahotit 15 hours prior to shipment and discarding or repairing failed units. a. !7 .. 2.12 and then add 15 hours to it..12. 2(A). this was not fouind to Ih true. a subtraction of time would be required.r to be truncated at 15 hours with a zero probability of failure before that time. 115. The corrected curve provides a more accurate prediction of the probability of failure. However. Wheit the curved becomes fairly perpendicular to the horizontal scale. For example..•Weimbull . 2C0. The amount of time to he added w(. extend the curved Weibull vertically through the time scale. The resultant. d. 240. W. Would you have expected the derived slope for fatigue specimens? c..5. vy eye. therefore. Where it intersects.1 life or 111 life. simply read the curve and subtract or add the time. curve is plotted in Figure 2.Problem 2-1: Fatigue specimens were put on test. 240. and characteristic life. to determine the time to failure for the 1/1(M unit (often referred to as the BI life). 4 4 4464 35 . Subtracting 15 hours frorm each of the failure times will adjust thecurve forthe absence of this time. b. one has to add the 15 hours to the time read from the Weibull plot.w 1 Up l)on questioning ite vendor. If you were quoting the B 1 0 life. and 190 hours. t . to determine distribution percentiles like the B. Data on Weibull paper that curves in the other direction (concave) would require addingtime to each point.3 hours.

. 11.0O .. L 4..-.: .4 // . o.12.6.~ U.11. .:.1< ....10.. ._. to Correction Applied 36/ x... * TOTRL OPERATING TIME-IHR) 000.k.' F0271853 Figure 2. i•' - -1 0/ 0.. \._.i~i N / ' - ' DISTRIB3UTIONP IIBLLL =0 0.. .970 9 944. ___. ". .867 SIZE = SPRE 261 F•ILURES = 71 To' CORRECTION = .

0 C'omm ent Failure Failure Suspension Failure Failure Suspension Failure Suspension L-8:15 *8:14 836 8:17 8M8 a.0) 14.0 6.6 1.. 220. C. Construct. q~.y lead to an answer to the problem? 1). a.Pro~blemi 21-2: Thlere were live failures . and characteristic life. 370. What value is needed to straighten the Weibull? Will the value found in "a" be added or subtracted from the failure values? Solutions to these problems are in Appendix J. 275.5. b.0 7.1 20. 130. and 1200 hours.of a part in service. 37 .5. 165.o) 8. What is the failure mode? Are there other clues which n. Se'rial Numiber KI1 W12 8*1 Time (hours~) 9. 0. Thle informatiton on these parts is.0 6.52. P~roblem 2-3: The following set of failure points will result in a curved Weibull: 90. a Weibull with suspensions included and determine its slope..

3 FORECASTING TECHNIQUES The observed failures and the population of units that have not failed are used to obtain the Weibull failure distribution..agement a clear view of the potential magnitude of a pro)blemI. . Monte Carlo simulation is only required when comn)lications arise in the risk analysis. and several examples are presented to explain further the techniques involved. as discussed in Chapter 2. This projection is important because it gives ma. In this chapter the use of the Weibuil probability d(ist ribution fiunction in predicting the occurences of a failure mode is explained.4 CALCULATING DISK Risk calculat ions are described in three sect ions: "* "* "* 3. if this prediction is made ftr different failure modes. management is able to set the l)riority for the solution o" each problem. In addition. 3.• : ~/ . The tech'niques used to produce the risk analysis can vary from simple calculations to those involving Monte Carlo simulation. is to predict the number of'occarrences of a failure mode as a function o" time. 3. etc. the f'orecast analysis is only as good as the failure data.5 Present risk Future risk when failed units are not fixed Future risk when failed units are fixed. It should he emphasized that.) 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. 1f" there is a population of N items and each has accumulated t hours or cycles. - PRESENT RISK The simplest case arises when there are no new units (no production) and no replacement of tailed units. day. 3. :. .. The tollowing additional input is needed for torecasting: a h) Usage rate per uni' per mont h (or year.CHAPTER 3 WEIBULL RISK AND FORECAST ANALYSIS 3. A risk analysis calculates the number of incidents projected to occur over some future period.. The additional input needed for risk analysis will be covered. These will he explained in the following sections.2 RISK ANALYSIS DEFINITION .1 'FOREWORD One of the major uses of Weibull analysi. The fit a data should be examined closely to ensure that they are from a single failure mode and will 'ihe Weibull distributio-. .

5 units have accumulated 3000 hours.ion :3.062 0. then that unit's contribution to the total future risk is: F(t+u) - 1 - F(t) F(t) (3.040 0.027 F(s) • N 0. plus the risk of the 2000 hcur units failing by 2300 hours. In general. &S FUTURE RISK WHEN FAILED UNITS ARE NOT FIXED A . .075. N. it a unit has accu ulated t hours to date without failure.m ' 'K" -" - .117 0.2) 39 :LU. 5 units have accumulated 4000 hours. it will be omitted from the population for the calculation of fture risk. TABLE 3.001 0. Assume that the population is subject to a Weibull failure mode with • = 3.0 and ? = 10000 hours. F(t). the expected number of failures over the next 12 months can be calculated. Since it is assumed that failed units will not he replaced. 5 units have accumulated 2000 hours. 2 . plua the risk of the 30 ho r units failing by 3300 hours. Table 3.0.e -0Q. is the Weibull failure distribution with the cumulative probability of failure by each time on the unitE as illustrated.1. etc. Assume that one of the 4000 hour units has just failed. The future risk will be composed of the risk o&' the 1000-hour units failing by 13 hours.58v) Sum .005 0.1.0 -wtFMt) F(1000) =*I.999 F(IO00) . Ifor a Weihull (listribution this becomes Ixl)ecled numixer of Ifailures (0 e ut/yj) N. 'WhaL is the cumulative expected number of failures from time 0 to now for this population?" Figure 3.1. and will accumulate u additional hours in a future peri.a-o"' = 10.075 The value of F(t) can also be read directly from the Weibull Cumulative Probability Plot.1. The question is. Given the same 25 units a• in Table 3.I: lquuit.001 5 5 0. Number(N) Time (t) PRESENT RISK of Units 5 5 5 on EachUnit 1000 2000 3000 4000 &M00 F(t) 0. Yearly usage of each unit wi I he 300 hours.310 0. N .) The cumulative expected number of failures in this case is 1..1. Therefore.1.1 summarizes the calculations involved.the expected numtl-r of faltures from this iwpulathion is the p)rolbaility of failure by time t multiplied by tile numilr of units.e -tal 1.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.135 Exampe of Calculation: = 1 . and 5 units have accumulated 5000 hours.008 0. (See Figure 3.

where F(0) 1 e lie ii If. no p~roduict~ion added. into tIheIfield. Washington. If a part is inspected and removed from service. otr Monte Carlo siitulitt itol met ho~ds. the chance of more than one failure over the soni time interval becomes significant. it no longer contriblutes to the fleet's risk. If it continues in service. In this case. IIn sOMe iinStlnCeSK a part's% service life will depend on decisions to be made in the future which will be dependent. thle methodology used in a Mont~e Carlo simulation is. D.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. 3. Then. "Weibull Renewal Analysis. S. the techniques of Paragraph 3. (n a Weihull distribution. However. 40 .i Ii g in I lit. as well as the addition of new units.t 1 11) lF(t) Table 3t.a poiwerful tool known as Monte Carlo simulation is useful. complex mathematical formulas.0. 639.known. even simple risk analysis can become complicated by the volumec of calculatitons involved. '~is tIn lirt. New York: Society of Automotive Engineers. 'WHITE~.657.7 -5(0.first I lh ours ofservice. a computer program automating the calcti~lat ions is4 useful. Of course. In this case. . prhI Iiiiuv oif' It( it follotws it Weibull li-tiliire dlist ribut ioin.*g iA . Monte Carlo simnulatiorn enables an analyst to build a computer model of the decision plan as it affects a i arl'o service life. the occurrence of' any one failure mode will not affect any other mode. the exp~ected number of failures may be calculated using published tabllest . it coIntinues to coIntrib~ute to the fleet's future risk. (1964).5 or less). C. random events such as the rxtensive wear If*a particular part and its replacement.2) is' (31 Fi. Since only the probability of this outcome may bx. occurring in each mode before the scheduled inspection i6 desired. described for the case (if t hree failure modes and a scheduled inspection. In cases where the probability of' failure is greater than about 0.. 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.(X)12) + N(00(K41) + 5(0. As an example. ItI F(l ) is miutch less tha 10I. the number (Itfailures. J. The effect of scheduled inspect ion on risk is straightforward.2 summarizes the future risk calculat ions for the population of 25 units.2506 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.(X)99) +." in Proceedings of tie Aerospace Resinbility and Maintainability Ctonference. It may include scheduled part inspections. and nol ret rof its.02316) =0.% Hence the expected number of failures from this pop~ulation over the next 12 months is: *Failures 3.01 54) + 5(0. 4000) hours. 29 June-I July 1964. at.6 can be applied. with tine failed unit. equaiot assiumiing p.4(0.5.

Rs Figur 4I -°- Aayi it eb : 100. 10O (HR) I*. . 41 : " . ..IWEIBtLLLOISTRIBUTION P = 3. I1 2i7 -...0 =10.. . . SF~D i " Figure 3.. * ../ 'W. .-I"\4 '" " "/ " "\ '•": " .. - - 1. OPERATING TIME .661000 3. II [:.. TOThL. .. • . . ) •.. L k a 3. _ -. . 12 -. Risk Analysis' With Weibulls q 271854. 2..000 uiiii- IX - I!•~l-kOl~ TOI~TAL OPRTN TIM (HR) /. ...l. "- . . • '.

.0154 0..- .0080 0. Scheduled Inspection 0 Mode 1 Time to Failure 0 Mode 2 Time to Failure Mode 3 Time to Failure FD 259949 Figure 3.0236 The following procedure is performed for each unit in the population.0353 F(t) 0.4) Scheduled Inspection 0Mode 2 Time to Failure Mode l Time to Failure Mode 3 Time to Failure // FD 259948 igure 3.0620 0." /.. Simulation Logic First Pass Advance the simulator to the first event. FlITURE RISK Each Unit's Risk F(t+u)-F(t) l-Frt) 0. See Figure 3.. iii / • / .2. The following equation is used to calculate the time to failure: Iime to failure = LR xln( 1 - random number (3.v .0022 0.0089 "Number of Units (N) 5 5 5 Current Time on Each Unit (t) (hr) 1000 2000 3000 Time on Each Unit at Year's End (t+u) (hr) 1300 2300 3300 F(t+u) 0.I I |7 .183 0. and regenerate a new time to failure for this mode.2..2.ii .0041 0...3.• - . .3.0010 0. Simulation Logic - Second Pass 42 . note the cause. I ' TABLE 3. if this event is a failure.% . generate a time-to-failure for each failure mode.1175 0.0266 4 5 4000 5000 4300 5300 0.0764 "0.0121 0. - / / *. See Figure 3.mm m •-.0012 0. Using random numbers Ihat are evenly (uniformly) distributed between 0 and I and the three Weibull failure distributions.

1009. 334. if all 1703 bearings ran to 4000 hours with 1000 hour inspections. are averaged to give the expected risk. and 1510 hours were observed. 423. The first two examples. there is a total expectation of failure for each bearing by 4000 hours of approximately 0. 3.5). A Weibull analysis similar to those described in Chapter 2 was followed to obtain the Weibull failure distribution for bearing cage fracture (Figure 3. 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. From this distribution plot we call see that the Blo life (time at which 10"I. Therefore.11.12. That is..052(1703)= 89 failures can be expected. The population of bearings within which the failures occurred is shown in Figure 3. .3.5) and read at 1000 hours that approximately 1. . S . The case study in Section 3.6 and applying Equationw 3. After many repetitions of this process.4.. The numlberof failures of each mode is recorded and the simulation is repeated.013) = 22 bearings would be expected to have failed. .2 with a monthly utilization of 25 hours or 12(25) = 300 hours in one year results in the calculations shown in Table 3.013 = 0.3% of the population is expected to fail. So. A more detailed example utilizing these principles is given in Section 3.052.. 2. . 0. Sections 3.1/ Continue this protess until the scheduled inspe4 Iion is reachtd. Additionally. . ] . the result.* . management wanted to know how many failures would be obsetred before this redesign entered the field. 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.. Enter the x-axis of the Weibull plot (Figure 3. The risk questions and solutions are: 1. the probability of a bearing failure by 1000 hours is 0. so a redesign was undertaken immediately.013 + 0. . illustrate the direct calculation of risk without simulation. . 43 A. ' 3.013. each using a different set of random numbers. . Thus about 12 more failures can be expected in the next 12 months._: -~//<. if it is assumed that each 1000 hour inspection makes the bearing "good as new" relative to cage fracture.013 + 0. .• . 1703 (0. This was much less than the Bil design life of 8000 hours..10 and 3. How many failures could be expected in the next year? Utilizing the methodology explained in Section 3.10 CASE STUDY 1: BEARING CAGE FRACTURE *. Bearing cage fracture times of 230. assuming failed units are not replaced. after the entire population of 1703 bearings reach 1000 hours each. S:3.013 + 0..9 CASE STUDIES V Several case studies in the use of the ideas developed in the previous sections are now presentýd.12 uses Monte Carlo simulation. of the population will have failed) is approximately 2430 hours. 990.

or about 13 fewer than for a utilization rate of 25 hours per month.8). how many fewer failures can be expected? Changing the utilization rate to 20 hours per month will chang? the calculation of expected risk.13 failures would be expected for each bearing.065 + .5.13(1703)=221. excluding air base D. a risk analysis for air base D only was requested. From comparison of the plots it seems that the bases are significantly d:fferent. 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). if there is a 2(00) hour inspection. Over the next 18 months. 56 failures can be expected using a 25 hours per month utilization rate.4. The new risk over the next 18 months is given in Table 3. by the time all 1703 of the bearings have been retired.5). at which time the bearing will be retired. Since the probability of failure. Now suppose no inspections were made until 4000 hours. indicates a decreasing failure rate phenomenon. About 42 failures. 3. Therefort. N. 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. It is shown in Chapter 7 that the two failure distribut ions may be proven statistically to be significantly different. the probability of failure by 2(00 hours is 0. are predicted. 44 S. the probability of failure by 4(X)0 hours is 0. 0.4 2) If the usage drops to 20 hours per month immediately.On the other hand.28. But a Weibull analysis of the failures at air base 1) (Figure 3. that is. Again utilizing the Weibull in Figure 3. by 4000 hours about 0. the expected number of failures with a 2000 hour inspection would be 0.6.0. was quite low by 4000 hours (the life limit of t lie part) for the fleet.065 = 0. A Weibull analysis of the fleet failures excluding air base D (Figure 3. set up the calculation as before (Table 3. Using the same approach as in the previous paragraph. *- -- . # <1.6).28(1703)=477 will have failed. Therefore.7) indicates a rapid wearout characteristic.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 "3.065. .

45 .ti N CD N M' 0 w ~0 10 NN a_ 0 ~ 1 00 0~~~~C a) (0 C 0 D ( I-N oww L- I CL U Ij 0 .

. = 6487 AS F L S I ZE = 17 0 3" .1 I . . . . -.0 .5. EIBULL DISTRIBUTION " -... R1-aring ('age Fracture 46 ":11. 445 .. ....2.9. s..•. _ FRILLIRES = 6 .S0. .. 40- I w --W 1T13 -. .. 7.. . 30 4.1 / -I- Wo.3 U. 48- - -. .o0= TOTRL OPERATING TIME (HR) FO 271855 Figure 1. 4. o D=2.. 2. L III l 2..8.

2225 0.0061 0.0366 0.0228 0.0070 0.0315 0. BLEED SYSTEM FAILURES BY AIR BASE Air Base A B C A D D A C D D C D D D D D D 9 p Hours at Failure 153 872 1568 212 1198 884 1428 806 1251 1249 1405 708 1082 884 1105 828 1013 64 32 / - 47 .0022 0.0158 0.3558 1. 0.0038 0.5193 0.0144 0.4301 0.0000 0.Fit TABLE 1.0093 0.0225 0.0481 0.0006 0.2963 0. ..0061 0.0315 0..0685 0.0121 0.3480 0.0106 0.0070 0. 0.0225 0.0034 0.0.0062 0.11.0228 0.0841 0. Number of Units (N) 288 148 125 112 107 go 110 .7731 0.9325 1.0761 0.4631 0.4691 12214 0.0186 0.0613 0.0185 0.0012 0.0118 0.0070 0.0258 Time on Each Unit at Year's End (t+u) 350 450 560 650 750 850 950 1050 1150 1250 1350 1450 1550 1650 1750 1850 1950 2060 2150 2250 2350 Total Risk F(t+u)-F(t).0121 0.0131.0034 0.0050 0.0925 N 0.0179.0094 0.613 TABLE 3. .0268 0.0481 0.6790 0.0422 0.0545 0.0151 0.0214 0.3.0685 F(t+u) 0.000 0.0020 0.0000 0.0029 0.0012 0.114 119 128 124 93 47 41 27 12 6 0 1 0 2 Cumrnt Time on Each Unit (t) 50 150 250 350 450 550 650 750 880 960 1060 1150 1250 1350 1450 1550 1650 1750 1850 1960 2050 BEARING RISK AFTER 12 MONTHS Each Unit'a Risk F(t+u)-F(t) I7-Xt).0012 0.0093 0.1197 0. 0.0422 .5859 0.0022 0.0613 0.0816 Sum .0049 0..0243 0.0545 0.0200 0. F(t) 0.4.0186 0.0002 0.0366 0.1148 1.3607 0.6473 0.0151 0.0268 0.

SAHUL SIZE = 2054 FAILURES 9 --- -TF - --- -- a.... (HR) 4. V.-. VT. 2 TOTAiL OPERATING TIME.. 2o . - - .a. IWEIGULI ISRBUI 0 =0. I U..6.e9g.-.- Z .s * .s FD 27185i Figure 3i.84v = 404444..L 7... Bleed System Failure Distribution Excluding Air Base D 48 .In a.3 W9 0. - - - -1 554 4044 *9..

2.U.- - 0 -n - 2. *1. 5 SAFI.4.7.j00 TOTAL OPERATING TIME (HR) FO 271858 \ .LE SIZE = 202 FAILLRES= 10 - - S..- I. S.10 .. A 6. Bleed Syvstem Failure Distribution At Air Base D . .Ii 0 =S.3. EIf IBiLL DISTRIBUTION = 2006.6.I 6 13 1 1 0. 22Im.\ Figure 3.7. i49 .

N (D\ 0 E0* 0CV) 04 C'J C/) (Im CI. *~ ClL 0) C-4- 500 .

0260 0.1415 0.5969 0.5286 0.1965 0.0000 0.2640 0.9218 0.4998/ 2.8700 0.6763 0.7130 5.9125 2.TAHI.7499 0.1186 0.7052 1.0018 0.3565 0.0028 0.0001 0.1241 0.50 1450 1550 1660 1750 1850 1950 2050 2150 Time on Each Unit in is Afonths (tfu) 500 600 '00 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 250 2600 F(t) 0.0041 0.8149 0.2227 0.0 0.0334 0..4446 2.4 0.5149 0.0112 0. Number of *Units (N) 0 0 2 0 0 2 2 10 26 27 22 24 24 11 it 20 8 4 2 3 3 1 Curre'nt Time on Each Unit W) 50 150 2.0000 0.0810 0..50 IN 450 50 650 750 850 950 1050 1150 1250 13.34.060 0.9218 4.3023 0.7188 0.3445 6.9792 lT-Frt) Total Risk F(t+u)-F(t) --TWIT- N 0.6732 0.9573 0.0007 0.0497 0.0199 0.0000.2352 'V vJ 51 51 .7752 2. .43380.0018 0.1688 0.387' 0.2939 3.0000 0.2856 0.0082 0.2287 0.0424 0.0004 0.0000 0.6259 0.0007 0.1675 0.tl.A540 3.8693 0.4802 0.4328 0.5774 0.2990 4.0397 0.3500 3.0061 0.0041 0.0000 0.0794/ O.''I / : N .9125 Sum .015! 0.0659 0.7618 F(t+u) 0.0058 0.0249 0.0882 0.7719 10.0061 0.0000 0.6080 0.5.:3.0147 0.8016 0. IBLEED) SYSTEM RISK AIT HR18 MONTHS 25 Hours Per Month) Each lnit's Risk F(t+u)-F(t) (lllilivzaion IbRu.:1 0.0984 0.0012 0.0532 0.56.I.

0012 0.3520 0.6328 0.8091 0. The failure distribution of each of these modes is illustrated in Figure 3.0002 0.5136 2.6732 0.1580 Tl'-F) 0. there is a scheduled inspection at 1000 hours.03M8 0. /\ .0112 0.0296 0.4422 0.6355" 0.8759 0.0301 3.7736 0.0532 F(t+u) 0.8323 4.0444 0.0058 0.3023 0.0044 0.7909 24 11 11 20 8 4 2 3 3 1 1250 1350 1450 1550 1650 .9. In addition. .9260 0.9127 3.0246 0.3300 0.2714 0.4015 0.0334 0.0273 0.0008 0.# .0002 0.0148 0.0000 0.2676 2. - .0000 0.0199 0.4802 0.9660 0. " .2027 I-FTt) 0.550 650 750 8150 950 10.0004 0.0688 0.7432 3.8201 2.1170 0.6296 8.084J 0. .0008 0.9738 2.12 CASE STUDY : SYSTEM RISK ANALYSIS UTILIZING A SIMULATION MODEL Assume a jet engine has four independent failore modes: * e * Overtemperature V:mne and Case cracking Oil Tube cracking Combustion chamber cracking.8323 Sum = 42.1465 0.7613 0.3877 1. BLEED SYSTEM RISK AFTER 18 MONTHS 20 Hours Per Month) Each Unit's Risk F(t+u)-F(t) Total Risk F(t+u)-F(t) N (Utilization Rate Number of Units Current Time on Each Unit Time on Each Unit in 18 Months (N) 0 0 2 0 0 2 2 10 26 27 22 24 (t) .0160 0.4775 0.0043 0.50 150 250 350 450 .7276 0.7064 0.0083 0.1192 2..0000 0.6.2222 1.0000 0.1675 0.0582 0.50 1150 (t+u) 410 510 610 710 810 910 1010 1110 1210 1310 1410 1510 F(t) 0.1750 1850 1950 2050 2150 1610 1710 1810 1910 2010 2110 2210 2310 2410 2510 0.5739 3.TABLE :1.2287 0.0000 0.2072 0. ' .0087 0..5556 0." 1) How many failures can be expected in each mode over the next 2 years? (Assuming a usage rate of 25 hours/month) How will lengthening the intpection interval to 1200 hours change this risk? 2) 52 / .0039 0.2656 2.0810 0.0020 0...0020 0.0493 0.2648 0.5384 0.0028 0.0000 0.0001 0.0000 0.3208 0.3871 0.1022 0.1186 0. At fail'ire or schieduled inspection the modes are made "good-as-new.5774 0.

•I .. * VANE AND CASE CRACKING =4.. . - -- - -- -.57 '1 2336.O....9..•-S =10193 OIL TUBE CRxCiKrI = 1.-•- - z ... *. 1-osm a.4 . I - a 20 COMBUSTION _CHAMBER CRACK [NG = 4. ° I ~II u*'s I I It1111•YlI WEIBLIL DISTRIBUTION . iI: 4.03 '7 3149 . *1. 1O 2. Failure Distribution Input to Simulation 53 -L I I\ ..19 L .VERTEE-ATURE 0 = 2..885 .9.- . 4 f." /1 . TOTAL OPERATING TIME (FIR) FD 271860 Figure 3. =12050 s : I # I ! loc. 4. !.

193 1Qn (1 - 0. 0.442 + 7.0.442. 10. therefore. l A Million Random Digits With 100. will he followed step l)y stel) to 'the first scheduled inspect. First-.193 [Qn (1 951 hours - 0.ulation. A Monte Carlo simulation baset on I hese groundrules is illustrated in Figure 3. one engine starting with 0 hours.149[Qn (1 - 0. Using Fquation :3.028 F.02$. and 0. 0.(X)7.603 = 3.336 [Qn (I = 1. Stop 1 (enierate random times to failure for each failure mode. Rand CorpoJration. four random n:umbers converted to the 0 to I range are 0. Generate another time to failure for F1. using the next random number. New F. using a table of random (Reference 1). 0.-Oil tube cracking - 0517 F4 = Combust ion chamber cracking I = 3.072 hours 12. * \* ~ * * .827 hours Ref.517. the scheduled inspecl ion was not reached.60'3.160 hours .ion at 1000 hours.'I'There is no easv solution to this problem willtout sin.0150 [Qn ( 10.876 hou 951hours on F failure 8.4: F1 = Overtemperature = = 10.10. 1955.088 hours Steps 2 & 3 rhe minimum of the times-to-failure and inspection time is 951 hours. 'I'l)rovide more det-ail. * numbers. The Free Press. 54 / .07 F2= Vane and case crac=king = 2. Stop 4 This failure was an overtemperature (F1 ) and is recorded as occurring 951/(25 hours usage) :18 months in the future.000 Normal Deviates.

) 0 NrU 0 00 co 00 C: 0 ca S E CL ca CM IO tIO ( ( 0 0 ~~U. 1iz 55 .O'm' 00 Vi E% Coto L.

other random numbers are drawn until all four failure times are greater than the initial age. This prcess can he continued for as many inspection intervals as desired. 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). F.. 56 . Increasing the inspection interval to 1200 hours increases the expected number of failures from 25 to 34. . and in Table 3.- '- . -. If any are less. For engines with greater than zero hours initially.8 for the 1200 hour inspection.12. by the end of 1981. . and the risk for the first 24 months is presented in Table :1. a delta of 9.4• . / %Y \ '/. The simulation in Figure :3. .¶iI Now.10 was run. which is the scheduled inspection. 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). F 4.- t' / "" \" / / _. Then.7 for the 0XX) hour inspection. F2. 10X) hours) is 0XX0 hours.11) and is repeated several times so that an average risk can be calculated.:. the minimum of (FI. the Monte Carlo process must be . modified. The above procedure is followed for each engine in the population (Figure 3. note that theth first. A plot comparing the two risks is presented in Figure 3.

01 * Fri ae) Co COS CD CC' 57S .N. 1~100 ~.E 9 E 69 99 10 0 C-4 a 0 0 9*D .

33 4.29 2. Chamber SJan SFeb Mar Apr May Jun Jul Aug Sep Oct Nov SDec I 29.77 2.32 3.13 1.44 5.TABLE 3.7.47 0.34 0.44 3.66 2.05 3.90 KEH = engine flight hours 58 : I .21 4.51 2.225 29.13 6.225 29.474 350.70 3.00 0.724 642.225 29.49 1.59 9.924 409.60 1.15 4.38 0.62 3.46 1.64 3.225 29.67 1.17 0.45 **SS****19815S Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 29.225 29.03 2.33 0.60 0.15 1.07 3.450 87.55 5.38 0.80 4.41 1.66 1.17 0.00 0.06 4.59 8.77 2.92 1.399 2.19 3.225 29.225 29.02 2.5 29.76 3.87 1.225 29.225 58.02 4.90 0.87 3.33 7.58 0.049 555.95 2.47 4.71 2.5 29.26 8.78 0.21 0.2215 379.174 701.225 29.03 4.225 29.68 7.83 6.00 0.95 1.824 526.15 1.225 29.28 3.949 672.225 29.27 281 3.95 5.22 1.274 584.87 0.16 4.90 4.225 29.47 1.22.22.499 613.47 4.675 116.31 3.225 29.07 2.250 321.225 29.149 438.67 391 4.225 29.699 * * 0.95 2.125 175.374 467.96 3.48 8.21 4. SIMULATION OUTPUT FOR 1000 HOUR INSPECTION * •Cumulative Incidents Month S.575 233..74 0.69 4.40 4.900 146.00 0.91 .74 0.225 29.# EFHO Curn EFH* Oil Tube Vane Case Over/Temp Comb.5 29.350 204.225 29.225 29.77 2.025 292.37 3.22.36 2.40 9.225 29.45 8.62 4.800 263.16 3.599 496.

