- V AFFDL-TRt-T843

LEVEL x

r THE STATISTICAL NATURE OF

b FATIGUE CRACK PROPAGATION

D. A. VIRKLER
B. M. HILLBERR Y
LL= P. K. GOEL
C* SCHOOL OFMECHANICAL ENGINEERING
PURDUE UNIVERSITY
C3 WEST LA FA YETTE, INDIA NA

APRIL 1978

4- TECHNICAL REPORT AFFDL-TR-78-43

Final Report - June 1976 to May 1978

Approved for public releae; distribution unlimited.

j 78 07 31 001
AIR FORCE FLIGHT DYNAMICS LABORATORY
AIR FORCE WRIGHT AERONAUTICAL LABORATORIES
AIR FORCE SYSTEMS COMMAND
WRIGHT-PATTERSON AIR FORCE BASE, OHIO 45433
 It

NOTICE

When Government drawings, specifications, or other data are used

for any purpose other than in connection with a definitely related
Government procurement operation, the United States Government thereby
incurs no responsibility nor any obligation whatsoever; and the fact
that the government may have formulated, furnished, or in any way
supplied 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 Information Office (01) 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.

MARGERYvE.tVARLr", R OBERT M. BADER, Chf

Project Engineer Structural Integrity Br

W!" THE COMMANDER

RALPH L. KUSTER, JR., Col, USAF

Chief, Structural Mechanics Division

1.1
"If your address has changed, if you wish to be removed from our
mailing list, or if the addressee is no longer employed by your
organization please notify, AFPDL/FBE, W-PAJB, ON 45433 to help us
maintain a current mailing list".

Copies of this report should not be returned unless return is required

by security considerations, contractual obligations, or notice on a
specific document.

AIR FORCC/S7$00/10Juy 10f?- 470

SECURITY %& %,ICATIOW OF T141S PAGE (l9ew Do,& Sheered)
1.~ -JEPORT DOCUMENTATION PAGE BEFORE__COMPLETINGFORM

S AFFDXTOR-43 -7 8

PURFUNAL/%VERSITV
THES LXAFTTETCA INDIO FTGU 11
I COIROPAAIONG OC.C NAMEPIIW
Aoo. ADDRESSM9

IS. :,95FCAI~~OW.
P OJET ASKN

STRCTUROLO MECHANICSL EIVISEEIONGSMO

11. CONTROLTIONG OFICAMET O tANDADDRESS

AppRoe forCpubliE reeae dsIributIo IM1D1

unlimited.A 7C

BLD. 410PPLEMENTARY
NO11S

FATIGUEFATRAN RELIABILTY

AICRAFOCE FROPGHTDNMC AOATION,

GMA HI
STTITIADISTRIBUTION WtHOD DIFFRENTATIO

conducved. Sixty-eligh replicate; constanr mpiudbrakprpgtion teststed

we ncedo2024-at3 aluminumInlloy.
DSRIUINSAEET(fte dibtre fiom
2,Ist deermirtio prora0

CRCrRPGTO
AM
pAreTIGEr lo-ora dsr
detif
dIst.KYWRibutionsiu wereveconsideI soeredtw-aand
butotrepraee
lc nosbrma it utotre
eblisrbtotL

~ /
codutd
dinvstributiion,tfree-faretiger grackdipribugtion, thoes
pAramteratistial
Sit-ih NOtestsOIT
were
relct con73n amplTtON
on 202C4-T3 almiu
conducted
NLSSFE
alloy.
~(
crac prpgto
Ditibto deemiaio
I
prgams
was

nl~d

h-vralsA~,N0rte7o
werev d/N n l/W"h olwn
 he generalized three-parameter gama distribution, and the generalized four-
parameter gamma distribution. From the experimental data, the distribution of
N as a function of crack length was best represented by the three-parameter
log-normal distribution.
Six growth rate calculation methods were investigated and the method which
introduced the least amount of error into the growth rate data was found to be
a modified secant method. Based on the distribution of da/dN, which varied
moderately as a function of crack length, replicate a vs. N data were predicted
This predicted data reproduced the mean behavior but not the variant behavior
of the actual a vs. N data.

I!

ILi IfAhf!- U5TP7 1

SECURITY Co.A5SSgpICA TIOM OF TIqS PAO5(Um Doet Enloe"E)

 FOREWORD

This report describes an investigation of the variability In

fatigue crack propagation under constant ampltiude loading sponsored
by APOSR-78-3018, and performed under Air Force Project 2307, Solid
Mechanics, Task 23070110, Variability in Fatigue Crack Growth.
Technical monitor for the project was Dr. J.P. Gallagher, formerly
of AFFDL/PBE. Me. M.E. Artley (AFFDL/FBE) assumed responsibility
for the project February 1978. The project period was June 1976 to
May 1978.

This program was conducted by the School of Mechanical Engineering

Purdue University, W. Lafayette, Indiana. Principal Investigator was
"Professor B.M. Rillberry; the graduate research assistant was
Mr. D.A. Virkler. Professor P.K. Goal was the statistician. Materials
for the test specimens were provided by the Aluminum Company of
America.

This report was submitted by the authors April 1978.

i.1

9,.
I
''1K

14
i-
I

_
___
i
 TABLE OF CONTENTS

"SECTION PAGE

I INTRODUCTION ............ .... ................. 1

II BACKGROUND .............. ... .................. 6

III OBJECTIVES OF INVESTIGATION ......... 11

2!
IV CRACK GROWTH RATE CAL -ULATION METHODS . .. 12

V STATISTICAL CONCEPTS .............. .... 21

VI DETERMINATION OF THE DISTRIBUTION ...... 55

Vii GROWTH RATE AND GROWTH PREDICTION . . . . . . 59

VIII STATISTICAL ANALYSIS OF PREVISOULY GENERATED

DATA ........ ..................... .... 62 4

IX EXPERIMENTAL INVESTIGATION .......... .. 73

X DATA ANALYSIS AND RESULTS .......... .. 83

XI DISCUSSION ..... .................. ... 153

XII CONCLUSIONS ..... ................. ... 186

APPENDIX A: Derivation of the da/dN Equation

for the Linear Log-Log 7-Point Incremental
Polynomial Method .. ............ ..... 189
APPENDIX B: Derivation of the da/dN Equation
for the Quadratic Log-Log 7-Point Incremental
Polynomial Method ........................ 191
2 . . . . . . . .
APPENDIX C: Derivation of C 193
APPENDIX D: DNDDPG Documentation ... . 195
APPENDIX E: CCDDP Ducumentation ....... 197
APPENDIX F: CGRDDP Documentation ...... 201
APPENDIX G: DNDDP Documentation ....... 202
APPENDIX H: DELTCP Documentation . . . . . . 204
APPENDIX I: DADNCP Documentation 205

twft
 TABLE OF CONTENTS (Cont'd)

SECTION PAGE

APPENDIX J: AVNPRD Documentation ........ 207

APPENDIX K: Random Order of Experimental Tests . 209

REFERENCES ................... ............ 210

I vi

,L
I-

LIST OF TABLES

5
TABLE PAGE

I Distribution of 016/a ...... 65

T. II Smoothing Effect of the Incremental Polynomial Method ...... 68

III Life Prediction Based on the Mean ....................... 69

IV Effect of Increasing ja .................................... 72

V Average Experimental Error ................................. 82

VI Average Goodness of Fit Criteria for the Distribution of

Cycle Count Data ........................................... 93

VII Distribution Rankings for the Distribution of Cycle Count

Data ....................................................... 94

VIII da/dN Calculation Method Results ........................... 102

IX Average Goodness of Fit Criteria for the Distribution of

da/dN Data ................................................. 115

X Distribution Rankings for the Distribution of da/dN Data ... 116

X1 Average Goodness of Fit Criteria for the Distribution of

Cycle Count Data Predicted froin the Distribution of da/dN... 123

XII Distribution Rankings for the Distribution of Cycle Count

Data Predicted from the Distribution of da/dN .............. 124

XIII Comparison of the Distributions Between Actual Cycle Count

Data and Cycle Count Data Predicted from the Distribution
of da/dN'. .................................................. 126

XIV Comparison of Actual Cycle Count Data with Cycle Count Data
Predicted from Constant Variance da/dN Lines ............... 128

XV Average Goodness of Fit Criteria lor the Distribution of

dN/da. Data ................................................. 138

XVI Distribution Rankings for the Distribution of dN/da Data ,.. 139

vii
Fl

LIST OF TABLES

t STABLE PAGE

I Distribution of AN/aa ...................................... 65

II Smoothing Effect ot the Incremental Polynomial Method ...... 68

III Life Prediction Based on the Mean .......................... 69

IV Effect of Increasing Aa .................................... 72

V Average Experimental Error ................................. 82

VI Average Goodness of Fit Criteria for the Distribution of

Cycle Count Data ........................................... 93
VII Distribution Rankings for the Distribution of Cycle Count
Data ........................................................ 94

VIII da/dN Calculation Method Results ........................... 102

IX Average Goodness of Fit Criteria for the Distribution of

da/dN Data ................................................. 115

X Distribution Rankings for the Distribution of da/dN Data ... 116

XI Average Goodness of Fit Criteria for the Distribution of

Cycle Count Data Predicted from the Distribution of da/dN... 123

XII Distribution Rankings for the Distribution of Cycle Count

Data Predicted from the Distribution of da/dN .............. 124

SXIII Comparison of the Distributions Between Actual Cycle Count

Data and Cycle Count Data Predicted from the Distribution
of da/dN .................................................. 126

XIV Comparison of Actual Cycle Count Data with Cycle Count Data
Predicted from Constant Variance da/dN Lines ............... 128

XV Average Goodness of Fit Criteiia for the Distribution of

dN/da, Data ................................................. 138

XVI Distribution Rankings for the Distribution of dN/da Data ... 139

I vii

ii
 LIST OF TABLES (Cont'd)

TABLE PAGE

XVII Average Goodness of Fit Criteria for the Distribution of

Cycle Count Data Predicted from the Distribution of dN/da .. 147

XVIII Distribution Rankings for the Distribution of Cycle Count

Data Predicted from the Distribution of dN/da .............. 148

XIX Comparison of the Distributions Between Actual Cycle Count

Data and Cycle Count Data Predicted from the Distrib!tion
of dN/da ................................................... 150

XX Comparison of Actual Cycle Count Data with Cycle Count Data

Predicted from Constant Variance dN/da Lines ............... 152

Vill

I Ike
 LIST OF ILLUSTRATIONS

FIGURE PAGE

1 Typical Raw Fatigue Crack Propagation Data ................... 3

2 Typical Logl 0 da/dN vs. Logl 0 AK Data ........................ 4

3 Schematic Representation of the Distribution of N ............. 88

4 Schematic Representation of the Distribution of da/dN ........ 9

5 Secant Method ................................................ 13

6 Modified Secant Method ....................................... 15

7 Incremental Polynomial Method ................................ 18

8 Typical Relative Frequency Histogram ......................... 23

9 Typical Relative Cumulative Frequency Histogram .............. 24

10 Golden Section Search Method ................................. 36

11 Typical 2-Parameter Normal Distribution Plo ............... 44

12 Typical 2-Parameter Log Normal Distribution Plot ............. 46

13 Typical 3-Parameter Log Normal Distribution Plot ............. 47

14 Typical 3-Parameter Weibull Distribution Plot ................ 48

15 Typical 3-Parameter Gamma Distribution Plot .................. 49 1

16 Typical 2-Parameter Gamma Distribution Plot .................. 50

17 Typical Data Chosen for Analysis from Overload/Underload Test

Data ......................................................... 63

Fit of the ,N/6a S18 Data to the 3-Parameter Log Normal

i astribut ion ................................................. 66

19 Effect of Increasing a ..................................... 71

20 Test Program ................................................. 75

ix
 LIST OF ILLUSTRATIONS (Cont'd)

F IGURE PAGE

21 Test Specimen ................................................ 77

22 Original Replicate a vs. N Data .............................. 84

23 Typical Replicate Cycle Count Data ........................... 85

24 2-Parameter Normal Distribution Parameters of Cycle Count Data

as a Function of Crack Length ................................ 87

25 2-Parameter Log Normal Distribution Parameters of Cycle Count

Data as a Function of Crack Length ........................... 88

26 3-Parameter Log Normal Distribution Parameters of Cycle Count

Data as a Function of Crack Length ........................... 89

27 3-Parameter Weibull Distribution Parameters of Cycle Count

Data as a Function of Crack Length ........................... 90

28 3-Parameter Gamma Distribution Parameters of Cycle Count Data

as a Function of Crack Length ................................ 91

29 Generalized 4-Parameter Ganmma Distribution Parameters of Cycle

Count Data as a Function of Crack Length ..................... 92

30 Typical Loglo da/dN vs. LoglO AK Data Calculated by the Secant

Method ....................................................... 96

31 Typical Log 1 0 da/dN vs. Loglo AK Data Calculated by the

Modified Secant Method ....................................... 97

32 Typical Log 1 0 da/dN vs. Log 1 0 AK Data Calculated by the Linear

7-Point Incremental Polynomial Method ........................ 98

33 Typical Log 1 0 da/dN vs. Log1 4K Data Calculated by the

Quadratic 7-Point Incremental Polynomial Method .............. 99

34 Typical Log 1 0 da/dN vs. Loglj AK Data Calculated by the Linear

Log-Log 7-Point Incremental Polynomial Method ................ 100

35 Typical Logl 0 da/dN vs. Log1 0 AK Data Calculated by the

Quadratic Log-Log 7-Point Incremental Polynomial Method ...... 101

36 Combined Loglo da/dN vs. LoglO AK Data Calculated by the

Secant Method ................................................ 104

37 Combined Loglo da/dN vs. Logl 0 AK Data Calculated by the

Modified Secant Method ....................................... 105

38 Combined Log 1 0 da/dN vs. Log 0 AK Data Calculated by the

Quadratic 7-Point Incremental Polynomial Method .............. 106

I
 FA
LIST OF ILLUSTRATIONS (Cont'd)

"FIGURE PAGE

39 Typical Replicate da/dN Data ............................... 107

40 2-Parameter Normal Distribution Parameters of da/dN Data as a

Function of Crack Length ........ .................. 109

S41 2-Parameter Log Normal Distribution Parameters of da/dN data

as a Function of Crack Length .................. .......... 110

42 3-Parameter Log Normal Distribution Parameters of da/dN data

as a Function of Crack Length ................................. 111

43 3-Parameter Weibull Distribution Parameters of da/dN Data as a

Function of Crack Length ..................................... 112
44 3-Parameter Gamma Distribution Parameters of da/dN Data as a

Function of Crack Length ..................................... 113

45 Generalized 4-Parameter Gamma Distribution Parameters of da/dN
Data as a Function of Crack Length ........................... 114

46 Replicate a vs. N Data Predicted from the Distribution of

da/dN ..................................................... 118

47 2-Parameter Normal Distribution Parameters as a Function of

Crack Length for Cycle Count Data Predicted from the
Distribution of da/dN ........................................ 119

48 2-Parameter Log Normal Distribution Parameters as a Function

of Crack length for Cycle Count Data Predicted from the
Distribution of da/dN ......................................... 120

49 3-Parameter Log Normal Distribution Parameters as a Function

of Crack Length for Cycle Count Data Predicted from the
Distribution of da/dN ......................................... 121

50 3-Parameter Weibull Distribution Parameters as a Function of

Crack Length for Cycle Count Data Predicted from the
Distribution of da/dN ........................................ 122
"51 a vs. N Data Predicted from the Mean and + 1, 2, and 3 Sigma
da/dN Lines .................................................. 127

-Z 52 Typical Replicate dN/da Data .................. 130

53 2-Parameter Normal Distribution Parameters of dN/da Data as a

Function of Crack Length ..................................... 132

54 2-Parameter Log Normal Distribution Parameters of dN/da Data

as a Function of Crack Length ................................ 133

xi
 LIST OF ILLUSTRATIONS (Cont'd)

F IGURE PAGE

55 3-Parameter Log Normal Distribution Parameters of dN/da Data

as a Function of Crack Length ................................ 134

56 3-Parameter Weibull Distribution Parameters of dN/da Data as a

Function of Crack Length..................................... 135

57 2-Parameter Gamma Distribution Parameters of dN/da Data as a

Function of Crack Length ..................................... 136

58 Generalized 3-Parameter Gamma Distribution Parameters of dN/da

Data as a Function of Crack Length ........................... 137

59 Replicate a vs. N Data Predicted from the Distribution of

dN/da ........................................................ 141

60 2-Parameter Normal Distribution Parameters as a Function of

Crack Length for Cycle Count Data Predicted from the
Distribution of dN/da ........................................ 142

61 2-Parameter Log Normal Distribution Parameters as a Function

of Crack Length for Cycle Count Data Predicted from the
Distribution of dN/da ....................................... 143

62 3-Parameter Log Normal Distribution Parameters as a Function

of Crack Length for Cycle Count Data Predicted from the
Distribution of dN/da ....................................... 144

63 3-Parameter Weibull Distribution Parameters as a Function of

Crack Length for Cycle Count Data Predicted from the
Distribution of dN/da ....................................... 145

64 3-Parameter Gamma Distribution Parameters as a Function of

Crack Length for Cycle Count Data Predicted from the
Distribution of dN/da ........................................ 146
65 a vs. N Data Predicted from the Mean and + 1, 2, and 3 Sigma
dN/da Lines .................................................. 151

66 a vs. N Data Showing Abrupt Growth Rate Changes .............. 154

67 Estimate of the Location Parameter of the 3-Parameter Log

Normal Distribution as a Function of Crack Length ............ 157

68 Typical Fit of the Cycle Count Data to the 2-Parameter Normal

Distribution .................................. ............. 159

69 Typical Fit of the Cycle Count Data to the 2-Parameter Log

Normal Distribution .......................................... 160

xii
 LIST OF ILLUSTRATIONS (Cont'd)

FIGURE PAGE

70 Typical Fit of the Cycle Count Data to the 3-Parameter Log

Normal Distribution .......................................... 161

71 Typical Fit of the Cycle Count Data to the 3-Parameter Weibull

Distribution ................................................. 162

72 Typical Fit of the Cycle Count Data to the 3-Parameter Gamma

Distribution ......................................... ..... 163

73 Typical Skewed Left da/dN Data .............................. 166

74 Typical Symmetric da/dN Data ............... 167

75 Typical Skewed Right da/dN Data .............................. 168

76 Typical Fit of Skewed Left da/dN Data to the 2-Parameter

Normal Distribution ........................................... 169

77 Typical Fit of Symmetric da/dN Data to the 2-Parameter Normal

Distribution ................................................. 170

S78 Typical Fit of Skewed Right da/dN Data to the 2-Parameter Log
Normal Distribution .......................................... 171

79 Typical Fit of Skewed Right da/dN Data to the 3-Parameter Log

Normal Dis tribution .......................................... 172

S80 Typical Fit of Skewed Right da/dN Data to the 3-Parameter

Weibull Distribution ......................................... 173

81 Typical Fit of Skewed Right da/dN Data to the 3-Parameter

Gamaa Distribution ...................................... 174
r
82 Typical Symmetric dN/da Data ................................. 179

83 Typical Fit of Symmetric dN/da Data to the 2-Parameter Normal

Distribution ................................................. 180

S84 Typical Fit of Synmmetric dN/da Data to the 2-Parameter Log

Normal Distribution .............................. .......... 181

85 Typical Fit of Symmetric dN/da Data to the 3-Parameter Log

Normal Distribution .......................................... 182
86 Typical Fit of Symmetric dN/da Data to the 3-Parameter Weibull
Distribution ................................................. 183

87 Typical Fit of Symmetric dN/da Data to the 2-Parameter Gamma

Distribution ................................................. 184

xtii
 LIST OF SYMBOLS

A chi-square tail area

a half crack length (in. or mm.).

af final half crack length (in. or mn.).

ai any discrete half crack length (in. or mm.).

aaverage half crack length used by the secant method

(in. or mm.).

a initial half crack length (in. or mm.).

0
B Weibull slope.

B2 curvature used in the Golden Section search method.

b scale parameter for the HLE 3-parameter Weibull distribution.

b0 least squares y intercept.

b1 least squares slope.

b2 least squares curvature.

C2 closeness.

CI scaling constant used by the incremental polynomial methods for

the mean of the strip data.
C2 scaling constant used by the incremental polynonial methods for
the range of the strip data.

C+ constant approaching positive infinity.

constant in the covariance matrix.

CV
c shape parameter for the HLE 3-parameter Weibull distribution.

D maximum deviation in the Kolmogorov-Smirnov test.

d convergence constant for the interior point penalty function

algorithm.

Xiv

I I
 da/dN fatigue crack growth rate (in./cycle).

da/dNi any discrete value of fatigue crack growth rate (in./cycle).

da/dN maximum fatigue crack growth rate (in./cycle).
max
da/dNMin minimum fatigue crack growth rate (in./cycle).

dN/da inverse fatigue crack growth rate (cycles/in.).

e Napierian base (2.718281828).

exp(X) e raised to the y power.

a e expected frequencies in the chi-square test.

S(c) cumulative density function of corrected data.

F -1 inverse cumulative density function for the generalized 4-para-

= g meter gat=&a distribution.

Sf~M) density function of a random variable.

SC(z) standard normal probability function.

Sg shape/power parameter for the gamma distributions.

H Golden Section search constant (0.618033989).

1H(z) gamma probability fun .tion.

2
I vairable used in d'eriving C

variable used In deriving C2

J iteration counter for the interior point penalty function
algorithm.

SK variable used in deriving C

k number of equiprobable intervals for the chi-square test.

L maximum likelihood estimator function.

In natural logarithm (base a).

log 1 0 logarithm (base 10).

m slope.

N cumulative load cycle count.

XI
 I

Nf final cumulative load cycle count.

Ni any discrete cumulative load cycle co-nt.

I average cumulative load cycle count used by the secant method.

NLS log (base 10) scaled cycle count data used by the log-log
incremental polynomial methods.
NS scaled cycle count data used by the incremental polynomial

methods.

n number of data roints in a data set.

number of data points lost at each end of the data by the

incremental polynomial methods.

nnp number of distribution paramete.

number of data points in the incremented strip in the

nSTRIP incremental polynomial methods.

o 1observed frequencies in the chi-square test.\

P objective function used by the interior point penalty function

algorithm.

P maximum applied load during the load cycle (lbs.).

max
Pmin minimum applied load during the load cycle (lbs.).

q, chi-square test class end points.

R R ratio.

R2 coefficient of multiple determination.

r iteration variable in the interior point penalty function

algorithm.

SE
SS
standard deviation of the errors.
standard deviation of the standard deviations. I
SSRES residual sum of squares.

T location parameter in the Golden Section search method.

TCSS total corrected sum of squares.

xvi
 II

St variable in the gamma probability function.

U variable used in deriving da/dN for the quadratic log-log

7-point incremental polynomial method.

u variable of integration in the chi-square tail area equation.

c
V covariance matrix.
Sv matrix.
inverse covariance

X variable to be plotted on the abscissa.

tmean of the errors.

mean of the standard deviations.

S
variable to be plotted on the ordinate. SY

Z Kolmogorov-Smirnov test statistic.

Sz
standeri zed distribution variate.
power
_" parameter in the gamma distributions.

Sa
hypothesis acceptance level.
a
0 shape parameter in the log normal distributions.

r gamma function.

asa
Euler's constant (0.5772157).

distance between two consecutive half crack length data points

orm.) S(in.
I
AK charge in stress intensity during the load cycle (ksi-square
root in.).

&K any discrete value of AK (kEe-square root in.).

4N number of cumulative load cycles between two consecutive

cumulative load cycle count data points.

AN/Ia change in cumulative load cycle count normalized by the change

in half crack length (cycles/in.).

I
LP

8+
change in applied load during the load cycle (lbs.).

convergence criterion constant.

positive constant approaching zero.

Xvil
 9 characteristic value of the 3-parameter Weibull distribution.

scale parameter (sometimes the mean) of a distribution.

V degrees of freedom.

T Pi (3.141592654).

CT shape parameter (sometimes the standard deviation) of a

distribution.

T location parameter (sometimes the terminus) of a distribution.

x random variable.
2
X chi-square test statistic.

Xc corrected random variable data.

Xi any discrete value of a ranidom variable.

Xmin minimum value of a random variable.

Xo expected minimum value of the 3-parameter Weibull distribution.

ddigamma function.

4' trigamma function.

hat (symbolizes an estimated variable).

Il

wl

i xviii
 ' t-

S~ SECTION I
! INTRODUCTION

.t Throughout the course of history, it has always been desirable to

be able to predict the life of a given design under expected service con-

ditions. Life prediction in metal structures has necessitated a need for

knowledge about the metal fatigue phenomenon. The metal fatigue process,

as it is known today, is complex and is still not fully understood. There

Ft
are many variables which influence the life of a metal structure, such as

the material, loading, and geometric characteristics of the particular

structure. This investigation involves only the determination of the ef-

I fect of material properties on life prediction.

One of the primary mechanisms by which metal fatigue occurs is the

propagation of microscopic cracks [1). The study of fatigue crack propa-

gation behavior has been widely conducted for some time in an effort to

kunderstand metal fatigue more fully. The information obtained from crack
p

propagation studies is then used in estimating the fatigue life of struc-

tures and components. Ideally, it is desirable that this estimated life

will exactly predict the actual life. Unfortunately, there are many vari-

ables which influence this prediction and some are not well understood.

One of the most important of these variables is how well the empirical

crack growth relationships obtained from experimental data actually re-

present the observed crack propagation behavior.

The raw data from a fatigue crack propagation test are the half

crack length, a, and the number of cumulative load cycles, N, needed to

grow the crack to that length from some reference initial crack length

for slightly increasing stress intensity level load conditions, called

constant amplitude loading. A plot of typical raw fatigue crack propa-

gation data is shown in Figure 1. The current interpretation of this

raw data focuses upon the fatigue crack growth rate as a functlon of an

applied stress intensity parameter, usually AK, the change of the stress

intensity during the load cycle. The fatigue crack growth rate is de-

fined as the rate of extension of the crack with respect to the number

of applied load cycles [23. Actual determination of the crack growth

rate requires an evaluation of the slope of the raw a vs. N data at

various discrete points, which results in the derivative of a with re-

spect to N, normally called da/dN. A plot of typical da/dN vs. AK data

is shown in Figure 2.

The importance of the fatigue crack growth rate as a variable of

interest is born out in the fact that the fatigue crack growth rate is

nearly independent of the geometry for the same stress intensity level

of loading [33. This allows crack growth behavior prediction based only

on the knowledge of the crack growth rate vs. the stress intensity level

of loading for a given material for any geometry chosen. Obviously, this

would be an important design tool if the crack growth behavior predictions

were accurate and reliable. These crack growth behavior predictions can

be used to predict the number of load cycles needed to grow a crack from

an initial crack length, a0 , to some nov crack length, ae, and the die-

tance a crack propagates, Ma, during a specified number of applied load

cycles. In addition, using various prediction techniques, the constant

i
amplitude loading crack growth rate behavior is used to predict variable

amplitude loading crack growth rate behavior [4].