73 10.900 146.225 29.225 29.77 16.53 1.40 1.22.65 3..39 8.924 409.68 6.21 1.00 0.225 29.25 6.40 1.949 672.00 0.58 6.43 2.225 29.40 0.174 701.499 613.'rTAIL 3.47 3.61 2.42 4.450 87.225 29.89 5AO 5.21 3.47 0.225 29.07 6.225 29.01 4.96 4.26 9.399 2.92 14.46 9.59 4.17 0.225 29.17 2.8.824 526.57 1.82 2.01 3.16 15.35 0.24 0.81 2.14 4.93 4.16 13.99 12.72 15.225 29.225 29.94 engine flight hours 4 4 59 Mj" .225 29.69 0.374 467.44 3.599 496.04 4.22.93 3.64 4.049 555.25 0.72 14.26 6.05 1.025 292.474 35G.36 2.724 642.84 1.69 Jan Feb Mar Apr May Jun Jul Sep Oct Nov Dec * EFH "Aug 29.07 6.31 3.5 = 379.225 29.274 584. SIMULATION OuT'itT FOR 1200 HOUR INSPECTION Cumulative Incidents Month )•980 EFtII ('urn EkI* Oil Tuhbe Vane Case thver/r.64 7.00 0.34 3.13 2.81 2.98 3.800 263.46 3.79 1.16 12.699 0.350 204.225 29.79 3.43 1.675 116.225 29.65 3.35 4.mmp Comb.08 1.16 4.31 3.225 29.93 2.225 58.28 2.62 1.55 3.08 0.32 4.12 2.21 0.226 29.02 4.225 29.86 5.92 15.225 29.00 0.99 5.125 175.58 6.149 438. Chamber Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec • *eS*ee*' 29.250 321.225 29.94 6.21 4.58 1.575 233.5 29.40 6.72 3.225 29.91 1.225 29.34 4.225 1981****** 29.31 3.

Problem 3-2 A turbine airfoil has caused unscheduled engine removals at the following times and locations.13 PROBLEMS Problem 3-1 A fleet of I() engines is subjected to a Weihull failure mode.3. How do these Weibulls compare? 60 'I.) 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.14).504 476 A) Location A A A A A A B C E G H G E C B B C G D E H J A ------ Generate a Weibull using the attached populations.- .13) and at Location A (Figure 3. The current engine times are as follows: Number of Engines 20 20 20 20 20 A. The Weibull has a slope of 3 and a characteristic life of I(XX) hours. overall (Figure 3.) Engine Time 150 hrs 200 250 300 350 What is the expected number of failures now? B. Time at Failure 684 (hours) 821 812 701 770 845 855 850 806 756 755 741 681 667 649 603 600 596 57b .

30. 0010 00. :101> ~0.-0 CLL 0 0) 0-0 CcU 0*00)1 0 . 0. oCM CO 00.

15. 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. and the three Weibull failure modes: a. /i i ii =2. 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.4 = 1126. The following population of 5 engines is at risk: I engine at (X)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.735 total operating hours. that make modes a and c "good-as-new. hours Assume two scheduled inspections.9." while not helping mode b. the next 24 months? from each population? (Use 30 hours/mo.) Problem 3-3 Given a control failure mode with • = 1. and the population of nonfailed units Figure 3." 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 3.587 hours Z b. A usage ra'e of 25 hours per month is assumed.26 and iq 19. at 1000 hours and 2000 hours. c.q = = 0.76 96. B) Would it be advisable to drop the 1000 hour inspection? Solutions to these problems are in Appendix J. I engine at 200 hours. I engine at 7M) hours. .H) How many failures can be expected in the next 12 months?.038 = 4996. I engine at 500 hours. 62 . hours =7.

co 'J o 0m v m 0 II60 .4I 00 oro..Go. 0 rC-S C.

O £fl UC) 00 0 co Nq CC.I-C. a-. 0 co CD4 cmJ CJ IC a'a 0~0 0 c 64 .

oDC Ml.Ci > C'. 0 st IC coo4 co- 00 cm' C F) co co 0 a0 65 .

793 0.669 0.s..811 0.r% 0.164 0.605 0..615 0.8!1 0.601 0.013 0.194 0.556 0. n-rI.609 3.422 0. TO 1.323 0.668 0.104 O.198 0..895 0.821 0..213 0.012 0.133 S0.750 0.655 0.457 0.428 0.8 0.169 0.850 0.210 0.971 0.159 0.491 0.124 0.010 0.t.689 0.070 0.300 0.180 0.624 0.604 C.980 0.-------- 66 .855 0.526 0.32 0.340 0.230 0.996 0.676 0.447 0.888 0.838 0.i34 0.I45 0.%03 0.085 0.635 0..34 0.393 0.612 0.748 0.5A2 0..580 0.945 0..480 0.r~z.083 0.429 6.717 0..932 0.551 0.4.636 0.168 0.7611.984 0.105 0.608 0.315 0..620 0.298 0.699 0.337 0.249 0.815 0.701 0.-tr.596 0.612 0876 0.202 0.- * .961 0.7f9 0.511 0..827 0.277 0.577 0.469 0.833 0.71.845 0.311 0.476 0.842 0.058 0.494 0.990 0.837 0.507 0.329 0.688 0.083 0.364 0. ~ .503 0.221 0.574 0.506 0.054 0.0 0.408 0.480 0.703 .420 0.630 0..431 0. 0.12 0.521 0.163 0.138 0.643 0.214 0.354 0.987 0.774 0..544 0.857 0.9.249 0.638 0.123 0.747 0.367 0.765 0.037 0.084 0.363 0.847 0.265 0.3.854 0.963 0.205 0.950 0.974 0.270 0.717 0.668 0.2.412 0.- .909 0.918 0.675 0.772 0.257 0./ \ TABILE :.933 0.805 0.490 0.376 0.222 0.389 0.418 0. e.384 0.551 0.943 0.495 0.519 0.982 0.709 0.928 0.365 0.655 0.090 0.041 0.826 0.394 0.706 0.600 0.673 0.167 0.224 0.794 0.238 0.390 0.785 0.166 0.203 0.861 0.497 0.649 0. uve.198 TABLE OF UNIFORM RANDOM NUMBERS FROM 0.36 0.687 0.402 0.658 0.851 0.761 0.502 0.

the Weibayes confidence level is 631". 4. If the Weibayes line is always constructcd immediately before the first failure. a Weibull equation is determine-". (See Appendix E. In this case.r ~I" CHAPTER 4 WEIBAYES 4. while unknown. i.. the confidence level is actually much higher than 63'. 67 t - . the denominator in equation (4. an equation may be derived (Appendix E) using the method of maximum likelihood to determine the characteristic life. i. N .1 FOREWORD - WHEN WEIBULLS ARE IMPOSSIBLE At times a Weibull plot cannot be made because of deficiencies in the data. this may be a strong or weak assumption. Typical situations would be when: (1) There are too few or no failures. (3) A test ). r is the -"imber of failed units and V is the maximum likelihood stimate of the characteristic life.lanfor a new design is needed.. A Weibayes line can be plotted on Weibull paper. The Weibayes line based on assuming one failure is conservative. the slope/shape parameter t is assumed from historical failure data or from engineering knowledge of the physics of the failure. confidence that the true Weibull lies to the right of the Weibaye. r = 1. If Weibayes analyses are consistently done long before the first. Given #.. a second assumption is required.e. Weibayes is defined as Weibull analysis with an assumed # parameter. Weibayes displays conservatism since the conridence level.minent. Therein. The first failure is assumed to bw ir-. Deperoding upon the situation. With #iassumed and q calculated from equa4tion t4. is at least 63'. (2) The age of the units is unknown. The plot is used exactly like a Weibull distribution. Since the assumption requires judgment.) The exact confidence level of the Weibayes lower bound is unknown because it depends on the time to the first failure.0 (otherwise.1) would be zero). this analysis is regarded as an informal Bayesian procedure. 4. line. Weibayes is never preferred over Weibull analysis but is often required because of weaknesses in the daLa. (4. and only the number of failures is known. with at least 63 '. Weibayes analysis has been developed to solve problems when Weibull analysis cannot be used.2 WEIBAYES METHOD lit a Weibayes analysis.1) Where t1 is the time or cycles on uniti. failure.1).3 WEIBAYES NO FAILURES In many Weibayes problems no failure has occurred.

.. However. a Weibull shift method may be employed. (See Section 6. (3) With nonserialized parts. . what is the probability that a new failure mode is present? A Weibayes test may not discover a new mode. . again Weibayes may provide a solution. but not the times to failure.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. :h• • . + i" tI N:p 4. •/ . If the physics of failure are known.4 WEIBEST - NO FAILURES %-4 . so it is appropriate to critically question the assumptions required by the Weibayes method in each case since the answers to these questions vary for each application. .. A %.X':. should a range of slopes be tried? (2) With a redesign.. How valid is the assumption? 4. In fact. The result was called a Weibest line.• - "..substantiate the redesign? -. the validity of the results depends on the validity of the assumptions. .0 (see Figure 4. ..693 failures would be assumed instead of 1. . Grderuck of Pratt & Whitney Aircraft/Commercial Products Division is credited for much of the original development of the Weibest concept 68 'N K. if the time on the components or systems is known. it may be impossible to determine the time on the failed unit(s) or the success unit(s). 4.5 UNKNOWN FAILURE TIMES Sometimes the number of failures is known. Three redesigned compressor cases have been tested in engines to 1600. /. some assumption must be made to obtain success or failure times.) 4. Typical questions to be raised are: (1) How valid is the assumed slope. the maximum and minimum # may each be used to determine the sensitivity of the analysis to the assumption. Mr. Y. Weibayes lines may be calculated for any confidence level (Appendix E).6 WEIBAYES WORRIES AND CONCERNS A 4 The Weibayes method is required when there are deficiencies in the data or when the data are not available. txý ~ V . _ Kl "- .%4 In the early development and use of this analysis* 0. ...•. for Weibayes with r 1. For example.0. it may be reasonable to assume that the same distribution of times applies to the nonserializcd parts. "If the times ton the failed units are known but the times on the successful units are unknown. Is this enough testing to .. 2900 and 3100 hours without failure. The Weibull method is always preferred over Weibayes.. : .. a library of Weibull failure modes may provide an estimate or a range of estimates.3. The Weibest line is a 504. -1 t . Joseph W.1). This is less conservative. lower confidence bound on the true Weibull characteristic life versus 63'. if the failed part is nonserialized and the component or system has been through overhaul.. e? If this assumption is shaky. there is more uncertainty in assuming a value for 0. 1. Of course. In this application..:- .

EIBULL DISTRIBUTION 0 = 4.247 SAPJLE SIZE = 275S FAILURES 15 - - - -\ U. - * . - -!/ . rMU -4 ' IS1JJ/ 4 io 111 'BIN TIME SINCE OVERHAUL (HR) Figure 4. Compressor Vane and Case FD 271a61 69 .404 "U.1. ..- .li --- -- .-- 4 .971 T= 2132..

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.1 and the slope from the Weibulls in Figure 4. • : •" " "• i'Z .. ~~~. We may state with 63"1 confidence that the failure mode for the redesigned units is to the right of this line and.. assuming the first failure is imminent.3. The assumption of slope. DF1 and DF2. As more time is put on these units without failure. This provides some confidence that the accelerated test provides the same failure mode experienced in service. + 5o 2 7. * The Weibayes line is plotted in Figure 4. = 10. Based on the physics of the failure. it was decided to replace these pumps after 175 hours of operation.2.Assunuing 1• 5. 1 50027]" L - 646 hr (4. the Weibayes characteristic life is calculated. had more service time but no failures.4).5.3 •Redesigred/Semce.3). 11 / Ieeignnediervice = ('B/M/SVC = 113B/M/Rig) iiRedem. This line is also plotrted in Figure 4. There were not enough spare pumps to immediately replace the DF1 and DF2 units. [r '.material vane and case. therefore.3) This Weibayes line is plotted on Figure 4.6) 646.5. 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. a Weibayes line can be estimated for the redesigned pump in service.082 hr I'roblem 3) One batch of turbopumps (DF3) produced nine service failures involving fire in flight.2) . Notice that the bench test Weibull has the same slope as the field failure Weibulls. an accelerated bench test was designed and two more turbopumps failed in a much shorter time (Figure 4. It is possible that the redesign has eliminated this • failure mode but that cannot be proven with this sample of data...2 hr + 140 hr) 646. P'roblem 2) There have been 38 turbopump failures in service (Figure 4.in/Rig (2186. significantly better than the bill-of. Two other batches of these pumps. How long can replacement of DFI and DF2 be safely delayed? 70 : "• . 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. Is the redesign successful? What service experience should be expected? Using equation 4. = 1 = L'(1600):' + (2900)) + (3(M)03'] .1) mav be used to calculate the characteristic life for a Weibayes solution. .. From a Weibull analysis. equation (4.3 hr = (15. the Weibayes line will move further to the right and more confidence will be gained that the failure mode has been eliminated. ~~(4. is based on an established Weibull failure mode and is valid. in this case.0 aind given the tinies on the three redesigned units. = 3406 hr ".

. .862 dTs ~~~~F 1-:. •.P =. •: 1...." TIM SIC .. EB IREIBAYES DISTRIBUTION = 4.-ac S726 .971- i. .% q.. °• 86> :::.50.9- °So. _ _ - -- = 4. ...I: .......247 SAPESIZE = 2755 FAILURES= 15S. P= 413....j 41 i • !t .. go...971 1 12 OS.. ...x•-:. ....12 \ ia(sEauto fNi . * %.- - "W.°...q• : 120 116 I" •Of ' _ Oa --.... • • -o~itd '"... w 0 ... .{•' F-1-" . OVRAL(R ~~ ~ ~~ ~sg ~ nAclr ~ ~ ~~ ~"2_-.- d 40.. 'V1 = 3406..

.624 "" = 2186. U-I 'I•.UI.. '. :- ' " iL. ..._ (EIBULL -." ' . .7. SIP . -.. - TOTPIL OPERATING TIME (HR) FD 271864 Figure 4.. - - .- B/M FIELD EXPERIENCE 1...3... DISTRIBUTION - 0 = 2.> 4 • ''4 o 0.2 LE SIZE = 800 F~AILURES =38. - i .I *.S. " - -. Weibull Evaluation of BIM Design 72 • .--.

L - -. - -- I - -- B/M LOW LUBRICITY TEST EXPERIENCE (2 FAILURES) . Pm.. DISTRIBUTION WE BULLD I I SI BT ION EIBU.I *• • m. - - TOTAL OPERATING TIME (HR) S/- / ! : PID 271865 '. • - B/M FIELD EXPERIENCE (38 FAILURES) . Weibayes Eval'uation of New Design in Accelerated Test 73 . -1- A. 'N Figure 4..4..

. .5. I.L REDESIGN PROJECTED FIELD EXPERIENCE 2. L 6. 4. .0828 (2 FAILURES).PI . . 4.. .-EIBULL DISTRIBUTION I •i . L.. 2.46 .. Hydrzulic Pump Failures 74 -' Sn . 1I - . .- ILU. • : . -. 11.1.S.Z-vTEST REDESIGN LOW C ITY EXP -- ii I-V tO.6. l. am. TOTAL OPERATING TIME (FIR) FD 271867 Figure . 4.1. _ . -.I *Ii I I I I I II1._ B/M LOW LUBRICITY TEST EXPERIENCE O0. _. 3.IO. 4. S. 0 0 B/M FIELD EXPERIENCE (38 FAILURES) i M 1/ V 3.

'3 rofil al 175 hours (morresponds to a risk level of 1. 4. 150. Therefore. 125.. A Weilwyes analysis.6. If most disk LCF .V.7 as indicated in Figure 4. is this success data sufficient to increase the predicted design life? Solutions to these problems are in Appendix J1. many pumps were utilized to their full life without premature replacement based on the Weibest Analysis. :. (Assume the two new replacement bolts are installed in the rebuilds of the high time engines. the Weibest B.1 life for the compressor disk is 1000 cycles. 901 and 40 hr. . 110.6. it is not known which engines had the failed parts. md l)F2 pumps were replaced when they reached 700 hours. If low cycle fatigue "iilure modes usually have slope parameters between 2 and 5.'lThere were no failures in ID)I and ID)12 even ihough symphins or the t. This did not create a supportability problem as these pumps had acquired very little time.(9!t3 ifailures were imminent.0. Weibest and Weibayes lines move to the right with time as long as no failures are observed due to the increase in success time. confidence) line shown in Figure 4. Five disks have accumulated 1-510* cycles and rive have 2000 cycles wiirhout any failures.o '•i 75 ~amsana ~ *~wrnamnsnwn r~- . predict the number of expected failures. using I he exist ijg Weilmll slope of 0i 4. produced (he Weihest (5(0'. and after rebuild the engine will accumulate 100 hours in the next year..alure mode were presenl..failures have a 11of 3. The I )1. "I.7 time eventually exceeded the pumps' overhaul time of 1000 hours. and assuming the (0.97 I.2 The design system predicted B.n -- - - - - o . The bolts are not serialized and as the failures were discovered after the engines were overhauled.4. In this case.) Problem 4. DFI.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. The same risk level was applied to the Weihayes line and a 7W0 hr safe period was recommended.

2. 1.S20 SAWILE SIZE =387 FAILLIRES =24 CONF LEVEL =0. 6.17S 0. 4. S. . 3.1000.8.L OPERATING TIM1E (HR) FO 271868 * ~Filure 4. 2. asw~ TOTP. Of) cc m J . 40.300.- 1. 4. 3. Whzen Should We Pull the Sus~pect Batch? 76 .6. 2. i.- 10. 4. 2.915 = 488..19.17.7.5.' WEIBLLL DISTRIBUTIONP = 4..8.90 _ __ 70._ - 73 ii wU sag Z) MO 2 no low~ fy 20 aK..

test plans developed by K. R. 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. where the test is passed if the success criterion is met. a part or system will have at most a 10'"1 chance of being accepted as sat islactory when in fact it is not. The conesponding table entry is multiplied by the characteristic life to be demonstrated to find the test time required of each unit. It also gives a success criterion. N. etc. It is assumed in the reliability testing section that the Weibull slope parameter. Fertig and N. Test plans can also be generated with a 1-failure success criterion. Substantiation testing demonstrates that a redesigned part or system has eliminated or significantly improved a known failure mode (0 and 17 are assumed to be known). iMann. 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. A mwasure of confidence is usually built into statistically designed test plans.1 illustrates the time-to-unbalance distribution. 5.CHAPTER 5 SUBSTANTIATION AND RELIABILITY TESTING 5. and three redesigned systems are available for testing. guaranteeing thait if the failure mode in question has not been fixed or the reliability requirement has not been achieved.1 is used to answer this type of question. The system is redesigned. and K. a two-failure success criterion. Reliability testing demonstrates that a reliability requirement has been met. with 0 (the Weibull slope parameter) equal to 2. 22. Table 5. 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.em be tested to demonstrate that this mode of unbalance has been eliminated or significantly improved? The Weibull plot in Figure 5. Thus. But all of these plans require more testing than the zero-failure plan.165-177. there is a low probability that the test will be paseed. How many hours should each syi. unbalance. 77 inmt. is known. (1980) Life-Test Sampling Plans for Two. and q equal to 500 hours. W. If the failure distribultion is known to be Weibull but 4 is unknown. It is entered with the value of 4 and the number of units to be tested. MannI may be used.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.2 ZERO-FAILURE TEST PLANS FOR SUBSTANTIATION TESTING A ball and roller bearing system has a Weibull failure mode. #. Tehnometris.Parameter Weibull Populations.- . Fertig..

10m0. 2.1.. - - 1BFALL AND ROLLER BER•RING UNBALANCE z DISTRIBUTION * 0 tO.. EIGULL DISTRIBUTION o00 HOURS 1= 1 h U B . '2. FD 271869 78 . . 5. * .65000. I.. . 3. 4.9. Ii i 'I Y 3I .0*. 17OTPL OPERATING TIME (HR) SFigure 5. SJO..3. 4. - - - -- - - " --. IBall and Roller Rearing Unbalance Distribution 3. -I--'.

If there is a constraint on the amount of test time accumtlated on each unit.442 0.780 0.300 0.256 0.787 0..0 Rapid Wearout + .164 0.682 0.677 0.596 0.649 0.303 0.384 0.733 0.826 0.824 0.486 0. confidence).761 0.0 and the ratio 300 test hours per system 500 hour characteristic life =0.463 0.. 79 .230 0. Table 5.277 0.936 0.027 0.192 0.477 0.406 0.292 4.5 5. If all seven systems are in balance at the end of the test.589 0.2 is entered with the known value of d and the ratio of the test time to the characteristic life being substantiated.745 0.802 0.619 0.662 C.288 0.772 0.536 0.616 0.331 0.065 0.556 0.576 0.333 0.128 0.90 0.871 0..565 0.215 3.058 0..5 Infant 1.0 .670 0.727 0.621 0.083 0.+ 0.657 0.358 0.254 0.551 0.916 0.504 0.490 0.0 and a sample size of three.0 Mortality Random 0.635 0.641 0. suppose in the previous example that no more than 300 hours could be accumulated on any bearing system.580 0.781 0.598 0.053 0. In the ball and roller bearing system example.2 is used to determine the number of units required for the test.693 0.006 0.728 0.876 X 500 hours = 438 hours.506 0.421 0. The resulting test plan is: test seven systems for .633 0.758 0.801 0.575 0.589 0. the zero-failure test plan to subrtantiate the ball and roller bearing system fix is: test three systems for 438 hours each.+ 0.144 0.697 0.439 0.711 0. 732 0.548 0.046 0 2.876.619 0.761 0. Thus.660 0. The corresponding table entry is 0.2 is seven. Table 5.597 0.679 0.TABLE 5.008 0.767 0.692 0.037 0.948 0.650 0.719 0.376 0.613 0.690 0.438 0.2 is entered with # equal to 2. Table 5.480 0.701 0. 5.486 0.6 -Table The corresponding entry in Table 5.339 0.5 -----.759 0.1 is entered with # equal to 2.0 Sample Size 3 4 5 6 7 8 9 10 12 14 16 18 20 25 30 40 50 1.539 0.577 0.693 0.637 0.379 0.528 0.403 0. then the unbalance mode was either eliminated or significantly improved (with 90'..530 0.319 0.149 0. If all three systems are in balance at the end of the test. The number of hours that each system should be tested is: 0.00 hours each.900 0.838 0..436 0.5 Gradual Wearout 0.801 0.842 0.1540 4.115 0.452 0.884 0.739 0. For example.663 0.895 0.854 0.092 0.565 0.275 0.(Brick Wall) .212 0.524 0.582 0.128 2.757 t.013 0.598 0.0 3.460 0.016 0.506 0. / 7 .077 0.425 0.8.608 0.021 0. The characteristic life to be demonstrated is 500 hours.505 0.181 0.204 0.108 0.415 In the balh and roller bearing example.003 0.943 0.56 0.1 CHARACTERISTIC PLANS LIFE MUILTIPLIIERS FOR ZMRO-FAIILJRE Tl'tSTr / CONFIDENCE LEVEL: 0.329 0.358 0...526 0.385 0.237 0.574 0. then the unbalance mode was either eliminated or significantly improved (with 90'( confidence).927 0.624 0.722 0.808 0.832 0.517 '3.358 0.240 0. 0 0.556 0.5 + 0.147 0.480 0.002 0.386 0.679 0.461 0.

.80 0.0 Gradual ------Wcarout * 2302. Form 2: The B10 life (or B1 life. 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'(.2 REQUIRED SAMPLE SIZES FOR ZERO-FAIIURE TEST PI.06 0.2 illustrates the requirement of at Icast 99"1 reliability (at most 1' unreliability) at 1000 hours. How many combustors must be tested. chance of failing before reaching its 1310 life.1278 35977 18420 10660 6713 4498 3159 2303 288 86 36 19 11 7 5 4 3 4.05 0.. A turbine engine combustor's reliability goal'.5 2.ndom 0.K /: avu'No .70 0. etc.TABILE 5.X`'.01 01.. (This is equivalent to the pereent failing being at most 300.0 4. See Figure 5..000 5.ANS CONFIDENCE LEVEL: 0.5 Rapid Wearout -4+ .02 0. .. . . 001"s ***** 0** * 82377 '* * 43519 ** 25373 95898 * 15900 66214 ** " 35094 350A4 * 7282 21025 72812 * 644 1440 3218 7196 156 28.was 99.60 0. .. By definition.5 5.) is required to be at least X hours or cycles..0 3.5 Infant 1.. .5 5757 40703 2559 1477i 1440 7196 922 4119 640 2612 470 1777 M60 1272 285 948 231 729 58 129 26 47 23 15 10 14 3. . . ..(Brick Wall) ..3 ZERO-FAILURE TEST PLANS FOR RELIABILITY TESTING This section contains Z.1 life. or B.. ~~. Success was defined as a comhbustor having no circumferential cracks longer than 20 inches (out of a possible 53 inches).90 4 6 0..10 0.09 0.5 + 2303 815 444 288 206 157 125 102 86 73 26 15 10 7 1 5 4 4 3 3 2.90 1. a 1 chance of failing before reaching its BI life. etc.04 0.3. aftcr a certaiin number of hours or cycles of life.00 24 17 14 12 1 10 9 9 8 8 6 5 4 4 3 3 3 3 3 231 1. • .).07 0.+ ***** **0** ***** **** ***** ***** ..ero-failure test plans for demonstrating a reliability goal when the underlying failure distribut!ion is Weibull with known slope parameter #.$.50 0. reliability at 1800 cycles under service-like conditiots. The number of cycles required to reach a 20-inch crack was known to1 follow a Weibull distribution with # equr. The Weibull plot in Figure 5.20 0.l to 3.*..40 0. the unit has a 10'.0 Ratio Mortality R. to demonstrate this goal with a high level of confidence? First.. .03 0. the reliability goal is re-expregsed as a characteristic life goal.0 ----* * 8.08 0..30 0. and then the test plan is designed.5 519 948 225 143 90 57 27 14 9 6 4 3 37 18 10 6 4 3 53 23 12 7 4 3 74 30 14 8 4 3 116 77 58 47 39 33 29 26 24 12 8 6 5 4 4 3 3 3 [ 5 4 3 3 9 6 5 3 3 Indicates sample size exceeds 100. 80 J . and how many cycles must each accumulate..

3 10...1c000. .t TOTAL..UJ- -.5.t . .tj. 1 . .2. 3 4.- - -4 - I - Hil I F W.99 Reliability at 1000 hr 81 /1 . 6 . OPERAiTING TIME (HR) FO 271870 Figure . 800. Weibull Plot for 0.S.

. 0 TOA PRTN I TIME(R AV 2137 .S 7.w Of- C)I I-I C-) D..S . I.3.. I 10D 2.. . B1 l n 1 82 ie I~~ I .2 . i-Ai0 .2.L769ico..~~~~ ~ Il ~ ~srto f.

-T -.F'ig ire .- iA - I.es I ht requiruiment tl'a 7" HJ) hf)ur B I0o life..- 4 0. U-4- 0.. - - .4.--t.I U. Illustration of 70(X) hr BIO Life 83 ! . I. U. 14 ' BIO81 = 7000 HFIl TOTAL OPERATING TIME (HR) FO 265617 Figure 5.9_ a.i1 il IIst rat.

assuming 8 2 or i. In the combustor reliability example..1 and 5. the B10 life requirement of 2000 hours with 8 that q is at least 6161.9 cycle characteristic life goal. then substituting t R(t) 0..99 at 1800 cycles (•1 .90)1" 2 .9 '7 be The 0. the 99% reliability goal at 1800 cycles was reexpressed as an 8340. Multiply the table entry by the characteristic life requirement to find thc test time required of each unit. 6161.2) gives 17 = 2000 and 2000 1-1n(0. or.2 can be used to design the test plan. See Figure 5.99)).2) gives 1800 ' - i _in(0.613.Reliability requirements assuming either of these two fornis can be expressed as a minimum characteristic life requirement. If the requirement is.99 into equation (5. for example.1 with # equal to 3.. = 8340.90 into equation (5. Ten combust3rs were available for this reliability demonstration test.igninp Test Plans Once the minimum characteristic life requirement has been calculated.6 Thus. reliability at time t is a function of q: r(t e (. Similarly.2) can be used to express either form of reliability requirement in terms of rl.5. enter Table 5.5.3). then substituting t = 1800 and R(t) = 0. giving t '7 == I-In R(t)I"" (5. To find the test cycles required of each combustor.0 and a sample size of 10. that the reliability of the turbine engine combustor must be at least 0. if the requirement is a B10 life of 2000 hours.1) This expression can be rearranged algebraically. = 2 is equivalent to the requirement 84 . at least &840. (iven that the time-to-failure dist ribution is Weibull. The corresponding table entry is 0.2) Equation (5. Tables 5.6 hours. with a known J1.6.99 reliability requirement is equivalent to the requirement that See Figure 5. D.9 cycles.

'I ..9340. A Reliability of (1.9. .' -- . 5.5.9 U.3 I S=8340..99 at 1800 Cycles Is Equivalent to On &140. -91 I TOTAL OF'-RWTING CYCLES F=D 271872 F'igure. .9 Dz3 J M U.-_.- . Cch CharacteristicLife When 8=3 II II " " . EIBf.L DISTRIBU11T -.I./ 8340.

.....U -.. . .6 hr Characteristic Life When 0= 2 86 . 0.5.. EIBLLL DISTRIBUTIONTI S=2 6161..6. -.. . - -- - - - - __ 7- AV271873 Figure .1• . A 2000 hr B10 Life Is Equivalent to a 6161. . so. SI U. .6.. loa ' mm.- . ....6 -. - - s .--- 0.

..4 TOTAL TEs r TIME Two reliability test plans were constructed in Section 5..3 :i KI....09 The corresponding entry in Table 5.2 is 3159.113 Total Test Cycles 3. Also. the test plans are simple and easy to use.. the lower is the total test time.. .2 limit the probability that substandard reliability units will pass the tests. # was assumed to be 3.i :i I Py tfie chari-I eris. and the ratio of the test time to the calculated characteristic. 5... the greater the number of units on test.I I lhe ciniistu.. life requirement. If no combustor develops a circumferential crack 1rnger than 20 inches... Plan 1.159 X 750 10 x 5. I life reillreleiiw l Pit 83...14... enter Table 5. Per Combustor 7..369.6. they are only designed to limit the acceptance of substandard reliability items.113 = = 2. They do not control the probability that units from a high reliability design will pass the tests.. oft: 0.. However. The situation is reversed for # less than 1.. How many combustors are required if each can accumulate at most 750 test cycles? To answer this.. In the comhustor example... This efficiency is realized for every test plan in this section where exceet7s 1. Plan 2 demonstrates the same reliability as.. 'Thus. the total test time is constant.159 10 Test Cycles..130 Note that.t. 87 .0. but requires fewer total test cycles.. relialili:. The resulting test plaa requires testing 3159 combustors for 750 cycles each..9.. the Weibuli siope parameter.. with 90'e confidence...hr exai plel. I hen Ihe test is passed....2 with the assumed value of 13.. If no combtustor develops a circunlfereptial crack longer than '20 i(nches. .I()09 cyc'les en I ry itf u.50 5. I ile 5... niillily i tig I Ie 'it.. Number of Combustors Plan 1 Plan 2 3... When 0 is 1. i(" .fiailure test plan o11 (lemnonstrate 99'.'at 18(K)cycles reqtuires.. in terms of total test cycles... In this case.3 to demonstrate that a characteristic life was at least 8340. and the ratio of the test time to the calculated characteristic life requirement is: 750 test cycles per combustor 8340....51130 4 cycles. the zero.9 ('ycle's gives a Itst I ill... it is more efficient to run the smaller number of combustors for a greater number of cycles.9 cycles characteristic life 0. regardless of the number of items on test. then the test is passed. S5 ADVANTAGES AND LIMITATIONS OF THE ZERO-FAILURE TEST PLANS The test plans introduced in Section 5. This is generally the most important goal in reliability testing..9 cycles. l:est:ing 10 c(ombustors for '511: cycles each.250 51.

or the size of the test can be increased until both requirements are met. - \ - *i' .8.\ " -• ' •. suppose the characteristic life requirement were dropped from 2000 hours to 1250 hours. In this case.2.. -. The plans assume that the unit's time-to-failure distribution is Weibull with known slope parameter t.2 with I1 = 2. instead of 14.6 through 5. hours/1254) hours = 3.8.. requires ji confidence.. a zero-failure test plan is desirable.7.65. To see this.- ' . Only five units would havc to be tested 1000 hours. the statistical assumptions inherent in the test design cannot be validated. (if )'. to demonsi rate that a characteristic life is at least 2044) hours with W (Enter failure.close to the 901" requirement. 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. In the preceding example. wit hout hours I1000 he tested units that 14 2.0. 88 . The risk is that unless some units are run to failure.2 with # = 2. assuming Table 5. (-nter Table 5. test plans were introduced to demonstrate that a lower limit characteristic life has been achieved. . with failed units. For example. with 90'e confidence. For example. The second option is c(enidered in Sections 5.5 and the ratio I(XX) test hours per unit/200() hours = 0. The minimum characteristic lite requirement can he reduced enough to guarantee a suitably high probability of acceptance for high reliability units.5 to find 14 units.5 curve of Figure 5. to get the five-unit requiremen. To find the probability of acceptance of the 40(M)-hour characteristic life design. the Weibull slope parameter # can be estimated and compared to the assumed value of jI. enter the J1 2. The zero-failure test gives these high-reliability designs only a 65".5 and the ratio (1000 tst hours per unit/1250 hour characteristic life requirement) = 0.6 NON-ZERO-FAILURE "lST PLANS In Section 5. . 4(X0 hours (characteristic life of interest) R . ithe experimenter is interested in guaranteeing a high probability acceptance for high reliability units.6-5.) There are two remedies for this problem.7 with a ratio of 40000. (For example.7 assist the experimenter in determining how much to reduce the minimum characteristic life requirements to meet high reliability requirements. enter the I = 2. the probability of successfully completing the zero-failure test mentioned earlier in this section for units whose characteristic life is 4000 hours is 0.) 5.2. chance of passing.. from Figure 5. probability. (See Figure 5.'P.iN some inst ances.5. The curves in Figure 5.7.) Another requirement might be that designs with characteristic lives greater than 4M(X) hours pass the test with at least 90'. is 0. Very high reliability often makes a demonstration test-to-fa'lure impractical. .7 with the x-axis ratio of: R or R 2. Reliability demonstration tests that terminate successfully with no failure have one other advantage. The chanmes of passing are 88% .5 curve in Figure 5.ours (minimum characteristic life requirement) - The probability of successfully completing the test.

CD 0 CD0 a00 CVJ 101 -a CO) '1a . CLC 0 89c ..

' A. When an item on test fails. t -." " - As discussed in Section 5. Nancy IL. Ray K Schafer. 5. Chapter 6. do not control the risk of rejecting units with acceptably high characteristic lives (called producer's risk). B. =aa.1. is the characteristic life to be demonstrated. . and avis usually set at 0. Singpuwwala (1974). it is not replaced. C.!".A ..05 or O.r. the test is passed. -. ill is sometimes referred to as t'e "design" characterist. i 'I SMann.a. The equations to be introduced require the definition of some standard mathematical notation. should be at least a 1. \ "• ' / '+I. John Wiley and Sons. They only control the risks of passing the test with a characteristic life below the lower limit (called consumer's risk). The methods can be Found in "'Methods for StatisticalAnalysis of Reliatilityand Life Data-1 . Summation Xi X15 + X2 + + X. pp 312-315 and p 328.- *' --. B.c life.".5.95.7 DESIGNING TEST PLANS These tests will have the following structure: A. 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. these test:.°. Put n items on test for t hours (cycles) each. - This section describes methods for calculating r0 and n satisfying the two constraints: A. If roor fewer failures occur.a° -. The probability of passing the test with a characteristic life as low as no should be no more than aw(minimum life requirement). Methods for StatistitLd Analysis of Reliability and Life Data.9 or 0.SS. and Noser D.--.att..<. The Probability of passing the test with a characteristic life as high as q.-. • .. vis usually set at 0. 90 " " " -• • . New York.%- .

. School of Industrial Engineering ared Management. .)' Eqtvtions (5.r0 . Kenneth K and Lynn L Jones (1979). Dr.. Oklahoma) built an interactive program that includes the ability to solve equations (5.) (5. Factorial n! n.B..' is at least as great as a.3) and (5. E.l I-e-(Uv00 Pi = I -e-(t/' 1) t is the test time per unit n is the number ofunits on test r.4) genera!:y require a computer 'or their solution. the following equations should be solved for r0 and n to satisfy the two "'probability of passing" requirements.4) generally cannot be solved for a combination of n and r0 that satisfy the target probabilities a 0 and a.. .(n 2) .4).3) and (5. a characteristic life equal to P0 Gro is set by the experimenter) a..3) and (5. 'Case.. 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 "0o is the probebility of passing the test with. The next section discusses the method recommended by Dr.. Final Technical Report. is the probability of passing the test with a characteristic life equal to 1. "An Interactive Computer Program for the Study of Attributeii Acceptance Sampling. and the corresponding true probability a.. Certain computer packages are available that solve these equations.) )r~ (( P . 1).(n . Number of suhsets of size r from a set of n ' (n ~ n! (n -r r / Assuming that the time-to-failure distribution of the items on test is Weibull." Oklahoma State University. Oklahoma 91 . Case' for solving equations (5. (Gil is set by the experimenter) Equations (5.3) "= r where p0 = I' r. 2. Stillwater. Case of the Oklahoma State Uriversity School of Industrial Engineering and Management (Stillwater.4. K.. r 0) r (5. Some authors recomm2nd solving the equations so that the actual probability ao of passing the test with il = n is no greater Ihar. with known parameter #.3) and (5.1 C.. exactly..

Through redesign and material changes the vane's durability was improved. ('alculate e-(t/1 p = I I -. assume that the turbines are 'Ct. Average no and n. and 5." Oklahoma Stat.-t' • P0 ="1 2. Select as the final value of r0 that which has the a-ratio nearer the desired ratio b = p0/pl. with 90`' confidence. Set r0 (= 0 Find the values of n that satisfy equations 5. University. How long should each be tested to demonstrate. During this test. Bearing failures followed a Weibull distribution with # equal to 1. n 2 &9 PROBLEMS Problem 5-1 A turbine engine exhaust nozzle control bearing was failing prematurely due to fatigue.8 RECOMMENDED METHOD FO'" SOLVING EOUATIONS describes the recommended mrethod.determined by the worst vane in the set - followed a Weibull distribution with # = 3 and i = 1300 cycles. Case's final technical report' following steps: 1. A significant percentage Of the engines in servict were being removed for vane repair or replacement prior to their scheduled turbine maintenance. to get the final sample size: n + n. Final Techunical Report. and the environment in which it operated was improved to give the bearing a higher expected life. 5.n 0 and nj.3 and 5. 92 .4. 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). Call them no and n 1 . It consists of the Dr. 4. The time to failure . Okdahoma. Twenty redesigned bearings were available ior testing. Calculate a =i--"P" and b = p0/pl. 8. School of lndusirial Engineering and Management. If a is greater than b. and repeat steps 3. increase r0 by 1. :1. 6. For the selected value of r0 . there are two values of n.5. respectively. "(enneth K and Lynn L Jones (1979). Stiliwater.5 (a common value for bearing fatigue) and q equal to R000 hours. 7. Continue the process until two contiguous values of r0 are found whose calculated a-ratios bound b. nip. The bearing was redesigned. calculated in step 3. "An Interactive Computer Program for the Study of Attributn Acceptance Sampling. f= 4. that the fatigue mode was significantly improved? Problem S-2 High pressurs: turbire vanes were eroding beyond allowable limits.

limited to running at most 5(XX) cycles each. Also.,ssume

that the time to engine removal for
3.

excessive vane erosion would st ill follow a Weihu!l distribution with ji

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', confiden•e 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 'This chapter 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) (2) (3 (4) (5) Tturhopiump Bearing Failures (earbox Housing Cracks ()pportutnistic Maintenance Screening Intervals Support Cost Model Vane andI Case Cracking.

..
N

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 nmight help resolve this problem. 6.3 INITIAL ANALYSIS
-

,
a

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. 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.

ti

* "

*. .
°9

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

*

I

TOTAL OPERATING TIME (HR)
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

.. r - 97 . FAIL By 600 TOT- ofi_ ' WEIBLLL DISTRIBUTION 0'= S. 4 * * .'. - JC- -. tO. -.. 00 Fiur 2 AugS S/Nn 650 ND UP: q'u~ -i.a J- -a-r -a-."* . S S/N 650 AND UP: I YEAR j 128 FAILURES INF= i zo - - ~~2 340 YEARS FAILURES IN 9 FAILURES NOW . 99%.. . . 4 .193 !SAMPJLE SIZE =353 HOURS -- !2. .USh-.936 9•= 462.- r r - .. 10.

Failures Failures Failures Failures Failures Failures Failure.27 332.12 More More More Motm More More More More More More More More More Failure. .40 M.33 More Failures In 5 Months 34.24 221.34 &-6.02 More More More More Failures Failures Failures Failures In In In In I 2 3 4 Months Months Months Months 27.18 34.f8x P5189 &=27 290.44 26.75 307.21 8&..14 0.1128.47 More Failures In 10 Months 110. Failures Failure. In 8 Months ftt.65 93.3.64 86.06 More Failures In 6 Months 44.Failure Distribution cumn% Prob of Failure Engine Fleet Size .06 11.11 323.-46Z2 N .01 147.90 281.94 331.50 46. Failures Failures Failures In In In In In In In In In In In In In 11 12 13 14 15 16 17 18 19 20 21 22 23 Months Months Months Months Months Months Months Months Months Months Months Months Months 340.73 62.l Moto Failure.WO. Risk Analysis TABILE 6.16 312.1 Cumulatim Failures I"ROJ('ECTED PUMP FAIoIJRES Forecast Future Failurrs 9.27 More Failures In 7 Months 65.05 17891 200.90 18.21 18&08 .17 12.07 Mote Failure.22 21.85 137.11 167.18 More Failum In 24 Months . In 9 Months 84.24 78.29 100.66 242.0 More Failures In 0 Months &12 7.26 1&.58 298.32 321.l-p -- 1 Movement of Fleet Aging 25 Hours/Mo Time (Hr) FO 236M3 Figure 6.97 118./i /• // .33 98 . - . o .94 158.

j7. . 99 . a quick inspection revealed 27 cracked units. and it was questioned whether the same failure mode would appear in the new installation. t-he lleet. The second Weibull based on seven failures was also a good approximation of the final Weibull (Figure 6. At the outset. 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. .. a statistical factorial experiment was designed to determine the cause of the swelling of the plastic cages for corrective action.50 and up with 175 hours or more) revealed 15 more bearings considered to be imminent failures. the only difference being that the earlier curve had a steeper slope (1I0 rather than 4. there was considerable discussion as to whether the data should be grouped together or a separate analysis should be completed for each aircraft type..ring that tourl ip)U)S No.41 an(d up with more than 175 hours of* lime be repllaed i. All such cages were removed from service. Coordinating with the turbopump manufacturer.. Note that the original three-failure curve is a good approximation of the final plot. 6W.6) as indicated on Figure 6.) 6. The change resulted in cages of lower density. "/ -. the alcohol diffused through the plastic causing it to swell and crack. . there were sufficient spare turbopumips to allow this toIbe accotmplished withbout groajding aircraft.6 FOUR MONTHS LATER - FINAL WEIBULL PLOT Inspection of the turbopumps replaced in service (Number 6. It was found that a process change had been made in the manufacture of the plastic cage to reduce costs. 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. that is. Of the 27. (Alcohol is a d( icing agent added to jet fuel. fuel. Also. Fort unately. this would not have been possible wit hout t he knowledge of the relatively low risk between 0 time and 20W hours.4). This type of crack would usually be discovered during an inspection for oil leaks. the plastic manufacturer. When these lower density cages were subjected to the combination of heat. it was decided that separate analyses should be run. and it was questioned whether each crack was one of a kind or whether they were related events.000 hours of operating flight time. Because of the different missions of the two aircraft. By this time the mesults of the statistically designed factorial experiment were available. This action was effective as there were no more field failures.4).8 INFORMATION AVAILABLE FOR ANALYSIS . . the inference from either curve would be substantially the same. 6.7 EXAMPLE 2: MAIN GEARBOX HOUSING CRACKS The main gearbox housing on some engines develrped cracks in the field.1. a rapid wearout problem. The Weibull analysis is presented in Figure 6. causing the bearing to fail. 6. "F. . Although this slope difference sounds large. four housings were from the new aircraft. the bearing manufacturer. Cracked housings were"beingdiscovered at a rate of 1/20.BJased tin t h)is ii 'ulvsis. This was a ruggedly built gearbox housing. in fact. Once the field was alerted to cracked housings.5 for both Aircraft 1 and Aircraft 2.. and. In addit ion. and alcohol. this identical gearbox was being introduced into a new aircraft. it w'is recnommende1dhl lot P'rJeci lngin .

TOTAL OPERATING TIME (HR) FD 2716"7 igu'l.- . . L2 0.90 7 1.LLL DISTRIBUTION 0 = 4....L. .m.mA.. NO ADDITIONAL FAILURES WERE "o. .61S = 488..0.- - FAILUES = 24 CONF LEVEL = 0. .520 SRI4LE SIZE = 387 -... 1. FAILURES.6. ..1 . * .1 INSPECTION OF PUMPS RETURNED TO THE VENDOR FROM THE ONE TIME INSPECTION PRODUCED 1S MORE PUMPS WITH BEARINGS CONSIDERED TO BE IMMINENT EXPERIENCED fiN SERVICE AFTER THE ONE TIME INSPECTION. . oil S* .. . . f[r Atim.. 0.... ' . EI..epintor I111plp j I' 100 ..I Wriblill Ihol. . .

. S . / . 101 __ _ _ . .4925 ...9 . L .. .u. I el-ol • • -FAILURES . J1OJ. 4. 1.93 . .. CONF LEVEL = 0. . k 14.. 4 out.e... WEIBILL DISTRIBUTION 0= 1 . Main Gearbox Housing Cracks At the time of the analysis. . .9.5... - -.842 S= 711. _ FA ILLURES = 4 CONF LEVEL = 0.736 v= •757.of 213 engines in the field were found with cracks in the housings. 1 fleet. *S.-.9 AIRCRAFT 2 - 03 I-I W z ii ..10M. steeper slope). S.. with Aircraft 2 having failures occurring earlier and at a faster rate (i. ..526 in the Aircraft.... there were 23 cracked housings out of 1. Both curves represented wearout modes.s rEIBLLL DISTRIBUTION - P = 2. TOTAL OPERATING TIME (HR) FO 271878 Figure 6. SAPLE SIZE = 213 . 4.•~~~AMSIZE 15I26 = 23i .100..- . 2 . From Aircraft 2.

the correction factor often comes under question and the easiest way to avoid this argument is to present the data as they are obtained.0 GiG Flee . One way to correct for this type of analysis problem is to correet the data back to a common crack length. Strain gages and vibration pickups were phlaed on gearboxes of each Pircraft and data were obtained. datla cf this niattire indi'ate I hat there may Iih more t han otne mode of faihinre. stresqes which shortened the fatigue life of the gearbox. and the Weibull is distorted by the clustering of events discove..6 for both aircraft through 1982.Webul Fa•lre Distribution/ 2 Weebuis 70 '30 I.. However. .. The run time on the component at the time of the leak or crack is usually not well defined.!c analysis for forecasting future failures was requested. Ui RISK ANALYSIS A ris. -so S40 ~ 40 Aircraft I 20 Aircraft 2 0 80 ~ I 8 Calenar YearwI 82. . leaks and cracks of this type do not usually result in engine Iatihure and are not disco! !red until an inspection is performed.tl( eadditional it c I whic'h Elh 1 Idli .. Cumulative Main Gearbox Housing Cracks 102 -- •-•.- *I S" / \ . - • . FD 256642 Figure 6. . . . . -- - . . . However.. sWp . . This wotuld explain the steeper Weibull slope for Aircraft 2. . Or:linarilIy. . The r-sults of the analysis are prsented in Figure 6.ed at inspection. .27 hrIGIBIMOnth .. lu t lI h dala do iio! Ifall -m a straighlI lite. 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. eslxeviallv in Aircraft I. nIoetd.6. This analysis used methods discussed in Chapter 3. . . This would be especially true if the time between inspections is large but occurring at specific times. is... . • . it can be seen that Aircraft 2 has a lower characteristic lie than Aircraft 1. 100 Assumptieons 90 ..=/ . It was concluded thait Aircraft "2 was subject to more vibratory.

the engine is in the shop at 1150 cycles.2 FOLLOW-UI) ANALYSIS RESIITS 'I Date No.348 1. From a rikk vie-jwpoint this could be considered as safety margin. The answer is based on economic considerations and depends on how close the engine or its modules are to the next scheduled inspection. 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. 'T'his is typical and is discussed extensi'. later 24 moa. The (rackin•g originated in the coverplate of" the gearbox. 'I'his problem was tracked for t1%o years afier the origim:l analysis was complete. later 24 mo.lonal flailtires niaY occur. The coverpiat e was redesigned not Ioi resonate at these f'reqiceniies. For example. adldi.842 2.. of Engines No.f a fix. T'hle driving force was a coverlplate resonance at certn a engine speeds. Although the primary concern is the repair of the engine.0 2131 500 7:12 4 10 13 2. I hus eliminating the prol)!eni. Hofwever. tile Weihbull is stable. the qui. simple one. beflore any action is taken to incorporate an e.7 :1604. inn shop. it obviously 103 -~l - N i I "-m. repair after 1340 cycles of operation and is due for a scheduled inspection at 1350 cycles (one cycle being equal to about 0.l .2. If the engine is in the shop at 500 cycles.782 1.715 7578 3 6552.5 153. if an engine is in the shop for.805 711. If.8 hour of engine flight time). The sieepness in the slope would cause an overprediction of the expected number of cracks.8 Witlh the large numnher of cracks associated with Aircrnt I. TABILE 6. later 1526 1609 1949 23 41 62 1. later Aircraft 2 Original 12 mo.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.gineering change to correct the problem.736 1. however. 'l'hi:. an analysis must also be performed to determine the cost effectiveness of the change. The findinjgs tt the initial analysis considered 23 cracks for Aircraflt and tiutr cracks for Aircraft 2. %on also arises should the engine undergo its next scheduled maintenance while it is availablc it.8 7038. there would be no question that it should be inspected before re-installation in the aircraft. However. "'l'e a-alvsis was repeated several times for each aircraft as more information became availab)le. of Cracks 0(hours) Aircraft I Original 12 mo. 6.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.11 HOW GOOD WERE THE FORECASTS? BI ecatise of the tinme lag fronm |problem definition to the i ncorpo ration . it is not so obvious that the 1350( cycle inspection shou!d be performed. The fix for this problem was a f'airl. Resulls l'thlttle follow-on analyses are presented in T'able 6.10 DETERMINING THE FIX . 6. Aircraft 2's Weiiull has changed considerably..

shotIld not be inspected (coslly part rep)lac'ements are involved). There is, therefore, a breakeven 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 was determined. This provided the Air Force with an economio decision criterion for per•orming scheduled maintenance.

"30) cycles
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
*S.Ppeline
Transporl Mamtenance

Outputl lioutiel

[teen

mantea

j

Parts Unectheduled Engine Remova:. (UE.A

'NSO "Crl O

Sceduiled

~oal

$nw

Removals (SER)

Sa"
Reemm of Conputw Model "Oppofsmamei John Madden AFLCIXRS Maintenmae Engine Snuai'on" Peald by

nt~erwal

FD 256843

Figurr 6.7. Approac3h T'o Optimizing Scheduled Maintenancet

Inpection Interval

I

-Opportunistic Maintenanc Wdteval

I
I
""I

I Unec~bedWe

Engine ReMoval

,.,

TieLkw ------

"

u

" I

" i
FO 25d544

Flight Hours or LCF Cycles

Figure 6.8. Scheduled and Unscheduled Maintenance Interaction

105 "

-

x.

II lns.he.duhl(4 laihtt ui. n evenis. un.chedulled engine renuI.vals (I lols). 2)

i

1111 -nj.as Weihll.s.

are 4f,1"lour basic

tvp•.s:

I)

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

I I ,~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 12th lstage 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) (2) (3) (4) 6.18 How will the problem affect engines? What can be done to fix the problem? Can the problem be detected through inspection? What recommendations should be made to the Air Force?

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

1 m TOTAL OPERATING TIME (HR) S~FD SI~i•'u ' t.q.( 'na rakn Sto.- . . .._ SAMR-E SIZE Ii. . . I0 12t1h Vaine and Cas•e C'ractimA. t" u W.117 90.V 1.- 2. a20 I1z "2 '1U % I.jdoo I 2..9. I 2.. 4. 6.7 t. r .2S . 4 ._ FRILLRES 7 = - 1338 - 70. SO.- -7..- .6 .0 11. S. J.7 r * °. . 2787 li41E .8S. 3.1O . 4.T.O . 271879 i ~108• .5 6.- .~ ""EIBtILL DISTRIBUTION A = 4.

80•

701350 70

so~
TAC'$ Tram of Fwill

.50

40

4

130

j20
to 0 1901981 1982 Calena Yew 1983 1984

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
,4

Often it is of interest to determine how far from the "true" value an estimate of 4 or 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 nexp

!

•x/ 1.05Z,,/2,' -- 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
4,12

Confidence Level 99% 95%

90%

2.576 1.960

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)) 988,x-84r or, 833.7 ýS i-1102.9 " <1.05(l.645)) 958.88 exp 1.840(.5

t

111

"_.s WE
0

ILLL 01

I

-

um..

= 1.84 S= 958.88 -. ,FAILLRES SPORE SIZE=45 =45

-

-4S

-

I

I'iv

II III

..

3 11

i 1.o. 2

1610M

TOTAL OPERATING CYCLES
FO 271O80

Figure 7.1.

Weibu-!
. a

Test Ce
...

im.

.

F7
112
YCE

.gaa

;

,"-

TORLOERTIN

P0 27188

/

igu

7...bulT/tCa ,

112i

) the where ti. a 90"f.3d) Again using Figure 7.05) isR 1 5".579 + (1.663 7..890)) -Therefore.6 = -0.05) and F 1 (0.168 + ( -0. the-confidence interval is 0. Cis the probability of a life of ait.1.579) 2(.4).a 2.03 :1.1) 1 1/2 u + Z. t~~.88)) 1.579 - 2.0) - 0. 113 .1 2 I1Var(u)I 4.(0.05 and ti.4 CONFIDENCE INTERVALS FOR RELIABILITY Anot her p~roblem t hat appears in Weibull analyses it.3 (7.645) NTOiO.268) ) :s R (700) •5 exp (-exp(--0. 0. while F 3 (0.10 (t2) u J.579)J - O. t hat of obtaining confidence intervals for the reliability ait a given point in . AA 1.84 = -0. Tlhe reliab~ility at the point.* 7.268 U2=-0. rank associated with the it 1' failure. from Table 7.05) ]i' tI . Again assume a large sample with no censored oliservat ionls.579 (1. Then the confidence interval is: exp (-exp (U 2)) :5 R(tV) -•exp (-exp (ul)) Example 7.5 CONFIDENCE INTERVALS ABOUT A FAILURE TIME Engineers are often interested in a confidence interval for the time associated with a given failure.3b) (7. ul -0.95)]i (7A. Compute u.465 :5 R(700) :s 0.19131 7. The step-by-step procedure follows: A.890 -0. = u U.1imie. confidence interval for the reliability at 700 cycles is desired.u (n70 Qn(9158. Var (^u) = 1.. The procedure is as follows: 1. 00 [n1- F1(0. Ninety percent confidence intervals will be assumed for all confidence intervals about the Weibull line in this section. 0.' units.(7..1.Compute u = (Qn(t) = - 2 n(4n))# -0.645) V05 _06 4.95) are in Appendix B.168 + 1.95 are the failure times associated with the ith failure and F 1 (0.95) is the 95 *'rank associated with the i t' failure. least . This confidence interval can be approximated by equation (7. Tables for Fj (0.2 IVar(uN/ (ZO.5 Q - F. __ = exp (- exp (-0.3c) Z. and will he dlenoted bky [Wt). Compute Var (u) :3.1913 (-0.

J'5l1l.0. 0.005[i' = 1 10." Srinivasan and Kanotsky. 0.05 0. Biometrika.fws: FI (0. prwd'ced Iroumt 10 2.Example 7.o .258 (from Appendix B) = t1 . if he .95= 100 [Qn 10258"] = 10010. I hotirs) n hire fior In' hat imi Iaitilires.4 the confidence bands about a single reliability were calculated.(2) contains the basic information for their construction. 197Z 114 S. 1'2"An Approah to the Cmnstruction of Parametric Confidence Bands on Cumulative ritribution Functions. The ah'cidall foI. 4 " /.5) and F(x) is the estimate obtained by substituting maximum likelih!od estimates for the parameters.95) 0.3. 1 - e (7. 7'.63 hours 7. 59.298411/2 54.00-5) (froti Appendix 1) (IM) [n I 0. Equation (7.F(x) + K(n)).fideim'e iilherval a itiu I lie fiml f'ailihre (21. 2 lel. Simultaneous confidence bands can also be placed on the Weibull distribution fir complete samples.4 Il(H himirs.nce.05) 1[..2 can be used to cah'ulate 90".08 hours F1 (0. (F(x) where F(x) = K(n). confidence bands about. The resuls in Reference( ) have been extended t) the Weibull Distribution.i/2 7.2.. Vol. -<• -. the Weibull Distribution.6 CONFIDENCE BANDS ON THE WEIBULL LINE In Section 7. it S~)Ispvm we are given Ihe Weieilml int Figure 7.5) together with Table 7.