2I
~- . -,
 R VS. N
REPLICATE CA TEST 17
DELTAP=
P 4.20 KIP
164 DATA POINTS
AD = 9.000 MM
R .2k)

x50.00-

CD
z,0.00-
LLU tI
cc

;20.00

xxxxxj -]x
Szo~0
Figure 1. Typical Raw Fatigue Crack Propagation Data
Ir
 DR/ON VS. DELTA K PLOT
REPLICATE CA TEST 17
DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
lo"'--Pl'M:X = 5.25 KIP
136 DATA POINTS
R= .20

6j -

xx

(..) - x
_ x x1c
Ix

aM X 40

4 ~kx

10

7.0 10.0 13.0 16.0 20.0 26.0

DELTA K (KSI-SQURRE ROOT IN)

Flgurc 2. Typical LoglO da/dN vs. LOglo0 K Da ta

&. 4
 DR/ON VS. DELTR K PLOT
e- REPLICATE CA TEST 17
DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
l'"136 PMX DATA
= 5.25 KIP
POINTS
R =.20
6-

.32.
0L - x x
=10"- x x

~ x
C3 ,Ixc x
2x

Xj4_

7.0 10.0 13.0 18.0 20.0 25.0

DELTA K (KSI-SQU:RE RMOT IN)
Figure 2. Typical Log 1 0 da/dN vs. Log 1 0 AK Dara

4
I

There are several methods of numerically determining the crack

growth rate from the raw a vs. N data. It has been suspected that the

crack growth rate calculation method has a very significant effect on

the variance of the resulting growth rate vs. stress intensity parameter

data [2,5,673.

During the prediction of crack growth behavior, the crack growth

rate vs. stress intensity parameter data is integrated back to obtain

predicted a vs. N behavior. Considerable variation in this predicted

crack growth behavior has been experienced, thus hindering accurate life

estimates [2,5,6,7]. This variation is a result of variation in the raw

crack growth data, variation due to the crack growth rate calculation

method, and material variations.

This investigation will compare several numerical growth rate cal-

culation methods and attempt to find the method which introduces the

least amount of error into the growth rate vs. stress intensity parameter

data. It will also attempt to describe crack growth behavior in a sta-

tistical manner with the expectation that this statistical description

of crack growth behavior will reduce the large amount of error currently

present in life prediction.

5
 I

SECTION II

BACKGROUND

Metal fatigue has long been recognized as a random phenomenon [81,

but until recently, little effort was devoted to applying statistical

tools to fatigue crack propagation behavior. By fitting different equa-

tions to the crack growth rate vs. stress intensity parameter data, nu-

merous equations of fatigue crack growth have been suggested [93. How-

ever, due to scatter in the data, it has been impossible to select which

equation is the most appropriate. Also, when the original crack growth

data are predicted from these equations, the correlation with the original

data is generally very poor [81. Due to the large amount of scatter in

the crack growth rate vs. stress intensity parameter data, investigators

have started using statistical methods to characterize fatigue crack pro-

pagation behavior [5,6,7,8]. It can be easily shown that the amount of

data scatter is generally considerably greater than can be accounted for

by experimental inaccuracies [8j. It has been pointed out that the re-

maining scatter is due to the essentially random nature of fatigue crack

growth which is a result of the relative nonhomogeneity of the material

8, 10].

From a macroscopic viewpoint, it is often convenient to regard a

metallic material as a homogeneous continuum, and basing engineering

calculations on this assumption does not generally lead to serious error.

However, the scatter observed in fatigue testing of a metallic material

arises precisely because it is not a homogeneous continuum, when

__ _ _ _ _ _
 considered on a microscopic scale [8]. Consequently, it is important to

examine fatigue crack growth from a statistical viewpoint. In order to

include fatigue crack propagation scatter in the general overall charac-

terization of fatigue crack propagation behavior, this investigation

will apply statistical concepts to fatigue crack growth behavior.

In considering the crack growth from some initial crack length, a0 ,

to a new crack length, ai, there is a certain mean and variance associated

with the number of load cycles required for this amount of crack growth

which characterizes the statistical distribution of N at ai. A schematic

representation of this distribution of N is shown in Figure 3. In order

to statistically characterize the crack growth behavior, it is necessary

to determine the distribution of N from experimental tests.

The variance in N illustrated above can be due to random errors in

the measurement of a, N, and AK, to systematic errors in these measure-

ments, and to the statistical variation in the material's growth rate

properties. Through the use of accurate equipment, the random errors in

the measurement of a, N, and AK can be reduced to an acceptable level and

"measured by a separate test. Through a careful experimental set up and

procedure, the systematic measurement error can be reduced. From this,

the desired statistical behavior of the material's crack growth proper-

ties can be determined.

In considering the crack growth rate vs. stress intensity parameter

data, there is some statistical distribution associated with the crack

growth rAte, da/dN, at some stress intensity level, AKi. A schematic

representation of this distribution of da/dN is shown in Figure 4. In

order to statistically characterize the crack growth rate behavior, it

I7
 S- ~1
Ji

ii
t

[do
[ J/;
S~//

77
do //
I
,
A
IK dN
// AKK i
/
V /
[J

,AK, AK*

Figure 4. Schematic Representation of the

Distribution of da/dN

"ii
is also necessary to determine the distribution of the crack growth rate

from experimental teats.

The variance of d&/dN illustrated above originates in the variance

present in the original a vs. N data. The density of the raw data (es-

sentially, the distance between 2 consecutive data points, Aa) and the

crack growth rate calculation method both contribute to the overall vari-

ance of da/dN. In order to determine the variance of da/dN due to the

variance in the original a vs. N data, it is necessary to determine the

effect of both data density and the crack growth rate calculation method

on the variance of da/dN.

Once the crack growth rate vs. stress intensity parameter data has

been obtained, the next step is to be able to predict the change in

crack length for a given number of applied load cycles or, inversely, the

number of applied load cyclem for a given change in crack length. The

variance of this prcdiction is directly related to the variance of the

crack growth race. In order to evaluate the effectiveness of this pre-

diction, it is necessary to predict the original a vs. N data from the

crack growth rate data and then compare the predicted a vs. N data with

the original a vs. N data.

This a vs. N prediction can be accomplished by either of two methods.

The currently popular method is to numerically integrate the mean da/dN

vs. AK curvo to obtain predicted a vs. N data (2,4,5,6,7,9). However,

no adequate method for determining the resulting scatter in a or N exists

[5]. An alternate method uses the knowledge of the distribution of da/dN

and the fact that da/dN is an independent random variable to obtain a vs.
N stop by step. This method is discussed in detail in Section 7.3. Using

this method, the variances of both a and N can be readily obtained.

10
 - .X n A ..- .. - - _'o
t"

SECTION III

OBJECTIVES OF INVESTI'SXTION

The main purpose of this investigationwas to apply statistical con-

cepts and theory to the study of fatigue crack propagation behavior. In

doing this, there were four main objectives to be met. They were:

1) Determine the statistical distribution of N (cumulative

load cycle count) as a function of a (crack length).

2) Determine which crack growth rate calculation method

yields the least amount of error when the crack growth

rate curve is integrated back to the original a vs. N

data.

3) Determine the statistical distribution of da/dN (crack

growth rate) as a function of AK (stress intensity

parameter.

4) Determine the variance of a set of a vs. N data predicteo

I f
" ~from the da/dN distribution parameters.

I I.?

| -1
 SECTION IV

CRACK GROWM RATE CALCUIATION }MTHODS

Numerous methods of calculating the crack growth rate from the raw

a vs. N data have bean used by various investigators [2,5,6]. None of

these seem to be universally accepted, but rather each investigator

seemed to favor a different method. Since it was virtually impossible to

investigato all of these methods, six of the more important methods were

selected for examination. These methods are:

1) The secant method,

2) The modified secant method,

3) The linear 7-point incremental polynomial method,

4) The quadratic 7-point incremental polynomial method,

5) The linear log-log 7-point incremental polynomial method, and

6) The quadratic loS-log 7-point incremental polynomial method.

4.1 Secent lOgthod

The secant method is a finite difference method and perhaps the

simplest of the methods considered [2,5,61. Basicallyt the secant

method calculate# the slope of a straight line between 2 adjacent a vs.

N data points. It then approximates this slope as the slops of the

tangent line of the a vs. N curve at an average crack length, a', and

average cycle count, Ni" A schematic representation of the secant

method is shown In Figure 5.

12
 |U

ii

IID
I

I I I~m

S1-1 !
x
Icj

dl I

N, NII
j N(CYCLE COUNT)
1

Figure 5. Secant Method

13
 The average crack length, a1 , is givn by

_a at- +2ai+l
- (I)
1 2 (1)
Similarly, the average cycle count, Ni, is given by

N +N
1 2 (2)

The slope of the line connecting the 2 adjacent data points, which is

used to approximate the growth rate, is given by

da - (ai,+,a,) (3)
d-i (N +1 Nti
dN1 (N~+ N)

at a andN i.

4.2 Modified Secant Method

The modified secant method is really an extension of the secant

method. Basically, this method averages the growth rates obtained by

the secant method so that the da/dN data coincides with the original a

vs. N data. The beginning and end points are assumed to be equal to the

first and last growth rates, respectively. A schematic representation of

the modified secant method is shown in Figure 6.

The growth rate is given by

2
+

at a and Ni for 1-2 to (n-1) wheae n is the number of data points in the

data set.

The first growth rate date point is given by

da (a2 -a) (5)

and N1 . dNI (N2
at a

14
 ii-

CD -

LL
xx
I
'-Jt
z A1 ......- ~ --

N-I I I
SI I I
iI II
S1 I1
SI I I
SI I
SN,_, N, N1 ,,
N (CYCLE COUNT)
r
Figure 6. Modified Secant Method

15

?
The last growth rate data point is given by

da . (an " aa-1)

dN" ('n -N.1)

at a and N .
a n

4.3 Linear 7-Point Incremental Polynomial Method

The linear 7-point incremental polynomial method is the simplest of

the four incremental polynomial methods. In each of the incremental

polynomial methods, a polynomial is fit by the method of least squares

to a series of data points, called a strip, and the derivative of the

polynomial is evaluated at the middle point [2,5,61. This strip is

then incremented by one data point and the curve fitting and evaluation

process is repeated. The strip incrementation process is repeated until

All of the data points have been used. Any odd number of data points

can be used for the incremented strip, although 7 points are usually

used. The incremental polynomial methods differ basically in the poly-

nomial which is fit to the data.

Initially the strip data points are scaled in the following manner.

Two constants, C1 and C2 , are calculated as follows [5,6]:

C1 Ni+nMl +2 'i-n4 (7)
121
N N
C2 =2 i+ s Ci"ns (8)

where nS n
nstrip- - (9
(9)
"-

where nstrip is the number of data points in the strip. Note that C1 is

the center of the strip cycle count data and C'2 is the range of the strip 6

16
 7

cycle count data. The data scaling is then performed as follows:

f N- U ,
C2 (1, K"

where Ns is the scaled cycle count data. As a result of this scaling,

the strip cycle count data runs from -1 to +1. This insures that when

least squares curve fitting occurs, the scale of the data vill not in-

fluence the curve fitting, which is a constant danger when using least

E squares as a curve fitting technique.

After the curve fitting has been performed, the derivative of the

resulting polynomial is then evaluated at the midpoint of the strip, Ni.

This evaluation takes into account the scaling that was performed prior

f to the curve fitting. A schematic representation of the incremental

polynomial method is shown in Figure 7.

In the linear 7-point incremental polynomial method, the fitted

polynomial is s first order linear straight line. After fitting by

linear least squares, the fitted polynomial takes the following form:

a-b +0b1 n (11)

Substituting the scaling equation,

r b1C, rb
a b - -.... A + IZ--N (12)
Lo C2 j L2

Taking the derivative of a with respect to N,

dN C2

Obviously, for a straight line, the slope is independent of where the

derivative is evaluated at.

17
 Rj~t~4

XI

xx
t--

R j
I I I !
x-I

U x

I -,I

N N, N
N (CYCLE COUNT)
d

Figure 7. Incremental Polynomial Method

1f

S18 i -
 AI.,N

Z4

U I

- II
i-,
.-,

CD
wzx -- N

~ I
C I I
- I
I
I ._

Nx N +time
N (CYCLE COUNT)

Figure 7. incremental Polynomial Method

i-I,,,l4l
l ,4 . .ii l~ . . -I I

18
 4.4 Quadratic 7-Point Incremental Polynomial Method

The quadratic 7-point incremental polynomial nsthod has gained wide

acceptance as a valid crack growth rate calculation method [5,6]. In

this method, the fitted polynomial is a second order curve. After fitting

by second order least squares, the fitted polynomial takes the following

form:

a b 0 + bN (14)
IS 2S

Substituting the scaling equation,

bC b C1 b 2b Cb
a0 C
a b- - +
2
-

C22o
2 C2
22
~
C
N(5 2
2 2
Taking the derivative of a with respect to N and evaluating at the mod-

point, Ni,

b 2bC (12b
.[a - ::") +H-N(6
dNIC2 C2 2,

4.5 Linear Los-7oa 7-Point Incremental Polynomial Mothod

The linear log-log 7-point incremental polynomial method was used

to determine if the data could be linearized by a log 10 transformation

on both the crack length and cycle count data. This method is essentially

the same as the linear incromental polynomial method except for the log

transformations of the input data Just prior to the data scaling.

The fitted polynomial is linear and takes the following formi

log a b + b lNL( 17)

where RLS is the lto scaled cycle count data.

"19
i
t
4.4 Quadratic 7-Point Incremental Polynomial Method

The quadratic 7-point incremental polynomial method has gained wide

acceptance as a valid crack growth rate calculation method [5,6]. In

this method, the fitted polynomial is a second order curve. After fitting

by second order least squares, the fitted polynomial takes the following

form:

a b + bN N 2 (14)
0 1S 2S

Substituting the scaling equation,

Sa [ bO -1 + + I~lN + [:-jN (15)

2C bC 2bC
"2 C2 2 (25)

Taking the derivative of a with respect to N and evaluating at the mid-

point, N1 ,

b 2b C [2b
dN C 2 2(16)

4.5 Linear Log-Log 7-Point Incremental Polynomial Method

The linear log-log 7-point incremental polynomial method was used

I to determine if the data could be linearized by a log 1 0 transformation

on both the crack length and cycle count data. This method is essentially

the same as the linear incremental polynomial method except for the log

"transformations of the input data just prior to the data scaling.

The fitted polynomial is linear and takes the following form:

log a - b+A (17)

where NI is the log scaled cycle count data.

19
 b -bC rb

d 2 b N IC2
dNi C2

The derivation of this equation is shown in Appendix A.

4.6 Quadratic Loa-Loa 7-Point Incremental Polynomial Method

The quadratic log-log 7-point incremental polynomial method was used

to determine if a second order curve fit could improve the performance of

the linear log-log 7-point incremental polynomial method. This method is

essentially the same as the linear log-log 7-point incremental polynomial

method except that the fitted polynomial is second order instead of firct

order.

The fitted polynomial takes the following form:

log a - b + blNLS + b 2 NL.2 (19)

The growth rate, da/dN, for this method, evaluated at the midpoint, Nit
2
is given by - [b 2 (l8Ni) 2b 2 ClogN- +b 2 C1]2

dN1 C2Ni

2b,2C 1 +2b21loic2
bl]
(20)
C2I
The derivation of this equation is shown in Appendix B.

20
I

SECTION V
i
STATISTICAL CONCEPTS
i-

When used properly, statistics is extremely useful in quantifying the

results of many engineering experiments. In many applications, however,

statistics is used as a quick substitute for a thorough experimental

analysis and often times it is used without checking the underlying

assumptiona or else the results are misinterpreted. In an attempt to

alleviate these problems, the statistical concepts used in this investi-

gation and their use as tools in analyzing fatigue crack growth behavior

will be presented and discussed.

5.1 Histoarams

The first step in statistically analyzing any set of data is to see

what the data looks like. Histograms are statistically derived pictures
I

of a data set. They give a rough idea of the shape of the density func-

tion of the data. They also five a rough estimate of the average value

and the amount of variability present in the data.

The most common histogram used is a frequency histogram. The data is

divided into several classes and the frequency of the data in each class

is plotted against the limits of the classes [11]. This type of histogram

frequently takes the form of a bar chart. A slight modification of this

involves calculating the relative frequencies in each class by dividing

the frequency in each class by the total number of data points. The re-

lative frequencies are then plotted against the limits of the classes.

21
 This is called a relative frequency histogram [12]. An example of a re-

lative frequency histogram is shown in Figure 8.

Another convenient form of the histogram is called a cumulative fre-

quency histogram. This histogram shows the frequency of data less than or

equal to a specified value. It is calculated by cumulatively adding suc-

cessive class frequencies of the frequency histogram from the smallest

class value to the largest class value. It frequently takes the form of

a step chart. Again, the relative cumulative frequencies can be calculated

by dividing the cumulative frequencies by the total number of data points

so that the last value of the relative cumulative frequency is equal to

one. When the relative cumulative frequencies are plotted against the

limits of the classes, the resulting plot is called a relative cumulative

frequency histogram C12]. An example of a relative cumulative frequency

histogram is shown in Figure 9.

5.2 Distributions

Once a rough idea of what the density function of the data looks like

based on the histograms, the next step is to try to fit the data to several

likely distributions. Eighv different distributions were selected as

likely candidates for the distribution of fatigue crack propagation

variables.

5.2.a Two-Parameter Normal Distribution

The most widely used distribution in statistics is the two-pararvter

normal distribution [121. This distribution was selected as a candidate

for the distribution of fatigue crack propagation variables mainly for

this reason and for the sake of completeness.

22

la. j M .--. - - -----

 RELRTIVE
FREQUENCY HISTOGRAM
lawo- REPLICATE CA TESTS.
N CLRSS SIZE = 4771
R = 49.80 MM
50 DATA POINTS
R = .20
.ISMo -

.12W0

Lz->- .1200 -I----

rUj

C3
, L- .0900

I---
-J
, .0600 -
bc

p0300-
!11
! ~~0.0000
.200 .2400 N (CYCLES)
.sw .2000 .1oo0 .32,o
(X1O 6)

' Figure 8. Typical Relative Frequency Hietogram

23

L
I
RELRTIVE
CUMULRTIVE FREQUENCY HISTOGRRM
REPLICATE CR TESTS.
N CLASS SIZE = 4771
A = 49.80 MM
60 DATA POINTS
R =.20

C3
S2J
LJJ

LU-
1.00-

LLJ

-J

.200

0.000A
Mm0"' '."2M
N (CYCL.ES) ()(1 )

Figure 9. Typical Relative Cumulative Frequency Histogram

24
 The two parameters of the two-parameter normal distribution are the

mean, designated by ., which is the scale paramster, and the standard

deviation, designated by a,which is the shape parameter. The density

function, f(x), for the tvo-parameter normal distribution is given by

0< < (21)

t
The estimates for the mean and standard deviation are computed by (11,12]
n
F (22)
n

SWhere n is the number of data point& and th symbol * symbolizes an

estimated value.

The standard errors of the estimate# provide a measure of how good

these estimates are. The standard errors of the estimated man and

standard deviation are given by (13)

s. R, - (24)

r 2

Fn (to ' 1) - 2* 2 (25)

where r represents the Same function. The covariance of with S is

always equal to sero, due to their orthogonality (13].

5.2.b To-Parameter Log Normal Distribution

The two-paraucer log normal disLribution has been suspected of beinS
a likely candidate for the distribution of fatigue crack propagation

25
variables [5,61. Essentially, the two-parameter log normal distribution

states that the Iogi0 of the random variable X, i.e., log1 0 x, is normally

distributed.

The two parameters of the two-parameter log normal distribution are

p, the scale parameter, and 0, the shape parameter. Tho density function

for the two-parameter log normal distribution is given by (14,15]

-X> 0

The estimates for p and 0 are computed by using the following *qua-

tions [14]

n
E lolOgoli
i- 1 (27)
n n 21
E (logloix
E- -. )2
il 1 (28)
n

The standard errors of these estimates are given by the following

equations (13,14].

S. E. , - 4 n (29)

.. "Zn2 (30)

The covariance of with f is again always equal to zero, due to their

w

orthogonality (13).

5.2.c Three-rarameter Log Normal Distribution

With the expectation of a better fit of the data, the three-parameter

log normal distribution was considered as a candidate for the distribution !

of fatigue crack propagation variables. -The main difference between

26
 the two-parameter and the three-parameter log normal distributions to the

inclusion of the lccation parameter in the three-parameter log normal dLs-

tribution.

The three parameters of the three-parameter log normal distribution

are j, the scale parameter, 0, the shape parameter, and the terminus, T,

which is the location parameter. The density function for the three-para-

meter log normal distribution is given by (14,15)

f(X) " (x )/2(B elogx -(%-xr) - , 0 < 0 < (31)

The difficulty in using distributions containing a location parameter

is the estimation of that location parameter. The parameter estimation

methods used to obtain the value of the location parameter are presented

in Section 5.3.

Once the location parameter, T, has been estimated, g and are easti-

rated using the following equations [151.

! n
-l 1 g ,. , 0(32)

n 2
i l
[og1 )
~ 1-0 x -
,, (33)
n
To obtain the standard errors of the estimates and the covariance

values, the covariance matrix for the three-parameter log normal distribu-

tien is computed. The covariance matrix is a symmetric matrix and is given

by (161

27
 (34)
V CV 2{(0+ 1)GXP(i) -1] 2 iex{p[ ] 34

where

( :(35)

The standard errors of the estimates are given by the diagonal terms and

the covariances between the estimates are given by the off-diagonal terms

of the 3 by 3 covariance matrix.

5.2.d Three-Parameter Weibull Distribution

The three-parameter Weibull distribution has long been considered in

representing fatigue data [17]. For this reason, the three-paramiter

Weibull distribution was considered as a candidate for the distribution of

fatigue crack propagation variables. This distribution also includes the

location parameter as one of its three parameters and thus the difficulty

of its estimation arises. Two basic methods were used to estimate the

parameters (Section 5.3) and each method required different paramters.

Thus, two sets of Weibull parameters and their associated equations will be

presented.

The first set of the three parameters of the three-parameter Weibull

distribution include the characteristic value, 0, -,+Lch is the scale pare-

meter, the Weibull sLope, B, which is the shape yarameter, and the expected

minimum value of X) .0, which is the location parameter. The density func-

tion for these parameters is given by [18)

28/ 4
 S B (.Xo/ B- ex (36)

"" 0 B- < 0 Xo < X

In the method of estimating the location parameter used with this set of

parameters, all three parameters are estimated simultaneously.

The second set of three parameters of the three-parameter Weibull dis-

tribution include b, the scale parameter, c, the shape parameter, and the

terminus, r, which is the location parameter. The two sets of parameters

are related as follows.

b 0 - (37)

c B (38)

- X (39)

The density function for the second set of parameters is given by [19]

f(X) C(X-r)c lbex-C , 0 < c < (40)

L b - < '<
Y

As with the previous set of parameters, all three parameters are

simultaneously when the location parameter is estimated. Sestimated

To

obtain the standard errors of the estimates obtained by the method re-

ferred to above and the covariance values, the covariance matrix for the

three-parameter Weibull distribution is computed. The covariance matrix

is a symmetric matrix and is given by (13]

V - v 1 (41)

where

29
 b c 1
T2.. 2(l'y) 2 t FII 1 2]_:

c r (42)

b.2

where y is Eulers Constant (0.577215) and represents the digasmna func-

tion.

Once the covariance matrix is obtained, the standard errors of the

estimates and the covariances between the estimates are obtained from the

same terms in the covariance matrix as outlined above for the three-para-

meter log normal distribution.

5.2.a Ganws Distribution

Due to the nature of the fatigue crack propagation process, two im-

portant assumptions can be made. The first assumption, called the in-

creasing failure rate assumption, states that b- -use the crack growth

rate increases as the crack grows (under consr ?litude conditions),

the rate, or probability, of failure increases ie crack grows. The

second assumption states that the distribution of a fatigue crack propa-

gation variable is independent of the crack length and is a function of

the initial crack length only. If these two assumptions are made, then

it can be proven that a generalized gama distribution is a valid distri-

bution for any fatigue crack propagation variable [13,20].

30
 Generalized Four-Parameter Gamma Distribution. The four parameters

of the generalized four-parameter gamma distribution are the location para-

meter, T, the power parameter, o', the scale parameter, b, and the shape/

power parameter, S. The shape parameter, c, is given simply by [ 2 1]

c - ga (43)

The density function for the generalized four-parameter gamma distribution

is given by [21]

f(X)- 01 1 1 (44)
bs 0 f(g) b Z 0
g~

All four parameters are estimated simultaneously using the parameter

estimation methods presented in Section 5.3. To obtain the standard er-

roe of the cstimates and the covariances between the estimates the covari-

ance matrix for the generalized four-parameter gamma distribution is com-

puted. The covariance matrix is a symmetric matrix and is given by

[13,21,22)

v n
V-1 (45)

where
U 2
2Q-4 rQj4)

& {, r(a) (46 )

' ,.
11 2*(i) + i re'(1) -2ii
AZ I 2il[ k
.2 b r(j)

where *' represents the trigamma function.

31
 - --

The standard errors of the estimates are given by the diagonal terms

and the covariances between the estimates are given by the off-diagonal

terms of the 4 by 4 covariance matrix.

Three-Parameter Camma Distribution. If the power parameter, a, is

assumed to be equal to one, the generalized four-parameter gamma distri-

bution reduces to the three-parameter gamma distribution. The density

function for the three-parameter gamma distribution is given by 121,23]

g
bg r (g) [ 1bL1 J) b
g T- I (47)

The three parameters are estimated using the same method used for the

generalized four-parameter gamma distribution. The standard errors and

covariances are found by using the covariance matrix for the generalized

four-parameter gamma distribution (equations 45 and 46) and setting a

equal to one.

Generalized Three-Parameter Gamma Distribution. If the fatigue crack

propagation variable of interest is &N/Aa, then considering the fatigue

crack propagation process it would be expected that AN vould be zero for

&a zero 124]. From this, it is assumed that the location parameter, y,

is equal to zero which reduces the generalized four-parameter gaza

distribution to the generalized three-parameter gamma distribution. The

density function for the generalized three-parameter gamma distribution

is thus 123]
X >0
"f(X) a L . exp - ) (49)
(s) b b 0
bar

I
32
 The standard errors of the estimates are given by the diagonal terms

and the covariances between the estimates are given by the off-diagonal

terms of the 4 by 4 covariance matrix.