LSGO0* *..0) and .._____ w LI. 3.: • / I..0 S= 100 HOURS "SAMPLE SIZE = 10 FAILURES =10 - / - s-o __ i .wEIBULL DISTRIBUTION -. .4~S 0910. . ..q - 1(M) fir 10 Failures 115 . P = 2. .• 2. w w ____• ° sI.... --.... ... : : "J. Weihil/ IPlot Where( # : 2.. . TOTRLL OPERATINg...- O.. TIME (I-RS) FO 271881 Figure 7.. 2.. .2.s~~oo . 5: . . .30 2...

The values for "C" are found to be about 0 and 6.125 0.. suppose it is necessary to make the statement: the expected number of occurrences is 3.0 and read up to the point wi.p) .ax.. Therefore.4. x = 0. From Table 7.338 0..2.0 respect ively. These Imnds are illustrated in Figure 7.237 0.. _ S.213 0.307. Therefore.222 13 14 15 0.9. confidence.307 0.169 0.05 and 43.269 0.152 0.7 WEIBULL "THORNDIKE" CHARTS A graphical method often used to determine the cumulative probabilities of the Poisson distribution was named for F.) As an illustration.axis at 3. with probabiity 0.95 lines extending from the y . 1(x/2°°)'".197 20 25 30 35 40 0.'s.3.119 45 50 75 0.0 and qp = 2000 hours.7" .074 Kxample 7.420 5 6 7 8 9 o10 0.380 0.-.0. ).0 intersectE the 0./x !.106 0.0.204 0. and the actuai number of occurrences will be between x and y with 0.j. 7.284 0. :1.307 for all x. A random variable x has a Poisson distribution with a parameter 1W if P(X = x) = exp (.246 11 12 0.117 0. 1. if the expected numbpr from a Poisson distribution is :1.3. less than 6.! binomial when the sami 1 le size becomes large.5 Consider the Weibull in Figure 7. use the Tiorndike chart in Figure 7.. 0 _ x < ou. To find x and v.086 1oo 0.141 0. (The Poisson distribution also arises as the limiting form of th.TABILE. .4 enter the x . 116 / .540 0. with a sample size of 7.0 will occur. 0 = 2. with 90". Thorndike and is illustrated in Figure 7. 2. COI )N ID'El)IN(CE' BOUIINI)D ON TIl'II WII llIJi1 LINE K(n) Sample Size (n) 3 4 0.2 the critical value of KM(7) = 0.e 3..307 _sF(x) _< 1 -e-(XM20 + 0.90 probability.0. 7.

7.0 = 2000. -- - CIA 3.1. SA4MPLE SII7Z = 7 FAILURES =7 S.10000._ . 4 . 100. TOTAL OPERATING TIME (HR) FD 271882 Figure 7. 3. 2.. . .8.3. 2. 4.10 ..- W-EI•_LL DISTRIBUTION 0 =2. Example of Confidence Bands on a Weibuli Line 117 .

0001 S~0.3 0.2 0.8 0../ 0999C=6 7 8 9 10 0.3 25 S• S0.6 Vau 0.4 0.5000. Cumulative Sums of Poisson (Thorndike Chart) 118 I' A 1+ ..1 0. ~0.1 0.2 - 0.07: 0. -C-0i S0.8 2f o: 3uPie Value of Mu Prime FO 2W565 Figure 7.5 0.40.601ii! 0.01 - 0.99 C.4. I5 0 0.001 0.

The usefulness or this information can arise. it can be stated that the probability of observing six failures by this time in the redesigned rotor bearing is from 0. How often can this process continue? The Weibull Thorndike charts answer this question.5. E'xample 7.08. enter the x . is replaced again.9.0 and proceed to find the "C" values where 0. The menager wishes to boe 90'. The current rotor bearing has = 1.axis intersects the "C".ig "C or fewer failuires by for Iiw( Weihull 411 I o II fi-ailsaro.0 . 2. no more than two spare parts are needed per delivered part. for example..0 sad proceeding to the point where 0.lines. Six failures. Therefore. is replaced (made "good as new").:1.12 are. so using Figure 7.90 probability band can be placed abomut the number of failures occurring by ti 7 :.0. calculate thy (q~ is the characteristic~ life from the Weihull).5 1hrotigh 7.10. the probability of having three or fewer filtires per unit by 60MN hours is to be calculated.EwCUrring b~y how Itihenls' lIIItlIs ai~.8 In spare parts provisioning. the p~robability of having six or more failures is approximately 1. 11giv4en Iiime. 2. E'xample 7. ent -r the x . confident that ho will not run out of parts. 1.axis at till 2.0. suppose the number of spare parts to be provided are required for a part having a #4 3. Weibiall Thornd'ike charts for )i's of04.45. Several examples of this technique and other uses follow.0 (Figure 7. It this case. and then read the probability of having "C" or fewer faiilures. IS 41 The1y v Ilt aI-.0(A at d 0.10.000/4.9317 hours.9:3.0 and an inspection time/characteristic life ratio it/q) = 2.0I 1. 1Ks.5.7). A 0.0.00M hours.08 to 0. Is this unusual? With a t:7 ratio =22. Using Figure 7.7 Suppose 4-1. Hence the redesigned rotor bearing is not as good as the current bearing. have been observed in the test of the new desigr due to this mode. and 41.95 probabilities intersect.7.5 and now t/q .0.i IN1it.6 (Given Weibull parameters 4 = 1.III I his sionn waY a graphical lechnipieit sioniiitr so the Tlhorndik&e chart has been developed isriomi olt Thiwsi charts gieI*thw lirshabililIy of*Ii. Exam ple 7... Entering the Weibull Thornidike chart for 4= 2. 1.92 0.0.9 on the y . enter the x .5 and qj= 3000 hours.7.axis as 0.1.axis at 2. Figuires 7. when the inspection interval is two or more times the characterist Iic life of the Weibull failure mode of a part.0.. :'.u Ii~ III?V-ivenl I111Wv.8) the probability of having six or more failures is approximately ItC . etc.0. This yields I and 5.5 (Figure 7. determine the time Mt of interest (liossibl) the inspect ion time).01.0) and proceed uip to the point where the line -C z-3" is intersected. respectively.2.axis of the chart with the closest 0 to the one of* in terest. the part fail.0.5.. When this happens. E'xample 7.5. fails again.7 and q = 4.37 =4. a limiting failure mode whose # 119 p0~ .axis at 3.9 A new design rotor bearing has been testEd for 22. 3 Again using Figure 7. entering the x .0.90 =0. arid 5. entering the Weibull Thornidike chart for 4 = 1. To Ilse t hese chart. Then the probability of observing:1 or fewer failures can be read frobn the y . t/j Ohi4NN)/:N --.

coo 40 C4 csv 120 .

0 0 09 jo0.C 41 In U) I 04 Im 0) 0. oq~ 4.jw 121 .

)3 jo qoic - 122 .LO CC 77j C! 040 I-D 41 V) SS91 Jo 0 jo souejjfl-.

coo 0 InZ 1123 .

.a 0 0 U. C'. V.oueJJnooo JO 0Jd 0 124 . C R .0 . L o 0 Cl a '4 a.0 0 . I-0In40C'. - a SW1JO £0 . C 0 0 In a '0 o * E.) Ia C) R. 0 In C) 0 U In C) S C. Y V.

1/2 0 C. ool~ oqj 12 .I 8 na o0..

(IN ql ORp'-pU 0 swl m~mo o 0jo o qpdwu 0 ItciC i 126 .

ae cmI cs 0 cio dI owl o joqo0 0 o suwmmo 127. .

(possibly as many ats 10014N)) are not. 5. if* the pairt f'ailure times are knowni. 1300. q 1400.1:3). United Technologies Corporation. This is an estimate of the "true" Weibull distribution.531. and 1493 cycles iri a population of 2500. Calculate the mean time toI failure OXMTl'R mT'rF (now refer t(I Figure 7. Kit imat e the Weibull parameters 11and q. 2. The proce(Iur~ to estimate the "true" Weibull distribution can be used: Sleps land 2: St p3:: MTTF 116. however. while the others. and draw a horizontal line from this point.14. using the methods in Chapter 2. oiiten iii donle if the I inles onl each finiluireare known hut the timies on Ilite . failures + No. Suppose there have been four flange failures with times of 1165.8 SHIFTING WEIB. utp to thlis pointI Weiln IlI nilu tre (list rihu t ions have been est inlated coi~mbint 4ionWit Ill I hel IN ipiun I io ns of'1 tilfi led IInits. 01f course. P'lot the failure data oil Weibull probability paper. W hat call he In im oblservedl ftiu lures. :1. In many engines only the most important 4(M-r-A() parts are serialized.. (Calculate the proportion failed in the toItal population. suspensions)..5 + 1300 + 1393 + 1493 MTTF ~4 128 137 13. Commercial Products Division.7 cycles . failures/(No. Fred IDauser.spiiut iiiof Wuntiailed units are unknown? Tlhis p~roblem arises in the failure analysis of data from jet engines because all parts irn an engine are nilt serialized. failures(75 D)raw a vertical line through the MTTF.7. calculate the cumulative ".failed point = (1 e-Ill't")x 100. No. times to failure for each part No.ULLS In nIlI IlInht has IlKeeii don(. 6I. the engineer call generate at Weihull distribution from the failures only. See*Figure 7. that is. the times on the unfailed units are unknowvn. 4. At the intersection of the vertical and horizontal lines draw a line parellelto the failure distribution. Ani out line of the met hod is as follows: 1. Statistician. developed a met hod t0 "adjust" this Weibull if the number of untailed units in the lplpulation is known. 9. P~ratt & Whitney Aircraft Group. the time (In the individual p~arts cannlot be tracked. 1393.

Vol.slions) Th'iverf. I'iilurvs/(No.53 (same as the four failure Weibull). s.)! Given a sample of seven failures with 4 2. suppose two plots like those on Figure 7. "t "Discriminat:on 15. 129 /M .ppi. complete samples will be assumed. IExample 7." Dumonceaux and Antle.No. 0.•re." the log-normal or the Weibul. For example. The statistical test from Reference ) can be used to discriminate between these two failure distributions. faildures I No. 4= 9.18 are given. between the log-normal and Weibull Distributions.4.29 and 4 = 1500 be rejected? Plotting the hypothesized Weibull as a dotted line on Figure 7.I/2500)) I. Procedure: 1. If this dotted line does not.1 • 0 (.•u. Since the log-normal is the most frequent alternative in failure analysis.15 The estimated distribution has a is now = 2629 cycles. then consider the sample to be from a different Weibull distribution.00)1. I61 0ti) t0.6. The test can be set up in two ways: to favor the log-normal or to favor the Weibull. as in Section 7.6. the eye is unable to t3 discern which is "best.haracteristiclife 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. One would have to reject the hypothesis that this sample comes from a Weibull distribution with 4 = 4.16. this will cover most of the practical cases that arise.3 gives Figure 7. can the hypothesis that the sample comes from a Weibull distribution with 4 = 4. and extremevalue distributions are plotted on Weibull probability paper. 5 and I. In this case. 17 = 2000. Again.1973. normal. Now place the hypothesized Weibull on this same plot.0.29 and i1 1500 since the dotted line does not lie entirely within the bands.6. 7. :1. Using the Weibull estimates of il and q for the failure distributions.10 COMPARING THE WEIBULL TO OTHER DISTRIBUTIONS Figure 7. 2.I. hailed II v. but the ".Wll. Technometrics. lie entirely within the confidence bands.. S leps . calculate and plot the confidence bounds using the techniques of Section 7. as a dotted line. cuml '.: See Figure 7.17 shows what happens if failure data from log-normal.

. + L -- . Example of Shifting a Weibull 130 . -..137. . ONLY ESTIMATED -TRUE. ...W . .arl 0 S M.1. le ''o..." EF TOTL OERAINGTIME (HR) Figure 7. - .. a. - : - :FAILURES -WEIBULL OF . IWEIBULL DISTRUBTION 8X IA MTI .

. .SI OO .•. . .. 4. . UO.-r- DISTRIBUTION --JIE•IBLLL *L"-: =4 1400.__WW -. I. I.= MLE SIZE4 FAILURES=4 4- t .1 - - - . 8 9. 4.10.. ! .7 SP. 14. 40.14. LJ LL_ .I -_I__ Ia. Flange Cracking 131 .. . TOTRL OPERATING CYCLES FD 271884 Figure 7.=' .4: .. -- - - J UI 00.__ _ -_ -.

_ EIBLLL DISTRIBUTION Pg.. I. '..IIll ESTIMATED DISTRIBUTION .7 -.IN CYCLES FD 271S86 Figure 7. e. TOTAL OPERAT.$3 9 .100m.2629- SAMPLE SIZE=2SO0 FAILURES=4 MTTF w. - - QI. Flange Ciacking Estimated Distribution 132 J I I /• .8.9 :S3 0 = 9!3 1400 ...5...1.= ". SAMPLE SIZE= 4 FAILURES= 4 = S I H z a .=9.

S' I I's.16.HYPOTHESIZED HEIBULL -(P=4. m. - ... is. NOTE:CA - CONFIDENCE BANDS CRN BE VISUALLY "TRIMMED" AFTER PLOTTING Iv 01.29. -.t-1 4 11 O. Hypothesized Weibull 133 . . .0. S - .L * * * ~TOTAL OPERATING CYCES .nU. _. . JOD FO 271W87 4:m im Figure 7. 1 = 2000) 000 -- -.4. T = 1500) ESTIMATED WEIBULL FROM SAMPLE (D= 2.

Picking the Best Distribution 134 .17.True Distribution Appearance on Weibull Paper Normal Cun. % VI Time Log-Normal Cum % Time Extreme-Value (Largest extreme) Cum % Time FO 2S6466 Figure 7.

109) STD DEV =0.9 --0 COVERPLATE w S 50.373093 SAMPLE SIZE = 20 20 w w 0.999 0 W M 99.250 (4104.0 U SFAILURES LOG-NORMAL DISTRIBUTION MEAN=8.2 NEIBULL DISTRIBUTION '= 4523.99.0 - z a- cc -j 1.000 99.0 0.0 cc -j D 100 1.000 100. Comparing Weibull and Lognormal D)istributions for Coverplate Failures 135 .000 FD 271888 TOTAL OPERATING CYCLES Figure 7 18.000 10.0 U S w 1.001 " 100 1.000 10.8 SAMPLE SIZE = 20 20 ILURES 10.0 -- 363.

idence Level 90% 1.998 95 % 1. 6556. 2813.f At . CRITICAL VALUES FOR TESTING THE DIFFERENCE BETWEEN LOG NORMAL AND WEIBULL (FAVORING THE LOG-NORMAL) 80% 1.007 0. 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. if W "Wtabk the Weihull is rejected in favor of the log-normal. 2569.4. and 70(X) cycles.5178.028 1.Wcilmll.976 Con.7) and compare to the appropriate table value for the confidence level desired (Table 7. n is sample size p• -WnW.4) and.. 27. 4ah'ualdat.ll:'I.984 0. t. 5716.26.026 1. 3294.016. 3283. . TABLE 7. l'_1Il) ] -.020 1.. 2979. 2160. 3916.038 1.041 1. 4462.(2wet)+' 112 f(lllt .041 1. W .) l l tf(t -2) l .067 1. the log-normal is rejected in favor of the Weibull. calculate W (2 r e a) z " ( t . if W 2 Wi. 6378. It. 7 .Favorhig tI it.8. 5984.. • = 4523.974 Confidence Level 90% 1.005 0.991 0. 3853. 136 .980 0.993 0. CRITICAL VALU•S FOR TESTING THE DIFFERENCE BETWEEN LOG NORMAL AND WEIBULL (FAVORING THE WEIBULL) 80% 1.016 Nimuber of Failures 20 30 40 5') Favoring Ihe log-normal. 3503.015 0. f )(0t.. TABLE 7.18 occurred at 1989.008 0. The estimated log.3.8.995 95% 1.082 1. The estimated Weibull parameters are = 3.3) and.014 Number of Failures 20 30 40 s0 Example 7. 4294...12 The coverplate failures that went into the plots in Figure 7.)I I .019 1.044 1.

the log-normal failure distribution seems to describe this coverplate failure mode better than the Weibil distribution. . Adjust this 10 failure Weibull for the entire sample size. Solutions to these problems are in Appendix J.6). the Weibull can be rejected in favor of the log-normal. In conclusion. 232.0.141592) (2.3730932]+1/2 -. since the value W in the test that favors the Weibull a Wble.141592) (2. 4. 230. the log-normal cannot be rejected in favor of the Weibull.'4 with SO'.373093 1-1 :U-= 1. Are the Weibulls in Figures 7.1000 hours. confidence interval for Reliability at 1500 hours in problem 1? What are the 90'. 327. in this case. This same decision is reached in the test favoring the log-normal.015. Therefore. using equation (7. Favoring the log-normal. and 451 hours.. V = 2000 hours. confidence intervals about th3 first three failures in problem 1? Given a Weibull with parameters 0 = 1. V . what is the 90% probability band on the number of failures to be expected by 4000 hours? A 10 point Weibull of failures only was generated and is illustrated in Figure 7. usirg equation (7. 164. 137 .373093. A p 8.039 2 (11989 f (1989)1 . Note: failure times are 51. confidence intervals for # and i? = 1. 116. [700 f (7000)1 ) +120 compared to a table value of 1. 79.71828) 0. b. 2 W .20 and 7. i! FFavoring the Weilull.7).. W [2(3.008. what What is the 90'. PROBLEMS (iven a Weibull derived from 40 data points with / are the 90'. 6. 0.250 11..962 - (11989 f (1989)1 . These failures are of a non-serialized part with a total population size of 2000... Which fiailtre (list riI it ion fits thelata I Ixecr Mol h tests. reject the Weibull.5.71828) 0. S! :1. Hence. 197. cnfiheince? normal Ipratieters Atre. 7.19. 2. 17(X)0 f (700)o]1 -1)20 comlaretl to a table value of 1.11 I.21 significantly different? Assume the "Weibull in Figure 7.12(3.20 is true. 414. will be performed. since W s Wtable. favoring first the Weihull and then the log-normal.

- ... 11 At 0If....59 2S8 _ . -. 11..19.---- - . ._ HEIBULL DISTRIBUTION 0 = =1... 5 138 .. FL SIZE = 10 FRILLUES ='10 U. Problem No.4%..... . •- I S.serialized Farts. - - - - TOTAL OPERATING TIME (HR) FD 271890 Figure 7.. . Weibull of Non.. . - --- SM .

. True WeibulU. -i M L-1 4- . 6 F 71 139 .4 L Figure 7. FAILLRECZ=10 --. -.' W EIBLLLDISTRIBUTION SAMPLE SIZE = 10 --- -- N.20. Problem No. -. - ------- MTTF 0.

. .u A 7 2. *. Problem No.L .. Suspect Failure. . . 6g4OH TOTAL OPERATING TIME (HR) FD 271893 Figure 7. . . 6 140 I 140 . .1.3L•IL DISTRI1UTION 0 =2.*.21. Mode.97 U. .4 SIZE = 10 1 -FAI tiIl iiII I MII Ga-----44 an. . 6. .SSHPL t=895.

63. - 10. 11. A failure mode that is independent of time. 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. Censored data Confidence Data that contain suspended units. 5. usually computer-adaipted. The characteristic life of the Weibull distribution. Mean or average time between failures. whose outcome depends on the application of randomly generated numbers. Relative frequency that the (statistically derived) interval contains the true value being estimated. i. the hazard rate remains constant with age. A failure mode characterized by a hazard rate that decreases with age. 8. The degree of agrtEment among estimates calculated from random samples drawn from a parent population. 11"Beta) The parvieher of thle Weilhill distribution that determines itS sphae and Ihalt implies the failun'r nmnle clutra'terislic (infant mortalily. In other words.e. Random (failure mode) 141 / . The precision is usually measured by the standard deviation of the estimates. It is alsoc ealled tihe slope parameter because it is estimated by the slope of the straight line on Weibull probability paper.. new units are more likely to fail than old units. P 1 (Eta) - 7. A mathematical function giving the cumulative probability that a random quantity (e. An unknown constant associated with a population (such as the characteristic liWe of a Weibull population or the mean of a normal population). a component's life) will be less than or equal to any given value. in the sense that an old unit is as likely to fail as a new unit. regardless of the value of tf. the W-ibull slope parameter. - 12. Precision - 13. A mathematical model of a system with random elements.. 4. Hazard Rate Infant Mortality Monte Carlo Simulation MTBF Parameter - 9.2("% of the lifetimes will be less than the characteristic life. or wearout). - 3. Distribution - 6. random. Also called fixed error. 2.g. The instantaneous failure rate.I APPENDIX A GLOSSARY 1.

old units are more likely to fail than new units. Wei' ill Analysis. Weibull Plot 142 14744N . and predict the occurrences of future failures. . 7-pro age for the failure mode. Risk Analysis Suspension A prediction f ithe oumimber of fI'miures exiiected Iti otcur in some future tine period. Itt'liallilitv 'Tlhe pr bjlumilily thai. determine failure mode characteristics. that has not. - A test or operational unit.14. whlo olkrating in the manve*r inleldecd.. 15. zero probability of failure prior to t . in order to estimate reliability. 17. Procedure for finding the Weibull distribution that best describes a sample of unit lifetimes. Weibayes/Weibest 20. but no failures have occurred). 21. when operating time has accumulated. Wearout A failure mode characterized by a hazard rate that increases with age. 19. the Weibull slope parameter. It is known as the minimum life paramete7 in the three-parameter Weibull distribution: units have. t- 18. A plot of time-to-failure data on Weibull probability paper. a syslemn will iero4rnm its inlended funi'tion wit isl'actoirily IFor a specif ied interval of1"ime. A method for constructing a Weibull distribution based on assuming a value of #. failed by the mode under consideration at the time of the life data analysis. 16. It is used wht-n there are certain deficiencies in the data (for instance. ie.

3 Sample Size Rank Order i 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 11 6.1 14.6 39.3 96.2 72.4 57.4 94.6 75.1 50.1 60.4 50.4 50.5 76.5 79.5 32.0 62.0 22.6 13.8 36.5 34.8 84.1 83. MEDIAN RANKS Sample Size Rank Order 1 2 3 4 5 6 7 8 9 10 1 50.4 20 3.7 58.6 87.5 90.4 8.j 95.9 8.4 28.3 22.4 17. AND 95% RANKS TABLE B.4 36.8 96.5 8.8 64.'t 38.9 38.7 18.6 50.1 14.5. 5% RANKS.0 55.9 36.APPENDIX B MECIAN RANKS.0 a8.5 89.6 25.0 92.9 86.9 26.8 81.3 88.5 67.4 90.9 72.1 70.0 79.0 81.3 13 5.5 20.5 45.6 16.3 20.3 96.4 74.? 41.6 32.5 63.8 35.7 95.1 63.1 54.6 44.6 76.0 55.8 32.1 15 4.4 3.9 87.5 89.I.2 52.6 15.8 47.4 44.4 2L.6 89.4 60.5 53.2 25.7 40.2 67.2 13.2 10.4 16 4.9 43.0 74.9 67.1 29.0 25.3 86.2 .5 10 6.3 41.0 27.3 50.0 68.7 37.1 26.8 11.3 62.b 45.7 61.5 61.- .6 13.7 3 20.5 10.1 90.2 93.0 57.5 50.0 82.3 81.2 73.9 5311 59.2 16.4 85.2 70.4 67.7 %3.9 9.1 18.9 54.4 22.7 9.5 8 8.5 70.9 79.2 78.8 14 4.6 71.5 35.6 47.2 19 3.4 74.8 73.0 7 9.0 60.9 32.9 28.6 29.0 5 12.4 42 1 57.3 50.4 21.8 18.8 50.5 39.1 32.1 65.4 39.3 4 15.4 30.0 18 3.9 24.7 17 3.5 77.5 21.0 44.4 84.8 91.5 52.7 9 7.8 67.7 96.1 12.0 55.0 63.0 78.9 31.3 34.8 91.9 17.1 91.9 27.1 77.7 37.7 42.3 82.0 86.0 6 10.0 69.2 50.5 143 auw.8 12 5.0 56.4 77.5 46.0 2 29.5.9 30.6 69.7 93.5 20.3 65.5 83.4 74.2 71.8 46.4 64.0 42.

8 23 7.3 84.9 86.7 22 3.4 96.3 5.8 77.8 91.6 10.2 97.5 32.3 39.9 36.3 26.0 24 2.3 18.7 61.8 93.9 58.4 75.1 9.8 60.5 94.7 35.7 43.8 10.9 52.6 81.9 17.0 48.2 86.4 2 2.7 55.8 94.6 46.7 6.7 4.0 54.2 63.8 65.7 87.5 62.2 82.1 60.0 90.1 25 2.1 42.1 32.0 81.8 45.7 48.3 38.8 84.9 16.0 53.3 64.3 50.7 9.6 68.9 16.3 97.5 32.5 87.7 80.0 92.6 41.5 64.9 75.5 44.7 21.1 29.7 40.8 34.7 78.2 47.2 93.2 61.2 34.5 14.5 17.I.6 78.2 56.3 51.3 65.9 15.2 35.7 415.6 3) 2.4 12.4 5.4 93.4 72.5 34.4 63.3 48.2 5.6 84.4 20.9 57.0 83.0 20.0 -87.9 7.9 89.2 58.3 60.5 30.4 92.9 40.1 71.2 80.1 71.8 61.3 27 2.2 71.1 11./ TABLE B.6 57.6 79.0 93.9 25.0 70.3 10.1 44.0 50.3 68.6 59.3 30.9 76.3 29.4 18.1 38.7 69.3.2 51.1 37.0 .9 36.0 5.4 56.8 97.9 19.3 50.9 64.3.1 46.1 41.0 68.4 90.5 6.5 76.5 11.3 39.0 97.2 28.3.5 90.2 28.2 7.0 56.5 25.7 73.5 92.8 74.6 87.1 12.9 25.1 23.9 2.1 13.9 84.9 9.4 41.3 29.7 52.3 28.8 47.1 51.5 35.9 83.4 20.4 73.8 68.0 5.0 42.1 94.2 27.0 31.7 73.0 19.7 13.4 72.8 97.2 67.4 74.6 59.8 55.1 82.4 97.0 70.5 88.8 97.7 79.5 15.8 F9.4 15.7 24. MEDIAN RANKS Sample Size Rank Order 1 2 3 4 5 6 7 8 9 "10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 2S 27 29 29 30 21 3.8 43.6 6.7 20.6 54.5 27.7 7144 .8 12.2 26 2.8 39.7 26.1 15.0 23.2 88.4 17.0 54.8 62.4 45.2 21.30.1 74.5 89.5 2.4 31.2 58.8 6.0 24.1 96.0 89.8 35.1 7.3 22.6 67.5 8.1 46.5 85.4 22.8 26.1 65.5 50.2 64.6 79.2 38.8 12.7 50.8 67.6 21.8 70.1 97.8 37.4 79.5 31.4 82.4 78.6 77.

2 1 S~145 / .9 31.9 88.9 67.7 77.7 56.0 18.0 47.2 91.4 54.8 74.21 95.7 64.8 70.4 90.3 65. 98.3 8.2 34.4 17.6 32.I.6 97.8 90.9 61.3 20.9 83.6 92.3 52.1 29.2 95.3 39 40 96.8 87.9 71.4 77.7 51.4 78.3 10.2 41.7 65.8 98.7 71.4 95.3 79.6 7.3 13.2 81.5 31.2 95.4 28.4 22.0 26.1 89.0 17.6 60.4 90.6 85.2 39.6 41.6 84.6 48.2 45.1 76.8 42.4 50.0 88.2 24.2.7 60.3 14.5 15.4 27.0 50.1 63.3 50.9 9.8 46.4 7.7 94.3 38.1 33.0 52.8 72.4 61.1 68.0 78.3 44.3 59.7 7.9 94.0 38.0 57.5 11.5 14.8 75.9 86.2 40.4 33.0 35.9 85.0 35.8 20.1 5.2 25.1 62.9 67.7 24.3 92.8 21.9 36.2 53.7 45.3 47.3 48.1 67.5 59.9 34 2.4 51.3 43.8 97.1 66.5 19.7 4.8 29.0 52.3 66..2 60.5 71.4 43.5 12.1 45.1 11.9 85.1 6.6 16.8 38.1 61.0 24.5 98.2 16.0 21.5 44.5 88.4 74.8 39.9 76.3 53.9 56.9 4.1 42.4 85.0 60.7 4.9 88.9 79.5 23.9 55.6 57.9 73.1 17.5 76.5 20.6 73.0 68.0 53.0 22.9 69.0 14.9 26.3 98.9 69.6 58.2 97.4 82.5 51.1 83.8 42.S 43.3 21.8 93.1 74.1 72.1 9 8 12.6 47.2 33.8 4.6 9.9 44.8 15.1 23.9 64.3 58.3 43.6 96.0 36 1.2 11.3 87.6 40.2: 6.0 48.6 13.9 35 1.2 67.6 39.3 54.0 5.0 37.9 4.9 18.5 80.5 83.6 29.8 37.8 48.0 73.6 91.0 98.0 68.5 72.3 30.4 94.0 35.2 5.9 91.9 80.6 18.8 93.5 10.4 23.1 84.1 39 1.0 85.0 37 1.1 97.4 85.2 28.1 12.6 51.6 62.4 33.4 22.0 29.2 48.8 41.TABLE B.0 32.5 19.4 46.8 59.7 32 2.7 10.3 57.9 47.0 39.3 89.0 11.1 66.5 55.1 31.9 58.5 78.5 27.2 40.9 82.5 62.0 73.8 28.8 26.8 14.8 34.7 37.9 36.0 8.3 75.8 55.5 27.4 93.6 55.9 97.9 32.1 14.9 19.4 16.1 50.0 4.7 79.8 50.1 34.0 70.7 - 40 1.7 70.7 9.8 63.8 95.9 36.1 63.0 16.5 54.1 76.1 56.6 80.8 33 2.7 17.1 58.1 38 1.9 70.6 46.1 8.6 61.0 21.8 4.3 76.0 41.8 7.3 66.0 96.0 52.6 252 27.3 11.8 30.0 33.3 6.3 72.1 78.8 36.4 86.6 30.8 31.6 75.7 65.1 56.3 25.0 23.7 14.0 44.5 83.5 88.9 82.3 32.2 79.3 82.7 81.7 63.6 66.6 38.3 64.7 77.6 237 26.1 92.1 90.7 89.0 20.6 51. MEDIAN RANKS Sample Size S" Rank Order 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 h4 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 31 2.8 92.9 69.