Three-Parameter Gamma Distribution. If the power parameter, a, is

assumed to be equal to one, the generalized four-parameter ganma distri-

bution reduces to the three-parameter gamma distribution. The density

function for the three-parameter gamma distribution is given by [21,23]

f b >)0 (47)
bg r (g) b g >

The three parameters are estimated using the same method used for the

generalized four-parameter gamma distribution. The standard errors and

covariances are found by using the covariance matrix for the generalize..

four-parameter gamma distribution (equations 45 and 46) and setting a

equal to one.

Generalized Three-Parameter Gamma Distribution. If the fatigue crack

propagation variable of interest is &N/as, then considering the fatigue

crack propagation process it would be expected that AN vould be zero for

ta zero [24]. From this, it is assumed that the location parameter, y,

is equal to zero which reduceo the generalized four-parameter gam

distribution to the generalized three-parameter gamma distribution. The

density function for the generalized three-para"ter geama distribution

is thus [23]

r ae> (49)
(9
f() b1 I(,S) 'b T 0
g >

32
 The three parameters are estimated using the same method used for the

generalized four-parameter gamma distribution. The standard errors and

covariances are found by using the 3 by 3 submatrix for b, 8, and & from

the 4 by 4 covariance matrix for the generalized four-parameter gamma dis-

tribution (equations 45 and 46).

.Two-Para.ter Gamma.Distribution. If the power parameter, ct, is again

assumed to be equal to one, the generalized three-parameter gama distribu- J

tion reduces to the two-parameter gamma distribution. The density function
]
for the two-parameter gamma distribution is given by [11, 12, 23]

f(X) " l ex- , b a 0 (50)

bgr(g)
rbg *
The two parameters are estimated using the same method used for the gener-

alized four-parameter &am= distribution. The standard errors and covari-

ances are found by using the 3 by 3 submatrix used for the generalized

three-parameter gamma distribution and setting a equal to one.

5.3 Parameter Estimation Methods

Ii
Since the determination of the estimates of the paramete's is critical

to a proper fitting of the data to the two, three, and four-parameter die-

tributions, two different parameter estimation methods were used (14, 25].

The first method, a graphical method, was selected for its simplicity

[17, 18, 26, 27]. The second method, the method of maximum likelihood

estimators (MLZ), was selected because of its reliability, accuracy, and

Swidespread acceptance [14, 15, 19, 21-23, 28-321.

5.3.8 Graphical Method

This method was the first of the two methods attempted, due mainly to
its simplicity in use [17, 18]. This method was tried with both the

33

000 ME
 _I~~qgJ
. ..........-.... ..........- ' -~~~~~~~~~~~~~~
.......
. . -.... ---
... ---
. . .,. . . -

chr.o-paraimtr l. normal dsltributien and th. three-patpter Vooi|

dticributLon, The graphiial athod &Bevlvei R

plesllng ghe dosi enpeelu-
probability paper whose iaxis dals#l *rreipend to Ipostel diItutvilen
charadetroiLtte and then eiesit"L the value o1OINe liatlen paveaver

ouh that the reouictno plot of dita follow# A 14uaiflif Ithe (f1,1i

Ono$ ihe ultiImst of the leuation partamatf il oOW"In she S!tlmetel at

the other twv parameters are medisraphtiolly, film foily Ohm pura'

"mlrl sin be estimated graphisally, slil limit$ use of this mahed

IM
to two or thiss palamtw diellibusiLeia [I173
for t6. thi~aa-parsinser lug nermal 41airtbutteat a pies of V ve, N
yiewd a italight tins for datial hallt Coll 0hvl!.pltmiel 105 nFIAl

ditrivbution [141 Oarso

Y * 0~l(5.) (II)
K uot

Wore O(M) to th. equation for Otw standard niormal pUolpoliiiy *bej1p 9'ir

to Ftion by (Cil

0(0) toep (a)d*

wheor

u !',)
7( (5k)

where F(s) to the sumulative dmnasiy fNctOW# of Ofe u6ruelsd dot. 0

to the value olf tts dato
l rroutld for iliat value of il loiwiton plasater

by the following equation.

Is C,
three-parameter log normal distribution and the three-parameter Weibull

distribution. The graphical method involves plotting the data on special

probability paper whose axis scales correspond to special distribution

characteristics and then selecting the value of the location parameter

such that the resulting plot of data follows a straight line [17,181.

Once the estimate of the location parameter is known, the estimates of

the other two parameters are made graphically. Since only three para-

meters can be estimated graphically, this limits the use of this method

to two or three parameter distributions 27].

For the three-parameter log normal distribution, a plot of Y vs. X

yields a straight line for data that follows a three-parameter log normal

distribution [141 where

Y G(z) (51)

X - log 1 0 (Xc) (52)

where G(z) is the equation for the standard normal probability scale which

is iven by [11)

z1 2
C(Z) exp 02) dx (53)

where

z - F(Xc) (54)

where F()(C) is cc cumulative density function of the corrected data.

the Xc

is the value of the data corrected for the value of the location parameter

by the following equation.

Xc x- 0 (55)

34

-~- -- f
r

For the three-parameter Weibull distribution, a plot of Y vs. X where

Y In In (-() (56)

X - In (xc) (57)

yields a straight line for data that follows the three-parameter Weibull

distribution [17,181.

For both of these plots, F(X) corresponds to the median ranks which

are calculated by [181

Xc
Fx()( - 1 ! n (58)

To determine the value of the location parameter such that the re-

sulting plot yields a straight line, an iterative process which minimizes

some variable must be used. For the graphical method, the variable to be

minimized is the curvature of a second order curve fit using least squares,

thereby assuring a straight line. One of the fastest and most efficient

of the many minimization methods available is the Golden Section search

[26]. Smethod
In the Golden Section search method, the value of the curvature (the

variable to be minimized) is calculated at two optimal locations and,

based on these values, a certain area where the curvature minimum is know,

not to exist is excluded from the rest of the search. This process is

repeated until the area remaining to be searched is less than some toler-
ance level. The value of the location parameter in this area is then
taken as the estimated value of the location parameter. A schematic re-

presentation of the Golden Section search method is shown in Figure 10.

35
 L I I -
- I mX III I-
"I

LU
I II I
.-
c'JI
I"I I.
I I-
-
I.} I j:I

I I-I_.
I I I _ _ >

T (LOCATION PARAMETER)

IF Be (Xi) < B2 (X2)

THEN Xe BECOMES Xwx

IF Be(X,)> Be,(X 2)
THEN X, BECOMES Xm!N

IF B,(XI) = B2 (X2 )
THEN Xe BECOMES Xmnx
AND X, BECOMES XmjN

H = 0.618033989

j F.Figure 10. Golden Section Search Method

36
 i

5.3.b Maximum Likelihood Estimators Method

After the griphical method was perfected and usqd, the need for a

more statistical approach to the estimation of the rametera of the two,

three, and four-parameter distributions became evident (Section 8.1).

This led to the use of the Maximum Likelihood Estimators method to statis-

tically estimate the distribution parameters.

The Maximum Likelihood Estimators (MLE) method involves solving maxi-

mum likelihood equations through the use of a nonlinear programming algor-

ithm [14,15,19,21,22,23,28,30,31,323. Many forms of the maximum likelihood

equations have been determined by investigators for the three-parameter log

normal distribution, the three-parameter Weibull distribution, and the two,

three, and four-parameter gamma distributions C14,15,19,28-32].

Three-Parameter Log Normal Distribution. The maximum likelihood equa-

tion used in this investigation for the three-parameter log normal distri-

bution is [15]

in L(T) (T) +- ln (T) (59)

n
where il ln(x'r)) ((0)
n 2

and (T) (61)

Three-Parameter Weibull Distribution. The maximum likelihood aqua-

tion used in this investigation for the three-paramater Weibull distribu-

tion is [19]

-c n
L(b,c,,r) a n(In c - clnb) + (c - 1) E ln(Xi-?) - b" (X1-1') (62)
i-I imi

Note that the maxium likelihood equation is a function of all three para-
meters whereas for the three-parameter log normal distribution, the maximum

37
 J

likelihood equation is a function of Just the location parameter. However,

the scale parameter, b, crn be factored out of this equation and estimated

separately. The resultint. '- parameter maximum likelihood equation for

the three-parameter Weibu.;' ,istribution is [13]

L(c,T) -In c - In E ) + E (63)+

inl n [J
where the estimation of the scale parameter is given by 13]
n 1~l/c
- (i~)C
2 1/c(64)

The effect of reducing the numbet of parameters in the maximum likelihood

equation is to reduce the computing time, and thus the cost, of the maxi-

mization of the maximum likelihood equation.

Generalized Four-Parameter Gamma Distribution. The mAximum likelihood

equation for the generalized four-parawater gamma distribution is [21,30,

32]

L~bg~ar)- n lmvy + (got-l) [ E ln(x -r)] s n ln(1)

(65)
n rnrbg)
Ji b

The number of parameters in this equation can also be reduced by factoring

out the scale parameter, b. The resulting three parameter maximum likeli-

hood equation for the generalized four-parameter game distribution is [13]

L(g,aT) -n lno'+ (-) 1Ji ln(Xi-'r) - g 1 + In E (X -)r

I1
- Ingn]

(66)
- a InF(g)
where the estimation of the scale parameter is given by [13]

38
 n
bai / (67)

Three-Parameter Gamers Distribution. The maximum likelihood equation

for the three-parameter gaomma distribution reduced to eliminate the scale

parameter is 13]

n n

L~,,. . (S-1) [ . n(X,-')] - [Iu+ In

ini
(x,-)
Eln(n)] -
i-i
(68)
- n lnr(g)

"where the estimation of the scale parameter is given by (131

_____________ I__________________ (69)

La

Generalized Three-Paramter Gamma Distribution. The maxitum likeli-

hood equation for the generalized three-parameter gamma distribution is

(21,30,32]

L(b,ga) -n Ina + (gce-l) [E n

J ul=
ln(Xi) gn ln(b0 ) (0
( 70 )
" ~n
-t n lnr(g)

This equation can also be reduced to eliminate the scale parameter, b.

The resulting two parameter maximum likelihood equation for the Senor-

alized three-parameter gama& distribution is [13)

L(Sao) - n Irtv + (ga-l) [I n

F. n(Xi1 I + 1lnE
n
(xiJ' -n l(Sn)J
(71)
- n ln1s)

where the estimation of the scale parameter is given by [13]

39

_______________________________________________________________
 bn E (X1)J (72)
Jii

Two-Raraater amma Distribution. The maximum likelihood equation

for the two-parameter gamma distribution reduced to eliminate the scale

parameter is (13]

L(g) -(g-1) Z ln(Xi)] - gxil + in Z (Xi4 ln(gn)

(73)

- n lne(g)
where the estimation of the scale paramiter is given by [13)
na
b (Xi (74)
JI"

Interior Point Penalty Function. After some experience using the .

Graphical method to estimate the location parameter of the three-pare-

meter Weibull distribution, it was found that the iteration tended to go

to minus infinity in some cases. Since this was the global (overall) maxi-

mum of the function to be maximized, it became necessary to use a method

that converged on the local maximum, and not the global maximum. The

method used to achieve this requires the use of an interior point penalty
function which prevents the value of each of the parameters from reaching

either of its global limits [151.

The interior point penalty function, better known as the objective

function in nonlinear programming terms, for the three-parameter log normal
distribution using the maximum likelihood equation is (15)

F(i,r) - ln L(T)- {(, + c+ -, +)1 + (Xm

1 n.,. +)-1] (75)

where c+ is a large positive number (a- 1025),

40
 r is an iteration variable, and
8
0 ).
9 is a small positive number (-

The objective function for the three-parameter Weibull distribution

using the two parameter maximum likelihood equation is 15]

P(Tc,r) - L(r,c) - {(T + C+- ) + (1<in " r'+)

(76)
+ (c - -+)+ (10- c - +)-

The objective function for the generalized four-parameter gams dis-

tribution using the three parameter maximum likelihood equation is [15)

Plg,oT,r) - L(g,cv) - (g - 1 - e+) + (100 - - "

+ (w
-1 +- (100
+ )1 (77)

+ +
()T - + + (xC- . r )-T]

The objective function for the three-parameter gamma distribution

using the two parameter maximum likelihood equation is [15]

P(gs,,r) - L(g,) - (gU 1 - + -1 + (l00- g -(8)+(78)

+ (T' 4 c+ - e+ -1 + (xain T- +)

The objective function for the generalized three-parameter gamma die-

tribution using the two parameter maximum likelihood equation is [15]

"P(g,r) - ,L(Sg) - - 14
)- I + (100- g- )1

(79)

+ ( - C-+) " + (100 -a - +)1

41
 I
rhe objective function for the two-partmeter gamma distribution using

the one parameter maximum likelihood equation is [15]

S,+)-1-1

P(gr) a L(g) - (g 1- C + (100-gS - )" (80)

The algorithm used to converge the objective function towards the

local maximum likelihood is as follows [15].

1. Maximize the objective function, P(r,r).

2. Check for convergence to the optimum i.e. when

- 'r(r,. 1)I < e (81)

where g is the convergence criterion constant.

3. If the convergence criterion is not satisfied, reduce r by

setting

rJ.l -drj, 0 < d< 1 (82)

where d is a convergence constant.

4. Increment j and repeat.

The maximization of the objective function has been done by many non-

linear routines [191. However, the Hooke-Jeoves pattern search method [33]

has enjoyed particularly good success in maximizing KZ objective functions

and was therefore utilized in maximizing the objective functions for the

three-parameter log normal distribution, the three-paramster Weibull dis-

tribution and the two, three, and four-parameter Saim distributions [15].

5.4 Goodness of Fit CriteSra

Once the statistical parameters for each of the candidate distribu-

tions heve been estimated, the distribution which the date follows the

closest must be selected from the candidate distributions. A statistical

42
 method which is used many times to find out how well data fits a certain

distribution is the goodness of fit test. Several goodness of fit tests

have been proposed (121, but three of the more reliable and widely used

goodness of fit tests have been selected as criteria for the selection of

the "best" distribution. These three goodness of fit tests are regres-

sion, the chi-square test, and the Kolmogorov-Smirnov test.

5.4.a Regression

Regression in its simplest form involves fitting a polynomial to a

set of given data plotted on certain axes (341. In the case of fitting

data to a statistical distribution, the data can be plotted on a plot

whose axes correspond to certain characteristics of that particular

statistical distribution (Section 5.3.a). It is known that if data fol-

lows that particular distribution, then the data will follow a straight

line fit when plotted on these special axes. If a linear regression is

performod on this plotted data, it can be determined how close the data

does fit a straight line. This then provides a measure of the goodz-ess

of fit of the data to that particular distribution.

If a set of data follows the tvo-paraumter normal distribution, a

plot of the data with the X axis as a linear scale and the Y axis as a
normal probability scale vill follow a straight line (181. The normal

probability scale is described in detail in Section 5.3.a. A typical

plot for the two-ptrazter normal distribution is shown in Figure 11.

If a set of data follows the two-parmater log normal distribution,

a plot of the data with the X axis as a log 1 0 scale and the Y axis as a

normal probability scale (Section 5.3.a) will follow a straight line '18].

In this plot, the location parameter is not estimated and is assumed to be

43

- . . .1- . . . . .. ... ... ... . ..

 2-FRRRMETER
NORMRL DISTRIBUTION PLOT
99.9 REPLICRTE CA TESTS. /
z R=49. oMM
68 ,RTR POINTS 1
R .20
CL 98.0- x
x
>" //X
--j g0.0- x,

iL
C-
cz
9-4
60.0-
G:
/ /
/ x

"CO20

r-=
A= .8133

.0986
f
z 10.0- // = .91810
cc8 x
/ xA

2-/ S.E. (0- = 15220.

I' SLOPE =4.8,10 6
:1 /
0.1
-60000.0 -50000.0 0.0 30000.0 60000.0
DEVIRTION OF N FROM THE MERN

Figure 11. TYpical 2-Paremeter Normal Distribution Plot

44i
zero. A typical plot for the two-paramster log normal distribution is

shown in Figure 12.

Both the plots for the three-parameter log normal distribution and

the three-parameter 'eibull distribution have been discussed in Section

5,3.a. A typical plot for the three-parameter log normal distribution is

shown in Figure 13 and a typical plot for the three-parameter Weibull dis-

tribution is shown in Figure 14. The three-parameter gamma distribution

plot requires the data to be plotted on a plot where the X axis is a

linear scale and the Y axis is a gamma probability scale. The equation

for calculating the gamma probability scale, H(s), is [27]

PH(Z) -1-
r(&) j t -lce dt (83)
0

where

z - F(Xc) (84)
c
where F(X) is the cumulative density function of the corrected data whicb

is given by equation (58). Equation (83) was solved iteratively for H(z)

using the interval halving method [271. A typical three-parameter gamma

distribution plot is shown in Figure 15. The two-parameter Samms distri-

bution plot also requires the X axis to be a linear scale and the Y axis

to be a Samna probability scale. A typical two-parameter game distribu-

tion plot is shown in Figure 16. In each of the above plots, the data

are plotted on the X axis against the corresponding median ranks on the
Y axis.

Linear regression uses linear least squares which uses the matrix

approach to linear regression to fit a best fit straight line to the data

(34). An a result of this matrix approach to linear regression, a goodnmas

45
 2-PRRRMETER
LOG NORMRL DISTRIBUTION PLOT
99 .g_ REPLICRTE qR TESTS.
Z - R = 4-980 MM
J 68 ORATVPINTS
o:: R ).20 ,

0l 98.0- /
/x
- x
-1J 90.0-

SC13.

cc 60.0-
S/0.
i//

c.
C310.0-
C32.0- .. A = .94330
(J R= .9003

:/ S.E. (U) = .00363

,_j /
SLOPE 29.836
=3 0.1 /- -

N (CYCLES)

Figure 12. Typical 2-Parameter Log Normal Distribution Plot

r
t4
 2-PARRMETER
LOG NORMRL DISTRIBUTION PLOT
REPLICATE tA TESTS.
=- 491 80 MM
68 ORTVPOINTS
CK R #.20

0~98.0 /

/x
I- x
- 90.0

ar-

C3A = .9530
0: .oZ = .0901
3 . R = .94330
cc. / Xx

'= 0 S .4092
SA
.0- / S.E. (U) = .00363
/- B =.00089S
cc _.1 / S.E. (9) =.000152
:::/ = 29.836
C. 0.1-/

N (CYCLES)

I1Fiure 12. Typical 2-Parameter Log Normal Distribution Plot

t4
 3-PRRRMETER
LOG NORMRL DISTRIBUTION PLOT
.9-9 REPLICATE CA TESTS. /
=49. 80 MM
S"J 68 DATA POINTS
Lj- R =.20

0_ 98.0- /

_x

Q_
CC 60.0-
a: so.o-

, A = .9633
Z =.0593
= .98257
C3.-
oo
0 AT = 200844
CC x S.E. (T2 = 81.77
A U = 4.7294

2.0-
20 S.E. (U) .72900
,.., = .018307/
cc(z S.E. (6) =.026869
/ SLOPE 6.734

2 10 2
N (CYCLES)-TAU HAT

Figure 13. Typical 3-Parameter Log Normal Distribution Plot j

47
 3-PRRRMETER
WEIBULL DISTRIBUTION PLOT
99.s- REPLICRTE CR TESTS. ,
sg~~~o 68A DRTR
= 49.80
PI MM
INTS i
R :.2o0

- //
70.0

I50.0- S x

c- / x A= .9533

x S/.92634
(1. /

-x / T = 220968
1.0- // S.E. (T) =10120.0
"L I ~/ / B = '0848
/ S.E. (B) 21180
// C 2.0649
S.E. (C) 2.5563
10 2 SLOPE = 1.3013
0. -I i I a I
10 10 2
N (CYCLES)-TAU HRT

Figure 14. Typical 3-Parameter Weibull Distribution Plot

48

'W. ' " . .... "-'r m'' -

 3-PARAMETER
GRMMR DISTRIBUTION PLOT
9.9- REPLICATE CA TESTS.
A = 49.80 MM
68 DATA POINTS
R =.20
i, /
99.0- /
//
- S99.0
U 90 // x

a- 9s.o- /X/ x

-90.0-
X/
A= .9378
a:: Z = .0633
- S70.0 / R = .97 140 i
* 0.. A

* C 50.0-
0.0 SE(A T = 213396
S.E. (T) =10277.0
El: B^ = 7238.4
!: e~o-S.E. (8 = 2427.7
AG = 6.0467
/ S.E.SLOPE
'GI = 3.3663"
=4.8401

0.1- SLv L 1
S-6=0.0 -30.0 0.0 M000.0 60000.0
DEV. OF (N-TAU HRT) FROM THE MEAN

Figure 15. Typical 3-Parameter Gamae Distribution Plot

49
 2-PRRRMETER
GAMMR DISTRIBUTION PLOT
99.9- REPLICATE CA TESTS.
SECANT METHOD
DELTA A = .20 MM
68 DATA POINTS
,A-6- A = 22.10 MM //
99.0- R .20 /

- 98.0-

C1'
LU1
. 96.0- x/

a_.. 90.0-

-J 70.0-

rm60.0 o 8 8
CL Z .0710
CC R .98864
C:: A = 6.163=103
A 3
ix' S.E. (B = 1.047x10
/ x AG z 2.S36x10l
/x S.E. (G) = 4.313
/SLOPE = 2.902x1
0 .1-Il I I

-=0000.0 -40000.0 0.0 40000.0 80000.0

DEVIATION OF ON/OR FROM THE MEAN

Figure 16. Typical 2-Parameter Gaum Distribution Plot

i 0i
so
Ltl Ae~- .- . - ,
 of fit statistic, called the coefficient of multiple determination, R

can be calculated. The value of R2 is always between zero and one. The
~ closer the value of R2 is to one, the closer the fit of the data is to a
22
straight line. Therefore, by comparing the values of R for each of the

distributions, the distribution with the highest value of R2 is the dis-

tribution which the data follows the closest.

This value of R2 can be corrected for the slope of the least squares

line in an attempt to achieve a more precise measure of the closeness of

the data to the straight line. This corrected value of R2 is called the

closeness and is given the symbol C . The derivation of C is given in

SAppendix C.

5.4.b Chi-Square Test

The chi-square goodness of fit test is a statistical method fir de-

termining how close given data follow a certain distribution. Basically,

it

I) divides the data into an optimum number of equiprobable intervals,

2) counts the number of data points in each interval (called the ob-

served frequencies),

3) calculates the number of data points that should be in each in-

terval based on the estimated distribution parameters (called the

expected frequencies), and

4) compares the observed frequencies with the expected frequencies

"2
The test statistic, X , is a measure of how close the observed frequencies

i; are to the expected frequencies, and t us how close the data follows the

given distribution. X is given by

f . ... . ...
... . . ... . I . .. . . . . . .. . . I-.... . . I [" .. . . .. . ,_ -5--
 2 k (o -et) 2

where k is the number of equiprobable intervals,

0 are the observed frequencies, and

e are the expected frequencies.

The lover the value of the chi-square statistic, the closer the ob-

served frequencies match the expected frequencies and thus the closer the

data follows the given distribution. However, the chi-square statistic

can not be compared between distributions that do not have the same number

of distribution parameters, n because the degrees of freedom for the chi-

square statistic for distributions not having the same number of distribu-

tion parameters is not constant [131. Therefore, the tail area of the

chi-square distribution to the right of the chi-square statistic, called

A, is computed for each distribution by [131

A' AaX1 u
/2 exp(-u) du (86)

where v is the number of degrees of freedo- and u is a variable of inte-

gration. The .alue of A is always becween zero and rmne, with A equal to

... .ig a perfect fit. The lover the value of the chi--.quare statistic,

the higher the value of the tail ares, all other thln,' There-

fore, the distribution to be chuscn as the distribution whiJI the data

follows the closest is thi one which has the highest value of A.

The chi-square statistic may be cor.pared with a critical value which

foilown the chi-square distribution at an ac.eptance level of ca with v

Sdegrees 2
of freedom, y2 [121, where
S~a

v k -n - (87)

52
 Acceptance of the proposed distribution as the distribution which the data

follows should occur when [12)

2 2
Xb : (88)
wit the acceptance level to test ac-V
The tail area, A, may be compared with ac-

ceptance of the proposed distribution. Acceptance should occur when (131

A 2: c (89)
a

The end points for the classes for the two and three-parameter normal

distributions were found by dividing a standard normal curve into differ-

ent numbers of equiprobable intervals [35). The end points for the equi-
probable intervals for the three-parameter Weibull distribution were

given by [19]

- + - (90)

The end points for the equiprobable intervals for the two, three, and

four-parameter gamma distribution were given by [21)

-1 +() 1/ I
where F. is the inverse cumulative density function for the generalized
9
four-parameter gamma distribution.

5.4.c Kolmogorov-Smirnov Test

The Kolmogorov-Smirnov test is another statistical goodness of fit

test similar to the chi-square goodness of fit test. Basically, it cal-

culates the sample cumulative density function and compares it with the

theoretical cumulative density function of the given distribution by cal-

culating the maximum deviation, D, between the two cumulative density

&53
functions [11. The test statistic, Z, is a measure of how close the two

cumulative density functions are and thus how close the data follows the

given distribution and is actually equal to D.

The lower the value of the Kolmogorov-Smirnov statistic, the closer

the sample cumulative density function lies to the theoretical cumulative

density function, and thus the closer the data follows the given distribu-

tion. Therefore, the distribution to be chosen as the distribution which

the data follows the closest is the one which has the lowest value of the

Z statistic. The Kolmogorov-Smirnov statistic may be compared with a

table of critical values to determine if the proposed distribut~oihould

be accepted as the distribution which the data follows [36].

in HII' Tp H a
m'] I m I a lli mIN le me i

54
 SECTION VI

DETERMINATION OF THE DISTRIBUTION

Several computer programs were written to determine the distribution

of the desired fatigue crack propagation variables using the previously

mentioned statistical concepts. The four programs vritten to determine

statistical distributions of fatigue crack propagation variables are:

1) Delta N Distribution Determination Program (Golden), or

DNDDPG,

2) Cycle Count Distribution Determination Program, or CCDDP,

3) Crack Growth Rate Distribution Determination Program, or

CGRDDP, and

4) Delta N Distribution Determination Program (WLl), or DNDDP.

6.1 Delta N Distribution Determination Proaram (Golden)

This program, called DNDDPG, was written to determine the distribu-

tion of the aN/Aa variable computed from the input a vs. N data which is

supplied by program DELTCP (Section 7.1). Basically, it fits the data to

four distributions and computes a goodness of fit statistic for the com-

parison of the distributions. The four distributions fitted are:

1) the two-paramater normal distribution,

2) the two-parameter log normal distribution,

3) the three-parameter log normal distribution, and

4) the three-parameter Weibull distribution.

55
It uses the graphical method, including the Golden Section search method,

to estimate the location parameter for both the three-parameter log normal

distribution and the three-parameter Weibull distribution. The goodness

of fit criterion used is C2 (Section 5.4.a).

This program produces output which includes the input a vs. N data,

the computed a vs. N data, some of the test conditions, some of the in-

ternal program parameters, the frequency distribution array, the &N/Aa

data, and the distribution parameters and a partial analysis of variance

table for each distribution. The plots generated by this program are a

relative frequency histogram, a relative cumulative frequency histogram,

and a distribution plot for each of the distributions. Further documenta-

tion of this program is shown in Appendix D.

6.2 Cycle Count Distribution Determination Program

This program, called CCDDP, was written to determine the distribution

of the N (cycle count) variable from a set of replicate cycle count data

at one crack length level. Identical load and test conditions are re-

quired for the replicate data. This program fits the data to six distri-

butions. These distributions are:

1) the two-parameter normal distribution,

2) the two-parameter log normal distribution,

3) the three-parameter log normal distribution,

4) the three-parameter Weibull distribution,

5) the three-parameter gamma distribution, and

6) the generalized four-parameter &ain distribution.

It uses the Maximum Likelihood Betimators method to estimate the para-

meters of each of the above distributions except the two-parameter normal

56
* distribution and the two-parameter log normal distribution. Three good-

Snoess of fit criteria are calculated for the comparison of the distribu-

* tions. They are:

1) the chi-square tail area,

2) the Kolmogorov-Smirnov statistic, and

2
3) R from regression.

This program produces output which includes the input replicate cycle

count data, the test conditions, some of the internal program parameters,

the frequency distribution array, and 1) the estimated distribution

parameters, 2) a partial analysis of variance table, and 3) the goodness

of fit criteria for each distribution except the generalized four-para-

meter gamma distribution, for which only the estimated distribution para-

meters and the goodness of fit criteria are printed. It also prints a

comparison of the distributions and the resulting "best" distribution

based on the goodness of fit criteria. The plots generated by this pro-

gram are the original cycle count data plot, a relative frequency histo-

gram, a relative cumulative frequency histogram, and a distribution plot

for each of the distributions except the generalized four-parameter game

distribution. Further documentation of this program is shown in Appendix

| E.

E6.3 Crack Growth Rate Distribution Determination Program

This program, called CGRDDP, was written to determine the distribu-

tion of the crack growth rate (da/dN) variable from a set of replicate

da/dN data at one crack length level. This da/dN data is calculated by

the DADNCP program (Section 7.2). Identical load and test conditions are

required for the replicate data.

57
 This program is nearly identical to the CCDDP program (Section 6.2),

using the same distributions, the same parameter estimation method, the

same goodness of fit criteria, and having nearly the same output. The

main difference is the variable of interest being da/dN instead of cycle

count. Thus the required input is different and some of the output is

different in this respect. Further documentation of this program is

shown in Appendix F.

6.4 Delta N Distribution Determination Program (WMU)

This program, called DNDDP, was written to determine the distribution

of the AN/Aa variable from a set of replicate da/dN data et one crack

length level. The da/dN data used is the same as that used by the CGRDDP

program (Section 6.3).

This program is based on the CGRDDP program. One main difference be-

tween them is that the input da/dN data is inverted to create the vari-

able AN/a. The second main difference is the assumption that for the

gamma distributions is equal to zero, thus reducing the 3-parameter gamma

distribution and the generalized 4-parameter gamma distribution by one

parameter (Section 5.2.3). Along with the change in variable, there are

appropriate changes in the output. Further documentation of this program

is shown in Appendix G.

58
 SECTION VII

GROCJTH RATE AND GROW/TH PREDICTION

Since this investigation was not just interested in the distribution

of fatigue crack propagation variables alone, it became necessary to

write several other programs to aid in the analysis of the experimental

data. These supporting programs inclued 1) Delta N Calculation Program,

or DELTCP, 2) da/dN Calculation Program, or DADNCP, and 3) a vs. N Pre-

diction Program, or AVNPRD. Several others not mentioned here were used

tc .n the analysA6s and manipulation of the experimental data.

7.1 Delta N Calculation Program

This program, called DELM P, vas written to calculate intermediate

6a vs. &N data to be used by program DNDDPG (Section 6.1). Basically,

it calculates Aa vs. AN data from a set of constant amplitude a vs. N

data by one of five different methods. These methods are;

1) the secant method,

2) reject certain selectable data points and use the secant method,

thereby increasing Aa,

3) the quadratic 7-point incremental polynomial method,

4) reject certain selectable data points, recreate new a vs. N data,

j and then use the quadratic 7-point incremental polynomial method,

and

5) use the quadratic 7-point incremental polynomial method, recreat,

new a vs. N data, reject certain salectable data points, and then

use the secant method.

59
Further documentation of this program is shown in Appendix H. I

7.2 da/dN Calculation Program

This program, called DADNCP, was written to calculate the crack

growth rate, da/dN, by the six different methods presented in Section

4. These methods are;

1) the secant method,

2) the modified secant method,

3) the linear 7-point incremental polynomial method,

4) the quadratic 7-point incremental polynomial method,

5) the linear log-log 7-point incremental polynomial method, and

6) the quadratic log-log 7-point incremental polynomial method.

For each of these methods, the calculated da/dN data is integrated back

into estimated a vs. N data, which is compared with the original a vs. N

data, resulting in an average incremental error. By comparing these

errors, the da/dN calculation method which results in the lowest error

can be selected.

The required input for this program is a set of constant &a a vs. N

data. This program produces output which includes the input a vs. N data,

the test conditions, da/dN vs. 6K and actual cycle count data vs. esti-

mated cycle count data for each da/dN calculation method, and a summary

of the errors from each method with the resulting "best" da/dN calculation

method. Further documentation of this program is shown in Appendix I.

7.3 a vs. N Prediction Program

This program, called AVNPRD, predicts a vs. N data from the distri-

bution of da/dN (or dN/da) as a function of crack length and compares it

60
 with the original a vs. N data, The required input to the MM1wLdg. @A

the distribution of da/dW (or dX/da) so a funcion oforhk length an

determined by the CC=?DD (or DNDDP) program, This pro~ram seOloes a
growth rate at each crack length using a random number ganerator and the

distributcon parameters. Thi growth rate is then used to salulase AN

as a function ot crack length vhwih is used to predirt replicate lats
of a vs. N data. These predicted oets of a vs. N data are then vompared A
with the original a vs, W data auts.

This program produces output which inuludeo the tees *endLSImep the j
predicted a vs, X data, and a plot of all of the prodiseed a vs. 0 date,

further documentetion of this program to shown in AppondLu J,

; Ie|

"f !
AI

4
 [I
with the original a vs. N data. The required input is the knowledge of

the distribution of da/dN (or dN/da) as a function of crack length as

determined by the CGRDDP (or DNDDP) program. This program selects a

grovth rate at each crack length using a random number generator and the

distribution parameters. This growth rate is then used to calculate AN A

as a function of crack length, which is used to predict replicate sets

of a vs. N data. These predicted sets of a vs. N data are then compared

with the original a vs. N data sets.

This program produces output which includes the test conditions, the

predicted a vs. N data, and a plot of all of the predicted a vs. N data.

Further documentation of this program is shown in Appendix J.

16.

L-
 SECTION VIII

STATISTICAL ANALYSIS OF PREVIOUSLY GENERATED DATA

A considerable amount of crack propagation data in the form of a vs.

N data have recently been generated at Purdue University for center crack

specimens of 2024-T3 aluminum alloy [37]. From this set of data, there

were 30 different overload/underload tests which were conducted under

constant stress intensity conditions and at constant Aa. From each of

these tests, approximately 19 to 155 data points, for a total of 2076

data points, were collected after the crack had grown through the region

influenced by the overload/underload sequence. The data typically chosen

for analysis is shown in Figure 17. This large amount of data was col-

lected following the overload affected region to establish a final steady

state growth rate as well as establishing the steady state growth rate

for the next test [37,38,39]. From this set of test results, there are

2 to 7 sets of data at each of five different loading conditions.

The value of these data for statistical evaluation centers on the

accuracy with which the original a vs. N data were collected. In these

tests, the crack length was monitored and measured with a 100X micro-

scope mounted on a digital measurement traverse. The traverse has a

resolution of 0.001 mm (0.00004 in.) with a direct digital read-out. A

printer activated by a push button was connected to the cycle counter

and the digital traverse. In collecting the data, the microscope was

advanced an increment of 0.01 mm, 0.02 m, or 0.05 mm (depending on the

62
 6.000-

Ri
OLUL TEST 13

5.000-

" 4.000-

9-D
C.,
' 3.00o0
-.J
w

2.000-

0 11000 2000 33000 44000 36000

N (CYCLES)

Figure 17. Typical Data Cbosen for Analysis from

Overload/Underload Test Data

63
 growth rate). When the crack had grown this increment as observed with

the cross hair in the microscope, the printer was activated with the push

Irtton and the crack length and number of cycles were printed. The re-

sultir, data are very dense and appear to be fairly accurate. This large

amount of data was used to make a preliminary statistical analysis to aid

in the direction and scope of this inv .,.ation [401.

8.1 Distribution of AN/a

The first step of the analysis was t, determine the distribution of

the variable &N/Aa which was calculated by the secant method. This was

done by writing a pair of programs using many of the statistical concepts

presented in Section 5. These programs, called Delta N Calculation Pro-

gram, or DELTCP (Section 7.1), and Delta N. Distribution Determination

Program (Golden), or DNDDPC (Section 6.1), were run on each of the data

vets. The distributions were ranked from 1 to 4 (1 being the best) based

on the goodness of fit criterion, C2 (Section 5.4.a). The rankings Vere

averaged over all of the tests and the results are shown in Table 1. The

best distribution was the three-parameter log normal distribution fol-

lowed closely by the two-parameter log normal distribution. A plot cf

the fit of the AN/Ia data to the three-paramter log normal distribution

in shown in Figure 18.

Baged on these results and the use of the DBLTCP and DNDDIFG program,

the following conclusions were made.

1) The 2-paramster Welibull distribution was tried and re-

i-cted fio, all further analysis because of its poor per-

"forman.;.' providing a fit for the AN/As data due to it's

Lack of i location paramter.

64
- I I SW!
 Table I

"Distribution of Wt4/6a

DISTRIBUTION Ce S_, AVE. RRNK

E-PARAMETER 0.9668 0.0190 3.95

NORMAL

2-PARAML ER 0.9969 0.0025 1 .87

LOG NORMAL

3-PARRMETER 0.9984 0.0014 1.31

LOG NORMRL

3-fRRA1ETER 0.9932 0.0055 2.67

WEIBULL.

'I !6

k.. T - - ". . --
 PRRAMETER
3*-
AG NCRMRL DISTRIBUTION PLOT
POST OLUL TEST 21
DELTA A = .05 MM CONSTANT
79 DATA POINTS
R= .20/

iL

cr- 58.0 /c.

"0 1
0-

c 50.0- ixi

I
C320.0-

C3~C 100 .9989

U) =c -1
S2.o /-LOS(LG) =2007
LOG 3.30
.LOGS= .22
8 = 3.099
II I
0.1-

3-aae Noma "Distibuton

Log 7

(DELTA N/DELTR A3-Xo

S/ Figure 18. LO
lit of the SN/aa Data to the 6 = 2266

66I
 2) Include the other four dist-ibutions in the analysis of

other fatigue crack propagation variables.

3) Using the constant amplitude portion of overload/underload

data does not lead to a satisfactory statistical analysis.

Therefore, a statistically designed test program was needed.

4) The graphical method of parameter estimation te,3ed to be

unstable and unreliable for the data used. Therefore, the

Maximum Likelihood EstimaL'rs method of parameter estima-

tion was tried and used.
2
5) The use of C as a goodness of fit criterion was poor be-

cause it failed to distinguish between the distributions

very well. Therefore, the chi-square av. Kolmogorov-1Bmirnov

goodness of fit tests were tried and used,

8.2 Effect of Quadratic 7-Point Incremental PolYnOmial Mathoil

The second step of the analysis was to examine the effec

the quadratic 7-point incremental polynomial method vs. ;.sing t
using
iscant
*
method in calculating the variable AN/,a. This was done by runnt... 4 the

DELTCP program and changing the AN calculation method for each of the

data sets. Once the AN/Aa data was calculated for each data set, it was

run on the DNDDPC program to determine the effect of the AN calcula on

method on the distribution parawters. The most noticeable effect was

*a the decrease in the variance using the quadratic 7-point incremental

polynomial method as shown in Table I1. From this, it is evident that

the quadratic 7-point incremental polynomial method introduces quite a

smoothing effect in reducing the amount of data scatter and thus the data

variance.

67

KI
 fable II

Smoothing Effect of the Incremental Polynonial Hathod

Mq. iTF.I. VAR. (I.@'.)

DISTRIBUTION "w .TL.cAFm ') v

VM. BeCMT)

2-PARAMETER 0.419 0.0820

NORMAL

2-PRRRMETER
LOG NORMAL
LOGNORRL0.'444 0.1273

3-PARRHETER 0.647 0.2817

LOG N(5PMAL

3-PARAMETER0
WEIBULL 0.871 O.'4185

II Go

i!
St
 8.3 Life Prediction Using Estimated Distribution Parameters

The next step in the analysis was to see if the estimated distri-

bution parameters could be used for life prediction. Using the mean of

the &N/aa data (for the two-parameter normal distribution) and the over-

all change in crack leogth (af-ao), the final cycle count, Nf, was pre-

dicted and compared with the observed value of Nf for each set of data

and then averaged over all the data sets. The results are shown in Table

III. From the "a:latively low amount of error, it is evident that etatis-

tical methods using estimated distribution parameters could prove invalu-

able for life prediction.

Table III

Life Prediction Based on the Mean

AVERAGE PERCENT ERROR

l 1.011l 2.93

8.4 Effect of A&

The final step in thn analysis of tUie previously generated data was

* to determine the affect of the size of 4a. This was done by using the

DELIC? program to generate AN data with different values of ha. By

k ~694
rejecting certain successive data points (i.e. every i out of 2, every 2

out of 3, etc.), data with increasing values of ,a were generated. The

DNDDPC program was then run on each different aa set of data for each of

the data sets. Also, several tests at the same load conditions were com-

bined to give a large amount of data and then &a was increased as do-

scribed above. The results are shown in Figure 19 and Table IV. From

these results, it is obvious that the larger 4a is, the smaller the re-

sulting variance of the data will be.

70
 .46S00 -
POST OLUL DATA

-. 37b -
0

I-00

L.L

i "i
h-
10 0
z

. : 7SO

0.0000
0 .I0= .0700} . 14O0 .2100 .em .35M0
DELTA A (MM)

F igure 19. Effect of Increasing a&

.71
 Table IV

Effect of Increasing 6a

Vim. athA.lO M9) vm. mf.16b M9)

DISTRIBUTION M) OV"EARvf,e. M,.0E M)

2-PARAMETER 0.617 0.513

2-PARAMETER 0.660 0.539

LOG NORMRL

3-FARRMETER 0.851 0.412

LOG NORMAL

3-PRRRMETER 0.883 0.761

WEIBULL

72
 SECTION IX
EXPERIWNTAL INVESTIGATION

I
In an effort to answer the investigation objectives, it became

necessary to conduct an experimental investigation to provide adequate

data for subsequent analysis. Through the use of previously collected

data (Section 8), it became increasingly clear that any experimental in-

vestigation that would be expected to provide meaningful results would

have to be statistically designed. Through the use of some preliminary

theoretical and experimental testing, a test program was designed.

9.1 Experimental Test Program

Given the objectives of the investigation (Section 3), it was evi-

dent that replicate tests under identical load and environmental condi-

tions had to be conducted. It was also obvious that constant amplitude

loading should be used rather than constant AK (load shad) loading since

it would be much easier to control and replicate and also give a range of

AK levels. To be able to find the distributions of N and da/dN, the data

from each test had to be taken at consistent discrete a levels.

To determine the actual load levels to be used, several preliminary

tests using the same lot of the same material were conducted. To obtain

the desired growth rates (da/dNei 3 1 x 106 in./cycla and

da/dN S 5 x 105 in./cycle) and keep the teoting time within reason,

it was found that AP should be 4200 lbs. It was also determined to use

i
S an R ratio of 0.2 to stay well out of the compression region.

73
 A preliminary theoretical investigation was conducted to determine

where the data was to be taken. It was found that to get the desired

range of growth rates, the data would have to be taken over at least

40.0 un. It was determined that steady state conditions would not exist

until 9.0 = due to the crack initiation load shedding process. In an

effort to reduce data error as much as possible and still obtain a rea-

sonable amount of datathe initial Aa was chosen to be 0.20 m based on

the statistical analysis of previous data (Section 8.4). Since the,

growth rate would be too fast to operate the optical system and the

printer at the end of the test for the load levels chosen, A& would be

increased to 0.40 - and finally to 0.80 mm. The number of data points

taken at Aa w 0.40 mn and 6a a 0.80 -m were arranged so that when succes-

sive data points were rejected (to find the effect of increasing Aa), -

there would be no large gaps in the data. A schematic representation of

the test program is shown in Figure 20.

In order to obtain enough data to conduct a meaningful statistical

analysis, it was determined that there should be at least 50 replicate

tests [13]. However, since more specimens were available, a total of 68

tests were conducted, thereby increasing the confidence of the statistical

analysis results. The test conditions are listed below.

0 - 9.00 .

- 49.80 mm.

Rt a 0.20

Pmin - 1050 lbs.

a- 5250 lbs.

AhP - 4200 lbs.

i ~74 744
 **. SPECIMEN - CENTER CRACKED PRNEL
7.a . 5,1 NIP Ct&II1FNT

M m m ,Am_&m

,-,. - - _- -- -- - -
,. / m ,,
-- IN an "offs
II~lM "Iwo

owl mow amO

SiN~wluII~p''

O ollO....... ..
SIh Inl
0199 Fri

I
1> a
Sr&i~u g
IO e& llr 1!

iwS) N '
| I
1I

'" @ 9g Pau i
 i .00 SPECIMEN - CENTER CRRCKED PRNEL
SPm= 5.25 KIP CONSTRNT
AP =4.20 KIP CONSTANT
R = .20 CONSTRNT
(49.8 M R 8 11
7 DATA MOINTS
(4'.2 M")

2R~.0
DAA ONT
AM,
(36.2 Mvii

P (MM &A .20 MM

137 DATA POINTS

{g~o I)I
R xxxxXXxx
(9.0 Wv)j

0.00
0.0000 .0600 .1000 .1600 .2M0 .2600
N (CYCLES) X10
X1O

,igure20. Test Progra

I 75

i
 9.2 Test Specimen

The test specimens used in this investigation were 0.100 inch thick

center crack panels of 2024-T3 aluminum alloy. The specimen geometry is

shown in Figure 21.

Test specimens were obtained with a trill finish and polished to a

mirror finish in the vicinity of the crack path to facilitate optical

observation of the crack tip during crack growth measurement. The lot of

specimens was numbered in order as they were taken out of the shipping

crate so that true randomisation of the samples could be accomplished.

The fixture plate holes were drilled and reamed to the desired dimen-

sions. The stress raiser shown in detail in Figure 2L was machined with

an electro-discharge machine.

Before loading each specimen, the centerline of the specimen was

scribed at the stress raiser and a silica gel desiccant was applied at

'zhe stress raiser. The entire expected crack path was then sealed with

clear polyethylene to insure desiccated air at the crack tip. Loading

was then applied parallel to the direction of rolling of the material.

9.3 Test Iguipmont

The test machine was a 20 Kip electro-hydraulic closed-loop system

operated in load control. A function generator was used to generate a

sinusoidal voltage signal which, when superimposed on a d.c. set point

voltage, constituted the desired input to the system. During testing, an

oscilloscope was used co *itor the feedback signal (load) and the output

of the amplitule &,%,:rmenc system of the tosting machine to insure cor-

rect load levels aid sinusoidal loading. A digital cycle counter was used

to count the number of 3ppliad load cycles. Crack growth was monitored

with a zoom stereo microscope operated at a magnification of 150x

76
 6mm

1Z 5/16N01A. 9/16M DIA.

22" STRESS RAISER

DETAI01!L

Ila~
Thickness 0 .1000

Z2:5d .25d

i Figure 21.

~77
Test Sp*cmCtiM
 rigidly mounted on a horizontal and vertical digital traversing system.

A crosshair mounted in the microscope was used as a reference line during

data acquisition. A digital resolver system on the horizontal traverse

produced a digital output with a resolution of 0.001 mm (.00004 in.).

The direction of travel of the optical system prior to data acquisition

was never changed during a test to eliminate any hysteresis effects in

the traverse system. Both the digital traverse and cycle counter outputs

(crack length and number of cycles) were connected to a mechanical

printer. The printer printed both the crack length and the cumulative I
cycle count by the operation of a push button. A strobe light synchro- I
nized with the feedback signal was triggered at the point in the load

cycle when the crack was most fully open to illuminate the crack tip.

More detailed discussions of the test equipment can be found in refer-

ences [37,38,39].

9.4 Test Procedure

Since the scope of this investigation strictly involved the deter-

mination of the effect of material properties on fatigue crack propaga-

tion, care was taken to control as many other variables as possible. All

tests were subject to nearly identical environmental conditions of room

temperature (24 C) and desiccated air. Loads were controlled to within

0.27. of the desired load using the test machine's amplitude meaeurement

system. To prevent any effects from the order in which the specimens

were run, the specimens were randomized using a computar program which

utilized a random number generator. The tests were run in the random

order determined by this program. The order of tests is shown in Appen-

78 I

6--
 Crack initiation starting at the stress raiser was performed
starting at 6P - 15000 lbs. and shedding the load 107. no sooner than

every 0.5 ma (12.5 times the change in plastic zone radius due to the

load shed) to the desired test load level. Fatigue cycling was done

initially at 10 hz up to 5.4 mm (due to reduced frequency response of

the testing machine at high loads) and then at 20 hz. To make certain

that no load effects were present in the data, the test load level was

reached 1.0 mm before data acquisition (58 times the change in plastic

zone radius due to the last load shed). The load level was held constant

throughout the test (thus increasing AK with increasing crack length).

All tests were started at the same init'ai crack length (2a - 18.00 am).

The location of the centerline of the specimen was noted as a reference

to insure consistent crack length measurements throughout the test.

Cycling was continuous throughout the test to eliminate any time or

underload effects on subsequent fatigue crack growth.

The crack length and number of cycles were monitored continuously

for each test and discrete data points were taken as determined by the

Lest program. Data were actually taken by advancing the optical system

the printer push button when the

by the specified increment and pressing
crack tip had grown to the incremented position as determined by the

crosshair in the stereo microscope. The amount of error in the data ac-

"quisition process is given in Section 9.5.

9.5 Muuzemnt AcuracY

In an attempt to isolate the data variance due to the material prop-

erties, a measure of the experimental error was needed. This experimental

error results from the random error in measuring the cycle count and the

crack length.

79
 By using the test machine's amplitude measurement system which com-

pares a known input signal with the feedback signal (applied load), the

loads can be controlled to within 0.27%.

Error in the crack length measurement is due to two sources. If the

spatial relationship between the microscope crosshair and the scribed

reference line on the specimen is not constant, then an undetermined

amount of measurement error is present. This usually occurs when the

microscope is accidentally moved with respect to the specimen and can be

avoided by a careful experimental procedure.

The second source of crack length measurement error is the alignment

of the crack tip vith the microscope crosshair. This alignmnent process

consists of 1) defining the crack tip location, 2) defining the crosshair

location, and 3j comparison of the two locations to see if they are iden-

tical. If they are, then the printer button is pushed and a data point

is taken.

To determine how well the observer's eye performs this alignment pro-

coss, the following test was devised. A crack was initiated and the cy-

cling was stopped when the observer determined that the crack had reached

9.00 m. He then took 10 repeat measurements of the crack length, being

careful to always approach the crack tip from the same direction to pre-

vent any hysteresis effects. This series of 10 repeat measurements was

repeated at 9 other predetermined crack lengths. The mean and standard

deviation of each set of 10 repeat measurements was computed and the

error of the original data point was then calculated in terms of the

standard deviation. The results of the 10 sets of repeat measurements I

are as follows.

80

- .
 XE -0.001414 mm.
SE a 0.001390 um.

where
XE is the mean of the errors,

SE is the standard deviation of the errors.

Therefore, the average experimental error for each data point is 0.001414

mm. The average experimental error as a function of the crack length

measurement interval, Aa, is shown in Table V. It should be noted here

that the larger Aa is, the smaller the average experimental error is.

I
81

Il
 Table V

Average Experimental Error

&A INCREMENT (MM) _ __ __I

AVERAGE ERROR (PERCENT) ____
__

0.20 0.1

0.40 0.36

0.80 0.17
82

1I

ii
fI

!~
1*

i8

.. --- -....- _ . _ _ *1.

 SECTION X

DATA ANALYSIS AND RESULTS

As a result of the experimental investigation conducted as described

in Section 9, 68 replicate a vs. N data sets were obtained. These data

are shown in Figure 22. Using these data, an analysis was performed to

meet the objectives of the investigation (Section 3).

10.1 Distribution of N

The first objective to be met was to determine the distribution of NI

as a function of crack length. The replicate N data used was readily ob-

tained from the original replicate a vs. N data. Typical replicate cycle

count data are shown in Figure 23. The distribution of the replicate

cycle count data was determined at each crack length level through the

use of the CCDDP program (Section 6.2). At each crack length level, this

program calculated the distribution parameters and goodness of fit criteria

for the six distributions and then compared the goodness of fit criteria

between five of the distributions in order to establish the distribution

rankings. The generalized 4-parameter Samma distribution was not consid-

ered for the distribution rankings because it was expected to have an ex-

cellent fit to the cycle count data due to it's power parameter (Section
im
5.2.e). The distribution parameters, goodness of fit criteria, and the

distribution rankingse were then combined over all of the crack length

levels.