2 61.6 98.6 11.1 47.1 37. 5.6 7.9 6.0 49.3 25.5 985 1.8 24.3 17.1 62.4 22.5 93.9 3-5.0 47.6 71.8 83.8 ' 73.8 41.9 58.6 79.4 6.1 35.5 12.4 7.1 32.1 8.5 17.9 23.4 94.3 59.7 5.8 84.1 20.3 12.0 64.8 43.3 28.6 86.8 60.3 79.7 29.6 75.0 84.0 52.4 67..5 3.5 33.8 12.6 50.4 42.9 45.3 15.9 64.7 13.0 85.8 25.3 67.0 58.4 91.2 96.5 965 1.0 72.5 66.9 33.0 92.4 1.7 73.8 77.4 :.0 16.4 86.4 69.8 39.1 79.6 91.8 37.6 56.7 94.0 53.7 68.8 6.2 59.1 55.4 11.1 71.1 80.5 64.1.8 36.6 40.9 19.3 14.2 90.9 96.0 10.6 5.5 92.1 44.2 15.5 696A 71.6 66.7 54.4 1.6 27.9 47.0 35.8 20.4 8.1 26.1 31.14 .7 31.6 80.3 13.5 3.5 94.9 89.0 46.P 92.7 6.8 57.8 17.8 66.2 76.1 74.5 29.3 37.0 9.7 15.1 66.4 13.3 23.5 39.6 62.17 .8 76.9 89.0 75.9 82.2 52.8 72.4 9.19 s50 1.5 47.9 87.3 41.5 7.1 57.6 46.5 30.8 68.35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 41 42 1.3 54.2 88.0 56.4 3.3 84.7 10.7 58.3 80.141t47H .4 73.7 44.8 42.1 35.3 47.3 1.1.3 20.1 25.7' 90.8 51.1 43. 39.6 3.1 37.0 81.4 56.8 61.8 72.7 44.5 19.1 64.1 18.4 70.7 13.9 62.2 13.6 48.2 88.1 24.4 37.5 .5 50.6 20.4 23.7 91.6 37.3 63.7 93.9 47.7 &3.3 54.1 68.9 20.1 29.16 .7 70.4 83.0 52.4 61.9 56.8 11.0 72.9 69.3 66.7 18.0 39.6 96.9 94.7 16.4 84.4 885 90.0 15.5 36.3 40.3 3.2 68.6 30.3 71.1 64.5 21.0 22.2 74.9 41.0 41.0 63. Uize ituatk Order I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 .6 25.5 21.7 48.7 33.9 54. MEI)IAN RANKS SAample.2 11.7 16.2 59.5 55.6 11.4 3.1 27.2 1.9 7.7 43.2 19.5 81.7 87.2 28.7 27.7 45.5 56.8 86.4 81.4 96.9 78.1 23.4 54.?.7 84.5 83.5 3.4 41.4 96.0 24.'I'ABiE 11.6 88.9 59.1 53.0 22.3 1.8 32.0 52.9 29.2 19.0 17.0 45.8 66.8 80.4 36.6 27.8 51.3 26.4 3.2 79.0 50. 46.2 89.9 60.3 92.5 15.3 96.9 70.3 77.8 8.2 54.2 27.4 81.7 92.7 73. 46.6 42.0 74.1 1.0 6.4 40.9 52.2 21.4 48.3 45.0 39.4 1.6 23.7 45.1 94.0 53.4 28.0 35.8 31.7 35.4 10.Z 48.5 5.8 28.2 31.5 38.3 34.1 21.0 13.9 82.5 44.5 95.0 52.2 76.9 40.3 93.7 75.9 69.3 78.3 61.8 70.2 70.3 8.9 98.9 14.2 98.7 29.2 17.8 79.1 76.7 86.5 14.2 72.6 66.2 86.7 38.2 87.9 56.3 24.2 30.4 82.9 51.6 26.9 75.1 53.2 38.0 26.1 38.2 30.4 82.6 33.3 76.1 98.7 82.0 62.3 55.2 34.6 57.5 20.3 42.5 25.3 94.1 86.7 85.2 91.2 10.2 98.4 34.8 74.0 43.1 55.8 51.0 90.6 4.9 £5.9 30.9 86.8 98.1 96.0 96.2 60.5 69.9 79.6 86.6 62. 96.84.4 74.2 .4 .1 93.0 22.8 33.0 45.4 65.1 6El 62.5 63.8 78.2 9.9 91.5 1.5 9.4 64.3 7.7 5.7 15.8 33.5 16.0 8.4 78.0 50.6 146 .2 11.5 25.1 33.6 58.8 60.5 96.7 74.9 50.6 83.5 67.0 54.3 32.5 96.2 65.6 43.1 77.7 18.I .8 77.2 57.6 69.7 88.

8 33.6 14.2 54.1 62.5 39.6 19.3 16.4 7.5 36.27.3 9. .4 3U1 7.9 83.0 4.5 74.34.6 82.3 1.3 2.0 52.9 65.3 80.2 20.9 13 0.TABLE B.9 4.2 56.5 10.1 26.2 56.8 16 0.1 47.4 13.4 20 0.3 39.0 13.9 6 0.9 53.1 66.3 28.b 58.0 39.0 31.1 Sample Size 14 0.8 4 1.5 22.9 635 76.1 19.2 1&1 24.6 11.4 5.3 74.3 22 5.7 22.9 68.8 79.3 2.4 42.8 13.9 27.5 4.1 19.2 2.3 2.1 34.4 41.7 5.3 29.4 FIVE PERCENT RANKS Sample Size 1 5.3 39.2 16.2 5 1.0 36.5 58.3 29.8 4.2 28.5 52.8 47.4 12.6 9.1 41.7 16.5 19.2 9.4 67.3 65.5 34.7 78.8 9 10 11 12 13 14 15 16 17 18 19 20 11 0.0 71.6 13.3 86.9 17 0.6 10 0.7 i5 0.1 45.6 6.7 18.6 8.3 32.7.6 72.6 26.2 84.0 36.2 17.4 3.0 64.4 15.4 70.2 15.9 35.9 43.1 34.4 59.5 31. Rain Order 1 2 3 4 5 6 7 8 9 10 Rank Order 1 2 3 4 5 6 7 .1 10.4 77.6 39.7 2L7 25.1 34.2 54.1 4.9 42.8 25.9 14.0 70.7 7.6 27.4 58.0 57.0 63.1 10.6 21.9 .0 2 2.8 18 0.1 77.6 19 0.6 15.1 12 0.6 4.3 80.9 27.7 15A.8 .2 30.4 61.3 3 1.5.0 147 U471C' / /.8 18.1 b1.3 2.1 9.7 9 0.3.9 1.7 35.0 52.3 12.0 45.6 71.9 11.9 61.3 44.8 22.9 40.6 73.7 22.8 41.8 65.8 6.4 45.2.2 48.9 8.0 7.9 47.7 50.9 60.3 27.1 12.9 76.2 1.8 66. 22.6 11.0 47.5 39.9 50.6 7 0.5 22.8 24.2 1.7 24.0 81.3 68.8 6.0 24.9 56.8 47.5 3.2 44.3 49.9 30X1 34.1 49.1 8 0.3 60.3 32.

3 3.2 1M.0 52.4 7.7 64.) 11.6 28.8 52.5 26.2 27.8 56.5 13.:.1 8.5 31.2 73.9 59.2 67.6 0.0 70.0 43.1 90.2 5.5 52.6 38.5 3.9 30.8 20.7 ).5 24.2 74.9 8.7 4.8 15.7 8.9 79.7 4.0 22.4 30 90.6 43.8 10.4 9.9 12.5 11.5 10.4 67.8 4.9 48.7 76.6 64.5 17.9 50.6 6.7 22 0.2 7.).3 88.I 1.0 68.9 19.6 21.0 72.1 72.1 24. 58.2 16.0 81. II'I.1 2.1 48.3 :3.8 435 47.0 80.8 70.3 86.3 54.6 3.7 88.2 20.C 16.2 56.2 18.8 23.2 2.1 71.0 50.9 20.6 22.2 32.4 6:3.7 28.1 27 ).A 33..0 15.0 66. I 1.0 89.0 76.3 61.5 23.6 89.1 84.3.2 39.4 57.8 26.4 11.1 62.8 41.1 75.5 17.8 30.7 46.3 13.8 7.9 27.6 49.5 87.5 61.4 "L3 5A6 8.2 0.4 63.1 58.0 31.8 13.0 20.7 45.1 47.3 19.7 R).4 11.6 6.4 14.3 22.6 50.2 68.9 14.0 80.2 1.5 3.KS Itank (Irder 1 2 3 4 5 S 8 7 8 9 10 I1 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 29 21 0.7 .1 85.3 34.1 41..7 45.7 31. 1.2. 1 1.0 60.6 53.1 o.0 53.6 18.4 -.0 6.0 5.3 51.9 5.0 35.5 79.2 38.0 7.6 75.7 59.2 1.8 62.5 32.2 1.1 31.8 9.7 12.8 33.1 43.0 33.2 31.6 40.0 17.5 25.4 12.6 15.8 28.4 53.4 74.2 1.8 29.0 9.3 16.9 87.8 36.4 41.1 41.2 23.4 5.6 45.8 29.9 35.9 69.8 76.6 24.0 55.2 1.6 36.1 56.3 9.5 34.3 59.5 41.4 51.1 12.8 2J I 1.4 148 / .7 21.9 40.9 32.8 77.9 30.8 6.7 47.1 80.1 37.7 58.8 84.9 25.5 43.6 45.5 69.6 64.( ENT RAN .'I'ABILE 1.1 72.0 65.3 22.2 2.1 35.4 28.7 68.2 39.0 62.3 46. i'1.2 2:: 0.6 19.3 49.0 71.8 37.3 79.3 55.6 54.1 87.3 15.0 63.7 22.1 25.9 17.1 24.3 :16.1 89. • 0.7 78A4 83.

ANFS PER('ENT Rank Order 1 2 3 4 5 6 7 8 9 .6 86.2 14.7 36 0.7 24.7 91. 18.9 2.4 39.1 23.1 77.0 MA6 149 .4 53.7 8.0 28.3 88.9 :.7 82.1 41.8 26.3 40. 26. 2!or 3.1 1.r.6 76.3 459 48.4 .5 64.2 75. .3 21.9 10.9 44.7 36.1 16.9 2.1 5.8 7.8 66.TABIb.9 80.8 22.7 47.5 42.2 10.0 15.6 4.4 63.7 68.7 88.0 33 0.2 27.8 76.5 87.6 40 v.1 74.6 116 14.1 13.1 51.6 47.3 64.0 87. FIVE Sample Size .3 79.9 18.L2 .. 36A 28$At 40.4 91.1 0..8 67.3 61.1 31A4 33.2 40..3 56.7 11.6 54.7 32 0.1 2A.7 3.9 12.3 2014 "21'5 21-7 26 9 29.9 87.9 28.6 59.9 736 77.3 41.0 32.3 500 52.3 37.4 39 0.2 64.3 15.4 7:1.5 8.0 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 27 38 39 40 31 0.1 0.1 1.0 66-7 69.2 3M 0.8 51.5 10.5 65.5 25.6 52.0 2.2 25.50.7 79.8.3 2f.3 9.0 57.1 1.G 1'1.8 74.8 5.5 26.8 44.9 .5 31.8 81.9 21 3.4 33.4 11.9.1 86.8 37.3 18.7 5.8 84.7 30.3 69.7 48.56.6 30.A 6.6 58.0 86.9 58. 1 3.A 3.9 5.6 4.5 52.1 34.4 92.1 8.0 2.0 26.5 75.4 66.0 78.7 21.1 56.0 44.4 43.3 43.4: S.3 45.2 21.1 1.1 91.0 72.5 50.6 16.0 20.3 69.1 13.4 92.4 19.9 :3.9 54.5 5.8 92.1 204 22.7 15.9 7.3 8.7 31.0 2.0 17.6 21.3 24.3 62.5 32.5 16. B.1 39.4 78.5 4.5 34.2 61.4 37.7 72.5 80.8 58.4 60.4 30.!.1 0.1 71.9 48.5 35 0.3 54.2 49.6 1 18.1 8.1 25.8 83.3 59.7 74.0 38.1 1.2 53.4 27.8 39.I 0.4 49.8 29.8 20.5.0 18.6 40.2 1L.141W / ..7 30.5 47.4 7.6 15.2 1'1.1 49.3 46.7 38.1 82.1 2.5 9.9:1 31.9 78.3 78.): 6.1 27.9 22.1 2.3 46.7 32.0 12.9 72.0 37 0.9 14.2 60.3 70.2 58.1 52.8 .3 65.7 41.4 54.1 71.2.0 91. 8.9 57.5 75.5 51.7 E6.3 34.6 23.3 12.9 406 42.8 5.4 62.1 19.8 29.0 48.3 67.9 71.4 62.u.6 35.6 61.A 43.7 10.5 16.8 25.7 44.6 7.0 81.6 72.7 35.2 6.2 :15.7 97 11.5 13.2 32.2 3.5 57.4 4.4 63.7 65.9 10.1 66.5 L3.9 76.8 47.2 14.9 41.2 70.9 19.6 68.O 1.4 61.5 81-7 85.5 81.9 44.3 34 0.3 3.3 46.6 41.3 83.0 68.2 68.3 83.5 69.1 0.2 18.2 36.6 60.3 84.5 6.3 55.3 7.8 53.5 90.6 86.1 24.7 37.4 12.3 6.4 10.3 63.6 38.3 3.0 59.1 33.4 33.2 50.9 1!.1 92.2 16.

1 12.3 75.8 60.9 42.5 33.6 32.3 12.4 44.1 74.2 43.1 0.1 0.8 19.0 3.0 79.6 34.6 64.1 76.1 59.4 17.9 48.1 0.4 4.7 25.0 80.2 68.1 57.1 5.5 15.0 28.4 8.4 4.2 64.4 43.3 22.9 3.5 72.8 89.1 0.3 77.4 21.2 72.1 0.1 25.1 23.7 1.5 19.2 36.5 67.1 64.0 37.5 67.9 25.7 41.4 38.5 34.5 38.8 9.6 87.3 73.6 29.4 8.8 1.7 39.1 14.8 77.7 75.3 68.5 6.4 47.3 61.5 13.2 33.7 16.9 57.8 6.0 39.1 0.9 86.0 66.8 70.9 92.8 55.6 93.7 38.7 7.9 32.7 89.3 46.9 68.0 89.2 14.9 3.3 77.7 26.9 29.1 18.3 30 31 32 33 l4 35 36 37 38 39 40 41 42 43 44 45 46 47 46 49 50 59.1 37.8 21.7 49.8 17.4 45.8 3.8 57.5 79.6 16.4 6.5 14.6 31.6 2.6 1"1.9 75.6 29.8 21.6 94.7 90.0 93.9 82.5 48.8 7M4 81.3 5.9 11.8 18.3 87.0 58.0 16.9 13.5 41.0 54.9 61.5 85.3 62.1 11.4 28.8 52.5 0.8 61.4 15.1 4.1 0.8 .4 53.2 54.8 4.5 65.0 13.4 59.8 3.9 90.7 17.8 76.8 93.8 54.2 36.8 1.1 31.3 22.7 4.9 55.5 4:.9 20.0 4.6 36.1 51.9 6.8 64.5 49 0.4 55.6 2.7 61.5 62.8 32.3 49.9 84.5 82.5 68.9 33.4 40.2 69.1 0.3 51.3 45.2 38.5 80.2 10.3 30.1 15.9 30.2 7.1 17.3 78.9 90.3 21.7 76.4 32.4 25.7 8.5 52.5 36.2 66.7 42.9 10.4 50.6 61.1 82.8 14.3 16.0 50.7 58.6 27.5 68.5 7.3 45.9 71.6 24.8 49.1 43.5 28.3 86.3 11.7 19.8 44 0.2 87.1 83.5 74. I'PIECNT RANKS SAmple.2 5.1 19.6 59.4 50 0.9 41.9 81.6 51.2 32.3 6.3 93.7 35.4 93.7 1.6 39.0 28.7 66.1 0.13.7 1.4 53.6 80.1 63.4 16.8 4.9 65.4 30.3 4.9 73.8 7.2 44.2 4.8 2.1 56.8 58.9 49.8 12.4 52.6 86.5 62.5 71.6 13.5 20.2 28.0 4.5 37.2 22.17 0.3 24.7 9.7 53.9 30.9 67.6 48.8 31.6 85.7 11.3 78.3 30.3 43.8 22.3 41.9 401 42.4 5.4 b6.6 23.3 43 0.7 65.5 70.2 44.0 43.8 67.0 11.6 12.4 18.4 17.2 90.9 24.5 52.3 569 42 0.0 60.8 51.8 14.7 2.8 3.8 27.5 18.6 48.9 86.7 1.6 .3 9.4 46.TAIILI 1.4 25.46.1 85.0 69.2 18.1 0.18 0.2 12.1 5.1 63.4 .8 83.3 .1 7.0 9.2 60.5 47.5 26.1 150 .4 42.2 27.7 23.3 44.7 56.9 64.0 79.2 58.0 85 10.3 65.6 50.3 35.3 81.6 89.7 1.5 84.9 20.5 11.7 13.1 46 0.5 31.4 62.0 64. Size SRank Olrdaer I 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 is 19 20 21 22 23 24 25 26 27 2b 29 41 0.8 50.3 45.3 70.3 39.3 19.8 15.3 26.7 15.7 33.9 35.2 23.8 12.5 46.2 &7 10.0 31.2 79.3 37.0 82.4 46.1 52.9 18.5 58.4 53.2 42.4 93.9 8.8 56.0 19.0 15.2 70.3 8.8 24.7 2.9 7.0 47.2 67.7 61.1 72.3 28.6 6.2 34.9 4.4 90.7 82.4 35.4 20.8 1.3 30.5 57.6 29.0 26.1 16.9 74.5 9.4 62.3 56.0 27.9 21.8 22.8 77.9 26.8 35.2 63.4 74.1 93.4 46.5 27. FIVE.8 94.1 44.0 5.9 71.1 41.6 72.4 21.5 8.8 73.2 40.4 39.9 87.4 23.6 49.3 73.3 76.2 9.1 10.0 34.9 24.6 70.6 54.7 1.2.4 70.3 83.8 79.8 357 37.

9 94.3 9 28.7 99.8 97.5 95.9 67.1 84.8 90.7 55.5 46.5 52.8 83.0 92.4 87.1 2&0 32.2 29.6 80.7 93.6 39.4 60.0 81.9 21.4 47. 1 86.1 16 17.2 x :1i.6 17 16.6 59.9 47.0 89.8 99.6 41.9 72.2 89.4 54.4 63.7 99.7 64.0 97.9 51.9 36.0 65.2 85.9 54.6 72.3 69.3 97.0 33.0 65.4 65.8 83.7 75.9 91.6 80.2 34.7 90.7 84.2 77.8 39.6 1i 18.8 60.2 86.5 54.0 80.3 99.0 63.5 45.8 43.0 2 77.4 75.6 77.5 71.8 52.3 41.4 98.2 64.9 78.6 28.4 92.7 60.1 78.1 77.9 70.9 58.6 43.0 70.7 27.7 55.5 74.1 90.3 58.3 82.2 47.2 98.5 93.0 49.4 57.8 95.8 77.1 90.9 73.3 99.0 56.6 93.9 68.8 99.3 34.2 8K.3 40.8 74. Rank Order I 2 3 4 5 6 7 8 9 1o Rank Order 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 NINETY-FIVE PERCENT RANKS nSample Size 1 95.5 12 22.6 77.4 81.7 20 13.8 96.7 19 14.7 88.0 92.3 97.3 84.0 57.7 89.3 92.0 59.6 83.1 72.4 Sample Size 11 23.8 52.6 97.9 85.3 42.9 99.0 26.8 49.6 151 .3 94.9 99.8 36.6 60.7 .7 81.0 86.7 92.8 66.8 75.3 87.2 86.6 29.1 94.8 60.5' 13 20.2 96.6 14 19.7 66.6 38.1 63.6 18 15.5 99.TABLE B.4 99.1 7 :14.8 95.9 6 39.0 52.0 60.9 65.4 95.5 22.4 73.1 99.0 72.1 99.4 88.6 65.3 23.9 41.0 37.1 67.5 46.8 84.3 98.6 91.5 980 99.6 99.6 99.1 58.8 87.0 75.6 97.3 4 52.4 50.6 48.3.5 31.7 98.1 45.5 92.9 71.4 25.3 88.A 43.6 79.5 68.5 35.8 80.7 31.2 95.6 69.1 96.0 9&2 97.

4 94.9 86.8 58.2 94.6 28.2 62.1 98.0 .1 84.•a&mple Rank Order 1 2 3 4 5 6 7 8 9 10 I1 12 13 14 i5 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 21 1:1.5 61.3 84.8 21.2 73.4 47.2 47.9 82.2 50.3 79.1 69.9 22.9 58.1 86.7 67.4 99.1 81.2 34.9 31.3 68.1 16.1 72.5 41.2 19.7 92.1 45.9 91.9 89.9 51.1 W6.6 66. S.1 71.4 92.4 84.6 96.4 40.7 52.8 99.0 .0 24.6 48.1 31.8 96.6 27.6 44.9 29.9 95.8 40.1 24.6 83.8 32.1 77.8 152 .3 58.3 96.3 97.4 83.1 84.4 73.7 22 12.5 98.7 85.7 82.2 27.' 75.7 53.2 41.3 45.8 67.0 99.0 96.8 98.8 76.0 76.5 2.5 90.7 99.0 66.8 15.8 81.9 98.9 39.9 46.9 30.5 36.9 94.9 46.8 87.3 85.3 90.3 54.3 78.8 59.9 43.5 20.9 53.4 70.9 46.9 65.3 82.5 64.8 19.7 90.6 564 60.7 74.6 71.9 88.2 25.9 31.1 28.5 49.7 29.6 50.5 92.7 35.4 75.6 52.9 75.7 87.9 41.3 96.1 93.7 75.5 44.4 80.0 78.3 35.8 35.7 68.3 42.3.6 92.8 63.6 48.0 58.4 43.9 37.2 23.6 83.3 54.7 23 12.2 68.8 27.8 36.:1 15.0 71.0 74.1 95.3 76.1 32.9 71.8 36.8 27 2m 1HU5 10.3 62.8 25.6 99.7 80.9 962 99.8 29 9.2 56.3 99.6 23.3 62.4 79.4 90.0 87.3 70.0 93.3 63.5 93.6 59.2 95.3 20.6 86.5 56.9 49.7 NINITrY-IIVK IRKRCENT IANKS .0 90.4 54.:i.8 51.7 25 11.9 91.4 64.1 94.9 93.0 79.3 56.6 97.0 778 80.32.33.0 49.7 67.4 68.9 69.Sizt 24 11.5 96.8 72.2 20.2 40.1 63.6 89.5 5b.4 30.7 26 11.2 85.8 69.'TABIX B.9 43.8 52.1 98.141II" .5 01.1 77.1 97.4 88.7 94.5 96.5 14.2 17.7 39.1 77.5 99.3 64.2 37.4 66.2 98.8 30 9.6 54.7 96.7 99.3 74.0 87.4 99.1 65.8 26.0 88.8 16.0 60.7 18.9 38.2 61.5 79.7 19.8 82.6 96.2 60.1 38.8 47.

5 91.1 78.5 62.7 69.1 23.4 73.0 94.8 79.7 12.9 34.6 19.4 97.2 69.5 17.1 49.8 Js 7.9 47.8 99. .0 19.2 74.8 84.8 153 1.7 68. Rank Order 1 2 3 4 5 a 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 NINETY-FIVE PERCENT RANKS Smple Size 31 9.2 61.5 44.8 80.4 97.8 34 8.4 54.3 21.1 16.9 87.1 48.2 70.TABLE B.0 30.1 82.7 99.3.6 94.4 40.8 9A0 99.4 90.3 84.4 98.1 30.4 48.5 43.5 43.6 2&.6 38.6 83.4 93.2 56.3 44.3 65.1 39.0 64.5 42.6 55.8 97.1 58.5 39.8 55.9 18.7 72.8 77.5 11.0 99.0 92.5 96.4 18.8 95.9 38.8 39.6 29.6 96.8 99.4 94.5 81.7 87.2 68.7 85.0 58.0 60.9 87.8 32 8.2 35.8 81.1 51.7 46.4 86.2 61.7 41.8 96.2 83.4 55.3 74.8 99.6 59.9 89.3 22.1 65.5 51.8 31.0 86.7 47.2 91.8 .5 56.7 36.0 33.7 37.2 68.4 57.5 42.1 99.4 31.0 52.5 45.3 51.2 14.8 42.9 99.6 13.3 67.2 12.1 31.4 30.0 82.5 68.6 59.8 16.5 70.6 49.6 40.7 31.1 61.8 25.4 95.7 64.4 20.2 74. 81.0 97.4 53.5 98.3 11.2 88.2 97.2 77.9 56.3 89.2 94.5 32.6 83.8 34.3 70.5 92.8 63.9 7.33 8.8 85.8 35.6 59.3 62.9 99.6 23.0 76.2 66.2 41.0 90.8 39 7.7 92.8 37 7.5 27.1 22.2 92.9 96.0 94.8 .4 62.6 59.0 33.5 86.3 64.9 12.0 28.7 89.2 33.7 56.5 16.6 91.8 76.0 60. 81.9 18.4 26.3 89.8 73.9 52.4 25.3 14.9 20.9 13.8 65.9 49.2 17.5 81.8 88.2 74.6 41.0 46.3 52.8 70.2 38.9 23.3 58.1 75.6 88.6 86.3 97.2 50.9 99.7 54.6 49.6 62.6.1 .5 72.5 93.1 71.8 91.3 29.3 98.2 80.8 40 7.4 88.0 C5.1 27.0 81.3 83.7 97.2 58.4 80.6 78.0 99.6 89.7 45.0 56.9 92.4 71.6 65.0 91.9 36.9 60.5 66..3 98.6 73.4 41.6 27.0 26..2 93.8 53.9 66.9 25.5 84.2 94.1 97.2 18.2 32.7 79.0 75.6 38.3 33.3 96.7 69.4 80.3 46.7 77.7 93.9 86.2 67.6 37.6 50.6 44.1 40.7 90.4 73.2 93.0 63.9 76.4 77.0 48.9 72.2 27.6 67.6 24.8 84.5 15.6 98.5 73.1 83.3 91.4 87.2 2L4 24.6 82.4 13.7 51.5 52.4 36.5 63.3 46.6 21.0 96.7 89.1 53.1 65.8 Su.8 :25 8.3 37.8 94.3 87.6 59.4 78.9 99.0 99.4 79.9 84.2 73.8 21.6 69.5 76.6 35.1 96.1 61.3 85.2 11.0 34.0 75.2 24.8 71.6 46.67.6 97.6 99.8 15.3 30.6 95.9 28.1 75.0 47.6 63.8 66.6 43.3 47.5 97.M 7.1 79.