83

S...w.. . .i - -.
 REPLICPTE R VS. N DRT:
164 DRTR POINTS PER TEST
68 REPLICATE TESTS
DELTA P = 4.20 KIP
P MAX = 5.25 KIP
50.00- O = 9.00 MM
R .20 . .....

i ,s--X40.00-.'!i
z mm-.-: ..
,

z 30.00

CC

20.00-

10.00 1I LI
0.0000 .0650 .1300 .1950 .2600 .3250
N (CYCLES] )XIOX 1

Figure 22. Original Replicate a vs. N Data

o8-
 P VS. N
37.00- REPLICRTE CR TESTS.
DELTR P = 4.20 KIP
68 DATA POINTS
no= 9.000 MM
R =.20
36.8-

36.60 -

Z 34
CD
LJJ
-j
_.1_S36.20 - ~X

U:036.20" x NM( xx
--- x x x

36.00-

S~~~~36.80-j=- i
.2= S. .e26= .M
am ..
N (CYCLES) (X1O 6)

Figure 23. Typical Replicate Cycie Count Data

I
& S~85,
 The distribution parameters of the cycle count data as a function of

crack length were plotted for each of the six distributions and are shown
j
in Figures 24 through 29. The distribution parameters are normalized so

that their minimum and maximum values are equal to zero and one, re-

spectively. As a result of this normalization, these figures do not show

the actual values of the distribution parameters but are intended to re-

flect trends present in these parameters.

The goodness of fit criteria for each distribution were averaged over

all of the crack length levels. These results are shown in Table VI. For

these goodness of fit criteria, the best fit of the data to a distribution

occurs when the chi-square tail area is a maximum, the Kolmogorov-Smirnov

statistic is a minimum, and the closeness, R2 , is a maximum. Using these

relationships,

fit
an understanding of which distributions provide the best

for the cycle count data can be obtained.

I
The distribution rankings at each crack length level were combined

over all of the crack length levels. By convention, the lower tht value

of the distribution ranking, the better the fit of the data to the given

distribution. The mean rank and it's standard deviation for each of the

distributions and the number of times each distribution was selected as

the beot distribution were calculated during this combining process.

These results are shown in Table VII.

The 3-parameter log normal distribution provided the best fit frr the

cycle count data by a wide margin, as evidenced by the low distribution

ranking value, the low Kolmogorov-Smirnov test statistic value, and the

very large number of times it was selected as the best distribution. The

3-parameter gamma distribution provided the next best fit, while the 2-

parameter log normal distribution and the 3-parameter Weibull distribution

a 86
t

2-PRRAMETER
NORMRL OISTRIBUTION
NORMALIZED PARAMETER VALUES
DELTAP = 4.20 KIP X - MU HRT
P MRX = 6.25 KIP + - SIGMRI HRT
RO = 9.00 MM
NORTR = 68
R = .20
Sl~oo-

LU
Li

I-j

I-
W."

* j J1$O
C3 +

- R (CRRCK LENGTH IN MM)

LLA 400-

! Figure 24. 2-Parameter Normal Distribution Parameters

i; of Cycle Count Data as a Function of Crack
Length

871
 2-PRRRMETER
LOG NORMRL DISTRIBUTION
NORMRLIZED PRRAMETER VRLUES
DELTA P = 1.20 KIP X - MU HAT
P MAX = .25 KIP +- BETR HAT
A0 = 9.00 MM
NDATA = 68
Cl) 1..20

r.I
c,-

z-

I--
CM

0 4.

$-* x
1

x.00
-o.Oo 10.00 20.00 30.00 50.00
A (CRACK LENGTH IN MM)

Figure 25. 2-Parameter Log Normal Dietr-ibution Parameters

of Cycle Count Data as a Function of Crack LeAngth

88

low*
 3-PRRRMETER
LOG NORMAL DISTRIBUTION
NORMALIZED PARAMETER VALUES
DELTAP = 4.20 KIP -TRU HAT
PMAX = 5.25 KIP - MU HAT
RO= 9.00 MM + - BETS HAT
NDATA = 68
R + S1oo-
.20

cc
0- +
X:

z 4" X
X

C
0.3000.4004
+
+ +
LU 444.

.200-

0.00 10.00 20.00 30.00 00.00 60.00

A (CRACK LENGTH IN MM)

Figure 26. 3-Parameter Log Normal Distribution

Parameters of Cycle Count Data as a
I Function of Crack Length

IL 89
 3-PRRFIMETER
WEIBULL OISTRIBUTION
NORMALIZED
DELTIA P"= 4.20 PARAMETER
KIP VALUES
4b - TAU HAT
P MAX = 5.25 KIP X - B HAT
RO = 9.00 MM +- C HAT
NDATA = 68

x
AJJ
+
CLi

"c +
Cc

0I

- +

It

LIJ * --x +
III

0.00 10.00 20.00 30.00 0.00 60.00

A (CRACK LENGTH IN MM)
Figure 27. 3-Parameter Weibull Distribution Parameters
of Cycle Count Data as a Function of Crack
f Length

90
 3-PRRRMETER
GRMM DOISTRIBUTIiON
NORMALIZED PARAMETER VALUES
DELTA P = 4.20 KIP , - TRUHIT
P MAX = 5.25 KIP X - B HAT
AO = 9.QO MM + - G HAT
NDRTR = 68
u 1 .000 R= .20

w ~4+

x +
C3

400

0o000- I. I " "

0.=0 40.00 20.00
I 30.IX I).00 60.00
A (CRRCK LENGTH IN MM)
Figure 28. 3-Parameter Gamma Distribution Paraamters
of Cycle Count Data as a Function of Crack
Length

91

iRPM
 GENERALIZED 4-PARRMETER
GAMMA DISTRIBUTION
NORMALIZED PARAMETER VALUES
DELTA P = q.20 KIP 0 - TRU HAT
F MAX = 5.25 KIP - ALPHA HAT
RO = 9.00 MM X- B HAT
NDRTR = 68+ - 6 HAT
R= .20 +
LUJ

C-. +x

Ci)
co
4D +

xI

29. ( CRCK LENGTH IN PMM)

92
 -T

Table VI

Average Goodness of Fit Criteria for the

Distribution of Cycle Count Data
zz

CHI-SQURRE KOLMOGOROV- CLOSENESS

DISTRIBUTION TAIL RREA SMIRNOV TEST (R SQURRED)

2-PRRRMETER 0.8365 0.0995 0.93310

NORiMAL

S2-PARRMETER
2-PANRAMETR 0.8842 0.0857 0.95799
LOG NORMAL

3-PARRMETER
L-PARAM 0.8694 0.0699 0.98223
LOG NORMAL

3-PARAMETER
-PRAETE 0.8340 0.0882 0.93658
WEIBULL

3-PARAMETER 0.8602 0.0722 0.97160

GAMMA

GENERALIZED
4-PARAMETER 0.8075 0.0722
GAMMA

93
 Table VII

Distribution Rankings for the Distribution of Cycle Count Data

NUMBER OF
STANDARD TIMES BEST
DISTRIBUTION MEAN DEVIATION DISTRIBUTION

2-PARMETER 4.982 0.1348 0

NORMAL

2-PRRRMETER 3.147 0.6780 7

LOG NORMAL

3-PARAMETER
3-PARAM 1.221 0.5882 137
LOG NORMAL______I

3-PARAMETER 3.650 0.6338 3

WEIBULL

3-FARAMETER 2.000 0.4969 16 1

94
GAMA
 I-I

tied for the third best fit for the data. The 2-parameter normal distri-

bution finished last in the distribution ranking as it provided a very

poor fit for the data.

10.2 Crack Growth Rate Calculation Methods

The second objective to be met was to determine which crack growth

rate calculation method introduced the least amount of error into the

da/dN data. This was to be done by integrating the da/dN data calculated

by each crack growth rate calculation method back into a vs. N data and

then calculating the error between the new a vs. N data and the original

a vs. N data.

The DADNCP program (Section 7.2) was run on each of the 68 original

a vs. N data sets. This program calculates the da/dN vs. 6K data, into-

grates the da/dN data back into a vs. N data using Simpson's one-third

rule and the trapezoidal rule, and then calculates a step by step average

incremntal error, as outlined by Frank and Fisher [2), for each of the

six da/dN calculation methods. The da/dN calculation method which re-

sults in the lowest average incremental error is then selected as the

best da/dN calculation method for that data set. The logl0 da/dN vs.

logl 0 AK data are plotted for each of the da/dN calculation methods and

typical plots of these data are shown in Figures 30 through 35.

The average incremental error from each da/dN calculation method was

averaged over all of the data sets and the number of time each d4/d4

calculation method was selected as the beet method was computed. These

results are shown in Table VIII. The modified secant method had the lowest

average incremental error, followed closely by the secant method. The

modified secant method and the secant method were both selected as best

95
 4

DR/ON VS. DELTR K PLOT

SECNT METHO
e REPLICATE CP TEST 68
-DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
PM:X = 5.25 KIP
10- 136 DATA POINTS
R = .20
6

IjJ
0-. UK

z
3
cr. x4xk Xx

X~X
S~x

1I0

ERROR =Ve

7.0 10.0 13.0 18.0 20.0 5.0

DELTRK (KSI-SQURRE ROOT IN)

Figure 30. Typical Log 10 da/d14 vs. Log AK Data

Calculated ry the Secant )69od
f

OR/DN VS. DELTR K PLOT

"MlnIFIV WIIIT MEP

I-REPLICATE CA TEST 68
VELTAR = .20 MM CONISTRNT
=-- P z 4.20 KIP
10 PMAX = 5.25 KIP
,137 DRTA POINTS
ER ..20 R= 2 9

$ -

Zia

I 7.10 1. 30 60 2.

10 -
ERROR =2.49 Y

7.0 10.0 13.0 16.0 20.0 0S.0

-DELTRX IKSI-SQURRE ROOT IN)

Figuro 31. Typicl Logi0 da/dN vs. LoS1.0 AK Data

.. Calculated by tbo Mq.df Lod hcant Mothod

97
 DR/ON VS. DELTA K PLOT
LINEAR 7-OIlNT INCREMENTAL POLYNOMIRL METHOD
_REPLICATE CA TEST 68
DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
o PMRX = 5.25 KIP
131 DATA POINTS
R .20
6
3

LI=I

U-1
U

Cr.

0 1

r~ -

*7.0 DELTAK
10.0 (KS-SQA
13.0 ERROOT
16.0 IN
20.0 a.0
*DELTA K I(KSI-SQUARE ROOT IN)

Figture 32. Typical Lo810 da/dN vs. Lo050 AK Data

Calculated by the Linear 7-Point Incremntal
Polynomial Method

98
 DR/DN VS. DELTR X PLOT
QUADRATIC ?-POINT INCREMENTAL .YNOMIAL METHMO
REPLICATE CA TEST 68
DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
PMRX = 5.2S KIP
"131 DATA POlINTS
R = .20
6 -

(-j
=IQ~

S10 R' .998616

ERIROR 6.32X,

7.0 10.0 13.0 16.0 20.0 26.0

DELTA K (KSI-SQU:RE ROCIT IN)

Figure 33. Typical LoglO da/d4 vs. Loslo AX Data

Calculated by the Quadratic 7-Point
Incremental Polynomial Method

99
 DR/ON VS. DELTR K PLOT
LINERR LOS-LOG 7-POINT INCREHENTRL FOLYNOMIRL METHOD
REPLICATE CA TEST 68
DELTA A = .20 MM CONSTRNT 2

DELTA P = 4.20 KIP

APIAX = 5.25 KIP
131 DATA POINTS
R= .20

u.J
C-
=10 -

elQ:

2 -
x
x

R = .992196
ERROR = 8.78%

7.0 10.0 13.0 16.0 20.0 M5.0

DELTA K (KSI-SQURRE ROOT IN)
x

Figure 34. Typical Log 1 o da/lN vs. LoglO AK Data

Calculated by the Linear Log-Log 7-Point
Incremntal Polynomial Msthod

100
 DR/ON VS. DELTR K PLOT
OUARTIC LOG-LOG 7-POINT INCREMENTAL POLYNOMIAL METHOD
2- REPLICATE CA TEST 68
DELTA A = .20 MM CONSTANT
DELTA P = 4.20 KIP
PMAX = 5.26 KIP
131 DATA POINTS
R = .20

U S-)

S 1
x

R% .998314
13'.0ERROR z 6.30%.
16.0 20.0 5.0
DELTA K (KSI-SQUARE ROOT IN)

Figure 35. Typical Log0 daldN vs. Log 1 K Data

Calculated y the Quadratic eog-Log
7-Poiat Incrematal ?olynoilal method

101
 Table VIII

da/dX Calculation Method Results

DR/DN OVERALL AVERAGE NUMBER OF

CALCULATION INCREMENTAL ERROR TIMES BEST
METHOD (PER CENT) METHOD

SECANT METHOD 2.70 17

MODIFIED 2.58 51
SECANT METHOD

LINEAR
7-POINT INCREMENTAL 6.96 0
POLYNOMiAL METHOD

QUADRATIC
7-POINT INCREMENTAL 6.83 0
POLYNOMIAL METHOD

LINEAR LOG-LOG
7-POINT INCREMENTAL 9.41 0
POLYNOMIAL METHOD

QUADRATIC LOG-LOG
7-POINT INCREMENTAL 6.65 0
POLYNOMIAL METHOD

102 j
methods, with the modified secant method selected three times as often

as the secant method. From these results, it can be stated that the

modified secant method introduces the lowest amount of error into the

da/dN data of the six da/dN calculation methods selected.

10.3 Distribution of da/dN

The third objective to be met was to determire the distribution of

da/dN as a function of 6K. The first set of da/dN data selected for

analysis was da/dN data calculated by the secant method, with the anti-

cipation of also finding the distribution of da/dN data calculated by

the modified secant method and the quadratic 7-point incremental poly-

nomial method. Data were selected from the first two da/dN calculation

methods because of their ability to re-create the original a vs. N data

and the quadratic 7-point incremental polynomial method because of its

widespread use. The combined data from each of these three methods are

shown in Figures 36, 37, and 38.

The steps of analysis for the distribution of da/dN are very similar

to the steps of analysis used for the distribution of N. First, the re-

plicate da/dN data used was obtained from the da/dN vs. 6K data generated

by the DADNCP vrogram (Section 7.2) using the secant method. Typical re-

plicate da/dN data are shown in Figure 39. The distribution of the re-

plicate da/dN data was determined at each AK level through the use of the

CGRDDP program (Sectio 6.3). At each AK level, this program calculated

the distribution parameters and goodaess of fit crit -is for the six dis-

tributions and then compared the goodne i of fit crite. . between the dif-

ferent distributions to give the distribution rankings. Again, the gen-

eralized 4-parameter gamma distribution was not included in the

103
 REPLICPTE DR/DN VS. DELTR K DRTR
SECRNT METHOD

2
136 DATR POINTS PER TEST
68 REPLICATE TESTS
10 DELTA P = 4.20 KIP
P MAX = 5.25 KIP
AO =9.00 MM
R =.20

LU

-J 10

I z

5 "1

L. ,
a.),.:
7.0 10.0 13.0 16.0 20.0 25.C
DELTA K (KSI-SQURRE ROOT IN)

Figure 36. Combined Logl0 da/dN vs. Logl0 LK Data

Calculated by the Secant Method

10-
 rI
REPLICRTE DR/DN VS. DELTR K DRTR
MODIFIED SECANT METHOD
2
137 DATA POINTS PER TEST
68 REPLICATE TESTS
DELTA P = 4.20 KIP
P MAX = 5.25 KIP
AOD 9.00 MM
R: .20

LUJ
-_J 2

z 10_

I ,".,,.*.f
I.'":.-

I ' :" ""

0-05

Figure 37. CombndLg 0 ad j ogcKDt

iA "/.0 10.0 13.0

DELTA K (KSI-SQUARE ROOT IN)
16.0 20.0 25.0

i Fibre 37. Combined ILo$10 da/dN vs. Lo$10 AK Data

Calculated by the Modified Secant Method

i 105
 REPLICRTE DR/DN VS. DELTR K DATR
QUADRATIC 7-POINT INCREMENTAL POLYNOMIAL METHOD

2
131 DATA POINTS PER TEST
68 REPLICATE TESTS
10"- DELTA P = 4.20 KIP
P MAX = 5.2S KIP
AO =9.00 MM
5 -R =.20

LU
_J 2 -

z 0

ITI
Z -s

7.0 10.0 13 .0 16.0 20.0 25.0

DELTA K (KSI-SQURRE ROOT IN)

Figure 38. Combined Log 1 0 d~a/dY vs. Losio AK Data

Calculated by the Quadratic 7-Point

106
 R VS. OR/ON
REPLICATE CR TESTS.
SERCNT
DLAP METHOD
4.20 KIP :
DELTA A = .20 MM
68 DATA POTNTS
RO = 9.OOL MM
.O- R= .20

z
Cr.]

CD
Z25.20-

WX 10sm ~ -mmiOC XIMOCS

! 25.0oo

S2Q5. -OADN(O"R/CYCLEJwXio.4)

107
r

distribution rankings. The distribution parameters, goodness of fit

criteria, and the distribution rankings were then combined over all of

the K levels.

The distribution parameters of the da/dN data as a function of crack

length (essentially AK) were plotted for each of the six distributions

and are shown in Figures 40 through 45. The distribution parameters are

again normalized to show the trends present in the parameters.

The goodness of fit criteria for each distribution were averaged

over all of the AK levels. These results are shown in Table IX. From

these results, an understanding of which distributions provide the best

fit for the da/dN data can be obtained.

The distribution rankings at each 4K level were combined over all of

the 4K levels and again the mean rank and Its standard deviation for

each of the distributions and the number of times each distribution was

selected as the best distribution were calculated. These results are

shown in Table X.

Each of the distributions gives a fair but not outstanding performance

in providing a fit for the da/dN data. There were no significant differ-

ences between the means of any of the five distributions, espeqially con-

sidering the high values of standard deviation about the mean. The 3-

parameter Sama distribution did have a slightly lower mean than the

other distributions and it also had the lowest value of the Kolmogorov-

Sairnov statistic. However, the 2-parameter log cormal distribution was

the beat distribution slightly more often than the other four distribu-

tions, but again there were no significant differences between the five

distributions. These results lead to the conclusion that the 3-parameter

gaSs= distribution provides a better fit for the da/dN data than the other

108
!

2-PQRRMETER
NORMRL DI STRIBUTION
i.eoo - NORMAL IZED PA.RAERN.VLUES
II.TA P:.26.
5= 4.e KIP
KIP X - MU HA:T
kOTA A= M M88 + - SIGMA HIT
.ooo- i x, ,eo

Cr..

.800-
a-
o_

.I~o + x,+4
.+ .

4
+ .9
Z: .400 4

0.000-+

0.00 10.00 20.00 30 40.00 60.00

A (CRACK LENGTH IN MM)
Figure 40. 2-Parameter Normal Distribution Parameters
of da/dN Data as a Function of Crack LAngth

109
 2-PRRRMETER
LOG NORMAL DISTRIBUTION
NORMALIZED PARAMETER VRLUES
.8 K.IIP .. X -MU HAT
PAR R =.20K M +-BETA HRT
1O =g9.M MN
S1.000- NORTA
= .= 68 +l .
R: .20k+

Li

a-
to +

V-4.

+
.. . . . .+
.800O0 + S.0 + 600
4 +
Li ~I0+ +*

At+++ (41+ +4+4.

Z .200- - + 4.4. +

0::0 +*4t.4.4.

0.0.00
0.0 10.00 20.00 ~.0 '00 00
A (CRACK LENGTH IN MM)".
Figure 41. 2-Parameter loog Normal Distribution Parameters
of da/dN Data as a Function of Crack Length

110
 3-P;PRRRMETER
LOG NORMRL DISTRIBUTION
NORMALIZED PARRMETER VALUES
SCA~NT THEM
P
DELTAP0 1MKIP
r MAX .ez5.K0- TAU HAT
%LIAA .20MH
.9.00 MM _ -MU
X HAT
NORTA 68HAT
c1.o0 w R= X20 +

a- . +

++ 4

i~ r ew + **+
+. ++
0. 07 r e+. *+
i. r + +
M+ oL+ + +
-+
+i +
4.

0.600 ' oo+4 CRKLESHI

+ 4W11 +

.200-

3
+ N
+o~ M
+C~&LNT~
da/d
Lo4lra
3-Paramete
+f D+aa uci

111 +
+
itriuinPaaitr
fCakLnt
 3-PRRRMETER
'- WEIBULL DISTRIBUTION
NORMALIZED PARAMETER VALUES
DElTA P "(L
6K~r,,,;''-
= 4.eO TRU
T HAT t
L .A
20 "" X HAT :
11= 9 .06 im,
A
NOWTA = B +- C HRT

LU +x
x

cc~q + +A.* AA
z +. +

(:3+ + , x

+ +. X46

++ ,}* 42+ + ++

o.000- fZI+NC,2 , ,I.0

F~sur43. 3-Parameter Usibull Ditsribution Parametrsr

of da/dl Data asm a Function of Crack L~ngth

i 112
 - - 3-PARRMETER
GRMMR DISTRIBUTION
NORMALIZED PARPHETER
IZ&TA: F = 4.eO KIP
VALUES
- 4b -' HT .

R0T = H -M
AI M X - 8 HRT
OM.TR = 68
,.00= .20
R + T

LU

cc .800 + + .4"

.C.
S++ 4
t- 4
++ +

CO +
4+ + *XX
v,

0 +4 + + *+

l .400 ++

z~~
- .: + x x .

CC (CRKLxHI M
Fiu+ 44 -aaee sn Dsrbto aatr
odad
+ D+t asaVnt XnofCakln

113 W
--.-.-.--- ~ - + W---

... 0... 0..10. 0 ...

-00 ...... 0.00 50.0_
 GENERRLIZED 4-PRRRMETER
GRMMF DISTRIBUTION
1.L'0- NORMALIZED PARR"METER VALUES
P'ELTA,
MA = P 6.26
= .KIP
KIP'KIP t - TAU HA:T
SIMTA A= MZ MM &, -ALPHA HFIT
SRAG = g.00 MMq
NMAITI
0TR.
= .20= 138 + X --B G HAT
HAT

& a

c .800- + +

0-
e.. ,= ..
+ + *+ + 4

+x v

+,+ + x
""- + xt, .x

+ X +114

Fig~ure 45. GewraliRtd 4-Param~t~e Games D)Istr:ibuton

Parameters of da/dN Data as a FrunctiLon of
Crack L4x th

I-I
 IJ

Table IX

Average Goodness of Fit Criteria for the

Distribution of da/dl Data

CGI-SQUFAE KOLMOGOROV- CLOSENESS

DISTRIBUTION TAIL AREA SMIRNI1V TEST (R SQUARED)
2-PARAMETER 0.8494 0.0915 0.94997
NORMAL

2-PARAMETER
2-PNRAM 0.9011 0.0779 0.97647
LOG NORMAL

3-PARAMETER
3-PNRAM 0.8442 0.0834 0.96966
* LOG NORMAL

3-FRRAMETER
*REIBU R 0.8474 0.0777 0.95942
WEIBULL

3-PARAMETER 0.8389 0.0737 0.98662

GAMMA

GENERALIZED
4-PARAMETER 0.7946 0.0726
GAMMA

115i
 Table X

Distribution Rankings for the Distribution of da/dN Data

NUMBER OF
DISTRIBUTION MERN STANDARD
DEVIATION TIMES BEST
DISTRIBUTION

2-PARAMETER 3.684 1.6497 27

NORMAL

2-PRRPMETER 2.603 1.1943 37

LOG NORMAL

3-PARAMETER 3.360 1.5524 26

LOG NORMAL

3-PARRMETER 2.985 1.1925 19

WEIBULL

A-PRIRMETER 2.368 0.9646 27

GAMMAR

116

___ _ --
___ IIIIII
I four distributions, but its performance relative to the other d7stribu-

tions is not strong at all. Due to this poor performance by the da/dN

data in fitting a distributinn, no analysis of da/dN data calculated by

either the modified secant method or the quadratic 7-point incremental

polynomial method was conducted.

10.4 Prediction of a vs. N Data from the Distribution of da/dN

The fourth objective to be met was to determine the variance of a set

of a vs. N data predicted from the da/dN distribution parameters. The

da/dN distribution parameters were estimated by the CGRDDP progiam (Sec-

tion 6.3) as described in Section 10.3. The AVNPRD program (Section 7.3)

was run on the da/dN distribution parameters and 68 replicate a vs. N

data sets were predicted. These data sets are shown in Figure 46.

To obtain the variance of this predicted data, the CCDDP program (Sec-

tion 6.2) was run at 14 crack length levels of the predicted data. The

distribution parameters, goodness of fit criteria, and the distribution

rankings were then combined over all of the crack length levels run.

For this predicted data, neither the 3-parameter gamma distribution

nor the generalized 4-parameter gamma distribution would converge on

parameter estimates, implying that neither distribution would provide a

fit for the data. The distribution parameters as a function of crack

length obtained for the other four distributions are shown in Figures 47

through 50. The average goodness of fit criteria for the four distribu-

tions for the predicted data are shown in Table XI. The distrioution

rankings results for the four distributions for this data are shown in

Table XII.

117

SI " ' 1 I I
 PREDICTED
REPLICRTE R VS. N DATA
60.00- SECANT METHOD
137 DATA POINTS PER TEST
68 REPLICATE TESTS
DELTA P = 4.20 KIP

50.00-
OELTR A = .20 MM
MMKIP
PAO MAX 9=.005.25
SR =.20
.1
1" 40.00-

z 30.00
LUJ

20.00-
--

10.00-

0.00- 1
0.0000 .0650 .1300 .1950 .2600 .3250
N (CYCLES) (X1O 6)

Figure 46. Replicate a vs. N Data Predicted

from the Distribution of da/dN

1 1
118J

I'--_i4
k

2-PRRRMETER
NORMRL DISTRIBUTION
NORMALIZED PARAMETER VALUES
DELTA P = 4.20 KIP X - MU HAT
P MAX = 5.25 KIP + - 6IGMHRT
RO = 9.00 MM
NDATA = 68
1.000- R= .20
*,- ,+ x

cr. .800--+ . + X
a-. +- -

I--
= x i
EDx
3 x

Z .200 x

+ XI

0.00 10.00 20.00 30 .00 4O.00 50.00

A (CRPCK LENGTH IN MM)

Figure 47. 2-Parameter Normal Distribution Parameters

as a Function of Crack LenSth for Cycle
Count Data Predicted from the Distribution
of da/dN

i 119
 Ii

2-PRRRMETER
LOG NORMAL DISTRIBUTION
NORMALIZED PARRMETER VALUES
DELTA P = 4.20 KIP X - MU HAT
P MAX = 5.26 KIP + - BETA HAT
PO = 9.00
NOATA = 68 MM
to 1.000- R :.,,20 xx
xx
LU
LU x
x
az x
0- x
xx
xi
+ x
LI

-J 00

+i

La

0.0001 113
0.00 10.00 20.00 3O.00 40.00 60.00
R (CRACK LENGTH IN MM)
N!
4.d
I Figure 48. 2-Parameter Log Normal Distribution Parameters
Io : 8as a Function of Crack Lengt~h for Cycle Count
S~~Dat~a Preditel~d from t~he DistribuItion of da/adY

120
 3-PARAMETER
LOG NORMRL DISTRIBUTION
NORMALIZED PARRtMETER VALUES
DELTA P = 4.20 KIP 0 TRU 'HAT
MRX = 5.26 KIP X- MU HT
O = 9.00 MM + BETR HAT
NOATR = 68
R q20x ~x --
xxx

1.10 x
x
xx
x

x
tt

C3
Aw -

Ro-

111 ~ 2
ENT~ 4 50.00tloe
0.00.

10. (CROLEN(SO Mm)"I

Yliure 49. 3-1aresiater USI Nolou Distribution Paraieters

am a Funtio, .,,f Crack Lansth for Cycle Count:
Data Preddicted from t he Diteribution of da/dol

I i21
 3-PRRRMETER
WEIBULL D1ISTR IBUTION
NORMALIZED PRRRMETER VRLUES
DELTA P = 4.20 KIP ,b - TAU HAT
P MAX = 5.25 KIP X - B HRT
AO = 9.00 MM + - C HAT
NDATA = 68
R= .20
x
t"- +
+4.
+
L X X
+. x
x x
X+ x x
. ++
Zx

So x

(1) +

u.400- X
1%4.
.j

Z.200-
-= A ~ EA i
0.000

0.00 l'"" 0
"I (CRRCLENG -N. .MM)".

Figure 50. 3-Parameter Weibull Distribution Parmeteru

to a function of Crack Length for Cycle Count
Data Predicted from the Distribution of da/dW
 Table X1

Average Goodness of Fit Criteria for the Distribution of Cycle

Count Data Predicted from the Diustribution of da/dN

CHI-SQUARE KOLfOSOROV- CLOSENESS

DISTRIBUTION TRIL RAER SMIRNOV TEST (R SQUARED)_

2-FARRMETER 0.9087 0.073S 0.98515

NORMAL

2-FIRRAETER 0.9128 0.0722 0.98497

LOG NORMRL

3L-PRRRMETER
LOG NORMRL 0.8828 0.0730 0.98515

3-PRRRMETER 0.8919 0.0818 0.96884

NEIBULL

I
-i

123
 Table XII

Distribution Rankings for the Distribution of Cycle Count

Data Predicted from the Distribution of da/di

NUMBER OF
STANDARD TIMES BEST
I DISTRIBUTION MEAN DEVIATION DISTRIBUTIN

2-PARAMETER 2.643 0.7449 2

NORMAL

2-PARAMETER
S~LOG LOGPNORMALR
NORMAL 1.214 0.5789 12

3-PARAMETER
SLOG NORML.286 0.6112 0

3-PARAMETER
3.857 0.5345 0
WEIBULL

12L
~!

i 124
 The 2-parameter log normal distribution provided the best fit for the

predicted replicate cycle count data, followed by the 3-parameter log

normal distribution and then the 2-parameter normal distribution. The 3-

parameter Weibull distribution provided the worst fit for the data of

the four distributions which the data fit.

The next step in the analysis was the comparison of the distributions

of N between the actual cycle count data and the cycle count data pre-

dicted from the distribution of da/dN. The mean and standard deviation

of both distributions at the crack length levels used above were computed
and the resultc are shown in Table X111. At every crack length level,

there was no significant difference between the means but there waL a

very significant difference between the standard deviations of the two

distributions. In every case, the standard deviation of the predicted

cycle count data is much smaller than the standard deviation of the actual

cycle count data.

As a check on the analysis above, a vs. N data were predicted from

the distribution of da/dN in a slightly different manner than for the

predicted replicate cycle count data. The mean and + 1, 2, and 3 sigma

values of da/dN at each crack length level were obtained from the distri-

bution of da/dN. Using these 7 lines of da/dN data, a vs. N data was

predicted. The results are shown in Figure 51. A comparison between the

actual cycle count mean and + 1, 2, and 3 sigma values and the cycle

count values predicted from the mean and + 1, 2, and 3 sigma da/dN lines

at a single crack length level is shown in Table XIV.

From the above analysis, it can be concluded that predicting a vs. N

data from the distribution of da/dN using the method described in Section

7.2 yields low error in predicting mean crack propagation behavior, but

125
 Table XIII

Comparison of the Distributions Between Actual Cycle Count Data and

Cycle Count Data Predicted from the Distribution of da/dN

MEAN STMIf DEUIATIOI

11.000 mIni 55735 Gm 151

13.000 SOUR 91Sm3 5-33 2512
15.000 117466 118700 6?1S 4348
07.000 130M 139571 10903 4M9
19.000 156352 ISS15 11849 3208
21.000 170796 1713"9 12489 3249
83.000 188138 183670 8304 3284
2 000 man 194504 8547 339
7.000 20233 204083 a"4 3432
2,.000 211030 a12612? 9123 3407

31.000 218 MM0333 S3 3416

33.000 84M 237186 3637 3S30
35.000 231416 233249 1007 35n
36.200 334573 336533 10191 3574

126
 PREDICTED A VS. N DATA
80.00- SECANT METHOD
137 DATA PMINTS PER TEST
DELTA P = 4.20 KIP
DELTA A = .20 MM
P MAX = 6.25 KIP
0.00- RO 9.00 MH
R= .20

40'~.00o -3a -2a -la M +10 + +3o

p-

CD
= 30.00 I

10.00

0.0000 low A AM

N (CYCLES) (XIO )3

Figure 51. a vs. W Data Predicted from the Keau

and 1, 2, and 3 Sitga da/d2 Lines

127
*{
 Table XIV

Comparison of Actual Cycle Count Data vith Cycle Count Data Predicted
1 from Constant Variance 4./dN Lines .- ,t.

A a 31.000 151 AMAL P1c1

-3 SIGMA 18733 117430

-2 S.IGM'A 1,58
41941

-1 SUMR 2455N IF3192

MEAI 21701 813316

+1 SIGMA 23382 16468

4e33 SIMA20341

+3 SlIGMA 20f 420

awon

1,28
 II

yields high error in predicting crack propagation behavior at the extremes

of the distribution of N.
-I
10.5 Inverse Growth Rate

Due to the failure of the da/d14 data to fit any of the given distri-

butions satisfactorily, it was decided that the growth rate variable war-

ranted a further investigation. Looking back at the original experilmnt-

al investigation (Section 9), it can be seen that the dependent variable

of the data was N while the independent variable was a (i.e. N was mea-

sured as a was varied). Since the dependent variable, N, provided a very

nice fit to the 3-paramster log normal distribution, it was strongly sus-

pected that changes in the dependent variable, AN, would also provide a

good fit to one of the given distributions. Since A a was constant it

wau decided to use AN/Ia, or in differential terms, dN/da, as a variable

of interest for further analysis.

The analysis conducted using dN/da as the variable of interest was

the same analysis used for da/dN. The first part of this analysis was to

determine the distribution of dN/da. Replicate dN/da data were obtained

by inverting the replicate da/dN data calculated by the secant method

using the DADNCP program (Section 7.2). Typical replicate dN/da data are

shown in Figure 52. The distribution of the replicate dN/da data was

determined at each AK level through the use of the DNDDP program (Section

6.4). At each AK level, this program calculated the distribution para-

meters and goodness of fit criteria for six distributions and than com-

pared the goodness of fit criteria bptwean the different distributions

to give the distribution rankings. The location parameter for the 3-

parameter gamma distribution and the generalized 4-parameter gamma

129
 R VS. DELTR N/DELTR R
21.80- RELICRTE CR TESTS.
SECFINT METHOD
OELTA P = 4.20 KIP
DELTA A = .20 MM
68 DRTA P3INTS
no = 9.000 MM
1.60- R .20

A.X

1-

:- 21.0-
_J

x 3)OM xx x

,..0 21 .00-

CX I

20.80
.01000 100 .2000 .p000

DELTIR N/DELTA R (CYCLES/IN) (fb)

Figure 52. Typical Rhplicate dOdN Deca

130
distribution was assumed to be zero by this program, thus reducing these

two distributions to the 2-parameter gamma distribution and the general-

ized 3-parameter gamma distribution, respectively. As before, the gen-

eralized 3-parameter gamma distribution was not included in the distribu-

tion rankings. The distribution parameters, goodness of fit criteria,

and the distribution rankings were then combined over all of the AK levels.

The distribution parameters of the dN/da data as a function of crack

length were plotted for each of the six distributions and are shown in

Figures 53 through 58. The distributiods parameters are again normalized

to show the trends present in the parameters.

The goodness of fit criteria for each distribution were averaged over

all of the 4K levels. These results are shown in Table XV. From these

results, an understanding of which distributions provide the best fit for

the dN/da data can be obtained.

The distribution rankings at each AK level were combined over all of

the AK levels and the mean rank and its standard deviation for each of

the distributions and the number of times each distribution was selected
|I
as the best distribution were calculated. These results are shown in

Table XVI.

The 3-parameter log normal distribution provided the best fit for the

dN/da data, as evidenced by the low distribution ranking, the low

Kolmogorov-Smirnov test statistic value, and the large number of times

it was selected as the best distribution. The 2-parameter log normal and

the 3-parameter Weibull distribution tied for the second best fit for the

dN/da data, both having roughly the same distribution ranking and

Kolmogorov-Smirnov test statistic value and the same number of tims it was

131
 2-PRRPIIETER
NORIML DISTRIB''TION
1.2,oo- NORMALIZEDSECANT
PRRRMETER
METHOD VRLUES
DELTA F = 4.20 KIP
P IA = 5.26 KIt
DELTA A = .20 MM
X - MU HRT
+ - SiGMA HAT
AO =9.M0 MM
tNATA = 68

iU
LUJ
i O 1.800- x
=.=
x
a:
I- x

Q.600 *X

cz .200
4 -

-'J 400

Z .200 -,

0.000 '
0.00 10.00 20.00 30.00 40.00 60.00
A (CRACK LENGTH IN MM)

Figure 53. 2-Parameter Normal Distribution Parameters

of dN/da Data as a Function of Crack Length

13i
I
132
I
t

2-PRRRMETER
LOG NORMRL DISTRI['JTION
NORMRLIZED PARRQ TER VRLUES
.ELTA
P = 4.eO KIP
F MAX a 5.26 KIP X - MU HAT
0.T"A = .20MM + - BETA HAT
AO = 9.00 "N
u'~~
.0)- NIATA .20
f=x = 68 +
I .m .-
x

"c- .8W 41

to' *A + "+
+" +
400 4+
%+
+

cc- + + +

O:::+
+
+
+N + +
++ +
*
+ +
+,,+
- -9
, o0.
% + * 4.

4- .
+*

+
0.000
0.00 10.00 20.00 30.00 '90.00 50.00
R (CRACK LENGTH IN MM)

Figure 54. 2-Parameter Log Normal Distribution Parameters

of dN/da Data as a Function of Crack Length

133

0 IR W
 I

3-PRRRMETER
,.. NORMAL DISTRIBUTION
1' ,Q iRALIZEDsmwRI
PARAMETER
ME'T VALUES
MFLTA: P = 4.20 KIP
.20 qi
OELT - TAU HAT
OELT A = M11
FO = 9.00 .20 M X - MU HAT
uW1T. = 6 + - BETA HAT
x
LaJ
W + x
_ 800" w 0 +

+
z +
+
0- ,"+

.6W + + +
+
I.-.

o3 # + +

fO0- + + 4+ +*
+ +
++ .+
C~r. x(+ ,+
" -,
~++. 4,
+ x. ,..-+ 4
x +++

o.00
,& .,
+ + IV ") " I . . I i"
4 4A iF
0.00 10.00 20.00 30.00 '40.00 60.00
A (CRACK LENGTH IN MM)

Figure 55. 3-Parameter Log Normal Distribution Parameters

of dN/da Data as a Function of Crack Length

134
 3-PRAPMETER
F NWEIBULL DISTRIBUTION
NORMALIZEDPR.:ETEER VALUES
i" ',DeTA P = 4.20 1K1
OrT P. Kpi:- TAU HAT
:.1 .0 " X B HAT
"A= W + "C HAT
u i.ooo- R .- .4p8+ *

xxx +
cci
.81)- *X+
.x x x

; Kx ++
+++
+
x.x_* *+ _ +~+4 +. 4. + 4+

X +& +
-3 '" + +

.200 x + +

-- 0.00) 10.00 20.00 30.00) . 40.00 50.00)

A (CRACK LENGTH IN MM )

-j+*
F.igure 56. 3-Paramet~er Welbull Disetribut:ion Paramet~ers
of dEi/da Date as a lvnction of Crick Length

I135
 ' 1~
2-PRRRMETER
GRMMR DISTRIBUTION
NORMALIZED PRS ER VALUES
D&,TA P = 4.20 KW " T-
I TA
W Kr )6.2--8 HAT
I=AFl = 68 ~CATHAT.
000 - R =x20
S.
+ +.

LUJ

Sao- +

z X
+ +;.x
S.00

+444.+
M- _.~ + X+w
~ .0-,"x + + + "=
x --
+xXx xx + +4 + +
' + + +-,+ + ,,
.qO0~ '+
+4 +'
+Ax +. i + + .+
x.,, I x +" + _++*
-+
.+ ++

S~~++!I $

+4.-+

.0 0
01.00 1O,.00 20.00
(CRACK LENGTH3Q.00q
1N MM) 000

Figure 57. 2-Parameter Game Distributiont Parameters

of dN/da Data as a Yunction of Crack Length

136
 GENERRLIZED 3-PRRRMETER
GAMMR DISTRIBUTION
1.o- NORMALIZED PARAMETER VALUES
:' " "46 "- B H RTT
NWnA z 68 + - G HAT.

++

S++ x
+ x
CL 0 % + + "":'-

+ +
-.I + + *+ 4A
++

+ +
+4 + +,, ,
-A ++ x ++
+ + 4 ++
+

"+
+
*jP +
xx;:
+ '
+ .+
:::
+"

+
ma+-+
__~~~~~~ . - "m- ,i ~| +x ~ IN+.;Er
~ II1 t dilH OI! i ilIlle I,,J II

137 1
 Table XV

Average Goodueis of lit Criteria for the Di*tribution of 4K/da Data

CHI-SQURRE KOLfOGOROV- CLOSENESS

DISTRIBUTION TAIL AREq StIIPNOV TEST (R $0URRED)

2-PRRAMETER
NORMAL 0.8383 0.0992 0.94912

2-PARRMETER 0.7
LOG NORMAL 0.9011 0.0779 0.97647

3-PRAMETER 0.8877 0.0695 0.97622.I

LOG NORMAL

3-PARAMETER
0.8409 0.0790 0.95477
WEIBULL

2-PRGAMETER 0.7640 0.0813 0.93431

GAMMA

GENERALIZED
3-PARAMETER 0.7507 0.0800
GAMMA

1I
138

= =u i
 Table XVI.

Dlatributton Rankings for the Diatrtbution of dN/da Data

-NUMBER F
STARNRO . TIMES BM.I
DISTRIBUTION MEAN DEVIATION DIVTRIBUTIdN

2-PRA4METER L.338 1.1816 10

NORMAL

2-PARAMETER
LO
OML 2.610 1.2363 28
LOG NORMAL

3-PRARMETER 1.860
LOG NORMRL 0.8621 6

3-PARAMETER 2882 1.2594 27

WEIBULL

2-PARAMETER 3.309 1.2019 15

GRMMA

139
 I

selected as the best distribution. The 2-parameter Same distribution

provided the fourth beat fit and the 2-parameter normal distribution pro-

vided the worst fit for the dN/da data.

The next step of the analysis was to see if the improved fit of the

.4Na
XdaM to a distribution would improve the prediction of a vs. N data

from the distribution of dN/da. The AVNPIRD program (Section 7.3) was

slightly modified for the dN/da variable and run on the dN/da distribution

parameters, resulting in 68 predicted replicate a vs. N data sets. These

data sets are shown in Figure 59.

The CCDDP program (Section 6.2) was run at a few crack length levels

-of the predicted data. The distribution parameters, goodness of fit

criteria, and the distribution ranki,,gs were then combined over all of the

crack length levels run.

For this set of predicted data, the generalized 4-parameter Samae

distribution would not converge on parameter estimates, implying that it

could not provide a fit for the dN/da data. The distribution parameters

as a function of crack length obtained for the other five distributions

-are shown in Figures 60 through 64. The average goodness of fit criteria

for the fiv, distributions for the predicted data are shown in Table XVII.

The distribution rankings results for the five distributions fr ,this data

are shown in Table XVIII.

The 3-parameter log normal distribution provided the best fit for the

predicted replicate cycle-count data, followed in order by the 2-parameter

log normal distribution, the 3-paraueter Weibull distribution, the 2-para-

meter normal distribution, and the 3-parameter game distribution.

The next step in the analysis waR the comparison of the distributionq

of N between the actual cycle count data and the cycle count data predicted

140

I. .U-- _l-- ie-

 PREDICTED
REPLICRTE R VS. -N .DRTA
60.00- SECANT METHOD
137 DATA POINTS PER TEST ..
68 REPLICRTE TESTS ' ..-" ..
DELTA P = 4.20 KIP
DELTA A = .20 MM
50.00- P MAX = 5.25 KIP
RD =9.00 MM
R .20

x 40.00 '.

Cm
LU30
z . 001

S20.00-

10.00-

0.00- . .
0.U000 .0650 .1300 -IWO .26W .3260
N (CYCLES) (X1O 8)

Figure 59. Replicate a vs. N Data Predicted

from the Distribution of dN/da
 2-PRRRMETER
NORMAL DISTRIBUTION
-NORMALIZED PARRMETER VALUES
'~~~ DELTA
MA F= = .8 ,i40KI- .,. mU.RT.HAT
x+--- .MA ., :.,,
A: = 9.00 MM
NORTR = 68
R .20
SI-

So

zJ .40x0

So2wX - x

0.00 10.00 "f.0

000
A (CFRLEG Mm)

Figure 60. 2-Parmset. 50=m1 ftetribution ParaMterg

as a Function of Crack Length for Cycle Count
Data predioted from the Distributiou of dN/da

142
 -

2-PRRRMETER
LOG NORMRL DISTRIBUTION ..
KIP
P = 4.20 PRRRMETER-
NORMALIZED
S-D-ELTAl VALUES HPT
":X-MU

PRO MAX = 5.25

= 9.00 mmH KIP + - BETA HAT
SNDSTn = 68M

1.000- .20 xx
L& xx

xx

L x

x
E5

Boo
+ x.-
.n

C3

_J+ x4
+

+ +

D+
+

14,
ligre 1.2-Parameter LOS *o001 Digtrtbutioo parm~ear
as a Function of Crack Unksth for Cycle Count
Data ft.4Lted from tbe Distribution of UW/da
 ~~7i

3-PRRRMETER
LOG NORMRL DIlSTRIBUTION
I
NORMALIZED PARAMETER VALUES
DELTRA = 4.20 KIP 0 -- TRU HAT
P MAX = S.25 KIP X - MU HRT
RO = 9.00 mm + - BETA HAT
NDRTA = 68
R
.- 4.20 x

I-
w
U) * x

C x

Z .400
S~X +

++ + +x + '
to.
+

0.00 "0' (CRACK LENGTH'"N MM)

Figure 62. 3-Patrmetr Log Notmal Distribution Parameters

---a* a ruuei. of Crask Lawh for Cycle rout
.. t-ft'dticts
. from the-latr.,utola of 41/da

"L144
 3-PRRAMETER
WEIBULL DISTRIBUTION
NORMRLIZED PARAMETER VRLUES
DELTA P = 4.20 KIP - TRU HAT 4
P MAX = S.25 KIP X - B HAT
RO = 9.00 MM + - C HAT
NDATA = 68
R.:00-
R .20 #-
LU
LU +.
x
+3+
I.
+
i: 4 X x

x +

0: x 0 +I

C3'4
-9
I-,,- x e

0.2000
4

.. I1 I - I + II ..

0. D0.00
000O (CRRCK LENGTH0.00
IN WA] '40.00 60.00

7?U1re 63. 3-Patammter Veibull Distributiou Parameters

ma a lwanctU of Crack Length for Cycle Count
Deta lNedicted from the Distribution of dN/da

j 143
 3-PHRFMETER
GRMMA DISTRIBUTION
1.2oo- NORMRLIZEO FRRRMETER VRLUES
DELTA P = 4.20 KIPF - TRU HAT
P MAX = 5.25 KIP X - B HAT
A' = 9.00 MM + - G HPT
NORTR = 68
S1.o000 R .20 x + +.+... .
L,

X:

ci.800-
Q~. x
+ 4b

00 0
C3 6000

Nj
-.
J +
ci"

Z .200
+
+
x X XX
x x x
0.0004 x -- T
0.00 10.00 20.00 30.00 40.00 60.00
A (CRACK LENGTH IN MM)

Figure 64. 3-Parameter Gamma Distribution Parameters

as a Function of Crack Length for Cycle Count
Data Predicted from the Distribution of dN/da

146

AL!=| II II I I I I I I I l| i
 Table XVII
Average Goodness of Fit Criteria for the Distribution of Cycle
Count Data Predicted from the Distribution of dN/da

CHI-SQUARE KOLMOGOROV- CLOSENESS

DISTRIBUTION TAIL AREA SMIRNOV TEST (R SQUARED)

2-PRRMETER 0.9816 0.0652 0.98765

NORMAL

2-PARAMETER 0.9640 0.06114 0.98990

LOG NORMAL

3-PARAMETER
3-PNRAMALE 0.9319 0.0567 0.99169
LOG NORMAL

3-PARAMETER 0.9135 0.0617 0.98100

WEIBULL

[
3-PARAMETER
GAMMA
GAMMA 0.20-40 C.Y44t93 0.85993

147

II
 Table XVIII

Distribution Rankings for the Distribution of Cycle Count Data

Predicted from the Distribution of dN/da

NUMBER OF
STRNORRO TIMES BEST
DISTRIBUTION MERN DEVIRTION DISTRIBUTI1ON

2-PRRAMETER 3714 1.0690

NORMAL

2-PARAMETER
OML
LO 2.S71 0.9376 2

3-PARAMETER 1.571 0.7559 8

LOG NORMAL

3-PARAMETER 2.857 1.0995

WFIBULL

3-PARAMETER
G1MMA
AMA2.857
2.52
1.i(.J j 2

148
 from the distribution of dN/da. The mean and atandard deviation of both

distributions at the crack length levels used above were computed and the

results are shown in Table XIX. At every crack length level., there was

no significant difference between the means but there was a very signifi-

cant difference between the standard deviations of the two distributions.

In every case, the standard deviation of the predicted cycle count data
wamuch smaller that, the standard deviation of the actual cycle count

data.

Just as for the data predicted from the distribution of da/dN, a vs.

N data were predicted from the mean and + 1, 2, and 3 sigma dN/da lines.

The results are shown in Figure 65. A comparison between the actual cycle
count mean and + 1, 2, and 3 sigma values and the cycie count values pre-

dicted from the mean and + 1, 2, and 3 sigma dN/da lines at a single crack

length level is shown in Table XX.

From the above analysis, it can be concluded again that predicting a

vs. N data from the distribution of dN/da using the method describad in

L Section 7.3 yields low error in predicting mean crack propagation behavior,
but yields high error in predicting crack ropagation behavior at the ex-

tremes of the distribution of N.

14
C
FI
SZ49
 Table XIX

Comparison of the Distributions Between Actual Cycle Count Data and

Cycle Count Data Predicted from the Distribution of dN/da

MEANSTAtIDFO DEUIATIMf
CRACK LEM C
(M) AC1L. PREDICTED ACTUAL PREDICTED

11.000 a1 5SSW7 65S6 2000

13.000 SOU-S 91067 5832 2560
15.000 117486 118326 6719 4361'
17.000 139342 138578 10903 4977
19.000 156382 1546402 11849 4941
21.000 17076 170694 12483 3288
23.000 182198 182852 8204 3323
25.000 192978 193730 8547 5219
27.000 2025M3 203416 8944 3506
2..000 211030 211809 9123 3465
31.000 218688 219505 9325 3509
33.000 225459 226408 9637 3633
35.00C 231416 232,52 10037 3667
36.200 234573 235717 10191 3641

I
i

PREDICTED R VS. N DRTR

60.00- SECANT METHOD
137 DATA POINTS PER TEST
DELTA F = 4.20 KIP
DELTA A = .20 MM
P MAX = 5.25 KIP
SO.O0- AO 9.00 MM
R = .20

x 40.00- -3o -20 -l +la +2a +3a

I--=
CD
Z 30.00-
-J

CC, 20.00-

0.00

0.0000 .1000 .2000 .3000 .4000 .5000

N (CYCLES) (XMO 6)

Figure 65. a vs. N Data Predicted from the HeaL

and * 1, 2, and 3 Sigm dN/da Lines

151

.s - -...
 Table XX

Comparison of Actual Cycle Count Data with Cycle Count Data

Predicted from Constant Variance dN/da Lines

A a 31.000 IMi ACTUAL PREDICT

-3 SIM 187M2 11i3s

-e SIGMlA 195632 140221

-1 SIGMA 205562 173118

MEAN 217991 213635

+1 SIGIA 233627 264538

+2 SIGIMA 53299 328255

+3 SIGMA 2rd046 409051

152
 SECTION XI

DISCUSS ION

Throughout the course of this investigation, a few unique events took

place and many interesting observations were made. Some of these have

rather simple explanations r i others require a quite detailed discus-

sion. Hopefully, some important conclusions can be made as a result.

11.1 Experimental Investigation

The behavior of fatigue crack propagation experienced during this

investigation was much different than first anticipated. The most sur-

prising event that took place in almost every test was the sudden changes

In the magnitude of the crack growth rate. Both sudden increases in the

growth rate, as if the crack had just come upon "ome unusually weak alum-

inum, and sudden decreases in the growth rate, as if the crack was exper-

iencing some unusually tough material, were observed repeatedly, many

times one or two millimeters after a previous event of similar nature.

One of the more outstanding examples of this type 'f behavior is shown in

Figure 66.

It appears that the material is made up mostly of a fairly homogeneous

material with many smaller areas located in a random fashion which char-

acteristically have vastly different crack propagation properties than the

majority of the material. The size of these small areas seems to vary

considerably from as small as less than I millimeter in length to perhaps

as large as 5 or 10 millimeters in length. These small areas obviously

have a very large effect on the overall smoothness of an a vs. N data set

153
 R VS. N
50.00 REPLICATE CA TEST 49
DELTA P = 4.20 KIP
164 DATA POINTS
Ac 9.000 MM
R- .20

60.00-

CD
Z 30.00 -
LU
-J
Nr-

lOOQ xxxXX xx)OcxXW x

0.00-

0.0000 .0M60 .10 .1960 .2600 .Ms

N (CYCLES) (X1O0

Figure 66. a vs. N Data Shoving Abrupt Growth Rate Changes

154
and on the total amount of scatter, especially in the growth rate data.