0 39.7 69.1 81.0 33.2 15.6 84.9 74.34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 41 7.5 48.0 99.2 99.9 97.5 78.0 17.2 83.6 64.6 82.0 20.7 95.0 98.3 31.1 91.2 99.7 60.8 41.0 72.3 64.8 98.4 75.3.5 85.6 19.0 18.1 99.2 42./ .2 89.9 99.6 73.3 87.8 21.8 79.0 72.9 91.4 65.2 22.6 58.0 96.8 47.0 2.2 29.6 73.5 56.5 68.7 75.32.6 54.6 75.5 95.3 52.1 21.0 69.4 23.9 25.4 56.8 20.2 80.6 18.4 95.5 80.6 58.5 13.1 82.8 22.5 48.4 21.3 5.8 70.8 50 5.9 87.7 80.2 85.5 12.2 55.1 48.2 94.1 37.6 56.5 86.2 47.7 98.3 80.7 38.9 67.8 46 6.0 89.2 91.9 78.4 72.4 90.7 63.9 61.1 88.4 25.5 96.1 88.9 74.2 17.0 22.1 84.1 69.5 82.6 36.4 67.5 83.5 41.6 27.2 65.3 74.5 95.1 88.5 10.3 72.3 97.9 94.9 40.7 59.2 99.4 69.d 58.1 99.1 96.5 50.7 67.5 97.2 63.8 61.0 80.0 95.4 54.6 34.9 44.4 56.2 86.2 66.6 95.5 53.3 76.1 66.3 74.8 60.3 39.7 90.1 44.9 89.8 71.6 42.0 75.4 94.9 98.1 98.1 93.1 40.4 96.7 93.6 22.0 81.4 57.4 92.5 41.4 32.8 855 87.5 44.2 43.1 52.5 88.3 99.4 37.6 60.8 45.6 23.9 46.1 99.1 27.5 78.9 9.1 26.0 82.8 83.3 49.7 35.6 92.8 67.4 38.5 64.2 87.8 53.1 99.3 9.8 78.1 86.0 40.8 88.9 38.8 41.5 b7.8 7 • 154 / .8 95.7 32.7 15.9 23.3 90.8 10.8 48 6.0 49.7 43.8 49 5.0 11.7 65.4 26.4 62.4 46.2 77.4 10.0 69.2 50.3 12.6 .6 94.5 51.4 66.3 96.2 98.2 92.1 27.5 65.4 45.5 70.9 35.4 51.6 87.6 96.1 84.1 31.2 62.6 47.5 57.0 77.1 43.9 13.9 85.6 73.8 85.1 99.2 78.0 16.4 35.8 62.9 26.6 59.5 50.7 46.7 34.5 57.5 49.2 68.5 75.3 44.7 10.7 97.1 76.9 17.1 72.7 67.1 54.1 50.2 99.3 24.0 64.7 60.6 28.4 42.0 62.5 60.6 48.6 52.4 95.8 29.6 79.5 59.2 99.3 67.4 88.4 36.7 35.7 41.3 55.4 37.0 94.9 83.5 61.4 20.3 42.6 81.1 47.0 14.1 81.5 71.6 58.1 13.0 20.1 49.3 92.4 43.6 93.9 70.0 83.3 19.7 94.0 55.0 78.0 63.3.7 84.7 20.3 54.4 W4.1 99.2 99.6 62.7 61.2 70.1 9.2 38.5 73.8 14.3 79.6 47.6 29.3 71.TABLE B.5 58.0 45.1 71.2 95.8 86.5 33.5 15.7 37.1 68.9 50.8 91.0 37.6 81.0 q0.7 93.2 64.6 16.8 90.4 89.4 32.3 34.6 90.8 88.0 43.0 9.4 76.1 25.8 43 6.9 64.8 30.3 16.2 99.7 76.3 56.8 40.8 62.4 77.2 93.9 54.4 35.8 52.4 93.2 29.1 45.0 31.4 31.9 93.1 79.6 28.1 28.1 53.0 33.2 39.5 82.5 29.1 95.2 46.0 29.5 57.6 68.4 55.9 24.8 97.7 17.9 76.3 91.7 12.5 17.3 99.9 73.7 71.8 9.0 86.8 51.1 48.7 80.3 74.0 92.0 91.0 36.5 24.8 45 6.8 47 6.4 84.3.1 51.0 14.1 12.8 32.6 26.7 89.1 81.5 55.6 66.6 31.6 38.6 53.0 77.6 27.8 44 6.2 99.3 13.1 34.8 42 6.5 13.1 90.5 86.9 98.1 98.0 71.8 92.6 77.6 980 99.3 69.6 91. NINETY-FIVE PERCENT RANKS Sample Size Rank Order 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 X2 &1 .3 19.7 25.1 99.3 18.

4099 0.Median Rank of Y (The median -ank is expressed in decimal form. as follows: Y .5000 6. EXAMPLE AND STEP-BY-STEP PROCEDURE The median rank regression method will be illustrated with the censored data listed below. The "best-fit" line is found by the method of least squares.0o0o 2. through the data plotted on Weibull p. Cycles 1500 1750 2250 4000 4300 5000 Status Failure Suspension Failuie Failure Failure Suspension 7001 Failure Step 1: Calculate the median ranks of the failure times us.l)er.0946 0.3333 4. the failure times and median ranks (see Chapter 2 for the calculation of median ranks) are "transformed".APPENDIX C RANK REGRESSION (WEIBULL PLOT) METHOD OF WEIBULL ANALYSIS 1.Qn (cycles-to-failure) X = Qn(Qn ( I .2523 0. The median rank regression estimates of the Weibull parameters are: 1A eA 2. straight line.) Least squares is then used to estimate A and B in the equation Y = A + BX. to estimate the Weibull parameters beta and eta. First.8041 155 .2500 Median Rank (decimal form) 0. Cycles to Failure 1500 2250 4000 4300 7000 Raitk Order Number 1. ig the methods oi Chapter 2.5676 0. METHOD Median rank regression uses a best-fit. respectively. These estimates will be referred to as A and 1.1667 3.

8714 • X.3132 7.8714)2 B3= 7.2940 8. calculate the natural. 2 = 7.7743 -3.5356 -28.5. and 1.3088 -1. =.7187 8.5356 5 A B = 2.5.4099 0. where: A A= V - X where V is the average of the Y's and X is the average of the X's.5205 4.2353 -0. n In the abmve example.1763 0.5676 0.1092 -0.50 4000 4300 7000 A Aý S'tep 3: Calculate the least squares estimates A and B of A and B in the equation Y.8735 - (-3.8714)(40.54 and A= 8.5460 Y X = 8.5381 A B = 0.-I x -i) ?y.A + BX.(0.A537 X -2.4887 Cycles to Failure 1500 22.8735 40.2523 0.(Y = Step 2: For each failure.1092 .5460) (-3.6397 -0. n Yi Y = SXiYi Xi = = -28.5392 156 . (base e) logarithm of the cycles-to-failure Rn (cycles-to-failure)) and X =Rn(Rn(1 - Median Rank of Y)) Median Rank 0.7743) A =8.0946 0.5554) (-0.8041 Y 7.3664 8.

.1 shows the data plotted on Weibull paper with the least squares line overlaid.80 4=e4 eO = 5111.e-(t/51115)' Figure C. 157 i •a4-?1 N.Step 4: Calculate the median rank regression estimates of # and q: == • 1.25 The Weibull equation used to calculate the probability of failure before t cycles is then: F(t) = 1 .

a ~~oa . . 2. .aa. . I. .L --- . 4 - _- ..I .. -' .. - .. li.--- III - -.. . TOTAL OPERA-TING CYCLES FO 2V1MS Figure C.g : . -- g I .... . . .._ v 5111. . 2 rooo . *o .8 -U. 1.I.EIBLLL DISTRIBUTION 0 S. wa .. *S .xampie D~ata With Le'ast Squares Line 158 . - LEAST SQUARES LINE O.25 SRMPLE SIZE = 5 FRILURES= 5 1.

the likelihood function is: L where n = i f(xi) f(x) f(x 2) . and satisfies F(x) = . 2.. New York. When the time-to-failure distribution is Weibull. FOREWORD Weibull analysis consists of "fitting" failure data to the Weibull distribution by estimating -the parameters. x2 . John Wiley and Sons.R(x) where R(x) = reliability at time x. F(x) is often called the "unreliability" at time x.. expressed in mathemat-cal form." of obtaining the observed data. The likelihood of obtaining the observed data. The maximum likelihood Weibull analysis method consists of finding the values of 0 and 7 which maximize the "likelihood. When the sample is complete (all units are run to failure). ) sample size dF(x) x) ..- and F(x) = 1 . this Appendix presents an explanation and an example of maximum likelihood Weibull analysis. is a function of the Weibull parameters 6 and i.. .. F(x) is the probability that a unit will fail before it acquires x units of operating time. 159 X X . is a function of the Weibull parameters # and V. The rank regression (Weibull plot) method was presented in Chapter 2. Maximum likelihood finds the values of P and -9 which maximize this mathematical likelihood function.APPENDIX 0 MAXIMUM UKELIHOOD METHOD OF WEIBULL ANALYSIS 1. f(x) =I e wo and L n= S17 ) 71 .. x.. Details of the maximum likelihood method may be ound in Wayne Nelson's text. Applied Life Data Analysis (1982).e 4000 Note that the "likelihood" of the sample failure data x1 . beta (#) and eta (q).e-(X/e)O In reliability terms. THE LIKELIHOOD FUNCTION The likelihood function is the mathematical expression of the prpbability of obtaining the observed data..

L 3.- - -'0 Given the failure times x1 . xn. A In the complete sample case. ji.. the maximum likelihood 3 estimate. equate the resulting expressions to zero.. i e -r e MAXIMIZING THE LIKELIHOOD FUNCTION The maximum likelihood method finds the values of / and q which maximize the likelihood function.The general form of the lFkelihood function for censored samples (where not every unit has been run to failure) is: k l. 'r 2.. satisfies A x' IInx i-n l SII i. x . x2 . x 2. xr = known failure times Ti.Q nl - - Qnx. and simultaneously solve for 3 and 17.. When the time-to-failure distribution is Weibull. satisfies A N A x-Q NA r i=1 n=nx0 i 160 31734- >i7 .denoted i3. the maximum likelihood estimate of/ is found using The maximum likelihood estimate of il is: n where 13is the maximum likelihood estimate of/i.res. iterative procedý. iI f(x.)" j I (1- FC(T)) where r = number of units run to failure k number of unfailed units x 1. differentiate the logarithm of the likelihood function with respect to JRand r/. Tk = operating time on each unfailed unit. .... To find the values of #i and il which maximize the Weibull likelihood function.. the maximum likelihood estimate of 13.. When the sample is censored.

Cycles 1500 1750 2250 4000 4300 5000 7000 A Status Failure Suspension Failure Failure Failure Suspension Failure The maximum likelihood 8 estimate. The second term in the eluatnon./ A is. / is again found using iterative procedures. the maximum likelihood estimate of censored samples. sums the logarithms of the failure times only. UInit. .where N = total sample size (N --number of failures (r) + number of suspensions (k)). censored at times Ti are assigned the values xrti Ti. Analogous to the complete sample case. in N A) 1/ý A i 4L EXAMPLE The maximum likelihood method will be illustrated with the censored data listed below. is the root of the equation 7 G(8) GO• -r-.Zxnxi i-i i-i 5 ZQn 1 1 ' - i - "3 In -" 0n 0 Zx?' 161 314T _____ ____ _____ ___ . f. -Qnxi.

162j I .0255 -0.1754 -0.257 -0.? initial value of i.) o..1746 -0.0000 Maximurm Likelih(od Estimate of Beta The maximum likelihood estimate of .0007 -0.t. (A Fortran ul~hroutint.8.twL. 1. using a modified Newton-Raphson procedure was used to find the value of giving t(.1. estimates of # are listed below with the corresponding value of G•3).0248 -0. is A '1=4900.253 2.179 2.256 2.The Weibull plot estimate.) G('8) 1. was used as zh.0005 -0.800 1802 2. This and l i.182 2.

The lower bound is derived using two facts from statisticr: 1. one-sided lower confidence limit on the mean life 0 is: Zt. L DERIVATION OF THE WEIBAYES EQUATION If no tailures have occurred. t2 . The Weibayes equation for q is r (E. It is used when there are certain deficiencies in the data tfor instance. n is the number of' sispensions or unfailed units in the fleet.2) 163 M47' ~ ~ -i .(on as. The exponential cumulative distribution function is F(tI . t. the first failure is imminei. then a conservative 100 (I . i. the Weibayes equation with r lower confidence bound on the true value of qi. and the units are sosceptible to an... . t2 . -1I n (E.1) where / is the assumed value of the Weibull slope parameter.e. If no failures have ocwurred. t2. are the operating times tecumulated by units 1. If no failures have occurred in a fleet with n units having operating times t. . t1 . 2.. then t 1p. FOREWORD Weibayves is a method for constructing a Weibull distribution based.APPENDIX E WEIBAYES METHODS 1. n.-. but no failures have occurred). Chapter 4 describes several applications of this method.-. ti. . 2. represent failure times drawn from a Weibull population . t. r is assumed to be one.(0"'. the Weibull slope parameter.vith slope parameter #i and characteristic life Ut.1 e-/. when operating time has been accumulated. If ti. t2 I gives a conservative 63'. lower confidence bound on the true value of q.. represents a random sample from an exponential population with mean life a = i?. q* is the maximum likelihood estimator of the true value of 7.. q" is :. If failures have occurred then a conservative 63 and Weibayes is used. r is the number of failures.uming a value of ji. exponential failure mode. .

* is the maximum likelihood estimator of *the true value of in.levelis unkmiowui.the Weibull slope parameter.then a .1) is eq-.1I . whille assuming that 8. .s.0 i .3).0 -1. I hait he t Irue confidence . giving S>-( " "enO. Solving for twr'find: -Rn R~na = = .) *nserva ive I(W) (I . is known.'.. . If failures have occurred.sided lower confidence limit on I# -j is: or ' • . i = 63.).€.The [•the Weihayes lower bound on q (E.368)% confidlnce bound for q.es lower b~ound on I is a 100 (1 -0. . This is shown by finding the value of that maximizes d the WeibuU likelihood equations from Appendix D. and Weibayes is 1 usea .3) with denominator..•lo~r a e that with # assumed.0 e.• ~(V j 1. determining a lower bound f: r also determines a lower bound T I he Weibuli line. In awlt I lte units lire susceptible to it Weiulll failure Inode with known ji (i!ind unknown 17).onservat ive meatis o•| r)". ' • =0.10.) ' one.368 Thus. if no failtires have oct'irred in it fleet with Ii unifs having owrating times I . a conservative 90V lower confidence im~und on I can he calculated by setting (v = 0.)""with at least 90% confidence onth rgthadsie f re~ahty E.Fr xapea onevaiv 0%loercn6dec ""Not ."~ ~ IM ""C(.jl to the lower confidence bound in (E. For example. the Weihas.Qn a = 1. limt is lit least 'hu.2"t conservative lower The confidence level can! be increased hy decreasing a in the denominator of the expression on the right hand side of ndeouahty (E.1.

similar to those discussed in Appendix D.4.4 (maximum likelihood) are identical. demonstrating that the Weibayes equation yields the maximum likelihood estimator of V. 41 I" '' ii 165 .It . result in the following equation for the maximmn likelihood estimat or o1"q (assuming that tiis known): Sr Y (E. when failures have occurred and i is known. I (Weibayes) and E.4) Equations E."T'hesecalculat ions.

is calculated as: ton S -• I-.5 (150". and is normally measured by the standard deviation.g'A+ .. 50% are above 3.In this case.1 illustrates the meaning of the accuracy and precision of Weibull parameter estimates. The estimated values of # and 17were stored. The standard deviation of the 1000 0 estimates f 1.5). When the fleet experienced its 5th failure.0: the # estimates are quite accurate with 5 failures. the median d estimate is 3. the 3 estimates. Of prime importance were the accuracy and precision of the risk forecasts. "Accuracy" refers to the difference between the "typical" # estimate and the true value of 11.5. with the suspensions and failure times intermixed).APPENDIX F MONTE CARLO SIMULATION STUDY ACCURACY OF WEIBULL ANALYSIS METHODS 1. data containing hot h failures and suspensions. S999 where 1 f2 + 0j. . The more the # estimates deviate from their mean value. As this is unusual compared to other Weibull applications. 'rhe 11 and q estimates were stored. Median rank regression is covered in Chapter 2 and Appendix C.e.5 and 2.I life estimates. the higher is their standard deviation. A simulated fleet with 2000 operating units was introduced to a Weibull failure mode with d = 3 and il = 13.5 and 4..5) and the true value of d is 3.000 hours. a Weibull analysis was performed.5.#. and the B. FOREWORD Safety considerations in aerospace operations require corrective action based on very small samples of failure data. Precision refers to the variability in the # estimates. Details of the Monte Carlo simulation method are discussed later in this appendix. for a total of 1000 estimates of # and jj under the simulated circumstances. 1000. +0m 100o 166 . Two methods of Weibull analysis were considered and comparedi: median rank regression and maximum likelihood. Maximum likelihood methods are introduced in Appendix D. These two methods are among the most commonly used for estimating ji and I with multiply censored data (i. It shows that 294 of the 1 000 estimates of # were between 0. 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). and the process was repeated many times. In 392 out of I000 simulation trihls. of the estimates are below 3. Figure F. the fleet was re-created and reintroduced to the Weibull failure mode. 0 was estimated between 2.. (No estimates were below 0.1 is a histogram uf the ft estimates.5. Figure F. A Weibull analysis was again done at •he time of the 5th failure..

Five Failures 167 .5 4.5 12.5 6.5 10.40 35 - Sample Size Average 1000 30 True 25 - Percent ofI 20 - SOccurrences I- 15 -Precdslon 10 - 5o• 0.5 2.5 18. Bcla Etimats: True Beta 3.5 14.5 20.5 Beta Estimate FD 256570 FiOcurerFre.5 8.5 16.5 24.5 22.

Both methods of analysis produce estimates of comparable accuracy. Maximum likelihood estimates are noticeably more accurate for 0 = 5. 147W 168 . tend to overestimate 6.5 shows the B. maximum likelihood estimates have 44% and 68% less variability than median rank regression 2. B. BETA (0) ESTIMATES The accuracy of the maximum likelihood (ML) and median rank regression (MRR) estimates. so its accurate estimation is important.0.3: I. life standard deviations as a function of sample size (3 to 10 failures.2. I0(X) to 20XX) unfailed units). Sbetween estimates.1 LIFE ESTIMATES The HI life is frequently used as a design criteria. When the failure data are plotted on Weibull paper. calcuated with a small number of failures and a large number of unfailed units. the slope is generally too steep. In the three-failure case. 3.3. The standard deviations of the small sample # estimates are shown in Figure F. especially when the data contain as few as three to five failures.2 and F.4 illustrates the accuracy of the B.. 2.1 life estimates over the range of the study. Figure V. The following can be seen from Figures F. 3. The accuracy and precision of the # estimates improve as the number of failures increases. Both methods of analysis. Maximum likelihood # estimates are more precise than median rank regression estimates. Figure F. 4. is illustrated in Figure F.4. especially with very few failures. Its estimate from field data is often compared to the predicted or design BI life. median rank regression and maximum likelihood.

Uu* -6 0 9 o I E co Cf 0 IXI Li) -6 4a) co m Co 169 .

0 0 cm -00 0 I I) P a 0 E co Z. 0 > ) I I. IZ I II170 .

II Go OD ogo FC4 040 "Co | | •1! !| C 71 171 .

1 I c I 39 caI ccI LUI I I I* ..

:3.7.9 x 10"4 1. STANDARD DEVIATIONS OF THE CHARACTERIkTIC. 1.300 2.1 life.5 True Eta 100.000 8. 5..000 3 Failures 4. 4.000 27. The accuracy of the median rank regression B.0 X 1010 1. "TableF.200 3.000 Type Estimates MRR M11 MRR ML MRR ML MRR ML MRR ML - True Beta 0.8 x 10 2" 2.1 life in approximately 58% of the simulation trials.1 life in 52% of the simulation trials.6 X 10" 320. vs.1 life estimate does not necessarily improve as'the number of failures increases from three to ten.000 3.5. For # = 5.I life is more often than not less than the true 3.200 1.900 3. LIFE ESTIMATES Standard Deviations of the Eta Estimates 5 Failures 4.0 500.000 6.2 X 10 76.3 X 1012 8 9.2 X 109 6 8. and 3. 58% for MRR).000 15.1 life estimates improves as the number of failures increases.1 X 10 21. maxinum likelihood typically underestimated the B.000 5.300 U0 Failures 6. for • = 0. 4.175 Median Rank Regression Maximum Likelihood 173 / .life (ML underestimated the B.6 and F. and 3. Median rank regression estimates of the B. ETA (n) ESTIMATES The medians of the characteristic life estimates from simulated fleets with few failures are shown in Figures F.0 MRR overestimated the B.5 X 1016 3.0 &.5 indicate that: 1.7 x 1014 &8 x 1013 4.4 and F. TABLE F. For • = 0. 2.0 13.I. The precision of the B.3 X l0°' 1.1 life are typically conservative: the estimated B.003.1 life (ML overestimated the B.5. 1.000 5. the maximum likelihood method typically overestimated the B.1 contains the standard deviations of the characteristic life estimates.Figures F.1 life in .40% of the simulation trials).900 1.

.lX10 8 -= - True Rank Regression Maximum Likelihood =0 Eta (.) 1X1o 7-'-I ixio 6 0 / I 2 3 I 1 1 4 5 I 6 I 7 8 9 10 Number of Failures FD 256548 Figure 1-76.s -- A 174 N" . inL AccuravY Alhdian ('harac'tris•ics Ife•. ..tima•.

000 5 6 Number 7 of Failures 8 9 10 11 12 FD25669 Figure F.000 True q do t m-. ra ao m m • 3 ETA ) 10.000 _ _ _ _ _ _ _ _ "5.----m n riS• 5 . - -w - . Maximum 1. Eta Accuracy Median CharacteristicsLife Estimntes - B 175 .7. - Rank Regression Ukelihood .1000 sos- 100.175 -=True -.500.

3 = 3 additional failures. The standard deviations of the r. there are 5 .7 indicate that.. A Weibull analysis is done anid the risk forecast is made. I.117Mn. Methods fc. reflecting the presence of several extremely large estimates of q at each simulated condition.5 = 6 additional failures. underestimated. 6. 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. 176 . S. The precision of the characteristic life estimates improves as the number 2.1 RISK FORECAST ACCURACY Tah!e F. 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). of failures increases.constructirg risk forecasts from multiply-censored life data can be found in Chapter 3. The median rank regression (MRR) and maximum likelihood (ML) estimates are conservative: q is typicall".. The accuracy of both types of estimates improves as the number of failures increases from 3 to 10. estimates are extremely large. The entry under "0" months ahead indicates the cumulative number of failures expected 'to date' (at the time the analysis is performed).3 = 2 additional failures expected over the 12 month interval following the occurrence of the third failure. and maximum likelihood predicts an "average" of 6 . and 12 months into the future. up to 12 months into the future. The risk forecast is then compared to the actual number of failures csused by the fleet's additionas 12 month advance. :1. and 10th failures.2 contains the medians of the maximum likelihood and rank regression forecasts for 0. The fleet is then advanced 12 months further through the Weibull failure mode using an average military aircraft utilization rate. The rank regression method predicts an "average" of 11 .. and F. This procedure is repeated 100 times with 100 different simulated Ilvel s to assess the variability and accuracy of the risk estimates. 4. 5. 5th..Table F. / / . I arid Figt:res F. when 11 :1 and there are three failures in the fleet. Thus.

5 No. Ahead 0 6 12 0 6 12 10 0 6 12 Median Risk 3 4 5 5 6 7 10 12 14 Median MR9 Forecast 8 if 7 c 11 12 14 17 Median ML Forecast 3 4 4 5 6 7 7 8 11 il 12 3 4 4 5 6 6 10 11 11 5 10 6 6 8 10 12 15 a-s No. Failureq 3 Months Ahead 0 6 12 0 6 12 0 Median Risk 3 4 7 5 7 10 10 14 19 Median MRR Forecast 6 12 22 8 14 25 13 Median ML Forecast 3 5 8 5 8 12 1' 5 10 6 12 19 28 14 21 177 .TAPLE F. Faikures 3 Months Ahead 0 6 12 0 6 12 0 6 . RISK FORECAST ACCURACY = 0. FailureX Maonth.2.12 Median Risk 3 3 4 5 5 6 10 10 10 Median MRR Forecast 4 6 7 6 7 7 11 11 11 Median ML Forecast 3 4 5 5 6 10 10 i1 5 10 a=' 3 0 6 12 0 6 12 0 6 12 3 3 4 5 5 6 10 11 11 1-3 Net.

The accuracy differences between the two methods increases as 0 increases.45 0.003 0.) TABLE F.22 12 3.0 0. Maximum likelihoKo risk forecasts are only slightly conservative.002 0.5.Np From Table F.3 6 10 0 6 12 0 6 12 34.002 0.0 3.1 0.6 4.84 2.0 Months Ahead 0 ML Forecast 0.0 91. they overpredict the numlbr of failures. 5.3 O.0 56.004 0.98 2..5 MRR Forecast Months Ahead ML Forecast 3 0 6 12 22.3.4 0. Failures = o.8 4.0 52.2 3.91 5. Failures 3 I MRR Forecast 73.wes are always more accurate than forecasts based on rank regression Weibull analyses.3 0. RISK FORECAST STANDARD DEVIATIONS ii "No.0o6 0.0 67. and 12 months into the "future". (The risk forecast 0 months into the future is the expected number of failures to date.8 6.1 0.2.1 0.002 6 12 94. it is seen that: 1.0 112.0 6 12 10 0 6 2.43 #No. they are only marginally more accurate when ji =0.49 178 .1 0. 2.8 3.'. 6.2 RISK FORECAST PRECISION Table F. Forecasts based on maximum likelihood Weihull analy.3 presents the standard deviations of the risk forecasts 0. Rank regression risk estimates tend ito be very conservrtive . Maximum likelihood methods are much more accurate than rank regression with # 5.0 3.59 1.2. 3.

0 19. and output segments.0 24. rocessing.5. Month* Ahead 0 6 12 0 Fomeast 11.4 a 0.4 5. ML Foremst 0.4 10.4 6 12 10 0 6 12 "TableF..4 12.1.TABLE F.0 36. The 2. The rank regression forecasts . components of each segment are listed in Table F. S.5 4.0 MRR IL Forecast 0.0 14.006 1. Failurta 3 Montha Aherd 0 6 12 0 * 12 0 6 12 MRR Foreefat 14. the precision of theresulting risk forecasts im roves.3.0 4. Maximum likelihcod forecasts are far zhore precise than rank regression forecasts. RISK FORECAST STANDARD DEVIATIONS a=3 * No.iary substantially (up to 50 times more than forecasts based on maximum likelihood Weibull analyses).0 88.006 1.011 V7 5e 0.3 o.4.0 5.016 L.1 3.3 shows that: I.0 29. over the entire scope of this study.011 I 10 Failums 3 No.7 14.U O.00 U. 3.0 7.0 37.2 4.79 6. 179 . L MONTE CARLO SIMULATION The Monte Carlo simulation consists of input.0 4.0 4. As the number of failures used in the Weibul' 3nalysis increases.

in the fleet according to the Weihull failure modle input. and actual year-aheadifailure cou.ihes fopr each failitre modte considered.)1MUILATIOR 1. The processing segment includes constructing and aging the fleet. and forecasting the additional number of failures expected in the year ahead.•4.1 . and rapid wearout failure modles.4. wearout. New member addition 4. this p~rocess was repeated i0(X) i.-et.. I. respectively. For the rank regres. One-year-shead ris k forecasts 3. Random failure time generation 5. I'aranivihr tesliniates. and (4) the per-unit monthly usage rate.4 and the actual Weibuli parameters input to the simulator. estimating the Weihull parameters.)(XX) units in the fi. This proc~essq continues until 10 failures occur in the fleet. year-ahead failure forecasts. and the year-ahead I'-recarts are made. Tlhe forecasting error is simply the difference between the actual and forecast failure counts. "-"&5 '•& F*leet gist (COMP)ONENTiS OF MONTE (CARLO I. The fleet characteristics are (11. The fleet is 1AH~ . and 10 failures occur. Failure times are gen erated for each member ofl the fleet using random numbers and the inpu~t WeibuUl paramneters. 5. Weibull analyses are performed at these times. (2) the fleet age distribution. Weibull parameter estimates 2.•tiniatp.nits are S~members. one. Actual one-year ahead failure c.|tor. The fleet is i hen revonst ructed. with IO(XX) to "-.s again. and the process begin. the number of units in the fleet.sion estimators.4 (OM()I~TSOF ON1:('RI()1I8LTO aged util :1. Se Figure F.9. random. incorporating new generating random failure times. 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.TAHBLE F. Fleet ag~ingl 31. fibl~s) three. The simulator generates fikilares among thle units of one-half. Fleet ýnput Frovensing -mstruction 1. The error in the estimation method is reflected in the differen('e. Fleet characteristics and Weilull failure mode parameters arc input into the simod. (3) the production schedule. The age distribution is assumed to be normal. Prodluctikon schedule Usage rate True Weibull pararneters Random numlh 2. Parameter estimation 6 Risk forecasting[ Monte Carlo S~The simulation procedure is illustrated in Figure F. • 180' .S. Covst considerations limited this nui lhr to I(N) fr the nmaximum likelihlmd estimato~rs. and rive were chosen to represent infant mortality. Sliope paramete. I ..ounts Ouhtput Fleet age distribution . hetween the parameter e. oonlput.

|. 0 U~(3 IW. L +3 Cd *ZIA 181 .