The average growth rate also seemed to vary somewhat from test to

test, with some tests running slow throughout the whole test, while

other tests ran fairly fast throughout the whole test. This phenomenon

is the cause of the outlying data sets in Figure 22. As also noted by

other investigators [5,6,7], the variation in growth rate at the beginning

of the test during the slow growth rates leads to most of the variation

in N at the final crack length.

As a result of these observations, the conclusion is made that this al-

loy is a very non-homogeneous material, especially considering the random

nature of crack propagation behavior. It very rarely obeys the smooth

growth rate equations often used to describe its behavior and does so

only when it's behavior is considered at a very macroscopic level.

11.2 Distribution of N

The conclusion stating that the cycle count data follows a 3-parameter

log normal distribution can be considered very strong. The only occur-

rences where this was not so was at short crack lengths where the need

for the location parameter used in the 3-parameter log normal distribu-

tion was not near as strong as at long crack lengths.

Fromfligures 24 through 29, the distribution parameters tend to vary

quite a bit at short crack lengths but ten,, to follow smooth curves after

a - 15 mm. The scale parameter in the first two distributions where no

location parameter is estimated have very smooth curves, showing that

mean crack propagation behavior does follow smooth growth rate equations.

The same smooth shape of the location prameter in the last four distribu-

'tions also supports this stateu.nt. Essentially these location parameter

j 155
 curves define au area where crack propagation will never occur. In other

words, the number of load cycles needed to reach a given crack length

will never be less than the estimate of the location parameter at that

crack length. This is shown in Figure 67. On this plot, crack propaga-

tion data will never occur to the left of the location parameter line.

Note from Figure 26 that the scale parameter, a, tends to remain constant

after a - 20 mm., allowing the location parameter to completely account

for the increase of N with increasing crack length. From Figure 24, note

the smooth increase of the shape parameter, &, as a function of crack

length, thus supporting the expectation of higher variances at longer

crack lengths. Another interesting event is shown In Figure 29. The

power parameter, c, of the generalized 4-parameter gamma distribution was

always estimated to be equal to one, thus reducing this distribution to

the 3-parameter gamma distribution. It should be noted here that over

half of the computer time used to obtain all of the distribution para-

meters was used to estimate the parameters of the generalized 4-parameter

gam9a distribution. By eliminating this distribution from the CCDDP pro-

gram, much time and money can be saved.

Another interesting occurrence which appeared very often is shown in

Table VI. Many times the distribution rankings implied by one goodness

of fit criterion could not be supported by another goodness of fit cri-

terion and often three different distribution rankings were implied by the

three goodness of fit criteria. In other words, the goodness of fit cri-

teria were not consistent between themselves. This necessitated a some-

what subjective analysis of the goodness of fit criteria. The closeness,

R , tended to be very sensitive to the scales of the plot and the slope

of the linear least squares line. Thus, the closeness was rarely used

156

-E .m-.-- -- -
I
II
160.00- REPLICRTE R VS. N DPTR
I 164 DATA POINTS PER
68 REPLICATE TESTS
DELTR P = 4.20 KIP
TEST

P MRX-= 5.25 KIP

0.o00 AO= 9.00 MM
R . . ........ ...

io

Z 40.o0- location parameter line__--- ,

z / . .t~.:. /*
~I--
iz30.00- I "'
-J /

I /
'
CC)
20.00-
I "
10.00-

10.00

S0.00 - I I i
0.0000 .0650 .1300 .1950 ,2600 .3250
N (CYCLES) (X1O 6)

Figure 67. Estimate of the Location Parameter of the

3-Parameter Log Normal Distribution as a
- Function of Crack Length

157
 unless the slopes were approximately the same between the different dis-

tributions. The chi-square tail area tended to be undiscerning between

distributions that provided fairly equal fits to the data. Thus, the chi-

square goodness of fit criterion was used only when there were fairly
4 large differences between the distributions. The Kolmogorov-Smirnov

goodness of fit test provided a fairly reliable and sensitive test and

was used heavily in establishing distribution rankings.

A typical fit provided by each of the five distributions for the

cycle count data is shown in Figutes 68 through 72. As state previously,

the 3-parameter log normal distribution provided a reliable tight fit for

the cycle count data as shown in Figure 70. The 3-parameter gamma dis-

tribution did surprisingly well and although it was not selected as the

best distribution very often, it consistently placed a close second to

the 3-parameter log normal distribution. The 2-parameter log normal dis-

tribution did not do well due to the lack of a location parameter. For

the 3-parameter Weibull distribution, the location parameter seemed to

work alright, but the shape of the density function did not match the data

very well as shown in Figure 71. The 2-parameter normal distribution

provided a very poor fit for the cycle count data and should not be in-

cluded in any further investigations of the dictribution of N.

11.3 Crack Growth Rate Calculation Methods

Of the six da/dN calculation methods selected, both the secant method

and the modified secant method contributed low amounts of error into the

da/dN data as shown in Table VIII. The modified secant method calculated

da/dN data which could be integrated back closer to the original a vs. N

data than the secant method could, perhaps because it calculates da/dN

158
 2-PRRRMETER
NORMRL DISTRIBUTION PLOT
REFLICRTE CR TESTS.
1 = 38.20MM /

CL 96.0x
]R 60 DATA POINTS
R .20/
//

-
I-I ~-x
-j 90.0-

90.0-
8o-

Cr. 60.0 /7R-.9"1

z

M = 4.9649

C,/ x AC=201
02.0
.02S.3' S.E. (U
C/ x
/ Am = 16701
CC /S.E. (T) = 1429.4
E /
S/ SLOPE =5.3u10"
S 0.1- I I I
-60000.0 -30000.0 0.0 30000.0 60000.0
DEVIRTION OF N FROM THE MERN

Figure 68. Typizal Fit of the cycle Count Data to

the 2-Parameter Normal Distribution

I 159
 2-PPRRMETER
LOG NORMRL DISTRIBUTION PLOT
5_s.s REPLICRTE CR TESTS.
RA=38.20 tiM
Lj 68/DPTS POINTS
IS/ - .20
Li/ x

rI
F- I

- 0.0-

cck
_j-

,v 50.0

:z
""0 = .9790
c- "Z = .0793
]z 100 /R = .95118
/x
v/x = 6.380S
w
1 .o .E. (U = .00355
6.-Wx A^ =.000655
,.-cz S.E. (B) =.000146
_J"mSLOPE = 30.662
S0.1

N (CYCLES)}
I-i

F7igurec . Typl.ca, Pit of the Cy,;Ie Count Data to

.-hes 2-ParaAster Log Normal Distribution

160
 3-PRRRMETE -
LOG NORMRL DISTRIBUITION PLOT
99.9- REPLICRTE CR TESTS. /
z R = 38.20Mt /
Lj 6b DRTI POINTE /

9C.
980 R=2

'-/

CC R= .9664
Z = .0609
zl 10.0 163
20.0-

z A
CC XS.E. -T 85.92
K/ .E O 4.7197[
U S2- 1oo = .77266
853 2
B =.016310
CS.E. (B) =.026363
SLOPE 7.143

N (CYCLES)-TRU HAT
-----.

Figure 70. Typical Fit of the Cycle Count Data to

the 3-Parameter Log Worual Dirtribistion

161
 3-PRRRMETER
WEIBULL DISTRIBUTION PLOT
999 REPLICATE CA TESTS.
A = 38.20 MM
9.0- 68 DATA POINTS
.20
90.0R
/
70.0-

LU 60.0-

C-)

S/ x R= .9098
c .c 7 xZ .0803
x R = .93932
/
/ A
_jx T =207836
0
1.0. 6.E. (TI = 7505.7
0- / B- 37085
/0.5 S.E. () = 19332
SE AC = 2.0614
/.E. (C) = 2.3206
SLOPE = 1.2771
0.1- 1 I I i
uS2 10 2 6 10'
N (CYCLES)-TRU HAT

Figure 71. Typical Fit of the Cycle Count Data to

the 3-Parameter Wibull Distribution

162
 3-PRRRMETER
GRMMR DISTRIBUTION PLOT
99.9- REPLICATE CA TESTS.
R = 38.20 Mm /
68 DRTA POINTS
R .20 1
99.5-x

z/
/x

C 95.0-/
99.0-

U= X/ .06.02
n 95.0--2

c/x
CC: Z= .0642
@ 70t.0-R .97401

~Cr60.0- 4 199349

CD:8 6458.1

SG= 6.4069 H
S.E. (G) =.:32
SLOPE =b.3,10"

-gO0.0 -30000.0 0.0 m000.0 o00o0.c

DEV. OF (N-TAU HAT! FROM THE MEAN

Figure 72. Typical fit of the Cycle Count Data to

the 3-Parmeter Cans Distribution

L .
16
163

,
dota at the original crack length levels instead of between the original

crack length levels.

None of tht incremelital polynomial methods calculated da/dN data

which could be inLes~ated b,' cven close to the original a vs. N data.

This it. no doubt due to thea 3-oo-.,Lng effact of these methods which tends

to reduce the sudden changes 'i growth rates. This is shown in Figu-e 38

where the number of extreme da/dN data points for the quadratic 7-point

incremental polynomial method is much less than Lhe number of extreme

da/dN data points for either the secant method of the modified secant

method (Figures 36 and 37). This is also shown in Figures 30 through 35

where the incremental polynomial method data Aullow a narrow band .line

while the recant method and modified secant method data folloa a more

broad band line. Note also from Figures 32 through 35 ch3 waviness of

the data shorinr. the large chnpges in growth rate noticed during data ac-

quisition.

If crack propagation data iVee always very smooth data, then the in-

cremental polynomial methods would'introduce a very small amount of error

into the da/dN data. But as stated pMeviously, the sudden changes in

growth rate are inherent in the crack pio.pazation process, and any attempt

at modifying these changes will distort thl resulting data and prevent it

from becoming a true representation of crack 'Wopagation behavior.

Of the four incremental polynomial mathods usod, botb Am qwaratic

7-point versaou. and the quedratielo$-lao 7-p0int-vrsU 'db the best job,

followed closely by the linear 7-point version. The linear log-los 7-

point version does a very poor Job 09 sbovp in Figure 34 and Table VIII.

The use of the tog-log transformation failed to give any improvemont over

the conventional incremental polynomial methods in the ability to reproduce

164
 aa
the original a vs. N data. The use of the second order polynomial fit

over a straight line polynomial fit improves thejperformance of the in-

cramental polynomial methods, especially when using the log-log transfor-

mation.

The amount of variaeibn-i. the 41yete tncreamniXl r cast error over

all of tharetest wa faildy 4eall, usualLy- les thsui; L2 per cent error,

indicating fairly consistent results over all of the experimental tests.

11.4 Distribution of da/dN

"No outstanding positive results were achieved for the disVrztion Of

da/dN. Each of the distributions provided roughly the same qu lity of fit

for the data, with the 3-parameter gamma distribution doing a Olightly

better job than the other four distribugions. ,

The da/dN data varied quite a bit as a function of AK as uhown in

Figure 36. As a result, different distributions would provide la fit for

the da/dN data at different AK levels, depending on the general O aed

skewness of the data at a given %X level. There were several ccasions

when the da/dN data was skewed lefe, as shown in Figure 73, syhtric, as

shown in Figure 74, or skewed right, as shown in Figure 75. COO.

When the data was either skewed left or symmetric, the 2-parameter

normal distribution provided the best fit. for the da/dN data, as shown in

Figures 76 and 77. When tbq data wa,sApwd righ, either o e. log nor-

ral distributions," the "44:parameter tivqrution,

4'bIoll or the 3-parameter

gamma distribution provided a fit for the da/dN data. Typical fits of the
skewed right da/dfl 8ate tb 'th$e f6ur disttibtrtions ', in Figures

78 through 81. Due to the large variations in the da/dR data, each of the

distributions is needed to provide a fit for the wide ranp of density

165
 RELATIVE
FREQUENCY HISTOGRAM
.r40- REPLICATE CA TESTS.
SECMT METHO
DR/ON CLRAS SIZE = 6.8174104 R = .20
DELTAP = 4.2 KIP A: 9.10 MM
68 DATA POINTS DELTA A = .Mo MM

>- ASMc

11

I-
GemJ
,L i

0.o0oo 0.0m .o

DON (1IN/CYCLE} (X1C -61

1iure 73. Typioal Skowed Left da/di Date

166
Fo

RELRTIVE
FREQUENCY HI STOGRRM
REPLICATE CA TESTS.
SECANT METHOD
OR/ON CLASS SIZE = 1.4.440 R = .20
DELTA F = 1.20 KIP A - 11.50 MM
68 DATA POINTS DELTA A = .20 MM
.2000 -

C--
>-.1 0-
Lu

L-

I-

0.01o

.1000 .150 .2m0 m0 =

OR/ON (IN/CYCLE) (MXO-)

FV1pro 74. Typical Syetrtc da/d3 Data

167 '
 RELRTIVE
FREQUENCY HISTOGRRM
REPLICATE CR TESTS.
SECANT METHD
DR/ON CLASS SIZE = 1.278xO R = .20
OELTR F = 4.20 KIP R = 34.10 MM
68 DATA POINTS DELTR A .20 MM

>- .100
II Lii
O
-AJ

c.~0

zz
.or
-04 -- - -

DAM/N (IN/CYCLe) () 0 4m*3

Irip"r 75. Typtool Skewd Iijht da/di Deat

 2-PRRPMETER
NMA DISTRIBUTION PLOT
~ - FLICATE CA TfSTS*
tMELTAA .20 MM
ck: 408 DATA PO~INTS
L&J~ 9.10 MM
'I-

Zi =.04

/~ =R .951759

~ 2.0 76.E. 40 = 2.'4B410,

/ A a = e.oe5ito

1' 5LOP = '4.5Olu1Oe

firms 76. ?Ipil.. lt of Vmmw Left as/do va& to

 .2-PPRRMETER 1

NORMOL DISTRIBUTION PLOT

"~~IV4CATE -f.A TESTS. I -

pSECIWT MtTMO
DOELTA AR= .20 MM/7
68DATA P13INTS
o~1*.o .20

X/1X

Cso-./L

CO

W S.E. 4O= .5o

11io 4

-ii SLCPE = 2.433%10~

EVITu1ON O DA,'DN FROM HE'MEAN

Iftlre
7. ypial o St t7 i &a/40 Dosto
the 2.?armater Notus iDstributiou

170
 2-PPR iETER
LOG NORMRL DISTRIBUTION PLOT
99.9-REP-LCRITC, ;TEM..

K A:I: 34.0
to mm:
a ge.o- aRA,.i
it =20 /
/x.-,
/X

- 80.0 t . .

oa-

CZ. .0616

C3 2.-A
.97

10.0- R .98632
: I'--/X
xA :-4.871
2.0/ S.E. (UA 1.276x10
.1-._ . -,/ A" 8 I -

C./ S.E. (B) 1.886410

/ SLOPE = 8.67402
: 0.1- / ..
510 14 2! 5

DR/ON (IN/CYCLE)

Figure 78. Typical tit of Skewed RiSht da/dl Data to the

2-Parameter Log Noual Distribution

171

I
 3-PRRRMETER
LOG NORMAL DISTRIBUTION PLOT
REPLICRTE CR TESTS.
SSECANT
U DELTA A METHOD
= .20 MM
i n,, A068DATA POINTS
= 34.10 MM
SLLJ /X
n. 98.o- R = .20

-j
-J 90.0-

I::.
a-
ca

Z = .0662
o~ =.o.99085
C320.0-
r A
z 10.0 "T = 3.764il0
c::
I,.-- S.E. = 4.373i1O=
/(/X- :-5.0

to-
JS.E.
/
,* *"
=
-
.696,10-'
.2
S/ S..A) = 2.9S1O710e
ac 9 S.E. (i) = 2.92iDte
::SLOPE = 6.21599
w
: 0.I*~- - ,1 ORI (NC .LE.(UHR= 6.26969l
; 2 S ~104
t OR/D)N ( IN/CYCLE]-TRU HRT

ief~re 79. 'Typical Fit of Skewed Rig~ht da/dN Daet to the

3-Pat.erUS Log Noarim Distribution

172
 3-PRRRMETER
WEIBULL DISTRIBUTION PLOT
99.9- REPLICRTE CR TESTS.
99.0-SECANT METHOD
9.0 DELTA A = .20 MM //X
68 OATR POINTS
90.0- A = 34.10 MM
R = .20
70.0-
z
WJ 60.01

Lj 30.0-
a-

/x

c_- 5.0-
fl -Jx
/"xI=.g3 R .063

x R= .94181
x T = 7.229,10
c 1.o-c S.E. fJ = 9.290x10
x10 -
B = 7.00SA
o -S.E.(B = 4.080%10
/
AC
A
= 1.961
S.E. CC) = 1.S05
SLOPE = 1.23502
0.1- 1 I
10 2 S 10 S5 2
OR/ON (IN/CYCLE)-TRU HAT

Figure 80. Typical Fit of Skeved Right da/dR Data to the

3-Parameter Weibull Distribution

173
 3-PRRRMETER
GRMMR DISTRIBUTION PLOT
999_ :YPLICRTE CR TESTS.
SECWT METHOD
4
DELTA A = .20 MM
68 DATA POINTS /
A=34.10 MM /
99.- R= .20 /
/

99.0-1
i... / x

L) 9.0 :
LAJ

goA .9992
c Zz .0/61
CC=
70.O0 6.610
Y: Z6- .06.

AT = 1.439x10'
q /.. a = 1.47Sm1O
.0 S.E. f
= 4.570z10 7"

G =4.898
A
S.E.S (G) =2.280
w
/ SLOPE 2.499410s
0.1 1 i .
O
6.0,1O
-. 0.0 61.010 i.e-iO
DEV. OF (OR/DN-TRU HRT) FROM THE MERN

FIgure 81. Typical Pit of Skeved light da/dR Data to the

3-Parameter Geama Distribution

174
function shapes.

The large amount of variation in the shape of the da/dN data as a

function of &K is shown in the plots of the distribution parameters as a

function of crack length (Figures 40 through 45). Note that there is a

lot of variation in the shape parameter (the pluses) and the location para-

moter (the diamonds). The variation of the shape parameter reflects the

changes in the amount of variance and the shape of the data. The variation

of the location parameter reflects the changing skewness of the data. As

the skewness goes from right to left, the estimate of the location para-

meter decreases rapidly. Also, from Figure 45, it can 'he seen that there

are many occurrences %-here the estimate of & was not equal to one, thus

implying the necessity of the inclusion of the generalized 4-parameter

gama distribution when analyzing da/dN data so that a wide range of den-

sity function shapeo can be accommodated for the da/dN data.

11.5 Prediction of a vs. N Data from the Distribution of da/dN

The results of the prediction of replicate a vs. N data from the dis-

tribution of da/dN were less revealing than anticipated. When comparing

Figure 46 with Figure 22, it becomes apparent that the variance of the

predicted a vs. N data is much less than the variance of the actual a vs.

N data. However, the mean of the predicted a vs. N data is very close to

the mean of the actual a vs. N data. The implication of this is that

crack propagation behavior is not being accurately modeled by a randomly

selected value of da/dN from the distribution of da/dN. In crack props-

Sation behavior, as discussed in Section 11.1, the growth rate at a given

AK level is not independent of the growth rates at previous 6K levels, as

evidenced by periods of up to 10 ma. of uncharacteristically fast or slow

175
 I
growth rates. However, the independence of growth rates is assumed in

the prediction of a vs. N data from the da/dN distribution parameters,

resulting in very smooth a vs. N data. This smooth a vs. N data lack&

the areas of sudden fast and slow growth rates discussed in Section 11.1

which occurs frequently in actual a vs. N data. Thus, the combination of

many smooth a vs. N lines of the same mean behavior results in the Teduc-

tion of variance noted above. To accurately predict crack propagation

behavior, some means of quantitatively describing the interdependence of

adjacent growth rates must be found.

When the distrihution of the cycle count data predicted from the

distribution of da/dN was analyzed, neither gamma distribution would

converge on its parameters as the estimate of the shape/power parameter,

S, tended to approach its upper global limit. The 2-parameter log normal

distribution provided the best fit for this date because the location

parameter of the 3-parameter log normal distribution was estimated to be

zero at most crack length levels.

When the distribution of the predicted cycle count data it compared

with the distribution of the actual cycle count data, as shown in Table

XIII, it can again be seen how the mean of the predicted cycle count data

is very close to the mean of the actual cycle count data while the stan-

dard deviation of the predicted cycle count data is much loes than the

j standard deviation of the actual cycle count data.

When a vs. N data are predicted from constant variance da/dN lines,

the spread of the predicted data is much wider than the spread of the

actual data, as shown in Table XIV. This occurs because either all very

slow or very fact growth rate da:t is used at the + 3 sigma da/dN lines,

thus causing either a very lonE or very short number of cycles. The

176
actual data, however, rarely has any growth rates on the order of + 3

sigma, and even more rarely has repeated growth rates on the order of + 3

sigma. On the average, actual data tend to have repeated growth rates

within + I s igma.

Froe Figure 51, it can be seen that the constant variance lines tend

to get further apart when going from left to right, indicating that the

distribution of N is skewed right. Since the distribution of N has been

determined as the 3-parameter log normal distribution which is a skewed

right distribution, the prediction of a vs. N data from constant variance

da/dN lines supports this conclusion.

11.6 Inverse Gro th Rate

As anticipated, an improvement in the fit provided for the dN/da

data over the fit provided for the da/dN data was obtained. The 3-par-

mater log normal distribution was able to provide the best fit for the

dN/da data without serious competition from the other four distributions.

This improvement is partially due to the inversion of the growth rate

variable. Since N was strongly log normally distributed, it was antici-

pated that AN would be log normally distributed also. Another reason for
this improvement was the exclusion of the location parameter from the

gamma distributions, Lhereby severly decreasing their ability to provide

an adequate fit for the dN/da data. The fit provided by those distribu-

tions which estimated a location parameter was significantly better than

the fit provided by the gaim distributions. Quite a large raWe of

values were estimated for the location parameter (from - 1.6 x 10"1 to

4.9 x 10S) and the absolute value of the estimate of the location pare-

water was always greater than 900, ULadicin8 no tendency to opproch -

177

__ I
 zero as acaumed in equation 48 (Section 5.2.0). Thus, this assumption

has not proven valid.

The value of the location parameter of the 3-parameter log normal dis-

tribution assumed negative values in many instances, indicating that

skewed left and symmetric dN/da data was present as well as skewed right

dN/da data. This was expected since the simple inversion of the da/dN
variable does nothing to change the skewness of the density function of
the data. The only effect of this inversion is to change the direction

of the skewness and to alter the shape of the density function slightly.
1.
A histogram of typical symmetric dN/da data is shown in Figure 82 and

plots of the fit of the dN/da data to each of the distributions are shown

in Figures 83 through 87. Note in Figure 85 the ability of the 3-para-

meter log normal distribution to handle symetric as well as skewed right

data. Again, due to the large variation in the dN/da data, each of the

distributions is needed in order to provide a fit for the data.

There is a large variation in the shape parameter and location para-

meter again for the dN/da data, as shown in the plots of the distribution

parameters as a function of crack length (Figures 53 through 58). The use

of dN/da does not remove these variations from the data, although it does

reverse the basic trend of the mean as shown by comparing Figure 40 with

Figure 53. The msan value of ds/dN increaes as a function of crack

length while the mean value of dN/da decreases as a function of crack

length, both being expected for constant amplitude loading.

The use of dN/da distribution parameters in the prediction of repli-

I cats a vs. Ndata did not change the predicted data noticeably. As au&-

jt geated previously, the problem of predicting a vs. I0 data accurately lies

1 178
I
I

S~RELRTIVE i
FREQUENCY HI STOGRRM :
.24W- REPLICATE CA TESTS.
SECANIT METH-OD -
ON/DR CLASS SIZE = 10644 R = .20- -.
DELTR P = 4.20 KIP A = 21.10 mm
68 DATA POINTS DELTA A =.20 MM

LLJ .O600
>- I
LU

IiI

0.00w ---

.100oo .100
OELTR N/DELTA A ICYCL /IN) (Y10 0)

Figure 02. 'ypical gyovmrirc do/da Data

S. . . " , . . . .
 2-PRRRMETER
NORMFIL DISTRIBUTION PLOT
, 999REPLICATE CR TESTS.
SECANT METHOD
LU
U DELTA R = .20 MM
88 DRTA POINTS
cc A = 21.10 MM
LU - .
.. 96.0 R= .20 /X
/x
-X
90.0-

r 80.0-

CIO

cr. Z= .0131
C3 20 .0 -/ 10.0- R .9.98'457
. . 1
, 0-0
c:P co0 // 1.63SX10' IT322
0
2.0- / S.Em (0) = 3.94'f.10'
/A 3
cc S.E. (0) = 2.783x0
ZD 0.1- /SLOPE = 2.8065iO
.....