8 182 S . . Initial Conditions of Simulator .9.*•.. * p Fleeu / Wei~bull Failure Mode M... *. C-ý FD 256451 Figure F..

ta. based on the maximum likelihood estimates. Rank regression provides accurate estimates of "low" percentiles like the B. Maximum likelihood "lower" percentile estimates have a slight positive bias with small numbers of failures and a large number of suspended items.A life under the conditions simulated in Appendix. Maximum likelihood risk forecasts are more accurate and preci3e than are rank regression risk forecasts with small failure samples. Outliers in the data. Maximum likelihood does not provide a graphical display of the data. This helps to identify instances of: a. implying the failure mode is random or memoryless?) cannot be tested using these estimates. "Low" percentiles refer to percentiles close to the time of the first failure. Rank regression provides a graphical display of the d-. correction. a 183 ÷-N . Is = 1. Statistical hypotheses about P and q. (The slope on the Weibull plot is too steep. b. The Weibull distributin not fitting . are available for all commonly occurring forms of censored data. 1. even for small sample sizes.ell to the data (suggesting perhaps another distribution. c. More than one failure mode affecting the units. (e. 2.) This positive bias decreases as the number of failures increases. 4. when computed with few failures and a couple of thousand suspensions. e. Confidence intervals on the Weibull parameters # and q based on rank rekresaion estimates are not available. 5. Both rank regression and maximum likelihood tend to overestimate t with small failure samples. Data needing a t. Rank regression risk forecasts are conservative (overestimate the risk) and less procise. Exact or approximate (large sample) confidence intervals on the Weibull parameters #i and q.g. d. - 3. like the log-norMal). . F. Statistical hypotheses about t and q can be tested using these estimateu.. Batch problems.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.

along with the failure times of a Weibull analysis can be produced for quite large samples. % •interest. they can be used to solve many of the and examples in this Handbook. the same answers as given in the Handbooxk may not be achieved. and will only be different becautse of the histogram input in Slie case of t hese programs. .-' 7- .:I by the use of the Newton-Raphson iteration procedure. the user is prompted for i.. Hence. 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. N1 7 7577 •.. .)f which were developed I'm and widl run on a TRSS4 MNKslol I or Ill microcomputer. 5-. ont. and 48K memory. If a histogram of a suspension population is available. use of the programs beyond this Handbook can be male. however. If more than 100 fiailures are to be input. The programs can be used to generate estimates of the Weibuil distributior parameters for samples of size less than or equal to 100. 184 . both ... in t he case of the FORTRAN program. with no printout options. They take approximately lOK bytes of RAM. at least one d. will require a FORTRAN compiler.5 -'77-7' %• .77~-777-7 ".. for those example canes that ate run with histograms from this Handbook. _ . in BASIf' and one in FORTRAN... programs will. It should be noted that. . Of p 1 ecitl note is the maximum . The parameter estimates should be close. the dimension statements at *he front of both programs should be increased.ieters of a Weibull (listrihution by the rank regression and maximum likelihood technique. where the individual times on each suspended unit were used for the . hut. 5problems .% . 5. Botl. "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%.5. first the programs should be checked thoroughly by the user.ikelihood parameter estimation capability in both programno. Two programns are provided. estimate the parav. however.iput.a *4 The programs run in the "immediate" mode.. examples in the Handbook. These programs illustrate the solution of the maximum likelihood equation (Appendix 1). Of course.e. i.-.k drive.

IN(100)).+(.1=1 TO MM TI(J)=PE/2.ACE THE NUMBER OF 'ELEMENTS IN EACH INTERVAL OF" PRINT"THE HISTOGRAM IN CC 1-10 W/DECIMAL" PRINT"USE -99.SIc(' Pro..0 GOTO 410 IF A(I) = -99999. USE -99999.: NEXT I:PRINT IM=0 FOR .A(100).0: M=0: SX=0..J)=ABS(AW): NEXT J GOSUB 5000 rOR I=1 TO N: IU=IZ(i): X(I)=A(IU): NEXT I FOR 1=1 TO N: A(I)=X(I): NENT I BI=BN+I: DJ=1..0: SY=0.0: XY-0. . DATA ORDER MEDIAN RANK" FOR K=l TO N 185 .XZ(100) I)IM \'X( Iom. GOTO 430 GOTO 360 I=I-I GOTO 360 N=I-1 BN-=N+IM FOR .TII t100).0 PRINT"PT.IO()100).A. 10 20 30 40 .J): V(.0: XX=0..X( 100) IP 1) IM 0 50 60 7O 80 90 100 110 - t N4 120 130 140 150 160 170 180 190 200 210 220 230 240 250 260 270 280 290 300 310 320 330 340 350 360 370 380 390 400 410 420 430 440 450 460 470 480 490 500 510 PRINT"ARE YOlU INPIUTTING A HISTOGIRAM OF S1JSIPENSIONST?" PRINT"ANSWElR Y OR N" INPtUT A$ IF A$="N" GOTO 310 IP=I PRINT"AN INTERVAL SIZE OF 50 IS ASSUMED" PRINT"0.IZ(I 00).raim for Weihu!l I'arrnet'trE.V'(100).J-1)*PE IM=IM+IN(. TO INDICATE" PRINT'"THE END OF DATA (NEGATIVES INDICATE SUSPENSIONS. TO INDICATE THE END" M=0 M=M+I INPUT XI IF XI = -99.'I'T(P10).0: 13J=0.I=1 TO N: AW=A(.I): NEXT J PRINT"INPUT THE FAILURE DATA AND SUSPENSIONS IN" PRINT"CC 1-10 W/DECIMAL. GOTC 250 IN(M)=XI: GOTO 210 MM=M-1 FOR I=1 TO MM: PRINT IN(I).7w -: 7-76 1 Listing 1. ANSWER Y Ol N" INPUT A$ IF A$ < > "N" GOTC 170 PRINT"PLACE THE INTERVAL SIZE YOU WILL USE IN" PRINT"CC 1-10 W/DECIMAL" INPUT PE PRINT"PI..s/itiatitpnn I)IM ]SA(100)..K .0: YY=0.tIA(100)." PRINT"UNLESS A HISTOGRAM WAS INPUT)" 1=0 I=I+l INPUT A(M IF A(I) =0..

0 COTO 660 IF A(K) =0.Ibj."D.ST.30 640 650 660 *670 680 681 690 700 710 .I TO MMI IF TI(.S) OR N(O)" INPUT A$ IF A$="N" GOTO 840 NF=0: PRINT"PLEASE BE PATIENT.0 GOTO 2050 NF=NF+I XZ(NFi=A(I) NEXT I OT=.' 'mjlimuv 4) 52 *550) .J+D.0 COTO 840 IF A(K) ->0.M) < 0.i) < A(K) THEN IM~IMi IN(.3)/(IlNE-4) XI =LOG'(A(K)):YPz 1.01 GOSUJ3 3000: YB=AU *820 *821 83% 8:11 9:32 8413 8414 815 &36 840 2000 2020 2030 2040 2054) 2064) 21:30 186 ." ETA="."BK=".J)/(BIl-BK): GOT() 720 BJ=B.J=(BI-13.RQ IPRINT-DO YOU WISH TO DO MAXIMUM LIKELIHOOD ESTIMATION?" PRINT"ANSWER Y(E.D.001: NC=0: DX=-.0 PRINT "R= ".= (BE*SX-SY)/GM:AV=~AL/BE:AV=EXI)(AV):ST.001: DY= .720 7310 740 750) 760 770 780 790 *800 * -810 IF IPil AND K--I THEN D~J=(II1-BJ)/AH1-BK) 6W ! K~lGOTO 6:30 IF IS(-K)=IS(K-1) GOTO 6:30 D-I=(B1-B.'.0OG(IOG(YIl)):Y'X(Ký) =Y REM PRINT: "B.".ý'QR((XX-SX*SX/CM)*(YY-SY*SY/CM)) R=XN/DE RQ=R*R IF' RQ >1.BK. B4Ii.0001: NL'=l00: XB=BE: YA=-.J)/(B1-BK) IF A(K) < 0.IT'S ITERATING" GOSUI3 2000 END FOR I=1 TO N IF A(I)<0."B1 =".50IF II' (110T() 180) 540) FOR V.RA(K) M=M-II SX=SX+X1:XX=XX+X1*X1:SY=SY+Y:YY=YY+Y*Y:XY=XY+X1*Y NEXT K GM7=M BE~= ((M*YY-SY*SY)/(GM*XY-SX*S.J=".O(K).( IF (XX-sSX/(X.-RA(K)):Y=1.J: RA(K)=(RO(K)-.B13 PRINT K.J=".J: RO(K)-B.0 GOTO 670 D.1 I' grajr~ n f~pr WlEjibUi/I / ira 'n(( Fs irmiorl b.AV RmI).BJI./(l.0 590 610 620 6..A(K).R.J) 56CIA NEXT IM 1) J1 570 IS(K)=!Mi 580' BK=IM+K 581 IF IP=1 THEN BK=BK-1.1." R SQtJARF.1.ingji 1.0 THEN RQ=1.R.0 COTO 820 XN=~XY-SX*SY/GM DE.LIE PRINT "I3ETA=".Y) At.

0 : S2#=0.25 GOTO 4250 BB#=X2#*1.4 GOTO 4250 BB#=X2#*1.0/BL) PRINT"MAXIMUM LIKELIHOOD ESTIMATES FOLLOW" PRINT'BETA=".N MAXIMUM .0 : S3#=0.-"- Listing I.2150 GOTO 2130 IPRINT"ITERATION FAILURE" PRINT"BETA=".TL 2280 3000 3010 3020 3021 3022 3023 3030 3040 3050 3060 3130 3140 3150 4000 4010 4020 4030 4040 4050 4060 40M0 4080 409C 4100 4110 4120 4130 4150 41G0 4170 4180 4190 4200 4210 4220 4230 4240 4250 4260 RETURN SI#=0.6 GOTO 4250 BB#=X2#*.02 RETURN IF NC > NL GOTO 4300 X2#-BB# D2#fYA-YB IF ABS(D2#-DY#) < .0 GOTO 4290 IF (NC-I) < = 0 GOTO 4040 GOTO 4090 DX#-BB# DY#-YA-YB NC=NC+1 BB#=BB#1*.0 IF XB>15.0 THEN XB=0.IKELIHOOD=".NEXT I IF IP=O GOTO 2250 FOR I=I TO MM:SU=SU+IN(I)*Tl(I)[BL: NEXT I SU=SU/RN:TL=SUI(I." ETA=".1 FOR 1=1 TO N PO#=TT(I)IXB SI#=S1#+PO# S2#=S2#+LOG(TT(I))*PO# NEXT I IF IP=0 GOTO 3130 FOR I=1 TO MM : PO#=TI(I)[XB: SI#=Sl#+IN(I)*PO# S2#=S2#+IN(I)*LOG(TI(I))*PO#: NEXT I FOR I-1=TO NF: S3#=S3#+LOG(XZ(I)): NEXT I AU=(S2#/Sl#)-(S3#/NF)-(1.2140.0 GOTO 4250 IF BB#<X2# GOTO 4190 IF BB#=X2# GOTO 4240 IF BB#>X2# GOTO 4210 IF BB#/X2u > =.XB." !.00001 GOTO 4320 BB#=X2#-D2Vt*(X2#-DX#)/(D2#-DY#) IF BB# < = 0.75: GOTO 4250 IF BB#/X2# < 1. OR XB < =0. BASIC Program for Weibull Parameter Estimation (Continued) 21 :L5 2136 21:17 2140 2145 2147 2150 2160 2170 2250 2260 2270 GOS1B 40X)0: XB -BB# ON JK (oTro 2130.7% 7C.0/XB) RETURN JK=1: BB#=XB IF (ABS((YA-YB)/YA)-OT) < = 0.0: RN=NF: FOR I=l TO N: SU=SLJ+TT(I)[BL.YB RETURN BL=XB: SU=0.BL.02 DX#=X2# DY#=D2# 187 .2150.

TI'N . I r..1K 2: NC 2: RKi 'IRN IRIINT"FAIIEI) TO CONVEI{R.N.'lionation (Conttinued) 4270 4280 4290 43WN) 43110 4320 NC NC. !* .E" .1K -3: NC -1: RTUI'tRN JKA4: RETURN 50O) 5010 5020 5030 5040 5050 53W60 5070 5080 W90 5100 5110 FOR . BASIC' Iriogrntm loor Wrihull Parameter E.J=1 TO N: IZ(J)=J: NEXT J IF N=I RETURN NM=N1FOR K=1 TO N FOR J=l TO NM NI=IZ(J) N2=IZ(J+I) IF V(NI) < V(N2) GOT3 5090 IZ(J+I)=Nl: IZJ)=N2 NEXT J NEXT K RETURN 188 a-. W . "M Listing 1.

)Rii)ERI(10H)) DIMENSION YX(IXE)).isting 2.J= 1.'ANSWER Y OR N') READ (5. .MM) 798 FORMAT(8F10.NE.'INTERVAL OF THE HISTOGRAM IN CC 1 ./ IPRNT=I IPOP=0 ISUM=0 WRITE (6.4006) FORMAT (2X. ?/. F•Y)R''RA.?. TO INDICATE THE END') 4001 4002 4003 4004 4005 8801 4006 m=0 M=M+I READ (5. USE -99.EQ.TIME(I00).4003) FORMAT (2X.MM.i.0) CONTINUE WRITE (6.INT( 100). ?/.4005) PERINT FORMAT (F10.'PLACE THE NUMBER OF ELEMENTS IN EACH' ?J. -99.2000) 211 189 .'CC 1 ./. ANO) GOTO 701 IPOP=1 WRITE (6.2X.4002) ANS IF(ANS .V(I(X)) DIMENSION IZ(100)..TE (6.N Pr g#ram for Weihill Parameter Estimation (C'ontinued) COMMON /BiO(CK I/INT.' ARE YOU INPUT'IING A HISTOGRAM OF SUSPENSIONS?'.RANKMI)(Ii)).'PLACE THE INTERVAL SIZE YOU WILL USE IN'. ?'W/DECIMAL.K.'. 1 WR!TE (6.+(J-1)*PERINT ISUM=ISIJM+INT(J) 2001 CONTINUE C INTERMEDIATE PRINT WRITE(6. ANO) GOTO 8801 WR.1007ýXINT 1007 FORMAT (F1O.0) IF (XINT .'O.2X.2X.1) C INTERMEDIATE PRINT 701 WRITE(6..2X.A( I0xU).798)(TIME(J).TIME.IHIST DIMENSION ISt I00).10 W/DECIMAL') READ (5.KL= 1.'.'AN INTERVAL SIZE OF 50 IS ASSUMED..4004) FORMAT (2X.2X.MM) 3090 FORMAT (1014) ISUM = 0 DO 2001 J=i.X(AGW) DATA ANO'N'/.4001) FORMAT (2X..ANSWER Y OR N') READ (5.PERINT/50.MM TIME(J)=PERINT/2.10.3090)(INT(KL).4002) ANS FORMAT (Al) IF (ANS ..EQ.) GOTO 212 INT(M) = XINT GOTO 211 212 MM .M. ?/.

I)BK=BK. -99999.101) AMI FORMAT(FIO.AND.'-X.MM IF (TIME(. 0. ORD(V. IS(K-1))GOTO 1911 3901 DJ=(BNI-BJ))/(jBNI-BK) 3911 IF WAK))390. =0 SUIMX =0.X.990)) FORMAT (2X. 'MEDIAN RANK') DO 634) K IN ISI IM=0 IF ([POP .0 N.EQ.)GOTO 2 GOTO I 3 1=14i GOTO 1 2 N=I-i BNý N+ISITM DO 4 i-I=1.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.EQ. 101 INPUT' READ) (.0 SUMXY. tJ NLESS HISTOG RAM' .SIJMYY--o.'NDICATIES THE ENI) OF D)ATA (NEGATIVFES' ?.40X) 390 DJ-(BNI1.'IN COlS 1-10 WITH DECIMAL. .0) IF WADI -EQ.1.J=1I.1T.5.' ?.EQ. 0) GO TO 3911 IF(IS(K) .900.EQ.0 WRITrE (i. A(K))ISUM=ISUM+INT(J) 631 CONTINUE IS(K)=ISUM 632 13K=ISUM+K IF (IPOP .N 4 V(Jl)=ABS(AGJI)) (CALL..BJ)/(BNI1-BK) GOTO 630 990 190 .0 Ki 0.4X.'.J . I)DJ=(BNI-BJ)/(BNi-BK) IF'%K .2tXX) FORMAT (2X.EQ./.2X.0 SUNIXXý-o0. K ./.N.1.0./2XINI)ICATE SLt SIENS IONS).0 IF(IPOP .'ORDER'.EQ.'INPU'1T THE FAILUIRE D)ATA AND StUS PENS IONS' ?.EQ. 1) GOTO 3911 IF (IPOP .4X.0)GOTO 3 IF (AMI ..EQ.0 StIMY ~-.4X. 1 .'PT.W. -9999%.I.'DATA'. 0)GOTO 6312 DO 631 .

SUMYY.SUMXY 800 FORMAT (2X.7) GM=M BETA= (GM*SUMYY-SUMY*SUMY)/(GM*SUMXY-SUMX*SUMY) ALPLN=(BETA*SUMX-SUMY)/GM AVED=ALPLN.'THE FOLLOWING ESTIMATES ARE RANKED REGRESSION'. 1.LT.2X.0 IF((SUMXX-SUMX*SUMX/GM) ./IETA ETA=EXP(AVED) WRITE(6.5001) 500) FORMAT (2X.800O)SUMX.GT.N.'SUMY='./(1-RANKMD(K)) Y=ALOG(ALOG(YPRIME)) YX(K)=Y WRITE(6.N 400 *1 * * .7.7) R-0.'ANSWER Y OR N') READ (5.5) WRITE (6.0)R= 1.E20.' DO YOU WISH TO DO MAXIMUM LIKELIHOOD ESTIMATION?'.7.Fl2.' ETA ='.5) M=M+1 SUMX=SUMX+XXX SUMXX = SUMXX + XXX * XXX SUMY=SUMY+Y SUMYY=SUMYY + Y * Y SUMXY=SUMXY + XXX * Y 60CONTINUE CWRITE (6.5. FORTRAN Program for Weibull Parameter Estimation (Continued) 3J B.'BETA='.J RAN KMD (K) =(0RD ER(K) -.SUMY.0 781RSQ=R-R WRITE(6.DO 191 .3101) 3101 FORMAT(2X.TML.'SUMX='.SUMXX.Fl0.Listing 2.ORDER(K).RANIVMD(K) 300 FORMAT (16.7. 1.E20l.F10.A(K).'SUMXY='. ?'SUMXX =".'SUMYY='.IZ(l) 1 J1.IZ) DIMENSION A(1).ETA '3100 FORMAT (/. ?'ESTIMATFS') WRITE (6.F1O.3100)BETA.3)/AB N+4) XXX=ALOG.E20.J.4.IPRNT.4.F1O.(A(K)) YPRIME= .BETA) 90CONTINUE STOP END SUBROUTINE ORD(A.RSQ 3200 FORMAT (2X.7.'R-'.2X. ANO) GOTO 900 IHIST=IPOP CALL MAXL(A.3200)R.300)K.EQ. ?/.Fl0.0)GOTO 7871 XNUM=SUMXY-SUMX*SUMY/GM DENOM=SQRT((SUMXX-SUMX*SUMX/GM)*(SUMYY-SUMVV SUMY/GM)) R=XNUMJDENOM 781IF (R .J+ DJ ORDER(Kh13.2X. 0.N.4002) ANS IF (ANS .BML.E20.'R**2='./.E20.E20.

%MP TT(I)=ABS(T(1.BETA='.) N2rýIZ(.40.PB.0OW00 NLIM=100 X=BETA PB=0. I)WRITE(6.001 NCT=0 DIELX=.01 30 PRN= AUX(X) IF (IPRNT .INT(I00). 1)WRITE(6.206)X.EQ.ISIG GO TO 30 40 -IF (IPRNT .205)FPR.XX(100). A(N2)) GOTO 2 ~IZ(.LT.E20.' ITERATION FAILURE '.NM .j IF (N .RN=NFAIL *50O .EQ.NS.'HIST COMMON /BLOCK2/XX.J+ 1) =NI IZ(.NUM.NFAIL IH=IHIST NSAMP=NUM NFA IL =0 DO 1 I=I.EQ.5.TIME.3E20. 1) RETURN NM -N -I D)0 2 K -1.NUM SUM=SUM+TT(I)**BML IF (IHIST .TIME(IO0) COMMON /BLOCKfl'T.N DO 2 J= I.BMLTM!..E15.+l) IF (ANI) .ODO)GO TO 1 NFAIL=NFAIL+! XX(NFAIL)=T(I) I CONTINUE~ TOL= .ODO .M.7) IFLAG=1 RETURN BML=X SUM=0.M.T(1).BETA) DIMENSION 771 (l00).44 4 * N I -IZ(.* 110 111 DO 110 I=1.)' .001 DELY=.NSAMP COMMON 1131.J=kN2 2 CONTINUE RETURN END SUBROUTINE MAXL(T.PRN 206 FOMT2.Listing 2.MM SUM=SUM+FLOAT(INT(I))*TIME(I)**BML 192 . 4 FORTRAN Program for Weibull Parameter Estimation (Continued) 1 IZ(. C) GO TO 112 DO Ill I=1.50.EQ.7) GO TO (30.IPRNT.PRN 205 FORMAT(2X.) IF(T(1) .' LN MAXLIKELIHOOD= '.LT.OCKIIINT. 0.50).

DEL1.NMTNTIMELOOP) SUBROUTINE SLOPE( X .EQ.0 SUM2=0.MM. 0.TIME(100) COMMON /BLOCK/T.MM SUM1 =SUJM1+FLOAT(INT(I))*TIME(I)*-X SUM2=SUM2+FLOAT(INT(I))*ALOG(TIME(I))-TIME(I)-'X DO 15 I=1.4 5 X2=X DEL2'=YA-YB IF (ABS(DEL2-DELI) .) ~WRI'I'(6.INT(100).' BETA='.7. 0)GO TO 11 DO 20 I'=1.8.02 GO TO 9 2 IF(NCT-NTIME)b.6)13. YA .F1O.TIMV.996) 96FORMAT(2X.75 GO TO 8 12 IF((X/X2)-l.0/BMI.12 11 IF((X/X2)-.IHIST COMMON /BLOCK2/XX.' ETA='.0/X RETURN * END . 1.TML 995FORMAT(2X.0 .4)8.LE.OR.5.13. 15.2) RETURN N * * END 10 "I: A 20 11 15 FUNCTION AUX(X) DIMENSION T(100).XX(100).F20. X .6.YB JK=l IF((ABS((YA-YB).1.GT..E-06) GO TO 20 ~X=X2-DEL2*(X2-X1 )/DEL2-DELI) IF(X)8.2 1 Xl=X DELI =YA-YB NCT=NCT+1 X=X*1.LT.TOLJK.SIM*-(I.' MAXIMUM LIKELIHOOD ESTIMATES FOR THIS CASE FOLLOW') WRITE(6.10 10 IF(X-X2)11.N SUMI=SUMI+T(I)**X SUM2=SUM2+ALOG(T(J))*T(I) 4*X IF (IHIST .O SUM3=0.1 DO 10 I=-1.3.0)X=0.8 13 X=X2*.995)BIM.14.N COMMON /BLOCK I/INT.Listing 2.NFAIL ZZZ=NFAIL SUM 1 =0.3 -3 IF(NCT-1)1. FOKIl'RAN P~rogram for We'ibull P'arame'ter &Rstimation (Continued) 112 * SUM 7-SUM/RN TIML .Xl.14 193 .'YA))-TOL')6.0 IF(ABS(X) .NFAIL SUM3=SUM3+ALOG(XX(I)) AUX=SUM2/SUMI-SUM3/ZZZ-1.

?El4 9.4X.'CONVERGENCE FAILURE IN LOOP'.8/ JK=3 NCT= I CONTINUE JK=4 RETURN END TO 9 S(GO .r .8..YB 100 FORMAT(IH0.0}2 XI=-X2 DELI =DEL2 NCT = NCT+ I GOTO 9 6JK=2 NCT=2 '8 " i4 ..100)LOOPX.X2"1. i • S~RETURN S20 I €9 WRITE(g.'X ='.25 7 X . FORTRAN Program for Weibull iParaineierEstimation (Continued) V.'YB ='.4 194 .(% TO 4.I2/1H ..E14.W*V .4X.i ~ Listing 2.YA.E14.' YA='. J •'14 X -X2*I1.

.1195 4.APPENDIX I WEIBULL GRAPHS Ii . Op a 195 : - 7• ..

2 4 556789 1. * 30.5 35 78a911 196 D2M .0 40.0 06.11WEIBULL 99.2 * 40.0 20.0 /T 0.5 3 4689 1.5 0.5 99.0 90 Beta DISTRIB3UTION Etasample size Failures 0Key: i o 1i - 63.99.1 1 .

-.flwfl 197 . .--.-- a CIA- 0 ID 00 0 C) )Q 0 C c q co OtCYOA GO P o ( 6Og C) 0N 0 C WUBOJ9d GAfl9. 7 ti H H+ t I 'i iI t ---- I I f It I-IItf ! ++ 1 -tttt l -.7. .

0 * * "U 0 * a.0 I_ _ _ ____ _____.. -.5 -- 1-. .. 05. .0 Beta= Eta i "Sample Size -Falure3 i = 1 to3 80.... 3.i_•.0 90.5- - 3 4 5 6 FU 26WI 198 .. t .0 40..0 ~4.4- - o. -. .____ .0 " " -.-_ mz Lzz . .- - - 0 i1 - ii .2 A~- ~j _ - 60.0 63.. 2--- .0o - - -- "50.. _.._. 99.' t .. . I~ - 20. .0 4 70.. 1 !/• / WEIBULL DISTRIBUTION 4..Key: 95. . - - -- 01 -- ~ 1. .0 30.9 i .. 10. * - 0.. .

APPENDIX J ANSWERS TO PROBLEM'S i. CHAPTER 2 ANSWERS 2.5 j=200 hours b. CHAPTER 1 ANSWERS: None 2.2 25.0 54 hours. -.ý 74. .8 64.1 Probem 2-1 a. Yes.1 54. • - \\ >. From median rank table &-t Figure J.8 35.5 45.. B 1.1. 199 -.1 &3.6 % 16.7 93.3 h . # =3. Time Median* Rank 1 2 3 4 5 6 7 8 9 10 (hr) 85 135 150 150 190 200 200 240 240 250 Ranks 6. K. Wearout or fatigue failures usually have steep slopes c.

.4 • I J1 11 1 .5 WERROUT "= 200 HOURS a. SSWLE SIZE1. -ier -'- -1 20 A 'I - 'NK I ..-. -- 4 .0 54 HOURS U.. EIBULL DISTRIBUTION 4' I - j F = 3.. 1.= 10 FAILtJRES = 10 B1.. II Si' - I Fiur...

2 Problem 2-2: Answers a.2' 36.2 .3 7.1 .. A batch problem mow be suspect.32 7. '.9 59.75 q = 29ghours b.4 .9 - U.4 + 1.27 a 5. (N + 1) (previous rank order number) 1 + (number of items beyond present suspended items) 7 5 1. ew Rank SNew Rank .. Infant Mortality C.87 .87 1+2 . Median Rank 8..2 Rank.3.5 82.spension Failure Sizapension Fa.4 Cank Increment . Serial numbers are very close..3.&ure Fa&' ire New Rank 1 2 3.3 + 1.2. if 0.7.17 See Figure J. - '..3% 20. - . I- S~201 • _.87 .4 = 9-2= I+4 =2 + 1. ._• .5. 1 2 3 4 5 6 7 8 Rank Increment - Comment Failure Failure Suspension S.5. .2..5.4 .6 3 - 1.4' 5. .