-M.0 -4O=.O 0.0 iO=.O 80000.0

DEVIATION OF ON/OR FROM THE MEAN

Iigure 83. Typical Fit of Symmetric dl/da Data to the

2 - 1Lra m ot
er No r nm
l Di st r ib u t io n

18
 2-PRRRMETER
LOG NORMRL DISTRIBUTION PLOT
0.;-
Z
REPLICATE CA TESTS.
6~~ECA~NT NETHOD
JOELTRA = .20 MM
S fDlATAR POINTS "
A 21.10 MM /
EL R .20

-J 90.0-

: (:80.0-

0-

r.C60.0-

z
xi
- A 9530
S20.0 -.
w Z= .0821
10 R .98859
CC
o I-e = 5.205
2.0- / S.E. A(O 1.060.10"
B-
AB- 7.636x1
CC / S.E. (8) = 1.300
SLOPE = 10.45911
_ S0.1

DELTA N/DELTA A (CYCLES/IN)

Figure 84. Typical Fit of Symetric dX/da Dat* to the

2-Parameter LoS Motmal Distribution

1I I
' 181
 3-PRRRMETER
LOG NORMRL DISTRIBUTION PLOT
REPLICATE CA TESTS.,,,,
S $ECWT MEITMD/
DELTAR = .20 MM
T P INTS SDATA680
S-- A = 21.10 MM
C o R .20 ,

1_j
- 90.0-

dr.
XA= .A769
Lm .0693
0 20.0-.95

3 0.0- S~~S.E. f =-8.134xlo'

=3.7117x10,
= 5.38S
U, exO- S .E. (U = 1.721
= 3.307X1O-
AB
CC
..j/ /S.E. (B) = 1.139x10
-J SLOPE = 15.92019
S0.1-" -'
'
10
DELTA N/DELTA A (CYCLES/IN)-TAU HAT

Figure 85. Typical Pit of Symmsetric dN/da Data to the

3-Paormater Log Normal Distribution

182

*2 j _____ ____4
 I
3-PARAMETER I

WEIBULL OISTRIBUTItON PLOT

P LIC ATE CA TE S TS . SRE
99.0- DELTA A =METHOD
'SECANT .20 MM /
68 DATA POINTS
90.0- RA 21.10 MM
R= .20
70.0-
ZI

0_"_.'J
33.0- /

/X/
- 10.0-

co /x A= .9813
S5.0 Z = .0690
Ot x R = .96606

j= 7.999110S
= 1.0- S.E. ( = 2.438x10"
0." A = 9.377104'4
S.E. (B = 3.S66x10
AC = 2.7S2
S.E. (C) = 1.265
SLOPE = 1.70414
0.1-
101 2 5 102
DELTA N/DELTA A (CYCLES/IN)-TRU HAT

Figure 86. Typical Fit of Symmetric dN/da Data to the

3-Parameter Weibull Distribution

183
 2-PRRRMETER
GRMMR DISTRIBUTION PLOT
999_ REPLICATE CA TESTS.
SECANT METHOD
DELTA A = .20 MM
68 DATA POINTS
99.5- A = 21.10 MM /
99.- R .20 /

U-- 96.0- j
LL 90.0- /'

-- .0-
760.0-

S60.0 = .9911
aZ .0693
x" 2 .0-. ctR = .99281 I
xA r^ = 6.485x10
31
S.E. (5) = 1.101x1O:
/'x ^G = 2.5212101
x S.E. (G) = 4.287
.SLOFE = 2.802O40
I 'I I ]
-80000.0 -40000.0 0.0 40000.0 S0000.0
DEVIATION OF ON/OR FROM THE MEAN

Figure 87. Typical Pit of Symetric dN/da Data to the

2-Parameter Giam Distribution

II
not in which variable is used to predict the data but rather in the

assumption of independent adjacent growth rates.

When the cycle count data was predicted from the dN/da distribution

parameters, the 3-parameter log normal distribution provided the best fit

for the data as the estimates of the location parameter were all at anti-

cipated values. This is an improvement over the prediction of cycle

count data from the distribution parameter3 of da/dN, because the

estimate of the location parameter was near)y always equal to zero. The

use of this location parameter significantly improves the fit of the pre-

dicted cycle count data to the 3-parameter log normal distribution.

Again, the values of i assumed maximum global values in both gaumna distri-

butions. This is most likely due to a lack of significant variance in

the predicted cycle count data. When the distribution of predicted cycle

count data was compared again to the distribution of actual cycle count

data, the mean data was almost exactly predicted while the predicted 1
standard deviation was again much less than the actual standard deviation,

which can be seen by comparing Figure 59 with Figure 22.

The a vs. N data predicted from constant variance dN/da lines almost
I
exactly reproduced a similar plot made from constant variance da/d14 lines,

as seen by comparing Figura 51 with Figure 65. Thus, the dN/da data seems f
to support the conclusion that the cycle count data fits the 3-parameter

log normal distribution the best.

.I I
?285
 SECTION XII 4
CONCLIS IONS

The most significant conclusions of this investigation are summarized

as follows:

1) The 2-parameter Weibull distribution was tried on previously

generated fatigue crack propagation data and, due to its very

poor performance, was dropped from the remainder of the sttiasti-

cal analysis (Section 8.1)

2) Actual replicate cycle count data followed a 3-parameter log nor- I

mal distribution, with especially good fits at louger crack

lengths "Section 11.2).

3) The modified secant method introduces the lowest amount of error

into the da/dN data of the six growth rate calculation methods

selected (Section 11.3).

4) The large amount of variance present in the da/dN vs. 6K data

prevented a consistent fit of the replicate da/dN data to any of

the candidate distributions (Section 11.4).

5) Replicate dN/da data followed a 3-parameter log normal distribu- I

tion (Section 11.6).

6) The method of predicting a vs. N data from the da/dN or dN/da

distributLon parameters was not completely successful due to the

assumption of independent adjacent growth rates (Sections 11.5

and 11.6).

186

1
 SECTION XIII

RICOIUNDATIO0 FOt FUITIHE WORK

The use of statistical methods in describing and predicting fatigue

crack propagation behavior worked very well. However, accurate life

prediction was not achieved because a total statistical description of

the crack propagation process has not been determined. Only a minute

percentage of the total possible experimental and analytical work needed

to achieve this total statistical description was conducted under this

investigation. Based on the observations, results, and conclusions of

this investigation, the following topics need further investigation.

1) Experimental crack propagation data vith N as the independent

variable and a as the dependent variable is needed. From this

the distribution of a as a function of N and the distribution of

da/dN as a function of N can be obtained.

2) A study of the interdependence of growth rate data would be

valuable for use in the prediction of a vs. N data from the dis-

tribution of growth rate data.

3) The effect of data density on the variance and distribution of

growth rate data needs to be found to aid in more accurate data

acquisition and analysis.

4) A study of the sudden growth rate changes in the original a vs. N

data mentioned in Section 11.1 would aid considerably in the

understanding of the crack propagation process.

1mood
5) A more reliable and accurate mehod of establishing the distrL-

bution rankings is needed. The goodness of fit criteria used in

this investigation did not totally fulfill this nosd.

il

Ii

i Ii

11t1i
 r I

tt
APPENDIX A

DERIVATION OF THE da/dN EQUATION FOR THE LINAR LOG-LOG 7-POINT

INCRE1NTAL POLYNC34IAL )4T0OD

The fitted polynomial equation for the linear log-log 7-point incre-

mental polynomial method is given by

a o + b, NLS (A-1)

where NLS is given by

log N- C
"NL 10c 1 (A-2)
2
where C1 and C2 are given by the scaling equstions (equations 7 and 8,

Section 4.3). Substituting into equation A-1 for NLS,

olog 10 N - C
los10 a - b + b, . 2 12 .' (A-3)
L C2

Solving for a,
b + bI log
lI N-C
- I

a a 10 o C22
b C b log10 N
b . -
+...
0 C2 C2
- 10

bo C (1o10 N) C
- 10 2 I0
b C b

- 10 2N (A-4)

189
Taking the derivative of a with respect to N and evaluating at the mid-

point, Ni, j
b C b
da o C bl C2
10 N, 2 (A-5)
~2

190
 -4

APPENDIX B

IARIVhTION OF lIii da/dN EQUATION FOR THE QUADRATIC LOG-LOG 7-POIfT

INCRIMINTAL POLYNOMIAL MITliO

The fitted polynomial equation for the quadratic log-log 7-point in-

cremntal polynomial method is given by

log 10 a a b + bNL+b 2N.l' (B-I)

vhere N. is given by equation A-2. Substituting into Equation B-I for

NIS
+l I
0l910 N - C1lo189 0 N C1 _12
lo a b +b +-b-c (B-2)
Og 01 a 0 1 C2 .j C2

Letting
uU - 1Ogl -NC I
(-3)
C2

dU . I (1-4)
dN N ln(10) C2

Then

log10 a - b0 + bI U + b2 U2 (B-5)

Solving for a,
b + b1 U + b U2

a - 10 (1-6)

Taking the derivative of a uith respect to U,

191
 d-.10 b0 *.d b1U .'o b!.U2
dU dU 1B

where

d
101
o
bU
10
b2 ' bU
10
b 2 U2 "
2b Un(10)
i
dU LL 2
(B-8)
b U2 b U
+ 10 .10 b in(
1 )

Then
bI U b 2 V2
b
La 10
do
10 -10 in(10) L 2 b 2.U + b (B-9).

I'0:63 the chtin rule,

'S I
Asd da . dU (B-10)
dN dU dN

- 10 0 1b b10! in (10).. L2 b2 U +b LN- ln(10) .C2 J

2
b'0 b1 U bU
2 (b-1l) I
C 10 2 Ub ) 2(2b
10 10
C2 (1) L 2 N I(I)C

Substitutin$ for U and evaluatin at the midpoint,. Ni

b!1io$ 14 bC b2 (log11) - 2boC Io111 + b2 C1 2
b 1 C- J L 2 2 J

2b 2 loMg - 2b 2 C1 b
C-+
2(5+)

402.-
 APfPNDX C

DERIVATION OF C2

From linear regression, the coefficient of multiple determination,

R is given by 34]

R. R"
-1-I S(C-I)
TCSS(C1

Where SSRES is the residual sum of squares and TCSS is the total corrected

sum of squares. Since SSRIS and TCSS are wasured in the vertical direc-
tion, it was desirable to correct them so that their direction is normal

to the slope, m. Let the slope be given by

m- J/K (C-2)

Where J is the side of a triangle along the least squares line and K is

the side of the triangle perpendicular to the least squares line. From

basic geometry,

12 . j2 + K 2 (C-3)

Where I is the third side of the triangle. I is always in a vertical di-

rection. Substituting from equation C-2 into C-3 for J,

I- (. K)2 +

K2 (2 + (C-4)
()

193
 2
Solving for K

.2
K2
K (C-5)C-6

,2/'
Solving for the correction of the slope,

Plugging this into equation C-1 to obtain R corrected for the slope,

called C

C2 "R 2
(K 2 / 2
) (C-7)2

SSlES-: 1
"1- 1 - (C-8)

p 1
I2

IA
194.
 t I7

APINDIX D

DNDDPG DOCUM141ATION

This program consists of a main pogra-, and 25 subroutines. The main

program (DNWDPG) reads in the desired Aa vs. AN data and calls subroutine

DELTA to re-create the original a vs. N data. The program flow is then

transferred to subroutine CLASS which divides the data into contont Aa

data sets and theu calculates the histogram frequencies for each constant

A&a data set. The program flow is then transferred to subroutine STPLOT

which determines the distribution.

Subroutine STPLOT uses, directly or indirectly, the following sub-

routines.

I. Parameter Estimation Subroutines

r
1. GOLDEN
2. CRVFIT

11. Scaling Subroutines

1. PRUPLT

2. WBLhn -.

3. WGSCAL

4. LESCAL

5. INISC

6. ODSCAL

III. Plotting Subroutines

1. DOPIDt
 2. RF

3. L00PIX

4. WBLLT

IV. Output Subroutinme

1. RIM
2. RITPA,

.. f ral Purpose Subroutines

eV. -

.1. LSTSQR

2. R.ANK
3. OUTLIR

4. NWMTAB
5. SIPSN

6. MMAR
7. MIAXI

A listing of this program can be obtained from:

Prof. B. M. Hllberry
School of Mechaneal Engineering
Purdue University
West Lafayette, Indiana 47907
Phone (317) 494-1600 " - "

It
1W1

,r-

. .. , ... . .; .: :
 APPENDIX 3

CCDDP DOCUMINtAT ION

This program consists of a main program, 53 subroutines, and 18 func-

tion subprogram. The main program (CCDDP) reads in the desired replicate

cycle count data and writes it by calling subroutine RITDAT. The program

flow is then transferred to subroutine CLASS which calculates the histo-

gram frequencies for the data. The program flow is then passed to sub-

routine STPLOT which determines the distribution.

Subroutine STFLOT uses, directly or indirectly, the following sub-

routines and function subprograms.

I. Parameter Estimation Routines

A. Subroutines
1. MUILN
I
2. lILEW

3. MIZG

4. HLZGG
{I 5. IHl

B. Supporting Function Subprograms

1. 7W

2. 7W

3. FG

4. FOG

197

_ _
_ _II_
 II. Statistical Parameters Subroutines

1. NRSTAT

2. WBSTAT

3. G16TAT

III. Goodness of Fit Routines

A. Subroutines

1. CHISQR

2. KOLSHR

3. NRMCS
4. WBLCS

5. GAMCS

B. Supporting Function Subprograms

1. FNMR

2. FWUL

3. FGAM

IV. Output Subroutines

1. RITPAR

2. RrITUS

3. PAROUT

V. Plotting Routines

A. Main Plotting Subroutines

1. AVNPLT
F
2. HISPLT

3. MUMPT

4. LOGPLT

5. WELPLT
6. GAMPLT

198
 B. Supporting Plotting Subroutines

1. ODAXIS

2. LGOAIS

3. ,MAXIS

4. SCINOT
VI. Scaling Subroutines

1. MICL

2. WBLSCL
3. GAMSCL

4. LGSCAL

S. LNSCAL

6. INLNSC

7. ODSCAL

8. SCALEL
VII. Stress Intensity Calculation Routines

A. Subroutine

1. DELTAK
B. Function Subprogram

1. FAB
VIII. General Purpose Statistical Routines

A. Subroutines

1. NRWfAB
2. OUTMIR
B. Function Subprogram

1. PNORM

2. FGAMMA

199
 IJ

3. FPSI . .

4. FTRIGM

6. FG(

I: ~8. FGHNEG

9. FSER.- . - .
10. F'FRAC

IX. General Purpose Subroutines

1. TABIL.

2. INrHAV

3. INTrERP - "

4. INVINT
5. LSTSQR

6. INVMAT

7. RANK

8. MANCHA
9. INITR - ..

10. ITOR

11. LOG

12. MAXR

13. MAXI

A listing of this program can be obtained from Professor B. M.

Hillberry (Appendix D).

J
 D-
APNDIX

t
CGRDDP DOCIUfWTATION
This program consists of a main program, 50 subroutines, and 17 func-

tion subprograms. This program ts nearly identical to the CCDDP prosram

(Appendix E) and only, the main program (CORDMP) and 3 subroutines are
changed. These 3 subroutines are;

1. CLASS,

2. STPLOr, and

3. RITDAT.

This program requires an input of replicate growth rate data and has the

sameOut-put as the CCDDP program. Subroutines DELTAK, ?TOR, and MAXI and

function subprogram FAB need not be loaded for this program. A listing

of this program can be obtained from Professor B. M. Hillberry (Appendix

D).

201
 APPENDIX C

DNWDP DOCUMNTATI.N

This program consists of a main program, 49 subroutines, and 17

function subprograms. This program is neArly identical to the CGRUP

program (Appendix F) and only the main program (DNDDP), 9 subroutines,

and 3 function subprogram are chanU9d. The 9 subroutines that are

changed are;

1. CLASS,

2. STflOT,

3. lUG,
4. MZGG,

5. O1TAT ,

6. GAMCS,

7. RITDAT,

8. RITPAR, and

S. GA ) ,LT.

The 3 function subprosrams that are changed are;

1. FG,

2. FOG, and

3. FOAM.

This program requires an input of replicate grovth rate data and has the

sam output a the COPIDD program. Subroutine 1WMAT need not be loded

202;
for this program. A listing of this program can be obtained
from
Professor B. M. H411berry (Appendix D)..-
.

203L
 APPNDIX H

D6LTCP DOCUMENTATION
This program consists of a main program and 5 subroutinas. The main

program (DELTCP) reads in the desired a vs. N data and calls the proper

subroutine(s) to calculate the A& vs. AN data according to the desired

calculation method chosen. The Aa vs. AN calculation subroutines are;

1. RHOCV,

2. STRIP, and I
3. DELTA.

The main program then calls ssbroutine RITDAT to write the 4a vs. AN

data. The only general purpose subroutine required is subroutine LSTSQR. I

A listing of this program can be obtained from Professor B. H. Hillberry

(Appendix D).

______

204
 S1

APPENDIX I

DADNCP DOCtU3NTATION
This program consists of a main program, 22 subrouttines,

tion subprogram. The main program (DADNCP)

dota set and calls, directly or indirectly,

and 1 func-

reads in the disired a vs. N

the following subroutines and

il
function subprogram.

I. Growth Rate Calculation Subroutines - *Ii

1. DADN I
2. SECANT ... .

3. HOD9KC I

4. STRIP
5. EVAL ..

II. Error Determination Subroutines .

1. DELTA

2. I1Ml GR c '",'....--i

III. Output Subroutines

1. RMTAT

2. RtITRIB
i 3. U|IULT

IV. Plotting Subroutines

1. AVIWLT

2. LOWL

I. I ...
 _ .Scalift Subroutiti.

1. LOSCAL
- --V , -Stress Intensity Calcudlatias lodtines

VI CasalPupoeSubroutine s

1. llx

2. ZUITI

- 3. CHICK
4. ITOR
S. KMf

6. LOG

7. UJTSQR

A ),toting of this program can be obtained from Professor B. bI. Hilberry

(Appendix D).

206
 AVNPRD DOC1DINTATION

This prosram consists of a main program, 19 subroutines, and 8 func-

tion subprograms. The main program (AVNPRD) reads in the desired distri-

bution parameters and calls, directly or indirectl,' the following sub-

routines and function subprograms.

,. Random Number Generating Subroutine

1. RNIG4
II. Inverse Distriution Subroutines

1. INVDIS

2. INWRBJ
3. IW21N

"4. I3LN..
5. INYVIL

* 6. INWCAN

111. Prediction Subroutines

r. 1. PUD

2. 8ICAMT. - ..

3. MDS&C
4. STRIP
IV. Output Subroutine
1. IITDAT

207
 V. Plotting Subroutine
1. AVNPLT

VI. General Purpose Statistical ROUtiEne

A. Subroutine

1. MrAB

B. Function Subprograms

~1. Film
2. FCA)MA

3. FINGAM

4. FOM

5. FOMINT

6. FGHMG

7. FSER

8. FFYAC

VII. General Purpose Subroutines

1. TAJIRL

2. IYrHAV

3. IwR.RP

4. INVIWT

A listing of this program can be obtained from Professor 3. M. Hillberry

(Appendix 0).

to

S.r |
 APPENDIX K

RAN=DON= EXNIRIMNUL 79STS

OFS

The r&inoelsed order of the: pecimens useid during testing is ~

s8

1.131 -19., 141 -3.49 55. 127

2. 101 20. 60 38. Ill 56. 6

-i -~147: 6339.
:8521 102 57. 71

-4. 77 22. 50 40. 64 58. 58

'- 4 ;23?3 13 1-1A.2 .59. -78

-35 119 24. 55 42. 94 60. 106

- - . 112 -26.
J.14
21
3.18
4 -25 44. 130
6112
62. 31

-9- 42 *,'27. -7 45. 143 63. 125

1.0. 40 28. 2 46. 57 64. 32

11. 18 29. 70 47. 41 '65. 116

1.2. 132 30. 19 48. 9 66. 62
13. 118 31. 52 49. 74 67. 2
14. 1 32. 83 50. 33 6.121

15. 35 33. 1.49 -51. 144

16. 123 34. 80 52. 81

17. 37 35. 9? 53. 93

18. 139 36. 96 54.1 9:9

209
 REFENCES

1. Van Vlack, L. H., El0Ments of Materials Science, 2nd edition,

Addison-Wesley Publishing Co., 1964, pp. 160-165,

2. Frank, K. H. and Fisher, J. W., "Analysis of Error in Determining

Fatigue Crack Growth Rates," Fritz Engineering Laboratory Report
No. 358.10, Lehigh University, Bethlehem, Pa., March 1971.

3. Wilhem, D. P., "Fracture Mechanics Guidelines for Aircraft

Structural Applications," AFDL-TR-69-111, Air Force Flight Dynamics
Laboratory, Wright-Patterson Air Force Base, Ohio, February 1970.

4. Gallagher, J. P. and Stalnaker, H. D., "Developing Methods for

Tracking Crack Growth Damage in Aircraft," AIAA/ASME/SAS 17th
Structures, Structural Dynamics and Materials Conference, Valley
Forge, Pa., May 1976.

5. Hudak, S. J. Jr., "The Analysis of Fatigue Crack Growth Rate Data,"

E24.04 Committee Meeting, ASTM, Philadelphia, Pa., 1975.

6. Clark, W. G. Jr. and Hudak, S. J. Jr., "Variability in Fatigue Crack

Growth Rate Testing," ASTM 124.04.01 Task Group Report, Westinghouse
Research Laboratories, 1974.

7. Hudak, S. J. Jr., Saxena, A., Bucci, R. J., and Malcolm, R. C.,

"Development of Standard Methods of Testing and Analyzing Fatigue
Crack Growth Rate Data (Third Semi-Annual Report)," Westinghouse
Research Laboratories, 1977.

8. Pook, L. P., "Basic Statistics of Fatigue Crack Growth," N&L Report

No. 595, National Engineering Laboratory, 1975.

9. Gallagher, J. P., "Fatigue Crack Growth Rate Laws Accounting for

Stress Ratio Effects," ASTH Task Force 324.04.04 Report No. 1,
Air Force Flight Dymmics Laboratory, Wright-Patterson Air Force
base, Ohio, 1974.

10. Juvinall, R. C., 2%ineerin Considerations of Stress, Strain. and

Strenuth, McGraw-Hill Book Co., 1967.

11. Osele, B. and Housing, R. W., Statistics in Research, 3rd e4ition,

lova State University Press, 1975.

_ _ . ... .. . _ _ _
210

_ __I Hl
 .Walpole,
12. gulneert
R.andS.Scientists,
and Myers, Macmillan
R. H., Probabilit and Statistics
Publishing Co., Inc., 1972.
1r

13. Goal, P. K., Statistics Department, Purdue University, V. Lafayette,

Ind., 1977.

14. Aitchison, J. and Brown, J. A. C., The Lholrmal Distribqttou,

Cmbride UniveraIty Press, '1957.

15. Wingo, D. R., "The Use of Interior Penalty Functions to Overcome

Lognormal Distribution Parameter Estimation Anomalies," Journal of
Statistical Computation and Simulation, Volume 4, 1975, pp. 49-61.

16. Lambert, J. A. "Estimation of Parameters in the Three-Parameter

Lognormal Distribution," Austrailian Journal of Statistics,
Volume 6, 1964, pp. 29-32.

17. Wirsching, P. H. and Yao, J. T. P., "Statistical Methods in "

Structural Fatigue," Journal of the Structural Division, American
Society of Civil Engineers, Volume 96, No. ST6, 1970.

18. Lipson, C. and Sheth, N. J., Statistical Design and Analysis of

Engineering Experiments, McGraw-Hill Book Co., 1973.

19. Zanakis, S. H., "Computational Experience with Some Nonlinear A

Optimization Algorithms in Deriving Maximum Likelihood Estimates

for the Three-Parameter Weibull Distribution," Special Issue of
Management Science on Computational Methods in Probability Models,
1976.

20. Lienhard, J. H. and Meyer, P. L., "A Physical Basis for the
Generalized Gama Distribution," Quarterly of Applied Mathematics,
Volume 25, 1967, pp. 330-334.

21. Harter, H. L., "Asymptotic Variances and Covariances -'f Maximum

Likelihood Estimators, from Censored Samples, of the Parameters of
a Four-Parameter Generalized Gama Population," Report 66-0158,
Aerospace Research Laboratories, Office of Aerospace Research,
U.S.A.F., Wright-Patterson Air Force Base, Ohio, 1966.

22. Parr, V. B. and Webster, J. T., "A Method for Discriminating Between
Failure Density Functione used in Reliability Predictions,"
Technometrics, Volume 7, 1965, pp. 1-10.

23. Johnson, N. L. and Kots, S., Continuous Univeriate DistributioMs,

Uoughton Mifflin Co., 1970.

24. Hillberry, B. M., Mechanical Enginearing Department, Purdue

University, W. Lafayette, Ind. 1977.

25. Stacy, S, W. and Mihram, C. A.,

"Parameter Estimation for a General-
ised Gema Distribution," Technomecrics, Volume 7, 1965, pp. 349-358.

211
24. iehbko, C. ., AA lotaydastonm to Computer-Aded -eOaan,P Prentice-
Hall, Inc., 1968.

27. Wilk, (. B., GOanadesikan, R., and Hayett, 14. J., "Probability
Plots for the Gems Distribution," Tachnometrics, Volume 4, 1962,
pp. 1-20.

28. Hrter, H. L. and Moore, A. H., "Local Nazism Likelihood Estimation

of the Parameters of Three-Parameter Lognormal Populations from
Complete and Censored Samples," Journal of American Statistics
Association, Volume 61, 1961, pp. 842-851.

29. Cohen, A. C. Jr., "Estimating Parameters of Logarithmlic-Normal

Distributions by Maximum Likelihood," American Statistical Associa-
tion Journal, Volume 46, 1951, pp. 206-212.

30. Harter, H. L. and Moore, A. H., "Maximum Likelihood Estimation of

the Parameters of Geama and Weibull Populations from Complete and
from Censored Samples," Technometrics, Volume 7, 1965, pp. 639-643.
31. Mann, N. R., Schafer, R. E., and Slngpuralla, N. D., Methods for
Statistical Analysis of Reliability and Life Data, John Wiley and
Sons, Inc., 1474.

32. Harter, H. L., "Maximum Likelihood Estimation of the Parameters of

a Four-Parameter Generalized Gamsa Population from Complete and
Censored Samples," Technometrics, Volume 9, 1967, pp. 159-165.

33. Hooks, R. and Jeeves, T. A., "Direct Search Solution of Numerical

and Statistical Problems," Journal of the Association of Computing
Machinery, Volume 8, 1961, pp. 212-229.
34. Draper, N. R. and Smith, H., Applied Regression Analysis, John
Wiley and Sons, Inc., 1966.

35. Dahiya, R. C. and Garland, J., "Pearson Chi-Squared Test of Fit vith
Random Intervals," Biometrika, Volume 59, 1972, pp. 147-153.

36. Massey, F. J. Jr., "The Kolmogorov-Smirnov Test for Goodness of Fit,"

Journal of American Statistical Association, Volume 46, 1951,
pp. 68-78.

37. Alsos, W. X., "The Iffects of Single Overload/Underload Cycles on

Fatigue Crack Propagation," M.S. Thesis, Purdua University, 1975.

38. Hillberry, B. M., Alsos, W. X., and Skat, A. C. Jr., "The Fatigue
Crack Propagation Delay Behavior in 2024-T3 Aluminum Alloy Due to
Single Overload/Underload Sequences," AFFDL-TR-75-96, Air Force
Flight Dynamics Laboratory, Wright-Patterson Air Force Base, Ohio,
1976.

212
 ||,

39. Skat, A. C. Jr., "Evaluation of Extended Crack Closure Wu Fatigue

Crack Delay Prediction for Single Overload/lInderload Sequences,"
H.S. Thesis, Purdue University, 1975.

40. Virkler, D. A., Hillberry, B. M., and Goal, P. K., "An Investigation
of the Statistical Distribution of Fatigue Crack Propagation Data,"
ASTH Meeting on Fracture Mechanics, Norfolk, Va., March 1977,

21

OU.S.0ne~nmt Printir0e e, 78 -- M5-wi0,a0