. .. . . . . .7 I w.. .75 = 29 HOURS SAMPLE. . . Infari Mortality 202 . . 2-2...: I. .RRE VERY CLOSF MAYBE A HEAT TREAT THAT */~! S...__ I I THWE SEIA NUMBERS OF FIRST TO LAST FAILURE-.EIBLLL DISTRIBUTION P = 0. . .SIZE= S FAILURES =5 INFANT MORTAL I TY - - - - - - - - - • - IL. ... TOTAL OPERATING TIME (HR) FO 27189? Figure' J. u.2. . I'ruhlu'n.

200 .425 920 104.9 79.8 139.200 90) 90) t.0 44.8 1.la (t3where t2) .) 55.k) (95) 1.3 20.4.200 -SeeFigures J.ti) tJ = 90 t 2 .3 and J.8 444.95 96.edi-in renn Using the formit.8 w 185 = 1858S Time 90 hr 130 165 275 370 525 1.7 Rank 1 2 '1 4 5 6 7 8 Time 90 hr 130 165 220 275 370 525 1.8 i 203 • .8 49.8 289.7 S220 9. 185) (185185 -_1.8 91.200 table "From m.2.015 .185) -(185 -185 -(1.9 67.0 44.119.42 .9 67.3 20.0 55.2 hr Median* Rank 8.1 32.185 t3 = 1.01.9 79.8 84.8 91.8 194.1 32.3 Preblem 2-3: Answerm Median * Ranks 8. Time - 80.200"1.

IBULL DISTR I[BUT IONhT SD0 = CAN'T BE ESTAB. . 9 = CAN'T BE ESTAB.. N'V 1--44-~ -4-4 W.T T .. r FAILURES = 8 .. '. low.A ..SAP. FOATLUTIME TOTa. "". =CI =10 (H8 2O . 'LE S IZE =8 T.. St8U 2204 DSRIUTO T BE iES AB.

'p.. Overall Population 205 . 'i .EIBLOL DISTRIBUTION . 1 00..2 ior =279.7. FAILURES =8 SAMPLESIZE=8 8 =80. /4 . Problem 2-3.4.o.1 -~ -" .I . S.T I. '. Figure J. isl .HR a. .

24 1.042 0.74 2.704 -1.067 0.118 F(t) .& CHAPTER 3 ANSWERS 3.0155 0.16 0.087 0.1 (a) Problem 3-1 'T'oday NumI7Lb(er Engines 20 20 20 20 20 Time on Each Engine 1.909 = 4. atdditional failures 6.0266 0.35 ff 6.704 Therefore..84 S=1.8 2W6 .N 0. N 0.533 0.909 (b) In Six Months Number Engines 20 20 20 20 20 Time on Each Engine 300 350 400 450 500 F(t) 0.533 0.008 0.31 0.042 F(t).50 200 250 300 350 F(t) 0.0266 0.0033 0.84 1.062 0.

13 po:ihlation: P"Ioint i)alta Mean Order Median Rank "47 48 58 59 60 S61 62 72 7:3 74 75 76 85 86 87 88 89 90 91 93 100 476.000 681.14409 0.02442 0.0M 667.000 649.04545 0.03741 0.07958 0.16407 0.841 29.6.000 756.3:14 2.352 0.0()0 596.000 60:3.067 2.388 21.000 755.769 24.240 4.18271 then.000 770.137 17..05839 0.000 855.09361 0. overall using Figure 31.RM F. .000 701.000 701.-.455 13. they "seem" different.02134 0.000 812.133 3. . 20 .(XX) 576.12525 0.532 31.06582 0.000 850.000 Mean Order 1.7 3..242 33078 0.14 population Point 3 4 6 7 8 9 L.16294 0. i/ .000 741..17247 0.08828 0.09698 0.000 845.01330 0.697 19. Figure J.19930 The associated Weibull Plots are in Figures J.000 806.346 5.016 14.5 and J.000 770.151 27.000 600.576 16.7 illustrates the total population Weibull with confidence bounds (from Chapter 7).000 821. .05348 0.687 4.079 22.000 821.128 5.06218 0.570 7.02937 0. the two Weibulls are significantly different..000 812.10640 0.011 8.000 1.452 6. since the Location A Weibull lies outside the All Locations Weibull.453 9.11582 0.000 845. However.460 26. using the Figure 3. .000 684. .12884 0.13467 0.0M) 504.07088 0.894 11.2 Problem 3-2: Turbine Airfoil Unscheduled Engine Removals (a) First.558 Median Rank 0.ta 384.15.

078 IEIBULL DISTRIBUTION 21 - - l l/ I I I II I i l!I II l 25 448 1i148 -• - . •. TOTAL OPERRTINQ TIME (HR) Figure J.--. - --.S*LE SIZE = 179 5 = 5. -.634 V=1107. Problem ...5.FAILURES -L = U.s .4. U....-Overall Population -V7 208 .

6. -....FAILUMS=6 6V W E113LLL DISTRIBUTIONr S.....610 = 981. 1~. U I I I fl li I " 9... .-. - ... . . "- - -- .~' .- 'I.P.I 0.01 hi 10 TOTAiL OPERATING TIME (HR) _ _ __ _ _ _ _ _ __ _ _ __ _ __ _ _ _ __ _ i_ \ . Problem 3-2. . ~ -. lwation A Only 209 . . iii___ FD 272253 Figure J. ..S. ... V.. .8337 S*VHLE SIZE = 31 -.. .

%ý. J.TWI S WZ W W. n. FAILURES =21 V. S..'P a.S2 210 .

0 4.0 4.0 46.9 1.0 1.450.0I 8.050.0 850.0 0.0 1.0 1.0 1.510.0 2.0 0.F(t) 0.0 6.0 0.310.0 64.0 148.510.0 1.0 1.0 152.610.310.0 0.0 1.0 2.0 1.6 0.0 7.0 1.1 0.0 .0 0.0 510.0 55.3 5.0 1.0 1. 16.0 0.0 1.250.4 3.0 H(t+360) -F(t I -.0 104.0 3.0 140.0 1. in 12 months Number of Units (N) 9.0 117.110.8 0.0 - 8.0 1.0 .110.0 7.0 2.3 0.0 96.0 1.0 39.0 0.5 0.3 5.0 0.0 7.0 2.0 155.010.7 0.550.410.0 1.9 1.0 7.910.0 1.0 2.2 0.9 1.1 0.0 97.0 1.0 1.3 0.0 2.0 1.0 1.0 158.0 1.0 30.0 0.0 6.0 L.410.0 1.0 1.) 710.0 1.0 1.3 0.0 1.0 1.0 1.0 55".0 0.0 1.9 7.010.0 1.850.0 1.0 2.0 85.0 1132.0 110.0 t 50.(t).210.0 !.0 9. 0.0 1.0 1.0 1.0 125.110.0 8.4 0.0 8.5 0.050.0 3.8 5.1 2.0 8.0 250.0 150. i 7.0 1.950.8 4.0 1.7 1.0 1.710.0 2.0 1.0 (.0 1.0 2.0 72.810.150.0 79.% 8.0 1. 6.0 1.250.1 0.7 5.0 14.0 1.0 1.650.0 1.0 610.0 8.0 FM) t 0.0 450.0 !.0 91.6 211 .0 1.0 ('umulative IInits 9.610.0 6.0 Pt 136o) 0.350.0 7.0 1.0 0.0 650.6 0.0 750.0 2.0 810.0 0.N I 7.2 0.8 0.0 1.150.(b) For the entire popla)htion.0 I 1 360 410.0 1.2 0.0 6.0 T7.750.0 i.0 1.0 1.0 2.0 350.0 9.0 9.0 1.0 0.9 0.0 910.0 2.0 157.0 950.4 0.

670.0 1.0 0.150.0 1.0 6.470.670.0 0.270.0 7.0 1.3 0.0 2.450.150.0 1.0 1.0 6. 720) .0 1.0 6.0 64.170.770.0 650.6.0 1.0 125.2 7.0 1.0 1(.9 8.9.0 1.0 1.0 7.0 1.0 1.0 2.0 (I f 720) Fi) Hi.) 9.0 6.0 1.570.950.970..0 1.0 1.0 132.0 8.0 250.0 1.)0.0 1.0 85.7 0.0 1.0 72.0 1.170.0 1.9 1.0 1.0 1.9 1.0 0..0 6.550.0 450.2 0.0 6.0 91.0 0.0 7.0 350.0 1.0 1.1 0.0 2.0 770 0 870.0 1.0 1.0 148.0 2.0 1.0 1.0 1.0 8.350.0 2..0 7.8 212 212 II~uo ____ .0 2.470.0 950.0 H.0 1.0 1.0 1.0 Fit) 0. .0 8.0 6.: 5.0 1.050.0 158.0 4.770.870.0 2.0 1.0 1.0 8.0 16.0 8.4 0.0 79. in 2.0 3.970.0 0.0 2.0 5.0 1.0 1.0 0.0 1.0 1.0 2.5 0.0 152.7 7.0 1.070.050.0 1.250.0 1.0 1.0 = 128.2 0.1I.0 7.0 155. For the enwire p4pthilai i.0 9.0 14.0 97.0 2.0 1.0 1.9 7.0 1.) 1.0 0.1 monihs .0 110.0 140.270.870.0 7.0 6.0 30.0 9.0 150.0 1.0 117.0 2.2 0.6 8.9 1.6 0.0 8.Number of nlits (N) •.0 850.250.0 9.0 1.0 2.0 46. IlInits I1 720 50.0 1.7 0.0 Cumulativ.0 1. 0.0 7.0 157.370.750.0 1.37(.0 970.I 0.0 2.7 0.0 7. .0 1.0 1.0 1.0 2.0 2.0 4.4 0.0 3.0 1.0 39.0 1.0 1.0 55.0 550.0 104.0 1.0 1.0 750.0 0.1 5.850.0 1.6 0.0 7.1.070.0 1.570.

N 0.0 1.0 1.0 1.0 0.0 0.0 1.0 1.0 0.950.0 F(t 43.0 0.7 1.0 1.0 0.0 1.510. in 12 months: Number of Units (N) .0 1.For the population at Ical.0 1.0 14 9 15.0 1.0 1.0 1.610.050.0 2.4 0.0 950.0 18.0 1.0 1.0 2.0 0.0 1.7 1.0 1.0 1.1 0.0 1.010.110.0 1.0 710.0 1.0 2.2 0.0 1.0 1Vro0.ts 0.0 1.9 1.0 910.0 1 (.0 0.0 1.- - -- .0 1.0 0.210.0 0.1 0.310.0 2.0 0.0 1.0 2.0 1.0 1.0 50.0 1.550.0 610.(t).0 2.0 21.0 1.0 0.0 1.350.0 1.0 1.0 0.0 0.0 0.0 0.850.0 0.210.0 0.0 0.4 0.0 2.0 510.0 2.0 1.150.0 1.0 1.010.0 1.0 350.0 0. .0 0.0 25.0 G(t) 0.0 1.0 1.0 F(t__ 9.0 1.0 2.0 7.0 2.0 1.0 1.410.010.0 250.0 2.0 750.0 19.0 1. 0.0 3.0 1.0 9.60) 0.0 1.0 1.0 810.0 850.0 L.0 1.0 2.0 ('umulative llnu.0 5.0 2.0 0.0 1.0 2.050.750.0 1.0 0.0 0.0 0.110.0 1.810.0 11.ion A.0 23.0 1.0 550.0 2.7 IV Ni - g 213 .0 0.0 1.0 450.0 1.5 0.0 tI3 410.0 1.250.0 2.0 1.0 1.0 0.0 1.0 1.0 1.0 2.650.0 2.0 1.0 2.0 23.0 1.310.0 2.0 1.0 1.0 2.0 0.0 2.410.0 1.0 1.0 1.0 2.0 17.0 650.710.0 1.0 2.7 1.0 "3.0 1.450.1 0.0 2.0 1.

0 0.0924 _- N.0 14.0 1.0 1.0 250.0 0.0 0.0 0.0 0.0 2.0 1.0 1.0 Cumulative Units 0.0 18.0 1. in 24 months: ' Number of UInits (N) 0.9 0.0 2.0 25.0 0.0 550.3 0.30 Number of failures = (0.0 2.0 2.0 1.000 F(4.0 1.0 870.0 1.0 0.0 2.IXX).0 2.470.0 1.450.770.000 hours.0 1.0 1.750.0 0.0 2.0 9.000) + F(1.0 1.0 2.0 15.0 750.0544) (1.970.0 1.9 i.370.0 1.0 1.6 G(t) 0.0 23.0 FYt) 0.0 3.0 7.1 0. Number of failures = (0.0 1.0 1.0 1.0 2.0 1.000) + F(I.0544 1 -.0 0.0 1.0 1.570.0231 + 0. .0 1.0 450.0 2.0 19.0 970.170.770.0 1.0 2.370.308) • 121 /21 214 .0 1.0 1.0 2.0 1.000) = 0.0 17.0 F(t f 720) 0.0 1.0 1.308) .0 0.150.0231 + 0.0 1.0 2.850.1253) (1.0 1.0 1. all of them will have failed and will have been fixed.0 1.0 1.0 1.0 1.0 t 50.0 1.0 1.6 G(t).0 0.0 1.0 1.050.0 2.000 F(2.0 1.650.0 2.570.170.0 1.950.0 350.0 650.0 2.0 1.0 t t 720 770.0 1.0 1.9 1.F(t) F(1.0 2.e "/p' equal I - e = (0.0 I.000) + F(1.0 1.0 2.0 1.9 1.0 1.0 1.0 1.0 1.000) = 0.050.0 1.1 0.9 1.0 M= 24.0 1.670.0 1.For the plpulation at iACation A.0 5.000) = 0.0 1.000) = 0.0 1.0 1.250.0 2.0 1.0 0.0 0.0 0.0 13.0 1.070.0231 = 0.0 1.0 850.550.350.0 2.N 0.2 0.0231) (1.0 2.0 1.0 I R).0 0.0 0.0 -1.0 1.0 0.870..0 1.0 1.0 1.0231 + 0.670. M3 Problem 3-3 (A) .3 0.000 hours) = F(1.e.1253 Number of failures = (0.0 1.308) = 164 (D) Inspection at 1.0 2.070.0 11.0 1.0 1.470.270.0 950.0924) (1.0 1.0 0.0 2.0 1. makes units "good as new" P (failure at 4.0 2.5 0.270.0 1. = I.0 2.308) = 71 (C) t = 4.0231 Number of failures (B) t = 2.0 2.0 1.0 2.0 1.0 1.0 2.0 21.0 1.0 1. 1 0.0 0.

.000 hour inspection is 900 hours or 36 months from present Reset Mode A and C to "0" ((T lpA.66))rV 1.8.877 hours - -)Y"'-4 996 months from lint inspection 789 hour .1088) (1. . 4.9bours 40 months from preent 1.996 (in I ))h/&494 215 .149 months from present 4A5&82ourm - 190 months from present IF. - 96AW (Inl( -6W~~) - 28.11./.308) IMA(2.0544 + 0.1188 142 1 Start with the top row of random numbers.0544 i F(2. F...A.126(tn( - .._ .4 >. of interest (2) Engine at 200 hours .000 hour inspection is 40 months from last inspection Past 48 months pt. (1) Engine at 100 hours I.2 hours 1.760 mouths from present -.8A1. .96.8874 24. '~ ~' ' • (R) lnsixe'tion at 2.214 bous .(XX0) 0. - 1. and note that one's answer will vary depending on the random number string.(NN) hours P (failure at 4.- 96W(In )1471 219..31 months from lost inspection 2. 7.00) 0.(NN) hours) Number of failures U3 Problem 3-4 (0.

.126 ~(In 36 months from last inspe-tion 916 hours - 2. - ~.47 4.000 hour inspection is 40 months from last inspection (3) Engine at 500 hours F4 . 3D8 months fre"n last inspection .566 hour. hours -3.587 (in.126(n( 1 - 1.000 hour inspection is in 800 hours Reset Mode A and C to "0" F.. - 9. .- 1. ... - .96.. .575 months from Iment 4.122 hours - 36 months from present 1.12( (in (I- 774V))'- 1188hours 47 months from last inspection 2.025 hours - 21 months from present 1.000 hour inspection is in 500 hours = 20 months Reset Mode A and C to "0" FA " 96.()) .000 hour inspection is 40 months from last inspection 216 Il* e.319 hours 164 months from present F..99-W-- -O )" -.- 1.587 (In ) - 3.*.88 3.. . -- - °- °'.. = :2 months from present - 96. 122 months from prosent K.587 (fn 897'.96S hours .325 (In- 1.994 hours - 159 months from last inspection F.S*-4' . .

rto o F.530hous.000 hour inspection is ir.2 ~ 47 months from lastinspectioa 2...ii~(n t 1..(4) Engine al 70X) hours FA 96$.. 14 mouths from present 1.33 months actually Reset A and C to "0" FA .126 ( n In . ..270mo.ntba. . .87 (9...000 'our inspection is 40 months %iom last inspection Inspection before failure A) No Failures B) 5 engines fail without 1. .. z• ' _ _' " " : "* . .1"( (I( ))"" 231776hour -1.78montsfrompresent 263 months from present 4 months from preset J.480 hours 1. . - 1.l85hours. .072 hours - - 164 months frem present F..L" F.4.671 months from present F. ))' - 45. .--~w~-)'' hours -47 months from 1I-t inspection 2.. 4..- 92.n (---- ))' . .48houos - 3. - - 2..996 (In ( ))'" - - 4. .96. ' .000 hour inspection I additional engine fails before the 2.000 hour inspection is in 100 hours = 4 months Inspection before failure - Fc = 4.. . • .000 hour inspection 217 "•.996 (in (-. ! .• . .O7hours 1.) - 1.813 hours 1...587 (f.000 hour inspection -a 40 months from last inspection (5) Engine at 900 hours .126 (n . - 96.197 9. 300 hours Reset Mode A anc C to "0" = !2 months ) V...- 96.587 11'"(- 70.- 1- monthsfrom tinspectio -- F. .008 hours - - 1. - 7..

134 0.717 0. one would expect 3. .045 0.. of Failure at 100 Let T. TIoD = = Prerent Time Time in 100 Hours Hours in Future Engine 1 2 3 4 5 6 TP 40 90 100 110 0 0 = 0.540 0.-. 2.2 Problem 4-2 The predicted design B.134 0.681 0.12 = 5 Consider the case when 1 = 2: 10o' + 110 + 1252 + 150' + 90' +40' a 187.0015 0.249 F7V4 Thus.429 0.34.1 life %wsume = 1000 a -3. 950o9 + 405 1 1.. . . '.3 LI P . o - .997 0. one would expect 2. 218 .969 2.0.0 0 0..5 . '5 5 ".207 0.. CHAPTER 4 ANSWERS 4.082 0..1 Problem 4-1 Leti.249 0. --2.972 0.0.794 Consider the case when ." 4.0 PIow = Prob of Failure at Present P 10o = Prob. a 2 = 5: 2 u- 1001 + 110 + 125 + Engine T.18 to 3. 4.34 additional failures during the next year. b . 147. T___ I o P__ 1 2 3 4 5 6 40 90 100 110 . the number of expected failures during the next year increases from 2. As 0 increases from 2 to 5.969 ..990 0.293 0 0 140 190 200 210 100 100 0.207 0 0 140 190 200 210 100 100 0..134 Thua.18 additional failures during the next year.4245 So: = 3.767 ..644 0.249 0.

21 9. . . .X . .. 9. ".001 =1 i .1 ..7)3 =fi0.-:" '.. one can conclude that the present data is insufficient to increase the predicted design life...' .999)1/ L =384.999 3845.. /. . ... . 9. .4.'"9" .1 life = 384..6 (which is less-than the predicted design value of 1000).. .. ".. ' •. .7 (-In 0... . .999 -(L/3845. .7) r L = 3 = In 0.e-(L/3M'T) e-(L/38.6 Since the Weibayes analysis predicts B.7 = B. .• .. ••" . . ...rT5hen(K)). .. . .• o .1 life predicted from the Weibayes analysis of the data. .+ 5(2000)' I Let L Then: 0. 9.1 i- 9.

"' Thus.2 Number of Turbines Required Table 5. First. with vane erosion within the allowable limits. then the failure mode has been significantly improved. 5. 220 . this reliability goal to a characteristic life gal: substitute t = 2300 cycles.95 (95% succeeding. Mhe corresponding entry is 0. the zero failures test plan is: run 20 bearings for 711 hours each. The only constraint was that it should not exceed 5000 cycles. and the third plan requires 5 x 5000 = 25.95. 5.2 3000 4000 . 5% failing) at 2300 cycles. confidence. The table below shows the number of turbines required.1 CHAPTER 5 ANSWERS Problem 5-1 Substantiation Testing Enter Tlble 5.5000 0.5. the second plan requires 9 X 4000 = 36. 4000. R(t) = 0.5.237 x "000 hours = 711 hours. with 90%. The number of test cycles per turbine was not fixed.65 0.000 cycles. the test plan that satisfies the test requirements and that requires the fewest total test cycles is: test 5 turbines for 5000 cycles each. However. Test Cycles per Turbine 2300cycles 1/3 Ratio of Test Cycles to V1= 6190.2 cycles.9.3 Problem 5-3 In the terminology of Section 5. If all turbines complete the test. If no bearing fatigue failures occur during the test. convert.000 cycles) Therefore. then no more than 5% of the turbines will be rejected for excessive erosion prior to 2300 cycles.2.95.237.48 0.81 22 9 5 / / Either of these test plans will satisfy the requirements of the test. with 90% confidence.1 with a sample size o" 20 and l equal to 1. The results are: I -f I-n(0.2 Problem 5-2 The reliability goal may be stated mathematically at R(2300) = 0. 5. which means that the reliability of the vane system is 0. assuming 3000. The required test time per bearing is: 0. (The first plan requires 22 X 3000 = 66.95)I or qi= 6190. the plan to test 5 turbines for 5000 cycles each requires the fewest total test cycles.000 cycles. and 3 into'equation 5. and 5000 test cycles accumulated on each.

=14 X 0. nj. = 1.23 Since./ t "40 t..) 221' . V. Note that the additional requirement in this plan has more than doubled the number of and test time required.031 0.5 . 2 = 2 n -33. using the method of Section 5.677. . Step 6: The final value of n is n.. n..exp(-(lO00/2000)2-5) = 0. = ... for r.2.031 Step 2: Setting r0 = 0.. + n. 31 and nj 36.9.. was found to be 14.3.. so the process of increasing ro is stopped here.5) * #units The final test plan is: test 33 units for 1000 hours. For r= 2. Step 4: a is still greater than b. . a is greater than b...162 a 3 3X -2" 24. a is less than b..• .5 ) = 0. p1 = 1 .23 than the a-ratio for r...•4'4 ".9 %and Equations Step 1: .9. So.- . n. If 2 or fewer units fail while on test. no.4 0. no. For r f= 1. 2 S. Step 5: For r.4. giving a =4.. The a-ratio for r.162 = 1 .. the test is passed.7 *.exp(-(1000/4000)2. W00O hours.. I':.5 or n = 33 (rounding 33. is increased by 1. 0.16 Step 3: 5.5 and is closer to b = 5. +36 . (The zero-failures test plan requircd that 14 units be tested 1000 hours each. is 2. for ro = 0. the final value of r. the value of n satisfying equation 5. = 23 and n= 18. .5..aire fLw r nnd n...-31 ----. 2000 hours.. is 3. ...7 ý.. r. . = 2 is 4.. giving a =6. f 1.. the value of n satisfying equation 5. ..for r. 4(4X) hours. = 2. .! p.

149 - F3. 7.3d e R(1500) S e*(-'.21 t.(U) = 7.)/4II.4315)2 (1.072 0.05 = 0.3b Var. Using equation 7.645) (0.01 80..0364)/2 = -0.P J •-• 6.7V.3c U.".3 Problem 7.020 "- =1.5et 4k7V 7 4 d ý<1.)140 |..ot.2.4315) + (1. . b. 0.1 Problem 7-1 Using equat ion 7.3 FA. 364 (-0.4315) 40 = . 1.008 = P'.1177)) _< R(1500) < 0.0364).• :s 1.U40 1667 _< V _< 2. .OW .o. 1. c.7454 U2 = (-0.2. F3A.2 and 11.3 For n = 40.3a U = (In (1500) .oM F 2.113 0.4315) .. confidence intervEI for n.22 :s . 222 -~-o.168 = (-0.7464)) (-e.0 F2.-- .5 e"' .(1. 200In .1177 7. 20(Oe4 0.411 _< 7. CHAPTER 7 ANSWERS 7.In (2000)) . 4 1.622 Therefore. CHAPTER 6 ANSWERS: None 7.2 Problem 7.o.4315 - 1.4 t 00 10.84 is the 90"c confidence interval for •. = - 0.001 = 0.6 /.1) (0.2 Using equation 7.AMIf)/Ih 0 _ 1 2000 e•' 0 h.645) (0..5 7. 20.0.001 -1O. 0.1913) (-0. from Appendix Tables 11./2 = -0.3a.5 =2000D "Usi-agequation 7. . and d.199 is the 90' 7. .

6 1) I I *•• T/Y = 4000/1000 = 4.7). + 451) 1/10 (2261). 7.q. = 4.9 probability bands on the number of failures occurring at T = 4000 hours.6400 hours. We want 0. 226.7 I . 21N4)i III II.1 4000 hours is (0.11 .81 80.. C = 0 b.95..05.33: 592.33 Failure2: 148.5% of the populc. Thus. When p = 0.01 5 Time _< 354. C = 7.k54. a 0. 10/2000 x 100 = 0.~.:1 "- * t . .11I.74. Step 5: 4 Step 6: On graph.74 1 1 = confidence intervals on first 3 failures: Failure1: Failure2: 20.4 Use Weibull-Thorndike Chart Figure 7.4 Problem 7.tion failed On graph Figure J.8 Id 1. 223 S.9 q new .9 probability band on number of failures by T 7. Entering x-axis at T/9 a.21 : Time 5 486.36 :5 Time _5592..5 Problem 7.0 MTTF = 1/10 (51 +. When p = 0. Step 3: Step 4.5 Steps I and 2: Completed on Figure J.4141. 90' 200n1-0149 = 48M.• :--" 2(X(jfIn I-0.59 =258..

. .246 Lower Band 0 0 "x 100 Upper Band 0. '1 22 122 i /.42 - e x/895"4212.0 1.247 0.508 0.. Bands are: °] [F(x) .e -[X/895.. F(x) + 0.974+ 0.284 0. - / _ .974 . ..4212. .246] For = 2. .974_ 0..662 0.253 1 200 300 400 500 0. .505 0.997 1.333 0. 11.757 0.- *.plot on Figure J. .- -.265 0.754 "Now.6 to calculate hands on Weibull from F'igureI. b. n. -.408 0 0 0 600 800 1000 1200 2000 0. .•.10 n=iO Using F(x) k(n) 0..246. - .. ..°* 7.0 0.o. since "true" Weibull lies partly outside the confidence bands.016 0. .246 1 .246 _ 1 .1= 895. ..6 Problem 7. . . .- • bh••. . • . .6 I - r Use the techni(I'e of Section 7..0. ?. we must conclude tha' the Weibulls are significantly different.e(-X/•)O...

0 . 2 '...00 -4 " 4-0- - Iq 1 I - Fi ure . .59I = 258. 3.. . -rhe 7-5- -- - -- 4 -A - .- I FAILURES =10 SA.. .' . w I22 V is I 1. "-• :.-w! Sr I I I - "-EIBULL DISTRIBUTIONJ /P = 1.1" .. .._ LE SIZE = 10 . . S. -- .u.. .. .3D..

4/ 1' Jr~-- . * o.FA IlLU -E .J(.. .59 _ -" Sm. *.S SIZE = 10 .RES = 10 . Prole 7-5 // . "".EIBILL DISTRIBUTION _ 9= = 258.~jao 4 Figr .1'..D . . /..0 1. - 4.1. -4 I U .. s ..4# si-4 a d SO - ..* 2 ..--.

.V I 1 I- I ....± .4 = 2..- -- S TOT~AL OPFERATING TIME (HR) FO 272257 Figure J.4. 1. t. . 11. -. t. ..... _. W -11 - -W.10. S*F SIZE = 10 FAILURES =10 v8=95. . Problem 7-6 22V 9.' " --.I t WEIBLLL DISTRIBUTI ON P ". t.97- --- -- 4 u. Joy -~t -.

S~~228 i . F0ILRS E.=O WEIBULLL DISTRIBUTION I A I / j0 U. *tU.. . !.SGov ernment Pilntin g Office: 1984 -.. II.1 . / I II. -rbe ' . ~Figure J.. aoimn Prntn -L& 75-G/ -iis I -98 . .if I - - - /I1 FiurJ.7 9i9-0 02 /918i ..97 q =89S .4 SAMR...4 .. S# .. .i .E SIZE =..-I- -.- •-I - - "--" A. J . . 11 11.!b'- •- ' "D= 2. - / eL. ii I 1s10.1. -.. -67 S. Problem 7-6 /: .0.